Datasets
Models
Downloads: 1
| # | Origin | Rationale | Data | EEG processing | Methodology | Results and discussion | Reviewed by | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Title | Year | Authors | Journal / Origin | Preprint first | Type of paper | Lab / School / Company | Country | Pages | Domain 1 | Domain 2 | Domain 3 | Domain 4 | High-level Goal | Practical Goal | Task/Paradigm | Motivation for DL | EEG Hardware | Neural response pattern | Dataset name | Dataset accessibility | Data | Data - samples | Data - time | Data - subjects | Nb Channels | Sampling rate | Offline / Online | Preprocessing | Preprocessing (clean) | Artefact handling | Artefact handling (clean) | Features | Features (clean) | Normalization | Software | Architecture | Architecture (clean) | Design peculiarities | EEG-specific design | Network Schema | Input format | Layers | Layers (clean) | Activation function | Regularization | Regularization (clean) | Nb Classes | Classes | Output format | Nb Parameters | Training procedure | Training procedure (clean) | Optimizer | Optimizer (clean) | Optim parameters | Minibatch size | Hyperparameter optim | Hyperparameter optim (clean) | Data augmentation | Loss | Intra/Inter subject | Cross validation | Cross validation (clean) | Data split | Performance metrics | Performance metrics (clean) | Training hardware | Training time | Results | Benchmarks | Baseline model type | Statistical analysis of performance | Analysis of learned parameters | Model inspection (clean) | Discussion | Limitations | Code available | Code hosted on | Limited data | First Reader | Second Reader | Validated by Author(s) | Citation | ||||||||
| 2 | EEG-signals based cognitive workload detection of vehicle driver using deep learning | 2018 | Almogbel, Dang & Kameyama | IEEE Conference on Advanced Communication Technology | No | Conference | Waseda University | Japan | 4 | Classification of EEG signals | Monitoring | Cognitive | Mental workload | Improve State-of-the-Art | Driving Game (GTA) | Muse (InteraXon) | Raw EEG | Internal Recordings | Private | 1 subject x 24 sessions (12 High / 12 Low Workload) 15-30min each session (Sliding windows [from 30s to 180s], 1/256 overlap) | 216 | 540 | 1 | 4 | 256 | None | No | No | No | Raw EEG | Raw EEG | z-score | N/M* | CNN | CNN | Yes | 38400x4 (38400 = 150s @ 256Hz) | 7 Conv + 3 FC | 10 | ReLU | Dropout: 50% | Yes | 2 (Softmax) | N/M* | Standard | Standard | RMSProp | Other | lr=0.002 | 64 | N/M | N/M | No | Binary cross-entropy | Intra | Leave-One-Session-Out | Leave-One-Session-Out | Train: 92% Valid: 8% Test: N/A | Accuracy | accuracy | N/M | N/M | 95.31% | No | None | No | No | No | "This study does not impose in any way a direct comparison with the distinguished previous works because the used data, experimental conditions, classification targets are different in each, but rather explore and introduce the potential of using deep CNN architecture in classifying raw EEG signals without any pre-processing." | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | Almogbel2018 | ||||||||||||||||
| 3 | Automatic ocular artifacts removal in EEG using deep learning | 2018 | Yang, Duan, Fan, Hu & Wang | Biomedical Signal Processing and Control | No | Journal | Key Laboratory of Power Station Automation Technology, Shanghai University | China | 11 | Improvement of processing tools | Signal cleaning | Artifact handling | Novel | Motor Imagery | WirelessEEG (Neuracle) | Clean EEG / Ocular artefacts | BCI Competition IV - I; Internal Recordings | Public | Each subject has 200 trails of motor imagery and each trail lasts for more than 6s. Subject 1, 2, 6 and 7 from BCI Comp. dataset + 3 (internal recordings) | 1400 | 140 | 7 | 59; 32 | 100 | 1) Band-Pass Filter: 0.05-200Hz | Yes | No | No | Raw EEG | Raw EEG | min-max | MATLAB | SAE | AE | N/M* | Yes | 100x1 | 3 | 3 | L1 | Yes | 100x1 | N/M* | Greedy Layer-wise training | Standard | N/M* | N/M | N/M | N/M | No | RMSE | Inter | No | No | 16520 training samples 15458 test samples | RMSE | RMSE | Not mentioned | N/M* | RMSE is lower for proposed approach than for benchmarks, as is the accuracy on the surrogate MI task | Shallow SAE, ICA, K-ICA, SOBI | Traditional pipeline | No | No | No | "Compared with the classical OAs removal methods, the proposed method has many highlights. [...] In the future work, we are going to improve the training method of DLN or try replacing the SAE with other neural networks such as convolutional neural networks (CNN) to strengthen its fitting ability for the details of EEG." | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Yang2018 | |||||||||||||||||||
| 4 | An end-to-end framework for real-time automatic sleep stage classification | 2018 | Patanaik, Ong, Gooley, Ancoli-Israel & Chee | Sleep | No | Journal | Duke-NUS Medical School, Singapore University of California, San Diego | Singapore | 11 | Classification of EEG signals | Clinical | Sleep | Staging | Improve State-of-the-Art: DL for Sleep (CNN + MLP) | Reduce the time necessary to stage sleep recordings by using DL | Sleep | No need for feature engineering | N/M | Raw EEG | Internal Recordings | Private | Four datasets ≈ 1700 polysomnography records a total of 11,727 hr of PSG data / 1,403,164 epochs | 1403164 | 703620 | 459 | 2 | N/M | 1) Pass-Band Filter (FIR): 0.3-45Hz 2) Downsampled to 100Hz (polyphase FIR filter) | Yes | No | No | Spectrogram | Frequency-domain | N/M | TensorFlow | CNN + MLP (2 Stages) | CNN | Consecutive probabilities outputted by the CNN are aggregated by the MLP | N/M | Yes | 32x32x3 (spectrogram 2D x 3 channels) | CNN: 16 MLP: 1 | 17 | CNN: ReLU MLP: tansig | N/M | N/M | 5 (Softmax) Probability of each Sleep Stage | dCNN: 177 669 weights MLP: 445 weights | Standard optimization | Standard | Stochastic gradient descent with Nesterov momentum | SGD | Learning rate: 0.001 Momentum: 0.9 Learning rate decay: 10e-6 | CNN: 300 MLP: 1000 | Trial and error | Yes | No | Categorical Cross-Entropy | Inter | Train-Valid-Test | Train-Valid-Test | Train: 75% of DS1 & DS2 Test: 25% of DS1 & DS2 Validation: DS3, DS4 | Accuracy Cohen's kappa | accuracy, Cohen's kappa | NVidia GTX 1060 | N/M | Test set: ~89.8%. kappa=0.862 Validation set 1: 81.4%, kappa=0.740 Validation set 2: 72.1%, kappa=0.597 | Expert rescoring of 50 records | Traditional pipeline | t-test on Cohen's kappa (automatic vs. expert rescoring) -> stat. diff. for validation set 3 but not for 4 | No | No | "... our framework provides a practicable, validated, and speedy solution for automatic sleep stage classification that can significantly improve throughput and productivity of sleep labs. It has the potential to play an important role in emerging novel applications of real-time automatic sleep scoring as well as being installed in personal sleep monitors." | N/M | No | N/A | No | Yannick Roy | Hubert Banville | Yes | Patanaik2018 | |||||||||||
| 5 | Epileptic Seizure Detection: A Deep Learning Approach | 2018 | Hussein, Palangi, Ward & Wang | Arxiv | Yes | Preprint | UBC | Canada | 12 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve State-of-the-Art: DL for Epilepsy (LSTM) | Improve performance on seizure detection with DL, on real conditions (with noise) | Resting State, Eyes Open, Eyes Closed, Seizures. | Automatically learns features | N/M | Raw EEG | Bonn University | Public | Bonn University (A,B,C,D,E) 5x 100 epochs of 23.6s | 500 | 197 | 15 | 1 | 173.6 | 1) Artifacts Removed 2) Band-Pass Filter: 0.53-40Hz (before saving the dataset... "hardcoded") | Yes | Yes (dataset already cleaned) | Yes | Raw EEG | Raw EEG | N/M | MATLAB, Python Keras with TensorFlow backend | LSTM | RNN | N/M | N/M | Yes | 100x2 | 3 | 3 | N/M | N/M | N/M | 2, 3 or 5 | N/M | Standard optimization | Standard | Adam | Adam | LR: 0.001 | 64 | N/M | N/M | Added artifacts (EMG, EOG) and Gaussian white noise | Categorical Cross-Entropy | Inter | 10-Fold CV | k-fold | Train: 80% Test: 20% | Accuracy Sensitivity Specificity | accuracy, sensitivity, specificity | NVidia K40 | 2h | 100% everywhere. For Sensitivity, Specificity & Accuracy of the 2-Classes, 3-Classes & 5-Classes. Robust to artificial artfiacts | Compared with many other SotA using the same dataset. BNN, ME, SVM, ELM, LDA, SVM, KNN, ANN, etc. | Traditional pipeline | No | No | No | Compared to the state-of-the-art methods, this approach can learn the high-level representations, and can effectively discriminate between the normal and seizure EEG activities. Another advantage of this approach lies in its robustness against common EEG artifacts (e.g., muscle activities and eye- blinking) and white noise. | Unbalanced class distributions | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Hussein2018 | |||||||||||
| 6 | Development of a brain computer interface interface using multi-frequency visual stimulation and deep neural networks | 2018 | Perez-Benitez, Perez-Benitez & Espina-Hernandez | IEEE Conference on Electronics, Communications and Computers | No | Conference | National Polytechnic Institute, Mexico | Mexico | 7 | Classification of EEG signals | BCI | Reactive | SSVEP | Improve State-of-the-Art: SSVEP with CNN | Increase number of commands and reduce eyestrain in a visual BCI | SSVEP (with LEDs) | Just another classifier! | (Custom-made) | SSVEP | Internal Recordings | Private | 11 subjects x 5 stimuli | N/M | N/M | 11 | 3 | 250 | N/M | N/M | N/M | N/M | Spectrum | Frequency-domain | N/M | N/M | SAE (Sparse AutoEncoder) | AE | N/M | N/M | Yes | N/M | SAE: 2 Final: 2 | 4 | Sigmoid | L2 regularization Sparsity loss | Yes | 5 (Softmax) Diff SSVEP Freqs | N/M | 1) Train SAE 2) Train softmax on top of SAE middle layer | Pre-training | N/M | N/M | epochs: 50 lambda (L2): 0.16 gamma (sparsity): 1.0 rho: 0.1 | N/M | N/M | N/M | No | Mean Squared Error (SAE) Cross-entropy (softmax layer) | Intra | No | No | N/M | Accuracy | accuracy | N/M | N/M | 97.78% (not clear if on training set or something else!) | k-NN Naive Bayes, Bayes Kernel Decision Tree, Random Forest, Gradient Boosted Tree Rule Induction, MC-SVM, ML Perceptron | Traditional pipeline | No | Visualization of learned parameters | Analysis of weights | The analysis of the DNN first layer weights reveals that there are two main patterns containing information about the SSVEPs in the power spectrums of the measured EEG signals: (i) the weights reinforces the features of the spectrum at frequencies {fst}, 3/2 {fst} and 2{fst} where fst are the frequencies of the MFVS and the other (ii) the weights reinforces the features of the spectrum at low frequencies from 0 Hz – 20Hz. | N/M | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Perez-Benitez2018 | |||||||||||
| 7 | Deep Semantic Architecture with discriminative feature visualization for neuroimage analysis | 2018 | Ghosh, Dal Maso, Roig, Mitsis & Boudrias | Arxiv | Yes | Preprint | McGill, UdeM | Canada | 11 | Classification of EEG signals | Monitoring | Physical | Exercise | Improve SOTA | Study the add-on effects of exercise on motor learning | Hand motor task before and after an acute exercise | Does not require hand-engineered features | (BrainProducts) | Brain Rhythms (SMR) | Internal Recordings | Private | 25 subjects 4 x 50 x [3.5 sec (holding) + 3 to 5 sec (rest)] | 5000 | 625 | 25 | 64 | 2500 | 1) Band-Pass Filter: 0.5-55Hz 2) Re-reference to average. 3) Visual Inspection, noisy signal segment removed 4) ICA to remove eye blinks 5) Morlet Wavelet (wave:7, 1Hz reso) | Yes | Visual inspection to reject transient artefacts ICA for eye blinks | Yes | Frequency Bands (55) | Frequency-domain | Per-electrode spectral normalization | Brainstorm (MATLAB), Torch | CNN | CNN | 1) Base CNN that expects baseline and post-condition data in parallel 2) CNN that predicts class 3) Adverserial Component to penalize subject-dependent training | Base CNN: spectral-only convolutions | Yes | 64 x 55 (channels x freq bands) [x2 since the Base CNN is used twice in a single pass] | [On TF maps, on topo maps] BaseCNN: 2 + 1, 3 + 1 Discriminator: 2, 2 Adversary: 2, 2 | 7 | ReLU | Dropout, weight decay | Yes | 2 Prob of EXE Prob of CON | N/M | Standard optimization | Standard | Adam | Adam | [On TF maps, on topo maps] LR: 0.001, 0.001 LR decay: 0.0001, 0.001 Weight decay: 0.001, 0.03 | N/M | N/M | N/M | No | Negative Log-Likelihood (part 1) & KL-Divergence (part 2) | Inter | N/M | No | Train: 80% Validation: 20% Test: N/M | Accuracy | accuracy | N/M | N/M | 98.70% | N/M | None | No | Visualization of class activation maps (proposed method) | Class Activation Maps | "Importantly, the proposed novel method enabled us to visualize the features learnt by deep networks such as CNNs, which may in turn yield better interpretation of their classification basis." | N/M | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Ghosh2018 | |||||||||||
| 8 | Cascade and Parallel Convolutional Recurrent Neural Networks on EEG-based Intention Recognition for Brain Computer Interface | 2018 | Zhang, Yao, Zhang, Wang, Chen & Boots | AAAI Conference on Artificial Intelligence | Yes | Preprint | University of New South Wales | Australia | 8 | Classification of EEG signals | BCI | Active | Motor imagery | Novel Approach: Cascade & Parallel CNN and RNN | Compare Cascade and Parallel CNN + RNN on Motor Imagery Dataset (eegmmidb) to SOTA | Motor Imagery (see eegmmidb dataset) | To capture temporal and spatial information. | N/M | Motor Imagery | eegmmidb; Internal Recordings | Both | (eegmmidb) 108 subjects, 3,145,160 EEG (2808min) (Internal recordings) 9 subjects x 30 trials (6 per class) Internal recordings: 10s action, 10s rest | 3145430 | 2898 | 108 | 64 | 160 | Offline | 1) 2D Mesh (Matrix) 2) Sliding Window (clips) 3) Normalize | Yes | No | No | 2D Mesh Clips (of Raw EEG) | Raw EEG | z-score | N/M | Cascade / Parallel CNN + RNN (LSTM) | CNN+RNN | CNN + LSTM combined (serial or parallel) | To capture spatial and temporal resolution | Yes | 2D Data mesh (time signal x spatial matrice) | 3 CNN + 1 FC (1024) 2 LSTM (64) + 1 FC (1024) | 7 | N/M | Dropout (0.5) | Yes | 5 | 5 Motor Commands | 5 (Softmax) | N/M | N/M | N/M | Adam | Adam | LR: 0.0001 | N/M | N/M | N/M | N/M | Cross-Entropy | Inter | N/M | No | Train: 75% Test: 25% | Accuracy | accuracy | Nvidia Titan X Pascal | N/M | Cascade: 0.9824 Parallel: 0.9828 | (Major and Conrad 2017) : 0.72 - ICA (Shenoy, Vinod, and Guan 2015) : 0.82 - SR-FBCSP (Pinheiro et al. 2016) : 0.85 - SVM (Kim et al. 2016) : 0.80 - SUTCCSP (Zhang et al. 2017) : 0.79 - XGBoost (Bashivan et al. 2016) : 0.67 - R-CNN | DL & Trad. | No | No | No | A large-scale dataset of 108 participants on five categories is used to evalu- ate the proposed models. The experimental results show that both the cascade and parallel architectures could achieve very competitive accuracy around 98.3%, considerably superior to the state-of-the-art methods. | N/M | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | Zhang2018c | ||||||||
| 9 | A hierarchical LSTM model with attention for modeling EEG non-stationarity for human decision prediction | 2018 | Hasib, Nayak & Huang | IEEE EMBS International Conference on Biomedical & Health Informatics | No | Conference | University of Texas, San Antonio | USA | 4 | Classification of EEG signals | BCI | Active | Mental tasks | Improve SOTA | Novel Approach: H-LSTM with Attention for Decision Classification | Allow or Deny Access based on ID + Image (Guard) | No need for hand-engineered features | ActiveTwo (BioSemi) | Raw EEG | BCIT Guard Duty | Private | 1782 of 5297 sequences selected: 892 Deny + 890 Allow 18 Subjects (10s windows) | 1782 | 297 | 18 | 64 | 512 | 1) Downsampled to 128Hz 2) Band-Pass Filter: 0.1-55Hz | Yes | N/M | N/M | Raw EEG | Raw EEG | z-score | N/M | LSTM | RNN | Hierachical (from samples in first layer to epochs in second layer) Attention mechanism | First layer acts on samples Second layer acts on epochs | Yes | 0.5s epochs | 2 | 2 | N/M | L2 weight decay | Yes | 1 Allow / Deny | N/M | Standard optimization | Standard | Adam | Adam | N/M | N/M | N/M | N/M | N/M | Cross-Entropy | Inter | 3-Fold CV | k-fold | Train: 60% Validation: 6% Test: 33% | ROC AUC | ROC AUC | N/M | N/M | H-LSTM (w/ Attention & 0.5s epochs): 82.6% H-LSTM (w/ Attention & 2.5s epochs): 81% H-LSTM (w/ Attention & 5s epochs): 81.6% H-LSTM (w/out Attention & 0.5s epochs): 80.3% H-LSTM (w/out Attention & 5s epochs): 73.7% | SVM: 65% CNN: 69% | DL & Trad. | No | No | No | "Using the attention mechanism does help enhance the discriminate features obtained from these epochs, although it does not help model the EEG non-stationarity" "Consistent with the observation from LSTM performance, we observed an increase of performance with shorter epoch length." | N/M | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Hasib2018 | |||||||||||
| 10 | Deep EEG super-resolution: Upsampling EEG spatial resolution with Generative Adversarial Networks | 2018 | Corley & Huang | IEEE EMBS International Conference on Biomedical & Health Informatics | No | Conference | University of Texas, San Antonio | USA | 4 | Generation of data | Generating EEG | Spatial upsampling | Novel Approach: GAN for EEG Upsampling. | BCI Competition III - Dataset V | GANs previous great results on image super-resolution | N/M | N/A | BCI Competition III - V | Public | (BCI Comp. III - V) 1,096,192 samples from 3 subjects (Windows of 1s, 480/512 overlap) | 36397 | 35.7 | 3 | 32 | 512 | 1) Downsampling in the number of channels (from 32 to 16) | Yes | No | No | None | Raw EEG | z-score | N/M | WGAN | GAN | N/M | Convolutional layers with kernel dimensions that find the relationships between channels | Yes | 32 x 512 (channels x samples) | Gen: 6 Conv Layers Discrim: 4 Conv Layers + 1 FC | 6 | ELU | Dropout (0.1 - 0.25) | Yes | 32 x SR (Channels x Super Resolved) (upsampled data) | N/M | Pre-trained Gen fine-tuned w/ WGAN framework losses w/ gradient penalty weight of 10. Also, label smoothing technique | Other | Adam | Adam | a=10^-4, b1=0.5, b2=0.9 | 64 | N/M | N/M | N/M | Gen: MSE Discrim: Distance | Inter | Holdout | Holdout | Train: 75% Valid: 20% Test: 5% | MSE MAE (mean absolute error) Accuracy, precision and recall (for classification task) | MSE, MAE, accuracy, precision, recall | N/M | N/M | [Scale 2 - Test] MSE: 2.06E3 | MAE: 24.66 [Scale 4 - Test] MSE: 8.68E3 | MAE: 64.39 ~10^4 fold (MSE) and ~10^2 fold (MAE) compared to Bicubic Interpolation | Bicubic Interpolated Channel Data | Traditional pipeline | No | No | No | "Feature scaling techniques besides standard normalization decreased model performance. [...] It was notably difficult and time-consuming to train GANs for EEG data. [...] After testing different variants of GAN: WGAN appeared to be more stable during training." | "It was notably difficult and time-consuming to train GANs for EEG data" | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | Corley2018 | |||||||||||||
| 11 | Spatial and Time Domain Feature of ERP Speller System Extracted via Convolutional Neural Network | 2018 | Yoon, Lee & Whang | Computational Intelligence and Neuroscience | No | Journal | Duke University Sangmyung University | USA | 12 | Classification of EEG signals | BCI | Reactive | ERP | Alleviate BCI illiteracy | Reduce BCI illiteracy in P300 spellers by using CNNs | Rapid Serial Visual Presentation (P300 speller) | Uncover new unknown spatial/temporal patterns. When an optimal filter is applied, the convolution will magnify the feature of interest and reduce the others [25]. | B-Alert X10 (ABM) | P300 and oddball paradigm-related EEG activity | Internal Recordings | Private | 33 subjects, 2 to 4 pairs of sessions (offline + online) 12 trials / session (each trial = 10s + ERP stimuli) 20 times x 6 icons / trial (300ms) 33x3x12x6x20 = 142,560 samples | 142560 | 712 | 33 | 11 | 256 | Both | N/M | N/M | N/M | N/M | Raw EEG | Raw EEG | N/A | TensorFlow Python | CNN | CNN | - | Layer 1: spatial correlation Layer 2: temporal filter | Yes | 14 x 300 (channels x samples) | CNN: 2 FC: 2 | 4 | ReLU | Dropout (0.1 - 0.25) | Yes | 6 | 6 different icons (Power On/Off, Volume Up/Down, Channel Up/Down) | 2 | N/M | Standard optimization | Standard | Adam | Adam | N/M | N/M | N/M | N/M | N/M | N/M | Intra | No | No | Train: 50% (offline) Test: 50% (online) | Accuracy, sensitivity, precision, F1 score, ROC (+ ANOVA on metrics) | accuracy, sensitivity, precision, f1-score, ROC | N/M | N/M | Accuracy: 88.9 for high performing group, 68,7% for low performing group | No benchmark | None | ANOVA | Visualization of weights | Analysis of weights | A P300 is not visible in all subjects, but there seems to be a P700 that is pretty consistent across subjects. Spatial features seem to play a more important role than temporal features in the classification of an oddball task. | - | No | N/A | No | Hubert Banville | Yannick Roy | Yes | Yoon2018 | ||||||||
| 12 | ?? | Spectrographic Seizure Detection Using Deep Learning With Convolutional Neural Networks | 2018 | Yan, Wang & Grinspan | Neurology | No | Supplement | Well Cornell Medical College New York | USA | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve State-of-the-Art: Using CNN on Spectrogram for Seizure Detection | Existing dataset, no mention of any task. (Supposed: Resting Sate) | N/M | Raw EEG (Seizure) | CHB-MIT | Public | 130 EEGs with 177 total seizures, and 549 EEGs without seizures. >90% of seizures were <2 minutes long. 130 EEGs with seizure and 130 randomly selected EEGs without seizures were converted to the median power spectrogram (MPS). The training set consisted of 16,992 seizure containing images and 16,992 images without seizures (80% of total images). The testing set contained 4,248 seizure containing images and 4,248 images without seizures (20% of total images). | 33984 | N/M | N/M | -1 | N/M | N/M* | N/M | N/M | N/M | Medium Power Spectrogram (MPS) | Frequency-domain | N/M* | CNN (4 variants of VGG16) | CNN | N/M* | No | Images (1s sliding window of MPS) | 16 | N/M | Dropout: 0.5 | Yes | N/M* | N/M* | N/M | N/M | N/M | N/M | N/M | No | N/M* | Inter | N/M | No | Train: 80% Test: 20% | Sensitivity Specificity | sensitivity, specificity | N/M* | All four CNN variants achieved >98% sensitivity and specificity | N/M* | None | No | No | No | "Convolutional neural nets can achieve high sensitivity and specificity in detecting seizures within spectrograms. However, generalizability and overfitting remains a concern. Further evaluation with more diverse data sets, images grouped by individual seizures, and additional regularization techniques is warranted." | No | N/A | TBD | Yannick Roy | TBR | TBC | Yan2018 | |||||||||||||||||||||
| 13 | Generating target / non-target images of an RSVP experiment from brain signals in by conditional generative adversarial network | 2018 | Lee & Huang | IEEE EMBS International Conference on Biomedical & Health Informatics | No | Conference | University of Texas, San Antonio | USA | 4 | Generation of data | Generating images conditioned on EEG | Novel Approach: generating images confitioned on EEG | Using EEG from RSVP to generate images (target or non-target) | RSVP - 5 Images/s | GAN models. | ActiveTwo (BioSemi) | RSVP | Internal Recordings | Private | 10 subjects, 5 sessions (~1h /session), 880 Epochs (1s windows) | 880 | 14.6 | 10 | 32 | -1 | Offline | - PREP Pipeline (EEGLAB): bandpass (0.1-55 Hz), robust referencing, interpolating bad channels - Downsampled to 32Hz - Subset of 32 channels (visual cortex) | Yes | Yes | Yes | Raw EEG | Raw EEG | z-score | EEGLAB | cGAN | GAN | It's to generate the image, not the EEG data. Based on DCGAN | N/A | Yes | 32 x 32 (channels x samples) | Generator: 4 Discriminator: 4 | 4 | G: Leaky ReLU D: ReLU | N/M | N/M | N/A | N/A | 64 x 64 (image) | N/M | GAN Style. | Standard | N/M | N/M | N/M | 16 | N/M | N/M | N/M | N/M | Inter | No | No | Train: 704 epochs Test: 176 epochs | Visual inspection (making sure generated image is of the right class) | visual inspection | N/M | 2-3h | Accuracy: 0.625 | None | None | No | Occlusion of input EEG and visualization of generated image | Occlusion of input | We demonstrated the performance of the proposed cGAN model and showed that generation with raw or normalized EEG produced better performance than that with added noise. We also showed how this model could be used for investigating the EEG and image associations. | N/M | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Lee2018 | ||||||||||
| 14 | Cross-Participant EEG-Based Assessment of Cognitive Workload Using Multi-Path Convolutional Recurrent Neural Networks | 2018 | Hefron, Borghetti, Schubert Kabban, Christensen & Estepp | Sensors | No | Journal | Air Force Institute of Technology (Ohio) | USA | 27 | Classification of EEG signals | Monitoring | Cognitive | Mental workload | Novel Approach: Using a Multi-Path Convolutional Recurrent Neural Network (MPCRNN) to improve SOTA on cross-participant classification of cognitive workload | Tackle cross-subs variability in cognitive workload assessment | Multi-Attribute Task Battery (MATB) environment | Assumptions regarding brain activity are better matched by a deep representation that includes multi-path connections. | ActiveTwo (BioSemi) | None | Internal Recordings | Private | 8 subjects * 4 blocks * 6 conditions * 5 min (1s windows) | 57600 | 960 | 8 | 128 | 4096 | 1) Trimmed to 303s trials 2) Downsampled to 512Hz 3) Down-selected 64 electrodes 4) PREP Pipeline to identify and interpolate bad channels, calculate a robust average reference, and remove line noise 5) High-Pass Filter: 1Hz 6) PSD 3-55Hz (2s Hanning-Windowed STFT, 1s overlap) | Yes | Yes (manual identification of high-variance segments) | No | PSD - Frequency Bands (53) | Frequency-domain | [-1, 1] | Keras, TensorFlow | (multi-path, residual) CNN + (bi-directional, residual) LSTM | CNN+RNN | It combines a wide multi-path, residual, convolutional networkwith a bi-directional, residual LSTM. | 1x1 convolutions to act as cross-channel parametric pooling | Yes | 20x53x64 (time x frequency bands x channels) | [very deep, see schema / paper] | 8 | ReLU and sigmoid | Dropout + batch normalization + early stopping + L1 + L2 | Yes | 1 (Mental Workload) | MPCRNN: 6.2M | Standard. | Standard | Adam | Adam | LR: from 0.0001 to 0.000001 | 128 | N/M | N/M | N/M | Binary cross-entropy | Inter | Test: Hold-out 1 Participant Training: 7-Fold Cross-Validation | k-fold; Holdout | Train: 6 participants Validation: 1 participant Test: 1 participant | Mean Accuracy | accuracy | N/M | N/M | between 80-86% (depending on sequence length used as input) | Simpler DL architectures | DL | ANOVA + post-hoc Tukey Honest Significant Difference tests | No | No | We found that while increasing sequence length improves model accuracy, it does not improve generalizability since cross-participant variance increases due to cross-participant distributional differences. Furthermore, longer sequences reduce temporal specificity which decreases a model’s utility in a real-time environment. The only condition among our experiments across sequence lengths, architectures, and training methods which resulted in improved accuracy and decreased cross-participant variance was the multi-path convolutional recurrent architecture. | N/M | No | N/A | Yes | Yannick Roy | Isabela Albuquerque | Yes | Hefron2018 | |||||||||||
| 15 | Classification of auditory stimuli from EEG signals with a regulated recurrent neural network reservoir | 2018 | Moinnereau, Brienne, Brodeur, Rouat, Whittingstall & Plourde | Arxiv | Yes | Preprint | Université de Sherbrooke | Canada | 5 | Classification of EEG signals | BCI | Reactive | Heard speech decoding | Improve SOTA | Classify heard speech (vowels) from EEG | Auditory Stimuli + Imagined Speech | Can extract features automatically | BrainAmp (BrainProducts) | Raw EEG | Internal Recordings | Private | 8 subjects x 3 stimuli x 200 times each (2s windows, onset at 0.5s) (preprocessing removed 30%!) | 4800 | 9600 | 8 | 64 | N/M | Offline | 1) Pass-Band Filter: 0.1-45Hz 2) Re-sampled at 500Hz 3) Windows of 2s (stimulus at 0.5s) 4) Trials with Amplitude > +-75uV rejected 5) Re-reference to local average | Yes | Yes (amplitude thresholding) | Yes | Spike Train from Ben’s Spike Algorithm (BSA) | Other | N/M | Python | RNN Reservoir | RNN | The reservoir comprises 512 neurons placed in a three-dimensional grid where 80% are excitatory and 20% are inhibitory neurons | N/M | Yes | Spike Trains per channel | N/A | N/M | Leaky Integrate-and-Fire | N/M | N/M | 3 | "a", "i", "u" | 1 | N/M | Reservoir: unsupervised tuning Classifier: linear regression | Standard | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Intra | 5-Fold CV | k-fold | Train: 4/5 Test: 1/5 | Accuracy | accuracy | N/M | N/M | 83.2% (64 electrodes) 1 Electrode: 57.3% 3 Electrodes: 71.4% 10 Electrodes: 81.7% [Chance: 33%] | CNN (3 conv layers of 64 filters) | DL | No | No | No | "It’s hard to compare these results with the previ- ous work where many different experimental conditions (e.g. different type and number of stimuli) and preprocessing has been used. However, we show here that excellent classifica- tion results can be obtained with minimal preprocessing of the EEGs." | N/M | No | N/A | No | Yannick Roy | Hubert Banville | Yes | Moinnereau2018 | ||||||||
| 16 | Deep learning for detection of epileptiform discharges from scalp EEG recordings | 2018 | van Putten, de Carvalho, Tjepkema-Cloostermans | Clinical Neurophysiology | No | Journal | University of Twente | Netherlands | 6 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA | Use CNN and/or LSTM to classify yes/no discharges | Pre-Recorded EEG, no mention of any task. (Supposed: Resting Sate and sleep) | DL is promising novel approach and is able to learn from large data-sets | N/M | Raw EEG | TBD | Private | Training (41,381 epochs), Test (8775 epochs) For validation we used 7 EEGs (47,122 epochs) with 538 focal epileptiform discharges and 12 normal EEGs (n = 11,782 epochs). (2s windows, no overlap) | 97278 | 3242.6 | N/M | 19 | 125 | Offline | 1) Band-Pass Filter: 0.5-35Hz 2) Re-referenced to both a longitudinal bipolar montage and a source Laplacian | Yes | No | No | Raw EEG | Raw EEG | N/M | Keras | CNN RNN | CNN+RNN | Multiple designs | N/M | Yes | 19 (channels) x 250 (2s) | CNN: 4-9 Layers LSTM: 50-100 Units Both w/ 1-3 FC Layers | 12 | N/M | Dropout (20-50%) | Yes | 2 | Normal IED (discharge) | 1 (prob [0,1] of discharge) | 9142859 | Standard | Standard | Adam | Adam | LR:3e-3 Beta1: 0.91 Beta2: 0.999 Epsilon: 1e-8 | N/M | N/M | N/M | N/M | Categorical Cross-Entropy | Inter | N/M | No | Train: 41,381 Valid: 58,904 Test: 8,775 | ROC AUC Sensitivity Specificity | ROC AUC, sensitivity, specificity | NVidia GTX 1080 | 2h | AUC: 0.94 Sensitivity: 0.73 Specificity: 1 | None | None | No | No | No | "We foresee that deep nets may outperform humans both in classification accuracy and speed, leading to a fundamental shift in clinical EEG analysis in the next decade." | N/M | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | VanPutten2018a | ||||||||
| 17 | Cognitive Analysis of Working Memory Load from EEG, by a Deep Recurrent Neural Network | 2018 | Kuanar, Athitsos, Pradhan, Mishra & Rao | ICASSP | No | Conference | University of Texas, Arlington | USA | 5 | Classification of EEG signals | Monitoring | Cognitive | Mental workload | Improve State-of-the-Art: Using RNN to measure levels of cognitive load. | Extract features less sensitive to variations along each spatial dimension | Working memory / workload experiment. (showing a set of letters and then showing a letter asking if the letter was part of the set) Sets of 4,6,8,10 letters corresponding to mental workload 1,2,3,4. | ConvNets have demonstrated the ability to extract features that are invariant to changes in input patterns | Neurofax EEG-1200 (Nihon Kohden) | PSD | NIMHANS | Private | 6490 samples, from 22 subjects Each trial of 4.5s sliced into 0.5s and an image was constructed over each time slice. | 58410 | 486.75 | 22 | 64 | 256 | 1) From 4.5s Windows to 9 Windows of 0.5s | Yes | No | No | 192 Features: 64 chan x 3 bands Theta (4-7Hz), Alpha (8-13Hz), Beta (13-30Hz) (FFT) Converted into images. (32x32) 3D electrodes spatial to 2D. | Frequency-domain | N/M | Theano, Python | CNN + BiLSTM | CNN+RNN | Transforming channels and frequency bands into images of 0.5s windows, fed to an Hybrid CNN+BiLSTM | N/M | Yes | EEG Images 32x32 (0.5s windows) (mixing 3 freq bands + 64 channels) | 9 Conv Layers + 1 FC + 2 LSTM Layers of 64 units + 1 FC | 13 | Sigmoid | Dropout (0.5) + L2 (0.0001) | Yes | 4 Classes (Softmax) | 1.66 Mil | Standard | Standard | Adam | Adam | LR: 10^-4 Beta1: 0.9 Beta2: 0.99 | 30 | N/M | N/M | Gaussian noise (on image) | Cross-Entropy | Inter | Leave-One-Subject-Out | Leave-One-Subject-Out | N/M | Accurary | accuracy | NVidia K40 | 18h | 92.50% | SVM, Logistic Regression, Random Forest | Traditional pipeline | No | No | No | "Our implementation was different from the previous attempts and learned the robust representations from EEG image sequences using a ConvNet and BiLSTM hybrid network. Our proposed hybrid network demonstrated the significant improvements in finding better classification accuracy i.e. up to 92.5% over various existing LSTM models." | N/M | Yes | Website | No | Yannick Roy | Isabela Albuquerque | TBC | Kuanar2018 | |||||||||||
| 18 | A Deep Learning Approach with an Attention Mechanism for Automatic Sleep Stage Classification | 2018 | Längkvist & Loutfi | Arxiv | Yes | Preprint | Orebro University | Sweden | 18 | Classification of EEG signals | Clinical | Sleep | Staging | New approach: "Explore the advantages of using a model qith selective attention applied to automatic sleep staging" | Learn representations of sleep EEG with attention | Sleep (PSG) | Learn better features | N/M | Sleep | UCDDB | Public | 25 x 6-9 hours (est. at 7.5h) (30s windows, no overlap) | 22500 | 11250 | 25 | 1 | 128 | Offline | 1) Notch Filter: 50Hz 2) Down-Sampled to 64Hz 3) Band-Pass Filter: 0.3-32Hz | Yes | N/M | N/M | 28 Features: Relative Power: Delta (0.5 − 4Hz), Theta (4−8Hz), Alpha (8−13Hz), Beta (13−20Hz), & Gamma (20−32Hz), Entropy, Kurtosis, and Spectral Mean of all signals and fractal component of EEG. [+ EOG & EMG features] | Frequency-domain | z-score | N/M | SAE | AE | Attention Mechanism (static & adaptive approaches) | N/M | No | 28 | 1 | 1 | Sigmoid | L2 normalization KL term in cost function for sparsity | Yes | 5 | SWS, S2, S1, REM, Awake | 5 Classes (Softmax) | N/M | 1) Training AE 2) Training softmax layer on learned features | Pre-training | SGD with momentum | SGD | Momentum: 0.9 LR decay: 0.01 | 30 | Random grid search | Yes | N/M | MSE | Inter | 5-Fold CV | k-fold | Train: 60% Valid: 20% Test: 20% | Accurary | accuracy | N/M | 2-3h | 60-90% on each of the 5 classes. | DBN SAE (standard a) SAE (fixed a) SAE (adaptive a) | DL | No | Visualization of attention mechanism weights | Analysis of weights | "[...] Many of the used features try to capture the most relevant information for the current sleep stage and therefore mimic the standard Rechtschaffen and Kales (R&K) system [38, 18, 17] that is manually used by sleep technicians." | Unsupervised learning treats all features equally; that's why attention mechanism is useful | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Langkvist2018 | ||||||||
| 19 | On the Classification of SSVEP-Based Dry-EEG Signals via Convolutional Neural Networks | 2018 | Aznan, Bonner, Connolly, Moubayed & Breckon | Arxiv | Yes | Preprint | Durham University | UK | 6 | Classification of EEG signals | BCI | Reactive | SSVEP | Improve State-of-the-Art | Apply CNN to SSVEP with dry EEG headset | SSVEP | Want an end-to-end system (no need to extract features) | Quick-20 (Cognionics) | SSVEP | Internal Recordings | Private | 4 subjects, 4 classes 640 trials total (160 per class) x 3s | 640 | 32 | 4 | 20 | 500 | Offline | None | No | No | No | None | Raw EEG | N/A | Pytorch | CNN | CNN | - | Layer 1: temporal filter | Yes | N/M | 2 (Tried 7 in the end) | 2 | ReLU | L2 normalization Dropout (50%) | Yes | 4 | N/M | Standard optimization | Standard | Adam | Adam | N/M | 32 | Grid search | Yes | No | Categorical cross-entropy | Both | 10-Fold CV and Leave-One-Subject-Out | k-fold; Leave-One-Subject-Out | N/M | Accuracy | accuracy | Nvidia GTX 1060 | 4 min | Subject 1 - all data: 96% Subject 1-3 (individually, only 20 trials each): mean of 89% Across-subjects: 78% Leave-one-subject:out:59% | Traditional feature-based pipeline (Riemannian Geometry + classifier) RNN, LSTM, GRU | DL & Trad. | No | No | No | Repeating the convolutional layer block increased accuracy on the held-out subject. | N/M | No | N/A | Yes | Hubert Banville | Isabela Albuquerque | TBC | Aznan2018 | ||||||||||
| 20 | A Long Short-Term Memory deep learning network for the prediction of epileptic seizures using EEG signals | 2018 | Tsiouris, Pezoulas, Zervakis, Konitsiotis, Koutsouris & Fotiadis | Computers in Biology and Medicine | No | Journal | National Technical University of Athens | Greece | 14 | Classification of EEG signals | Clinical | Epilepsy | Prediction | New Approach: Using LSTM for seizure detection. (claiming they are the first ones but they are not, so its a Improve SOTA) | Apply LSTM for seizure detection on CHB-MIT | Resting State, Eyes Open, Eyes Closed, Seizures. | Expand from CNN to LSTM. (they claimed to be first, but they are not...) | N/M | Seizures | CHB-MIT | Public | 983h, 185 seizures (5s windows, no overlap) | 707760 | 58980 | 23 | 23 | 256 | 1) Selecting only channels that are stable across recordings (for cross-validation) 2) Kept 18 channels. | Yes | No | No | Time Domain: the 4 Statistical Moments, Standard Dev, Zero Crossings, Peak-to-peak Voltage, Total signal area, decorrelation time. Frequency Domain: FFT (PSD), DWT. Cross-Correlation: Max absolute coefficient. Graph Theory: Local & Global measures. (all on 5s windows) | Combination | N/M | Keras Tensorflow Python 3.6 | LSTM | RNN | - | LSTM Length: predicting seizures from 15 min before onset to 120 min before onset | Yes | Features x EEG Segments 643x[5-50] | LSTM_1: 1 (32 HU) LSTM_2: 1 (128 HU) LSTM_3: 2 (128/128 HU) + 1 FC (30) | 3 | ReLU | Dropout (finally discarded, because the shuffling of data seems to be enough) | Yes | 2 (Softmax, 1 hot encoded: preictal or interictal) | N/M | By shuffling the EEG segments that are used as input, the LSTM network is forced to learn more generic preictal patterns as each sequence consists of random, non-adjacent preictal segments that not only come from various locations with different time distances from the actual seizure onset, but also from the preictal activity of different seizures. | Standard | Adam | Adam | LR: 0.001 B1: 0.9 B2: 0.999 Decay: 0 | 10 | Manually trying 3 different configurations | Yes | Splitted minority class in smaller subgroups to balance classes | Cross-Entropy | Both | 10-Fold CV | k-fold | Eval: 3/24 Train: N/M (assuming 21/24) | Sensitivity (SEN) Specificity (SPEC) False Prediction Rate (FPR) Preictal Window | sensitivity, specificity, false prediction rate | N/M (they seem to say CPU) | N/M | [Segments] SEN, SPEC | [Events] SEN, FPR 15-min Preictal Window: 99.28, 99.28 | 100, 0.107 30-min Preictal Window: 99.37, 99.60 | 100, 0.063 60-min Preictal Window: 99.63, 99.78 | 100, 0.032 120-min Preictal Window: 99.84, 99.86 | 100, 0.02 | SVM Decision Trees Repeated Incremental Pruning to Produce Error Reduction (RIPPER) (LSTM outperforms all of them on all subjects) | Traditional pipeline | No | No | No | In theory, better EEG signal representation could be learned if the size of LSTM network was substantially increased, by adding more layers and memory units, to compensate for the increased input size of directly providing the EEG signals. However, the computational cost of training larger LSTM networks increases rapidly requiring more training time or using arrays of GPUs. Even if computational cost was not a problem, this approach would require even more EEG data to effectively train the millions of network parameters. | 1) Overal amount of Data 2) Number of Seizures | No | N/A | Yes | Yannick Roy | Isabela | TBC | Tsiouris2018 | |||||||||||
| 21 | Joint Classification and Prediction CNN Framework for Automatic Sleep Stage Classification | 2018 | Phan, Andreotti, Cooray, Chén & De Vos | Arxiv | Yes | Preprint | University of Oxford | UK | 11 | Classification of EEG signals | Clinical | Sleep | Staging | Improve State-of-the-Art (~New task: predict neighboring classes too) | Use one-to-many approach with a multi-task softmax to leverage neighboring data to predict sleep stage | Sleep | Re-using their previous network. (H. Phan et al., 2018) | N/M | Sleep events | MASS | Public | 228,870 epochs x 30s from 200 subjects | 228870 | 114435 | 200 | 1 | 100 | Offline | 1) Convert 20s epochs in 30s epochs (+5s before + 5s after) | Yes | N/M | N/M | Spectrogram (STFT) Hamming window 2s + 50% overlap Log Spectrum | Frequency-domain | N/M* | Tensorflow | CNN | CNN | Conv-Pool-Softmax | Layer 0: filter bank on spectrogram | Yes | 129 x 29 x {1, 2, 3} Bins x time, x channels 30-s epochs | 1 (1xCNN +Pooling +Softmax) | 1 | ReLU | L2 Dropout (20%) | Yes | 5 | Wake N1, N2, N3 REM | 5 x (1 + 2 * nb of neighbouring windows) | N/M* | Standard optimization | Standard | Adam | Adam | LR: 0.0001 | 200 | N/M | N/M | Randomly selected batch with balanced classes | Categorical cross-entropy | Inter | Leave-10-Subjects-Out (20-Fold CV) | Leave-N-Subjects-Out | Train: 180 subjects Valid: 10 subjects Test: 10 subjects | Accuracy, Kappa, Specificity, Sensitivity, F1-score | accuracy, Cohen's kappa, specificity, sensitivity, f1-score | N/M* | 1.36 hours | Multimodal acc.: 83.6 % | One-to-one and Many-to-one with same architecture and with a different ConvNet architecture without l-max pooling DeepCNN DeepSleepNet | DL | No | No | No | Increasing the number of filters in the Conv layer doesn't impact the performance much Adding other modalities (EOG, EMG) lead to significant improvements A context size larger than 3 leads to performance degradation Using recurrent layers might help | No | No | N/A | No | Hubert Banville | Yannick | Yes | Phan2018 | ||||||||
| 22 | Deep Convolution Neural Network and Autoencoders-Based Unsupervised Feature Learning of EEG Signals | 2018 | Wen & Zhang | IEEE Access | No | Journal | Xiamen University | China | 12 | Improvement of processing tools | Feature learning | Improve SOTA | Learn features for epilepsy detection using unsupervised learning | Resting State, Eyes Open, Eyes Closed, Seizures. | Learn features automatically | N/M | Raw EEG | Bonn University; CHB-MIT | Public | DS #1 - Bonn University (A,B,C,D,E) 5x 100 epochs of 23.6s DS #2 - CHB-MIT (first 10 subjects) 200 + 200 examples of 4096 (@ 256Hz = 16s) | 500; 400 | 197; 106.6 | 10; 10 | 1 | 173.61; 256 | Offline | 1) Common average reference 2) Bandpass 0.53-40 Hz | Yes | N/M | N/M | 1) Chose single channel with the most variance | Raw EEG | min-max | Scikit-learn Python | Convolutional AE | AE | Various (tried multiple classifiers on top of the encoder) | - | Yes | 4096 x 1 | 9 [YR: not sure how HJB got that 9] | 9 | ReLU | N/M | N/M | 2 | Seizure No Seizure (not explicit) | 4096 x 1 | N/M | 1) Training AE 2) Training standard classifier on learned features | Pre-training | Adam | Adam | N/M | N/M | N/M | N/M | N/M | Mean absolute error divided by input mean amplitude | Both | 5 and 10 -Fold CV | k-fold | N/M | Accuracy | accuracy | N/M | N/M | No aggregate is reported... (see paper, they report results per subject and per classifier) | PCA Random projection | Traditional pipeline | No | No | No | Less than 4 hidden units on the bottleneck layer led to a drop in accuracy as compared to standard dimensionality reduction techniques. Their approach is flexible to new datasets... | "It is very difficult to train multiple hidden layers [...]" | No | N/A | No | Hubert Banville | Yannick | TBC | Wen2018 | ||||||||||
| 23 | Deep learning with convolutional neural networks for decoding and visualization of EEG pathology | 2018 | Schirrmeister, Gemein, Eggensperger, Hutter & Ball | Arxiv | Yes | Preprint | University of Freiburg | Germany | 7 | Classification of EEG signals | Clinical | Pathological EEG | Improve SOTA Feature visualization/interpretability | End-to-end detection of abnormal EEG | N/M | Automated EEG diagnosis | N/M | Raw EEG | TUH Abnormal EEG Corpus | Public | TUH Abnormal Corpus 2740 + 277 = 3017 (x 16min) (they explored using [1, 16] min) (6s windows) | 482720 | 48272 | 2132 | 21 | 250 | 1) Select 21 electrodes common to all subjects 2) Remove 1st minute 3) Crop to recording to up to 20 minutes 4) Clip amplitude to +-800uV 5) Resample to 100Hz | Yes | N/M | N/M | Raw EEG | Raw EEG | N/M | Pytorch | CNN | CNN | Tried two architectures: shallow and deep CNNs | Shallow CNN tailored to decode band powers | Yes | 600 x 21 | Deep: 5 conv layers Shallow: 1 conv layer | 5 | ELU | N/M | N/M | 2 | N/M | Standard optimization | Standard | Adam | Adam | Used SMAC | N/M | SMAC | Yes | N/M | Binary cross-entropy | Inter | 10-Fold CV | k-fold | Train: 5480 (~90%) Test: 554 (~10%) | Accuracy, Sensitivity, and Specificity | accuracy, sensitivity, specificity | N/M | < 3.5 h | Accuracy: 85.4% (deep), 84.5% (shallow) Sensitivity: 75.1% (deep), 77.3% (shallow) Specificity: 94.1% (deep), 90..5% (shallow) | CNN and linear model with band-power features as input | DL & Trad. | Wilcoxon signed-rank test | Effect of spectral perturbations of the input on the resulting prediction | Input-perturbation network-prediction correlation maps | Perturbation visualizations showed that the CNNs used information related to changes in delta and theta bands. Suprisingly, shorter length EEG recordings yielded better accuracies. | "Still, to yield more clinically useful insights and diagnosis explanations, further improvements in ConvNet visualizations are needed." | Yes | GitHub | No | Isabela Albuquerque | Hubert Banville | TBC | Schirrmeister2017a | ||||||||||||
| 24 | Predicting sex from brain rhythms with deep learning | 2018 | van Putten, Olbrich & Arns | Scientific Reports (Nature) | No | Journal | University of Twente | Netherlands | 7 | Classification of EEG signals | Personal trait/attribute | Sex | New Approach: Detecting Sex from RS EEG with DL (CNN) | Predicting an individual's sex from their EEG | Resting State EEG. | No need for engineered features, and "have potential to detect subtle differences in otherwise similar patterns". | N/M | Raw EEG | Brain Resource Int'l Database | Public | 1308 subjects x 40 segments x 2s (2s windows, no overlap) | 52320 | 1744 | 1308 | 24 | 128 | 1) Downsampled to 128Hz (from 500Hz) 2) Band-Pass Filter: 0.5-25Hz | Yes | EOG regression | Yes | Raw EEG | Raw EEG | N/M | Windows 10 Keras, Tensorflow Python 3.6 | CNN | CNN | None | N/M | Yes | 256 x 24 (Samples x Channels) 2s epoch | 6 | 6 | ReLU | Dropout | Yes | 1 0: Female | 1: Male (2 from schema) | 9,051,902 | Standard optimization | Standard | Adamax | Other | LR=0.002, B1=0.9, B2=0.999, e=10^8, decay=0.00 | 70 | N/M | N/M | N/M | Categorical Cross-Entropy | Inter | No | No | Train: 1000 subjects Test: 308 subjects | Accuracy | accuracy | NVidia GTX-1060 | N/M | 81% (of correct classification over all subjects) | LR | Traditional pipeline | Permutation test | Visualization of learned filters through Deep Dream-like backprop on inputs | Generating input to maximize activation | While not all details of the features used for classification by the deep net have been revealed, our data show that differences in brain rhythms between sexes are mainly in the beta frequency range. | N/M | No | N/A | No | Yannick Roy | Hubert Banville | TBC | VanPutten2018b | ||||||||||||
| 25 | Deep learning with EEG spectrograms in rapid eye movement behavior disorder | 2018 | Ruffini, Ibanez, Castellano, Dubreuil, Gagnon, Montplaisir & Soria-Frisch | BioarXiv | Yes | Preprint | NeuroElectrics University of Montreal | Canada | 10 | Classification of EEG signals | Clinical | Sleep | Abnormality detection | New Approach | Using DCNN for Rapid Eye Movement Behavior Disorder | Resting State EEG. | Exploiting compositional structure in data | N/M | Raw EEG | Internal Recordings | Private | (118 + 74) = 192 subjects 148 windows of 1s per subject (1s windows) | 28416 | 473.6 | 192 | 14 | 256 | Offline | 1) Band-Pass Filter: 0.3 and 100 Hz [Hardware] 2) Notch Filter: 60Hz [Hardware] ((FFT) after detrending blocks of 1 second with a Hann window (FFT resolution is 2 Hz)) | Yes | N/M | N/M | Spectrogram Frames | Frequency-domain | z-score | Tensorflow | DCNN | CNN | Conv-Pooling-Dropout | N/M | Yes | 14 x 21 x 20 Channels x FFTBins x Epochs | 5 | 5 | ReLU | Dropout | Yes | 2 | Parkinson's disease Healthy | 2 | N/M* | Standard optimization | Standard | N/M | N/M | N/M | N/M | N/M | N/M | Random replication of subjects in the minority class | Cross-Entropy | Inter | Leave-Pair-Out (one subject for each class) | Leave-One-Subject-Out | N/M | Accuracy ROC AUC | accuracy, ROC AUC | N/M | N/M* | Net: Problem [ N ] ACC (AUC) DCNN: HC vs PD [2x73 / 2x1] 79% (87%) RNN: HC vs PD [2x73 / 2x1] 81% (87%) DCNN: HC+RBD vs PD+DLB [2x159 / 2x1] 73% (78%) RNN: HC+RBD vs PD+DLB [2x159 / 2x1] 72% (77%) | Stacked RNN Shallow CNN | DL | No | Maximizing network outputs for a given class | Generating input to maximize activation | Although here, as in [28], we worked with time-frequency pre-processed data, the field will undoubt- edly steer towards working with raw data in the future when larger datasets become available—as suggested in [21] | "We note that one of the potential issues with our dataset is the presence of healthy controls without follow up, which may be a confound. \We hope to remedy this by enlarging our database and by improving our diagnosis and follow up methodologies" | No | N/A | Yes | Yannick Roy | Isabela Albuquerque | Yes | Ruffini2018a | ||||||||
| 26 | Deep transfer learning for error decoding from non-invasive EEG | 2018 | Völker, Schirrmeister, Fiered, Burgard & Ball | IEEE International Conference on Brain-Computer Interface | Yes | Conference | University of Freiburg | Germany | 6 | Classification of EEG signals | BCI | Reactive | ERP | New approach: Exploring Transfer Learning for BCI. | Using CNN on 2 different BCI tasks, can it generalize? Transfer Learning across subjects and across tasks | 1) Eriksen Flanker Task 2) Online GUI to control intelligent robots | Enables transfer learning | N/M | 1) Error 2) Mental tasks (MI) | Internal Recordings | Private | 1) 1000 trials x 1.5s x 31 subjects 2) (3032 +/- 818) x 4 x 1.5s 1.5s / epoch (onset at 0.5s) | 31000; 12128 | 775; 303.2 | 31; 4 | 128; 64 | N/M | 1) Re-referenced to Common Average (CAR) 2) Resampled to 250Hz | Yes | N/M | N/M | Raw EEG | Raw EEG | Electrode-wise exponential running standardization | Python BrainDecode Scikit-learn | CNN | CNN | N/M | N/M (see BrainDecode paper) | No | N/M (Raw EEG windows) | N/M (see BrainDecode paper) | N/M | N/M | N/M (see BrainDecode paper) | N/M | N/M | N/M | N/M (See braindecode paper) | N/M | N/M (see BrainDecode paper) | N/M | N/M | N/M | N/M | N/M | N/M | N/M (see BrainDecode paper) | Both | Withing-Sub: Leave-One-Session-Out CV Between-Sub: Leave-One-Subject-Out CV | Leave-One-Session-Out; Leave-One-Subject-Out | Within-Sub Train: 80% Within-Sub Test: 20% Between-Sub Train: N-1 Sub. Between-Sub Test: 1 Sub. | Normalized Accuracy | normalized accuracy | N/M | N/M | Between-Subject Transfer Learning Flanker Task: 81.7% Normalized Accuracy GUI Robots Task: Poor results, because only 4 subjects. Between-Paradigms Transfer Learning Both failed. ~50% | rLDA (CNN outperforms rLDA) Also, best result ever reported on the Error Detection on Flanker Task* | Traditional pipeline | Paired t-tests | Input-perturbation network-prediction correlation maps | Input-perturbation network-prediction correlation maps | (1) As a next step, techniques including data augmentation and automated hyper-parameter and architecture search might help to improve the generalization of deep ConvNets. (2) For a generalization to new subjects, our data suggest that a training subject group of at least 15 subjects might be necessary for reliable error decoding on unknown subjects. (3) In the flanker task, our deep ConvNets achieved the highest to date reported average accuracy. | N/M | No | N/A | Yes | Yannick Roy | Hubert Banville | TBC | Volker2018 | |||||||||||
| 27 | DeepIED: An epileptic discharge detector for EEG-fMRI based on deep learning | 2018 | Hao, Khoo, von Ellenrieder, Zazubovits & Gotman | NeuroImage: Clinical | No | Journal | McGill University, Osaka University | Canada | 14 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA | Detect interictal epileptic discharges in noisy EEG data collected during an fMRI recording | Resting state EEG - Seizures. | Reduce the amount of time it takes to manually label interictal epileptic discharges | BrainAmp (BrainProducts) | Raw EEG | Internal Recordings | Private | 67 patients (148 studies) Average study time: 50 min (range, 18–72 min) (~1s windows) | 201000 | 7400 | 67 | 25 | 200 | Offline | 1) Bandpass 0.5-50 Hz 2) fMRI-induced artefact removal 3) Electrode-wise exponential running standardization [6] was applied with a decay factor of 0.999 4) BCG artifact removal (ballistocardiographic) | Yes | N/M | N/M | Raw EEG | Raw EEG | N/M | N/M | CNN (ResNet) | CNN | - | - | Yes | 25 x [16 to 256] | 31 | 31 | ReLU | Dropout on penultimate layer (50%) | Yes | N/M (different EID types) | EIDs | 128 (FC) going to softmax and triplet (real output N/M) | 999,920 | Standard optimization | Standard | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Softmax for multi-class classification Triplet loss function | Inter | No | No | Train: 30 subjects Test: 37 subjects | ROC curves Sensitivity False positive rate | ROC, sensitivity, false positive rate | N/M | N/M | Median sensitivity: 84.2% False positive rate: 5 events/min | Cross-correlation (template-based) method for finding similar EEG epochs | Traditional pipeline | One-Way Anova + Post Hoc paired t-test | No | No | In their tests, they asked experts to edit the outputs of the net and reject false positives; they argue that it's a necessary step and that it is not too time-consuming. | - | No | N/A | No | Hubert Banville | Yannick | TBC | Hao2018 | ||||||||
| 28 | Deep learning for hybrid EEG-fNIRS brain–computer interface: application to motor imagery classification | 2018 | Chiarelli, Croce, Merla & Zappasodi | Journal of Neural Engineering | No | Journal | G. d'Annunzio' University | Italy | 12 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA | Improving MI classification with DL in multimodal system | Motor Imagery | High performance on other tasks | (EGI) | ERD/ERS | Internal Recordings | Private | 40 trials (C1: 20 / C2: 20) of 5s 200 samples x 15 subjects (1s windows) | 3000 | 50 | 15 | 123 | 250 | Offline | 1) Bandpass 8-30 Hz | Yes | N/M | N/M | Power in the mu-beta range, averaged across 1-s | Frequency-domain | N/M | TensorFlow | Fully-connected NN | FC | N/M | N/M | Yes | 123 x 1, 16 x1, or 139 x 1 | 5 | 5 | ReLU | Dropout (0.75) | Yes | 2 | Right-hand MI, Left-hand MI | 2 | N/M | Standard optimization | Standard | Adam | Adam | LR=1e-4, B1=0.9, B2=0.999, constant=1e-8 | 90 | N/M | N/M | N/M | Cross-Entropy | Intra | 10-Fold CV (1000x) | k-fold | Train: 180 Test: 20 | Accuracy | accuracy | N/M | N/M | EEG only: ~70%, NIRS only: ~77%, EEG+NIRS: ~83% | LDA, linear SVM | Traditional pipeline | 2-way repeated measurement ANOVA + post-hoc analysis | No | No | DNN worked better than CNN, RNN not tested. | RNN was not tested | No | N/A | No | Hubert Banville | TBR | TBC | Chiarelli2018 | ||||||||
| 29 | Preference Classification Using Electroencephalography (EEG) and Deep Learning | 2018 | Teo, Hou & Mountstephens | Journal of Telecommunication, Electronic and Computer Engineering (JTEC) | No | Journal | University Malaysia Sabah | Malaysia | 5 | Classification of EEG signals | Monitoring | Affective | Emotion | Improve SOTA | Improving classification of preference (like vs. dislike), and overcoming intra- and inter-subject variability | Rating of 3D Stimulus (1: like very much, 2: like, 3: undecided, 4: do not like, 5: do not like at all) | N/M | B-Alert X10 (ABM) | Raw EEG | Internal Recordings | Private | 208 trials: 9s + [5-15]s, from 16 subjects (full trial as windows) 10 other subjects were for kNN (not counted) | 208 | 65.87 | 16 | 9 | N/M | Offline | 1) Notch Filter: 50Hz | Yes | Proprietary artefact rejection and interpolation | Yes | 45 features (PSD for each channel) D (1-3Hz), T (4-6Hz), A (7-12Hz), B (13-30Hz), G (31-64Hz) | Frequency-domain | N/M | R | DNN | FC | N/M | N/M | No | 47 | 2 | 2 | ReLU | N/M | N/M | 2 | Like very much Do not like at all | N/M | N/M | Standard optimization | Standard | Adadelta | Other | N/M | N/M | N/M | N/M | N/M | Cross-Entropy | Inter | 10-Fold CV | k-fold | N/M | Accuracy | accuracy | N/M | N/M | 63.99% | SVM Linear: 60.19%, SVM Radial: 59.67%, OneR: 59.00%, Adaboost: 58.65%, Random Forest: 57.74%, NNet: 57.71%, JRip: 57.21%, Naive Bayes: 56.79%, C5.0: 56.74%, kNN (k = 5): 56.29% | Traditional pipeline | No | No | No | "An initial study using kNN provided sufficiently good results in a 10-subject study. However, when expanded to a larger cohort size of 16 subjects, the results were not encouraging. However, the use of deep learning was able to observably overcome some of the difficulties presented by inter-subject variability posed by larger cohort sizes in EEG-based preference classification." | Intersubject variability | No | N/A | No | Yannick Roy | Hubert Banville | Yes | Teo2018 | ||||||||
| 30 | An Automated System for Epilepsy Detection using EEG Brain Signals based on Deep Learning Approach | 2018 | Ullah, Hussain, Qazi &Aboalsamh | Arxiv | Yes | Preprint | National University of Ireland King Saud University | Ireland | 18 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA | Improving ternary classification of ictal vs. normal vs. interictal windows | Resting State, Eyes Open, Eyes Closed, Seizures. | Automatic feature learning | N/M | Raw EEG | Bonn University | Public | Bonn University (5 sets x 100 x 23.6s) Each 100 samples --> 800 windows (512 points windows, 6.25% overlap) | 4000 | 197 | 15 | 1 | 173.6 | Offline | N/M | N/M | N/M | N/M | Raw EEG | Raw EEG | z-score | TensorFlow | Pyramidal 1D-CNN (P-1D-CNN) | CNN | No pooling | 1D convolution motivated by EEG being a "1D signal" | Yes | 8 EEG windows Raw EEG (1 channel) | 3 Conv + 2 FC | 5 | ReLU | Dropout (0.5) Batch norm | Yes | 2 or 3 | 2: Epileptic vs. non-epileptic 3: normal vs. ictal vs. interictal | 2 or 3 Classes (Softmax) | N/M | Standard optimization | Standard | Adam | Adam | LR=0.001, B1=0.9, B2=0.999, epsilon=0.00000001, locking=false | N/M | N/M | N/M | Overlapping windows (87.5% and 25% overlap) | Cross-Entropy | Inter | 10-Fold CV | k-fold | Train: 90% Test: 10% | Accuracy, Specificity, Sensitivity, Precision, f-measure, and g-mean. | accuracy, specificity, sensitivity, precision, f-measure, g-mean | N/M | N/M | 99.1 ± 0.9% (for 3 classes problem) The mean accuracy of the proposed system is 99.6% for all the sixteen cases/ Many results (see papers) comparing Binary / Tenary classifications. | Random forests, Naive Bayes, kNN | Traditional pipeline | No | No | No | "According to our knowledge until this date, DL approach has never been used for this problem. The mean accuracy of the proposed system is 99.6% for all the sixteen cases (shown in Table 8 last column), which figures out the generalization power of the proposed system." | Small datasets | No | N/A | Yes | Yannick Roy | Hubert Banville | TBC | Ullah2018 | ||||||||
| 31 | A Novel Channel-aware Attention Framework for Multi-channel EEG Seizure Detection via Multi-view Deep Learning | 2018 | Yuan, Xun, Ma, Suo, Xue, Jia & Zhang | IEEE EMBS International Conference on Biomedical & Health Informatics (BHI) | No | Conference | Beijing Laboratory of Advanced Information Network State University of New York at Buffalo | China | 4 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA | Use end-to-end model with attention mechanism to select channel and detect seizures. | N/M | "Explore inherent EEG representations" | N/M | Raw EEG | CHB-MIT | Public | CHB-MIT: 9 out of 23 subjects 4302 EEG fragments (windows length = ??) | 4302 | N/M | 9 | 23 | 256 | Offline | N/M | N/M | N/M | N/M | Spectrogram (STFT) | Frequency-domain | N/M | N/M | 2 SAEs | AE | Channel Encoders (SAE) Global Encoder (SAE) + Attention | N/M | Yes | N/M | 2 | 2 | N/M | Dropout | Yes | 2 | No seizure Seizure | 2 | N/M | [Not clear!] Unsupervised pretraining, followed by fine-tuning with softmax layer? | Standard | Adam | Adam | N/M | N/M | N/M | N/M | N/M | Cross-Entropy | Inter | Holdout | Holdout | N/M | F1-Score Accuracy AUC of ROC and precision-recall curves | f1-score, accuracy, ROC AUC, PR AUC | N/M | N/M | [F1-score] - Channel Attloc: 0.9781, Channel Attglo: 0.9785 [Accuracy] - Channel Attloc: 0.9651, Channel Attglo: 0.9661 | PCA+SVM (PSVM) SAEs + attention DNN + hard channel selection | DL & Trad. | N/M | Analysis of mean attention score values for a single subject | Analysis of activations | "To the best of our knowledge, this is the first work using attention mechanism for biosignal channel selection in healthcare." | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Yuan2018a | |||||||||
| 32 | Compact Convolutional Neural Networks for Classification of Asynchronous Steady-state Visual Evoked Potentials | 2018 | Waytowich, Lawhern, Garcia, Cummings, Faller, Sajda, Vettel | Journal of Neural Engineering | Yes | Journal | U.S. Army Research Laboratory Lab for Intelligent Imaging and Neural Comp. University of Pennsylvania University of California, Santa Barbara | USA | 21 | Classification of EEG signals | BCI | Reactive | SSVEP | Improve SOTA | Use ConvNet for SSVEP classification | SSVEP (12 classes!) | Automatic feature learning without domain-specific information | ActiveTwo (BioSemi) | SSVEP | Internal Recordings | Public | 10 subjects x 15 block x 12 trials x 4s (1s windows) | 7200 | 120 | 10 | 8 | 2048 | Offline | 1) Bandpass 9-30 Hz 2) Downsampled to 256 Hz | Yes | N/M | N/M | Raw EEG | Raw EEG | TensorFlow, Keras Original Stimuli (from 2015) on MATLAB with Psychophysics Toolbox | EEGNet | CNN | Filter banks (temporal convolutions) followed by spatial filters | N/M | Yes | 8 channels x 256 samples | 3 | 3 | ELU | Batch norm Dropout (0.25) | Yes | 12 | 12 different combinations of frequency and phase | 12 | 46,476 (45,900 trainable) | Standard optimization | Standard | Adam | Adam | N/M | 64 | N/M | N/M | N/M | Categorical Cross-Entropy | Inter | Leave-One-Subject-Out | Leave-One-Subject-Out | Train: 90% Test: 10% | Accuracy | accuracy | N/M | N/M | ~90% for 7/10 Subjects. 60%, 75%, 30% for the others. (chance = 8%) | CCA (Canonical Correl. Analysis) C-CCA (Combined CCA) | Traditional pipeline | Paired t-tests when comparing to baseline | Visualization of feature activations with t-SNE | Analysis of activations | "Although unexpected, these within-class clusters highlight the strength of the deep learning approaches to learn diagnostic features directly from the data." | Experiment did not includ a non-control state | Yes | GitHub | No | Yannick Roy | Hubert Banville | Yes | Waytowich2018 | |||||||||
| 33 | Deep Classification of Epileptic Signals | 2018 | Ahmedt-Aristizabal, Fookes, Nguyen & Sridharan | Arxiv | Yes | Preprint | Queensland University of Technology | Australia | 4 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve State-of-the-Art: Using LSTM for Epilepsy classification | End-to-end seizure detection | Resting State, Eyes Open, Eyes Closed, Seizures. | Automatic feature learning | N/M | Raw EEG | Bonn University | Public | Bonn University (5 sets x 100 x 23.6s) (full 4096 points / 23.6s as windows, no overlap) | 500 | 197 | 15 | 1 | 173.6 | Offline | None, but the Bonn University Dataset already has some preprocessing. | Yes | N/M | N/M | Raw EEG | Raw EEG | N/M | Keras | LSTM | RNN | N/M | N/M | Yes | 100 x 4096 (100 samples of 4096 segments) | Model 1: 1 LSTM + 1 Dropout Model 2: 2 LSTM + 2 Dropout + 1 FC | 3 | N/M | Dropout (0.35) | Yes | 2 | No seizure Seizure | 1 | Model 1: 16,961 Model 2: 116,033 | Standard | Standard | Adam | Adam | LR: 1e-3, b1:0.9, b2:0.999 | 4 | N/M | N/M | No | Binary Cross-Entropy | Inter | 10-Fold CV | k-fold | Train: 70% Valid: 20% Test: 10% | Accuracy, Sensitivity, Specificity, Precision and the Area Under the Curve (AUC). | accuracy, sensitivity, specificity, precision, ROC AUC | N/M | N/M | Accuracy: [Valid] 95.54% [Test] 91.25% Sensitivity: [Test] 91.83% Specificity: [Test] 90.50% Precision: [Test] 91.50% AUC: [Test] 0.9582 | None | None | No | No | No | "We experimented with various numbers of memory cells in each layer and obtained the best performance with a network configured with one single layer with 64 hidden units (Model 1) and with 2 hidden layers of 128 and 64 hidden units respectively (Model 2)" | N/M | No | N/A | Yes | Yannick Roy | Isabela Albuquerque | TBC | Ahmedt-Aristizabal2018 | ||||||||
| 34 | Emotion Recognition from EEG Using RASM and LSTM | 2018 | Li, Tian, Shy, Xu & Hu | International Conference on Internet Multimedia Computing and Service | No | Conference | South China University of Technology Lanzhou University | China | 9 | Classification of EEG signals | Monitoring | Affective | Emotion | Improve SOTA | Using rational assymetry (RASM) as features and LSTM as classifier on DEAP dataset for emotion classification. 2 Classes (Positive / Negative Valence) | Watching emotional movies (clips) | LSTM to capture temporal dependencies in emotions | N/M | Raw EEG | DEAP | Public | DEAP 895 Trials x 125 windows 63s each trial to 125 windows (1s windows, 50% overlap) | 111875 | 939.75 | 32 | 32 | 256 | Offline | None | No | No | No | RASM14 (STFT + Hanning Window --> 4 Freq Bands) | Frequency-domain | N/M | N/M | LSTM | RNN | N/M | "In our assumption, emotions change continuously, and this continuity is reflected in the temporal correlations of EEG signals. To explore the correlations, the classification method of Long Short-TermMemory networks (LSTM) is adopted." | Yes | 125 * 14 * 4 (segments * pairs * bands) | 1 | 1 | N/M | Dropout (0.5) | Yes | 2 | Positive valence Negative valence | 1 | N/M | Standard | Standard | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Inter | 10-Fold CV | k-fold | N/M | Accuracy | accuracy | N/M | N/M | RASM + LSTM: 76.67 (Accuracy) RASM + SVM: 65.62 (Accuracy) Zhang, 2016: 69.67 (Accuracy) Chen, 2015: 73.00 (Accuracy) Li X, 2016: 72.06 (Accuracy) | SVM Zhang [10] (DE + GELM) Chen [2] (Fusion feature + HMM) Li [6] (Wavelet energy + CRNN) | DL & Trad. | No | No | No | "Although the accuracy of our experiment is more than 75%, it is not good enough for applications. The task of the future work is to improve the recognition accuracy. More features will be tried especially those reflect the characteristics of EEG signals in frequency-space domain." | N/M | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | Li2018 | ||||||||
| 35 | EEG detection and de-noising based on convolution neural network and Hilbert-Huang transform | 2018 | Wang, Guo, Zhang, Bai & Wang | International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI) | No | Conference | Changchun University of Science and Technology Jilin Engineering Research Center of RFID and Intelligent Information Processing | China | 6 | Improvement of processing tools | Signal cleaning | Artifact handling | New Approach | Denoising EEG with Hilbert-Huang Transform after a detection of (yes/no) EOG artifact from a CNN classifier | N/M* | Nonlinearity of EEG | N/M | Raw EEG | Internal Recordings | Private | 2000 training data and 100 test data (3s windows) | 2100 | 105 | N/M | -1 | 1000 | N/M* | N/M | N/M* | N/M | IMF / HHT | Other | N/M | N/M | CNN | CNN | 2*2 convolution kernels | N/M | Yes | "characteristic matrix of the extracted instantaneous power" | 1 | 1 | Softmax | N/M | N/M | 1 EOG artifact yes/no (softmax) | N/M | N/M | N/M | N/M* | N/M | N/M | N/M | N/M | N/M | No | N/M* | Inter | No | No | Train: 2000 Test: 100 | Accuracy | accuracy | N/M | N/M | 80% | No | None | No | No | No | The results show that the method in this paper takes a little longer CPU time compared with the traditional wavelet de- noising [4] and HHT de-noising alone. But the signal-to-noise ratio after de-noising is obviously higher than the other two methods. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Wang2018a | ||||||||||||
| 36 | Data Augmentation for EEG-Based Emotion Recognition with Deep Convolutional Neural Networks | 2018 | Wang, Zhong, Peng, Jiang & Liu | International Conference on Multimedia Modeling | No | Conference | Shenzhen University The Hong Kong Polytechnic University | China | 12 | Generation of data | Data augmentation | New Approach: Data augmentation on Emotion datasets for Deep learning models | Data augmentation on SEED & MAHNOB-HCI dataset and evaluation using ResNet & LeNet. | Watching emotional films/clips | Data augmentation for deep models with many parameters | N/M | Emotions (Frequency Features) | SEED; MAHNOB HCI | Public | DS #1 - SEED: 630 EEG segments, from 14 subjects 14 subjects x 15 videos x 3 sessions x 4 min splitted into 3x62s each for 1890 segments total DS #2 - MAHNOB-HCI: Between 34.9 - 117s (avg at: 75.95) 188 negative, 208 neutral and 131 positive segments (1s windows, no overlap) | 117180; 40025 | 1953; 667.1 | 14; 30 | 62; 32 | N/M | Offline | 1) Downsampled to 200Hz 2) Band-pass filters: 5 freq bands 3) SFTF with non-overlapping Hamming window 1s | Yes | Manual removal | Yes | Differential Entropy (DE) per band | Other | N/M | MATCONVNET | ResNet LeNet | CNN | Data Augmentation Paper, it's not about these networks. | Data augmentation with Gaussian Noise of various std | No | n x l x 5 n: electrodes l: length (time) 5: Freq Bands DE | LeNet: 5 ResNet: 14 | 14 | N/M | N/M | N/M | 3 | positive, neutral, negative | 1) 3 2) 3 | 1) 4,000 2) 20,000 | Standard optimization with augmented data | Standard | N/M | N/M | lr = 0.1 | 100 | N/M | N/M | Gaussian Noise (augmented up to 30 times) | N/M | Inter | No | No | 1 - SEED) Train: 1134 1 - SEED) Test: 756 2) N/M | Accuracy | accuracy | N/M | N/M | DS #1) LeNet: [Pre] 49.6% | [Post] 74.3% DS #1) ResNet: [Pre] 34.2% | [Post] 75.0% DS #2) ResNet: [Pre] 40.8% | [Post] 45.4% DS #2) LeNet: N/M | DS #1) SVM: [Pre] 74.2% | [Post] 73.4% DS #1) PCA-SVM: [Pre] 49.8% | [Post] N/M% DS #2) SVM: [Pre] 42.5% | [Post] 44.3% | Traditional pipeline | No | No | No | By analyzing the experimental result, we find that the data augmentation method can effectively improve the performance of deep models. In future, we will seek to use other data augmentation methods, such as generative adversarial networks, to generate more effective samples of EEG data and improve the performance of EEG-based emotion recognition. | N/M | No | N/A | Yes | Yannick Roy | Hubert Banville | TBC | Wang2018 | ||||||||||
| 37 | A convolutional neural network for sleep stage scoring from raw single-channel EEG | 2018 | Sors, Bonnet, Mirek, Vercueil & Payen | Biomedical Signal Processing and Control | No | Journal | Université Grenoble Alpes CEA Leti, MINATEC Campus (Grenoble) Dijon University Hospital (Dijon) Grenoble University Hospital | France | 8 | Classification of EEG signals | Clinical | Sleep | Staging | New Approach: Sleep Stage Scoring (5 stages) with CNN on Single EEG Channel | Use CNNs on raw EEG data for 5-class sleep prediction | Sleep | CNNs have presented good performance in other domains and other EEG tasks. | N/M | Raw EEG | SHHS | Public | Dataset SHHS-1 (5793 polysomnographic records) 5,384,401 epochs of 30s ~ 5 years of data! (30s windows) | 5384401 | 2692200 | 5728 | 1 | 125 | Offline | None | No | No | No | Raw EEG | Raw EEG | No | TensorFlow | CNN | CNN | (no mention of pooling or dropout) | 1D convolutional layers | Yes | (3750 * 4) x 1 30s epoch + 2 preceding + 1 following 30s @ 125Hz = 3750 samples | 12 Conv Layers + 1 FC (256) + 1 FC (5 classes) | 14 | Leaky ReLU | N/M | N/M | 5 | Wake N1 N2 N3 REM | 5 [prob for each class] (Softmax) | N/M* | Standard optimization | Standard | Adam | Adam | lr = 3 ×10^−5, b1 = 0.9, b2 = 0.999 | 128 | N/M | N/M | Tried cost-sensitive learning and oversampling (didn't improve. didn't use it.) | Multiclass Cross-Entropy | Inter | Train-Valid-Test | Train-Valid-Test | Train: 50% Valid: 20% Test: 30% | Accuracy | accuracy | NVidia GTX980Ti | N/M* | 87% | Tsinalis [15] CNN: 0.75 Supratak [16] CNN-LSTM: 0.86 Liang [9] [...] : 0.88 Zhu [10] DVG, SVM: 0.85 Fraiwan [6] T-F, RF: 0.83 Hassan [38] EMD, Ensemble: 0.87 Hassan [11] EMD, [...]: 0.89 Hassan [12] PSD, RF: 0.88 Hassan [39] EMD, [...] : 0.83 Sharma [13] Iterative filtering: 0.88 Hsu [14] Energy, RNN: 0.90 | DL & Trad. | No | Visualization of synthetic inputs that maximize class probability | Generating input to maximize activation | "This study shows that it is possible to classify sleep stages using a single EEG channel and a convolutional neural network work- ing on raw signal samples without any feature extraction phase and with performance on par with other state-of-the-art methods." "Further research is necessary to address class imbalance. Ensemble learning [35] or CNN-specific methods [36] may prove suitable" | N/M | Yes | GitHub | No | Yannick Roy | Isabela Albuquerque | Yes | Sors2018 | ||||||||
| 38 | ChronoNet: A Deep Recurrent Neural Network for Abnormal EEG Identification | 2018 | Roy, Kiral-Kornek & Harrer | Arxiv | Yes | Preprint | IBM Research - Australia | Australia | Classification of EEG signals | Clinical | Pathological EEG | Improve SOTA | Detect abnormal EEG with a new end-to-end architecture | ? | Automatic interpretation of EEG from raw data | N/M | Raw EEG | TUH Abnormal EEG Corpus | Public | TUH Abnormal EEG Corpus Training set: 1361 abnormal/1379 normal sessions Test set: 127 abnormal/150 normal session Became: 14,971 / 15,169 windows for training (1min windows) | 30417 | 30417 | N/M | 22 | 250 | None | No | N/M | N/M | None | Raw EEG | N/A | N/M | 1) Conv+GRU 2) Inception Conv+GRU 3) Dense Conv+GRU 4) Inception Dense Conv+GRU | CNN+RNN | Conv filter sizes grow exponentially inside a given layer (e.g., 2, 4, 8) | - | Yes | 15000 x ? | 1) 7 2) 7 3) 7 4) 7 | 7 | N/M | N/M | N/M | 2 | N/M | Standard optimization | Standard | Adam | Adam | 500 epochs | 64 | N/M | N/M | N/M | N/M | Inter | 5-Fold CV | k-fold | Train: 90.8% Test: 9.2% | Accuracy | accuracy | N/M | N/M | 1) 82.31 2) 84.11 3) 83.89 4) 86.57 | CNN-MLP: 78.80 DeepCNN: 85.40 | DL | No | No | No | The ChronoNet architecture is a general-purpose architecture for time series - has been applied to speech data classication. | - | No | N/A | No | Hubert Banville | TBR | Yes | Roy2018 | |||||||||||||
| 39 | EEG-GAN: Generative adversarial networks for electroencephalograhic (EEG) brain signals | 2018 | Hartmann, Schirrmeister & Ball | Arxiv | Yes | Preprint | University of Freiburg | Germany | 7 | Generation of data | Generating EEG | Generate EEG signals | Generate EEG signals using GANs | Motor imagery | GANs are good at generating data | N/M | Raw EEG | Internal Recordings | Private | 438 EEG Signals They don't talk about the lenght of these signals. They show plot of -500ms to 2500ms (probably windows of 3s) | 438 | 21.9 | N/M | 1 | 250 | None | No | No | No | None | Raw EEG | Subtract mean then divide by maximum absolute value | N/M | Wassertein GAN (modified) | GAN | - | Conv layers instead of autoregressive model, as it worked well in the authors's other papers | Yes | Gen: 200 Discr: 768 | Gen: 14 Discr: 14 | 14 | Leaky ReLU | Gradient penalty | Yes | N/A | N/A | Gen: 768 Discr: 1 | N/M | GAN optimization with increasing resolutions | Other | Adam | Adam | "Equalized learning rate" lr = 0.001 beta1 = 9 beta2 = 0.99 | ? | N/M | N/M | No | Improved Wassertein distance | Inter | No | No | Train: 286 Valid: 72 Test: 80 | Inception score Frechet inception distance Euclidean distance Sliced Wassertein distance | inception score, frechet inception distance, euclidean distance, sliced Wasserstein distance | N/M | N/M | [Many values] | WGAN with gradient penalty | DL | No | Visual inspection of generated segments (time series distribution, spectrum distribution, examples) | Analysis of generated outputs | The metrics did not correlate with visual performance, and so the authors recommend using many metrics to obtain a balanced view | Mode collapse in GANs | No | N/A | No | Hubert Banville | Isabela Albuquerque | TBC | Hartmann2018 | |||||||||||
| 40 | Know Your Mind: Adaptive Brain Signal Classification with Reinforced Attentive Convolutional Neural Networks | 2018 | Zhang, Yao, Wang, Zhang, Zhang & Liu | Arxiv | Yes | Preprint | University of New South Wales, Tsinghua University, Michigan State University | Australia | Classification of EEG signals | Multi-purpose architecture | Make general framework for EEG classification | Apply a single architecture (reinforced attentive CNN) to EEG classification | 1 & 2: Motor imagery 3: Person identification 4: Pathology (seizure detection) | Skip time-consuming feature engineering and no task-specific classifier. | EPOC (Emotiv), N/M | 1 & 2) Motor Imagery 3) None 4) Seizures | eegmmidb; Internal Recordings; EEG-S; TUH | Both | DS #1 - eegmmidb: 20 x 28000 points (@160Hz) DS #2 - Internal: 7 x 34560 points (@128Hz) DS #3 - EEG-S: 8 x 7000 points (@160Hz) DS #4 - TUH: 5 x 12000 points (@250Hz) (windows of 1 point) | 560000; 241910; 56000; 60000 | 58.33; 31.5; 5.8; 4 | 20; 7; 8; 5 | 64; 14; 64; 22 | 160; 128; 160; 250 | None | No | N/M | N/M | None | Raw EEG | N/A | TensorFlow | CNN with attention + DQN | CNN | 1) Replicating and shuffling incoming samples 2) Attention mechanism trained with RL 3) CNN 4) Nearest-neighbour classifier | A) Replicate and shuffle operation intended to randomly unveil interesting spatial patterns | Yes | 1 x nb_channel | CNN: 3 DQN: 2 | 3 | ReLU & Sigmoid | L2 | Yes | 1) 5 2) 6 3) 8 4) 2 | N/M | Standard optimization (including reinforcement learning) | Standard | Adam | Adam | Learning rate: 0.001 | N/M | N/M | N/M | No | Cross-entropy | Inter | N/M | No | N/M | Accuracy, Precision, Recall, F1-score Latency Resilience | accuracy, precision, recall, f1-score, latency, resilience | N/M | 10 min | Accuracy 1) 0.9932 2) 0.9708 3) 0.9984 4) 0.9975 | Not clear what they were trained on (samples? features?): Linear SVM, Random Forest, kNN, LSTM, GRU, Adaptive boosting, LDA + 5 state-of-the-art papers for each (20 total) | DL & Trad. | No | No | No | Latency is comparable to other methods The number of channels used affects the performance. | - | Yes | GitHub | No | Hubert Banville | Yannick | TBC | Zhang2018a | ||||||||||||||
| 41 | Gated Recurrent Networks for Seizure Detection | 2018 | Golmohammadi, Ziyabari, Shah, Von Weltin, Campbell, Obeid & Picone | Arxiv | Yes | Preprint | Neural Engineering Data Consortium, Temple University | USA | 5 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA (their previous work) | Explore Gated RNN (LSTM & GRU), explore initialiazation and regularization of these networks | (see TUH dataset paper) | Improve their last results | N/M | Seizures | TUH Seizure Corpus | Public | TUH EEG Corpus (Train + Test | in sec) Seizures: 51,140 + 53,930 Non-Seizures: 877,821 + 547,728 (21s windows, no overlap) | 72886 | 25510.35 | 246 | 22 | 250 | None | No | N/M | N/M | LFCCs + First & Second Derivative of LFCCs | Other | N/A | N/M | 1) CNN + LSTM 2) CNN + GRU | CNN+RNN | 2D CNN to 1D CNN to bi-LSTM First LSTM output: 128 (1s data / epoch) Second LSTM output: 2-way sigmoid (classification of a 1s epoch) | 1) Gated units to avoid vanishing gradient. 2) RNNs to capture long-term dependencies. | Yes | 210 x 22 x 26 (Windows * Channels * Features) | 3x 2D CNN + 1x 1D CNN + LSTM | 5 | ELU | 1) L1 2) L2 3) L1/L2 4) Dropout 5) Guassian Noise | Yes | 1 (classification - sigmoid) | N/M | Initialization: The best performance is achieved using orthogonal initialization | Standard | Adam | Adam | N/M | N/M | N/M | N/M | No | MSE | Inter | No | No | Train: 928,962s Test: 601,659s | Sensitivity, Specificity | sensitivity, specificity | N/M | N/M | CNN + GRU - Sensitivity: 30.83% | Specificity: 91.49% CNN + LSTM - Sensitivity: 30.83% | Specificity: 97.10% Best Regulation: L1/L2 Best Initialization: Orthogonal | Compared CNN+GRU vs CNN+LSTM Compared 10 different initialization methods (see comments) Compared 5 different regularization methods (L1/L2, L1, L2, Gaussian noise, Dropout) | DL | No | No | No | LSTMs outperformed GRUs. We also studied initialization and regularizations of these networks. In future research, we are designing a more powerful architecture based on reinforcement learning concepts. We are also optimizing regularization and initialization algorithms for these approaches. Our goal is to approach human performance which is in the range of 75% sensitivity with a false alarm rate of 1 per 24 hours [11]. | No enough labeled data. Having certified specialist to label the data is very expensive, and hard to have people to do it. | No | N/A | Yes | Yannick Roy | TBR | TBC | Golmohammadi2017b | |||||||||||
| 42 | Optimizing Channel Selection for Seizure Detection | 2018 | Shah, Golmohammadi, Ziyabari, Weltin, Obeid & Picone | Arxiv | Yes | Preprint | Neural Engineering Data Consortium, Temple University | USA | Classification of EEG signals | Clinical | Epilepsy | Detection | Study the Impact of Number of Channels | Explore the impact of using/having from 2 to 22 channels with same network | (see TUH dataset paper) | Lower the number of EEG channels required (also save disk space) | N/M | Seizures | TUH Seizure Corpus | Public | TUH EEG Seizure Corpus (TUSZ) No more information about samples/time (1s windows) | N/M | N/M | N/M | 22 | 250 | None | No | N/M | N/M | LFCCs + First & Second Derivative of LFCCs | Other | N/A | N/M | CNN + LSTM | CNN+RNN | (same as there previous paper: Gated Recurrent Networks for Seizure Detection) | (same as there previous paper: Gated Recurrent Networks for Seizure Detection) | Yes | 210 x 22 x 26 (Windows * Channels * Features) | 3x 2D CNN + 1x 1D FC CNN + 2x Bi-LSTM | 5 | ELU & Sigmoid | Dropout | Yes | 1* (classification - sigmoid) | N/M | N/M | N/M | Adam | Adam | N/M | N/M | N/M | N/M | No | MSE | Inter | N/M | No | N/M | Sensitivity, Specificity | sensitivity, specificity | N/M | N/M | 22 Channels - Sensitivity: 39.15% | Specificity: 90.37% 20 Channels - Sensitivity: 34.54% | Specificity: 82.07% 16 Channels - Sensitivity: 36.54% | Specificity: 80.48% 8 Channels - Sensitivity: 33.44% | Specificity: 85.51% 4 Channels - Sensitivity: 33.11% | Specificity: 39.32% | No | None | No | No | No | The results presented in this paper use the Any Overlap scoring method [11] in which true positives are counted when the hypothesis overlaps with one or more reference annotations. False positives correspond to events in which the hypothesis annotations do not overlap with any of the reference annotations. This method of scoring is popular in the EEG research community. | - | No | N/A | No | Yannick Roy | TBR | TBC | Shah2017 | ||||||||||||
| 43 | Improving brain computer interface performance by data augmentation with conditional Deep Convolutional Generative Adversarial Networks | 2018 | Zhang & Liu | Arxiv | Yes | Preprint | Beijing Institute of Technology | China | Generation of data | Data augmentation | Generate EEG signals | Generate EEG signals using GANs for data augmentation | Motor Imagery (Left/Right Hand) | To increase amount of data available for training | N/M | Motor imagery | BCI Competition II - III | Public | BCI Competition II - III (1 Subject x 7 runs x 40 trials x 9 seconds) Used only 280 trials (140 training / 140 testing) Took 5s from each trial: 4s-9s (5s windows) | 280 | 23.3 | 1 | 3 | 128 | Offline | None | No | No | No | Continuous Wavelet transform (Morlet) Only keep 7-15 Hz (Time-Frequency Domain) | Frequency-domain | z-score | N/M | Augmentation: cDCGAN Classification: CNN | CNN | Conditional Deep Convolutional GAN (cDCGAN) + label information as input to both generator and discriminator | 2D kernel to accomodate input TFR | Yes | N/M | N/M | N/M | ReLU + Leaky ReLU + Sigmoid | N/M | N/M | 2 | Left Hand Rigth Hand | Same as input (not mentioned) | N/M | cDCGAN optimization Training CNN with real and artificial data | Standard | N/M | N/M | N/M | N/M | N/M | N/M | GAN [0.5 - 2x] (artificial EEG data) | N/M | Intra | No | No | Train: 50% Test: 50% | Accuracy | accuracy | N/M | N/M | No augmentation: ~83 % 50% of augmentation: ~84% 150% of augmetation: ~84% 200% of augmentation: ~85.5% | None | None | No | No | No | Data augmentation with GAN does help increasing accuracy when limited data is available. | Limited amount of data available per subject when training a BCI. | No | N/A | Yes | Hubert Banville | Yannick | Yes | Zhang2018b | |||||||||||
| 44 | Convolutional neural networks for seizure prediction using intracranial and scalp electroencephalogram | 2018 | Truong, Nguyen, Kuhlmann, Bonyadi, Yang, Ippolito & Kavehei | Neural Networks | Yes | Journal | University of Sydney Royal Melbourne Institute of Technology Swinburne University University of Melbourne University of Queensland University of Adelaide | Australia | Classification of EEG signals | Clinical | Epilepsy | Prediction | Improve SOTA | Use CNN to improve SOTA in seizure prediction | Ongoing recording with and without seizures | Test CNN on different epilepsy datasets | N/M | None (Seizures) | Freiburg Hospital; CHB-MIT; Kaggle: AESSPC | Public | DS #1 - Freiburg: 311h (59 seizures) DS #2 - CHB-MIT: 209h (64 seizures) DS #3 - AESSPC: 627h (48 seizures) (30s windows, no overlap) | 37320; 25080; 75240 | 18660; 12540; 37620 | 13; 13; 2 | 6; 22; 16 | N/M | 1) Removed Powerline: 47-53Hz + 97-103Hz 2) Removed DC Component (0Hz) | Yes | N/M | N/M | STFT (2D Freq x Time) 30s EEG windows | Frequency-domain | N/A | Python Keras Tensorflow | CNN | CNN | Batch Norm + Pooling 2 Dense | First, we keep the CNN architecture simple and shallow as described above (Ba & Caruana, 2014) | Yes | n x 59 x 114 (electrodes x time x freq) | CNN: 3 FC: 2 | 5 | ReLU Sigmoid Softmax | Dropout (50%) | Yes | 2 | N/M | We applied cost-sensitive learning by changing 300 the cost function in a way that the misclassification cost of preictal samples is multiplied by the ratio of interictal samples to preictal samples for each patient. We over-sampled the minority class. The cost-sensitive learning was for comparison. Though the two methods achieve comparable performance. | Standard | N/M | N/M | N/M | N/M | N/M | N/M | Overlapping windows (overlap % subject-specific to match classes) | N/M | Intra | Leave-One-Seizure-Out | Leave-One-Sample-Out | Train: N/M Valid: 25% of training Test: N/M | Sensitivity FPR (/h) | sensitivity, FPR | NVidia K80 | N/M | Measures (Epilepsy Specific): SOP of 30 min | SPH of 5 min DS #1) Sensitivity : 81.4% | FPR: 0.06/h DS #2) Sensitivity : 81.2% | FPR: 0.16/h DS #3) Sensitivity : 75.0% | FPR: 0.21/h | Compares on 3 Datasets Compares to 14 other SOTA (papers) | DL & Trad. | Wilcoxon signed-rank test | No | No | (1) Unbalanced Classes. (2) Comparing results with SOTA is complicated because each approach was tested with one dataset that is limited in the amount of data. | No | N/A | Yes | Yannick Roy | TBR | Yes | Truong2018 | |||||||||||||
| 45 | Semi-supervised Seizure Prediction with Generative Adversarial Networks | 2018 | Truong, Kuhlmann, Bonyadi & Kavehei | Arxiv | Yes | Preprint | University of Sydney University of Melbourne University of Queensland | Australia | Classification of EEG signals | Clinical | Epilepsy | Prediction | Improve SOTA | Use unlabelled data and data fusion to improve SOTA in seizure prediction | Ongoing recording with and without seizures | Leverage unlabelled data | N/M | Raw EEG | CHB-MIT; Freiburg Hospital | Public | DS #1 - Freiburg: 311h DS #2 - CHB-MIT: 209h (28s windows, no overlap) | 39985; 26871 | 18660; 12540 | 13; 13 | 16; 6 | 256 | STFT on 28-s windows with 50% overlap Removal of power line noise frequencies | Yes | N/M | N/M | STFT | Frequency-domain | N/M | Tensorflow | 1) GAN 2) CNN | Other | - | - | Yes | 1) GAN generator: 100 x1 2) GAN discriminator: n x 56 x 112 3) CNN: Same as discriminator | 1) GAN generator: 4 2) GAN discriminator: 3 3) Classifier: 2 | 4 | Softmax, Sigmoid | Dropout (50%) | Yes | 1) GAN generator: 2) GAN discriminator: 3) CNN | N/M | 1) Train GAN 2) Train 2 new FC layers on top of discriminator using labelled data | Other | N/M | N/M | N/M | N/M | N/M | N/M | Overlapping windows (overlap % subject-specific to match classes) | N/M | Both | Leave-One-Seizure-Out | Leave-One-Sample-Out | Train: N/M Valid: 25% of training Test: N/M | ROC AUC | ROC AUC | Nvidia P100 | N/M | AUC: 77.68% (CHBMIT), 75.47 (Freiburg) [6 and 12% less than benchmark] | CNN | DL | No | No | No | Although the performance decreased as compared to a standard CNN, the authors argue this can reduce the effort put into labelling the data. | - | No | N/A | No | Hubert Banville | TBR | Yes | Truong2018a | ||||||||||||
| 46 | Time Series Segmentation through Automatic Feature Learning | 2018 | Lee, Ortiz, Ko & Lee | Arxiv | Yes | Preprint | Princeton University | USA | Classification of EEG signals | Multi-purpose architecture | Improve SOTA | Detect changepoints/breakpoints in data (changes in signal) and apply to different types of time series data | Eye movements | Deep learning models for changepoint detection don't make assumptions about the underlying processes, as opposed to standard models | EPOC (Emotiv) | Eyes open vs. eyes closed | EEG Eye State | Public | EEG Eye State Dataset 117 seconds from 1 subject with Emotiv (14980 points, using windows of 25 points) | 600 | 2 | 1 | 14 | 256 | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Stacked Autoencoder | AE | - | - | No | N/M | 2 (encoder) | 2 | N/M | Tied weights in encoder and decoder L2 weight decay | Yes | 2 | N/M | Standard optimization | Standard | Stochastic gradient descent | SGD | N/M | N/M | N/M | N/M | No | Cross-entropy (or square loss?) | Intra | No | No | N/M | ROC Prediction loss (specific to task) MSE Prediction ratio | ROC, prediction loss, mse, prediction ratio | N/M | N/M | ROC curves... | Bayesian changepoint detection (based on Gamma or Gaussian priors) Pruned Exact Linear Time method Density-ratio estimation method | Traditional pipeline | No | No | No | Deep learning avoids typical problems in modelling changepoints. | - | No | N/A | No | Hubert Banville | TBR | TBC | Lee2018a | ||||||||||||||
| 47 | Investigating the Impact of CNN Depth on Neonatal Seizure Detection Performance | 2018 | O’Shea, Lightbody, Boylan & Temko | Arxiv | Yes | Preprint | Irish Centre for Fetal and Neonatal Translational Research, University College Cork | Ireland | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA | Use CNN to improve SOTA in neonatal seizure detection | Ongoing recording with and without seizures | Improve SOTA with CNN-11 based on their CNN-6 (2017) | N/M | Raw EEG | Internal Recordings | Private | 18 babies: over 800 hours of multichannel unedited EEG containing 1389 seizure (8s windows, 7s overlap) | N/M | N/M | 18 | 8 | 256 | Down-sample to 32Hz Filtered between 0.5 and 12.8Hz | Yes | No | No | 8 sec windows (1 sec shift) | Raw EEG | N/A | N/M | CNN | CNN | Conv - Batch Norm - Pooling Output not Dense layer but Global Average Pooling | "The 11-layer network can learn more simple features in the first layer (3 samples wide) and more complex features in the final layers (212 samples wide)." | Yes | 256x1 (8 sec x 1 channel) | 11 | 11 | Softmax | Batch norm | Yes | 2 Seizure / Non-Seizure | 28,642 | The network was trained for 100 epochs, after each epoch the validation AUC was calculated. | Standard | Stochastic Gradient Descent | SGD | LR: 0.01 Momentum: 0.9 | 2048 | N/M | N/M | Sliding Window (Shifted by 1s, 7/8 overlap) | N/M | Inter | Leave-One-Subject-Out | Leave-One-Subject-Out | "The training data contains less than 2% of the validation dataset" | ROC AUC | ROC AUC | N/M | N/M | AUC: 97.61% AUC90: 86.85% | CNN - 6 layers (O'Shea et al., 2017) SVM | DL & Trad. | No | No | No | This represents a substantial improvement over a shallower 6-layer CNN network which has a smaller range of receptive fields. These results represent the current best results for this task obtained using a single classifier. | N/M | No | N/A | No | Yannick Roy | TBR | Yes | OShea2018 | ||||||||||||
| 48 | Removing Confounding Factors Associated Weights in Deep Neural Networks Improves the Prediction Accuracy for Healthcare Applications | 2018 | Wang, Wu, Xing | Pacific Symposium on Biocomputing 2019 | Yes | Preprint | Carnegie Mellon University University of Illinois Urbana-Champaign Petuum Inc. | USA | 12 | Improvement of processing tools | Reduce effect of confounders | New approach: Reduce effect of confounders in medical data | Reduce the effect of confounders in medical data (e.g., gender bias in training data) | Students watching MOOC videos | Learn representations from scratch | Mindset (NeuroSky) | Raw EEG | Internal Recordings | Private | 10 students x 20 videos x 2min 10 confusing / 10 not confusing (??s windows) | N/M | 400 | 10 | 1 | N/M | N/M | N/M | N/M | No | No | Raw EEG | Raw EEG | z-score | TensorFlow | Bi-LSTM | RNN | Use of Confounder Filtering | N/M | No | N/M | N/M | N/M | Tanh | N/M | N/M | 2 | Confused Not-confused | 1 (sigmoid) | N/M | N/M | N/M | N/M | N/M | N/M | 20 | N/M | N/M | N/M | Binary Cross-Entropy* | Inter | 5-Fold CV | k-fold | N/M | Accuracy | accuracy | N/M | N/M | CF-Bidirectional LSTM acc: 75.0% | SVM: 67.2% K-Nearest Neighbors: 51.9% Convolutional Neural Network: 64.0% Deep Belief Network: 52.7% RNN-LSTM: 69.0% Bidirectional LSTM: 73.3% | DL & Trad. | No | No | No | The use of confounding filtering improves the predictive performance. | N/M | Yes | GitHub | No | Isabela Albuquerque | TBR | Yes | Wu2018 | ||||||||||
| 49 | HAMLET: Interpretable Human And Machine co-LEarning Technique | 2018 | Deiss, Biswal, Jin, Sun, Westover & Sun | Arxiv | Yes | Preprint | Georgia Institute of Technology Massachusetts General Hospital | USA | 9 | Classification of EEG signals | Multi-purpose architecture | New approach | Help experts generate high quality labels | Tested on Epilepsy data, could be used for different tasks | Features can be automatically extracted to help experts label the data | N/M | Raw EEG | Internal Recordings | Private | D: Using 140,000 of 390,486 x 16s sequences (unbalanced with 5 classes) 1) 20,000 (89h of EEG) : 80/20 - train / test. Patients in the testing set are not present in the training set (testing is per- formed on unseen patients) 2) 20,000 (89h of EEG) : 80/20 - train / test Patients in the testing set are also present in the training set (testing is performed on known patients). 3) 100,000 sequences from D | 140000 | 37333 | 155 | 19 | 200 | 1) Low-Pass filter: 60Hz 2) Computation of montages* (not sure what that means) 3) 16s windows | Yes | No | No | Raw EEG (None) | Raw EEG | N/M | Python Tensorflow | CNN CAE (Conv AutoEncoder) | Other | 1D CNN FC Layer only for training | One advantage of CNNs is the automated feature selection that happens during training. Without additional work, the model learns the features that it finds most relevant for its given task, from the raw signals. | Yes | 16x | Classifier: 6 Conv + 1FC | 7 | ELU | Dropout (20%) | Yes | 5 (softmax) | N/M | Co-Learning Supervised & Unsupervised | Other | Adam | Adam | N/M | 128 | N/M | N/M | Flipped Electrodes Left <-> Right side of brain, keeping references the same (Fz, Cz, Pz) almost 2x dataset. | N/M | Inter | No | No | Train: 80% Test: 20% | Accuracy | accuracy | Intel(R) Xeon(R) E5-2630 2.40 GHz 32 cores 256 Gb ofRAM 4 GPUs Tesla K80 | 13h | Before re-labeling | After re-lbl full | After re-lbl re-eval only HAMLET-CNN 39.36% | 40.75% | 68.75% HAMLET-CAE 38.46% | 39.06% | 67.97% CNN 38.89% | 41.58% | 68.75% MLP 21.04% | 23.14% | 14.06% | CNN MLP | DL | No | 1) Retrieval of closest labelled example to explain the decision on a specific input 2) Analysis of weights | Retrieval of closest examples, Analysis of weights | To summarize, first, we have introduced a novel tech- nique, HAMLET, for human and machine co-learning that is suited for creating high-quality labeled datasets on challenging tasks with a limited budget. This technique has benefits that can appreciated in many deep learning applications. | N/M | No | N/A | Yes | Yannick Roy | TBR | TBC | Deiss2018 | |||||||||||||
| 50 | Addressing Class Imbalance in Classification Problems of Noisy Signals by using Fourier Transform Surrogates | 2018 | Schwabedal, Snyder, Cakmak, Nemati & Clifford | Arxiv | Yes | Preprint | Emory University | USA | 7 | Generation of data | Data augmentation | Improve SOTA | Use FT Surrogates for Data Augmentation. (Tested with a CNN on Sleep Data) | Sleep Dataset (CAP) | Some EEG problemes are unbalanced. (e.g. Sleep stages, Epilepsy, etc.) For DL to perform well, we need data augmentation techniques. | N/M | Sleep | CAP Sleep | Public | CAPSLPDB: 94 out of 101 overnight PSGs x ~8h (30s windows, no overlap) | 90240 | 45120 | 94 | 2 | N/M | Low-pass filter: 13Hz (4th order Butterworth) Downsampling to 32Hz | Yes | No | No | Raw EEG (None) | Raw EEG | N/M | N/M | CNN | CNN | 1D CNN for each channel: 2xEEG + 1xEOG + 1xEMG | Yes | 30s Raw EEG | Conv 1D: 4 Conv 2D: 1 FC: 3 | 8 | Dropout | Yes | 6 | Wake S1, S2, S3, S4 REM | 6 (softmax) | N/M | N/M | N/M | RMS-Prop | Other | LR: 0.0016 Momentum: None Decay: 0.9 | 128 | Baysian Hyperparams Optim. | Yes | FT Surrogates | N/M | Inter | 5-Fold CV | k-fold | Train: 4/5 Validation: 1/5 Test: N/M | F1-Score Accuracy | f1-score, accuracy | Google Cloud | N/M | Accuracy (no augmentation): 67% | 73% | 51% | 64% | 75% | 70% Accuracy (FT surrogate): 83% | 86% | 38% | 75% | 97% | 46% Accuracy (IAAFT surrogates): 91% | 83% | 48% | 79% | 96% | 81% | (all internal, no external) No data augmentation FT surrogates IAAFT surrogates | None | No | No | No | Increases in the S2-accuracy seemed to be at the expense of stages S1 and S3 for larger values of α. Based on these results, we hypothesize that the effect of surrogate augmentation on an individual class accuracy does not directly depend on their conditional prediction accuracies, which are on the diagonal of the conditional confusion matrix (cf. Fig. 4(a)); instead, augmentation may introduce mixing between class labels indicated by a large off-diagonal element upon which the accuracy of one of the mixed labels will dominate. | Unfortunately, we were not yet able to evaluate and compare IAAFT surrogates with these results due to temporal and budget constraints. | Yes | GitHub | Yes | Yannick Roy | TBR | TBC | Schwabedal2018 | |||||||||||||
| 51 | EEG Classification Based on Sparse Representation and Deep Learning | 2018 | Gao, Shang, Xiong, Fang, Zhang, & Gu | NeuroQuantology | No | Journal | Zhejiang University City College, | China | 7 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA | Use CNN + Sparse coding on top of CSP features | Motor Imagery | N/M | N/M | CSP | BCI Competition III - IVa | Public | BCI Competition III - IVa 140 + 140 = 280 samples (length = 6s) | 280 | 28 | 2 | 118 | 100 | Offline | Band-pass filter 8-15Hz | Yes | No | No | CSP (32 CSP filters) | Frequency-domain | N/M | N/M | CNN | CNN | CNN's input is a sparse representation of CSP features | N/M | Yes | 28 x 28 | CNN: 2 FC: 1 | 3 | ReLU | N/M | N/M | 2 | Right Hand Right Foot | 2 (softmax) | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Binary cross-entropy | Inter | N/M | No | Train: 280 Test: N/M | Accuracy | accuracy | N/M | N/M | Accuracy Class 1: 98% Accuracy Class 1: 99% | Sparse representations (not clear what is the classifier) | Traditional pipeline | No | No | No | Performance of CNN+sparse representations is less afect when the number of training samples decreases. | N/M | No | N/A | No | Isabela Albuquerque | Yannick Roy | TBC | Gao2018 | ||||||||
| 52 | Use of features from RR-time series and EEG signals for automated classification of sleep stages in deep neural network framework | 2018 | Tripathy, & Rajendra Acharya | Biocybernetics and Biomedical Engineering | No | Journal | Siksha 'O' Anusandhan, India Ngee Ann Polytechnic, Singapore SUSS University, Singapore | India | 13 | Classification of EEG signals | Clinical | Sleep | Staging | Improve SOTA | Use DNNs on EEG + ECG for sleep stage scoring | Sleep Dataset (MIT-BIH) | They don't mention why DL. | N/M | Raw EEG (Sleep) | MIT-BIH | Public | MIT-BIH 18 records * 100000 points / 7500 points per window (30s windows) | 240 | 120 | 18 | 1 | 250 | 1) 5 Band-pass filters to 5 freq bands | Yes | No | No | 14 EEG-HRV Features (out of 19) (The dispersion entropy and the variance features are evaluated from the different bands of EEG signal) (the RQA and dispersion entropy features are evaluated from the IMFs of RR-time series) | Other | N/M | Matlab 2015a | SAE | AE | 3 DNNs EEG features + HRV features combined as inputs. Outputs = 2 classes (x3 DNNs) | N/M | Yes | 14 EEG Features ECG Features (30s window) | 2 AE | 2 | Sigmoid | L2 (N/M... Assumed from the formula) | Yes | 2 (softmax) 3 DNN Networks Classifying 2 classes each | N/M | Greedy Layer Wise | Pre-training | SGD | SGD | N/M | N/M | N/M | N/M | N/M | (See Formula) | Inter | 10-Fold CV | k-fold | N/M | Accuracy (Acc) Sensitivity (Sen) Specificity (Spe) | accuracy, sensitivity, specificity | CPU 2 GHz 2 GB RAM | 1 Instance: EEG: 4.89s RR: 0.03s | Acc Sleep vs Wake: 85.51% Acc Light vs Deep Sleep: 94.03% Acc REM vs NREM: 95.71% | Hayet and Slim [55] (ELM, Werteni et al. [56] (SVM), Adnane et al. [16] (SVM), Rossow et al. [57] (HMM), Redmond and Heneghan [58] (QDA), Song et al. [59] (Multivariate Discrim. Analysis), Prucnal et al. [12] (NN), Hasan et al. [11] (RUSBoost), Da Silveira et al. [13] (RF) | Traditional pipeline | No | No | No | The dispersion entropy values for delta (d), theta (u) and alpha (a) bands are found to be more discriminatory for the classification of the wake and sleep classes. | The limitation of this work is that we have used only 18 subjects. The performance of this work can be improved using more subjects from the diverse race. The number of REM sleep stage instances in MIT-BIH polysomnography database is less as compared to deep sleep, light sleep and wake classes. | No | N/A | Yes | Yannick Roy | TBR | Yes | Tripathy2018 | |||||||||||
| 53 | Emotion stress detection using EEG signal and deep learning technologies | 2018 | Liao, Chen & Tai | IEEE International Conference on Applied System Invention (ICASI) | No | Conference | Department of Information Management Chaoyang University of Technology | Taiwan | Classification of EEG signals | Monitoring | Affective | Emotion | New approach | Use CNN to classify Attention & Meditation from raw EEG | Listening to music | Exploring the use of DL for stress detection via EEG | Mindwave Mobile (Neurosky) | None | Internal Recordings | Private | 7 subjects x 10 min (1s windows, no overlap) | 4300 | 70 | 7 | 1 | 512 | N/M | N/M | N/M | N/M | Frequency Bands | Frequency-domain | N/A | N/M | CNN | CNN | N/M | N/M | No | 1s (N/M, assuming 512 samples) | 7 | 7 | RELU | N/M | N/M | 1 0: Meditation 1: Attention | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Grid Search | Yes | N/M | N/M | Inter | N/M | No | Train: 80% Test: 20% | Accuracy F1-Score | accuracy, f1-score | N/M | N/M | Accuracy: 80.13% | None | None | No | No | No | The F1-score shows that our system is better in predicting class 1 than predicting class 0. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Liao2018 | ||||||||||||
| 54 | Hierarchical internal representation of spectral features in deep convolutional networks trained for EEG decoding | 2018 | Hartmann, Schirrmeister & Ball | BCI Conference | Yes | Conference | University of Freiburg | Germany | 6 | Improvement of processing tools | Model interpretability | Model visualization | Improve interpretability of CNNs | Study most activating inputs. Study effect on internal representation of variations in the input signal | Motor imagery | End-to-end learning | N/M | Raw EEG | Internal Recordings | Private | 14 subjects x 1000 trials x 4s (4s windows) | 14000 | 933 | 14 | 128 | 5000 | 1) Downsample to 250 Hz 2) Common average re-reference | Yes | No | No | Raw EEG (None) | Raw EEG | N/M | Pytorch | CNN | CNN | See Schirrmeister et al. (2017) | See Schirrmeister et al. (2017) | Yes | 522 x 128 (samples x channels) | CNN: 5 FC: 1 | 6 | ELU | N/M | N/M | 4 | N/M | Standard optimization | Standard | Adam | Adam | N/M | N/M | N/M | N/M | N/M | Cross-entropy | Intra | N/M | No | Train: 80% Test: 20% | Accuracy F1-Score | accuracy, f1-score | N/M | N/M | Mean accuracy over 14 subjects: 88.6% (but this is not the focus of paper) | None | None | No | 1) Signal perturbation (amplitude & phase) 2) Most-activating input windows | Input-perturbation network-prediction correlation maps, Analysis of most-activating input windows | Analyzed effect of perturbations in phase and amplitude of input signals. Earlier layers focus on frequency-related information while latest layers focus on amplitude. | N/M | No | N/A | No | Isabela Albuquerque | Hubert Banville | TBC | Hartmann2018b | ||||||||||||
| 55 | Spatial-Temporal Recurrent Neural Network for Emotion Recognition | 2018 | Zhang, Zheng, Cui, Zong & Li | IEEE Transactions on Cybernetics | Yes | Journal | Southeast University, Nanjing, China Nanjing University of Science and Technology, China | China | 9 | Classification of EEG signals | Monitoring | Affective | Emotion | New Approach: Stacking 2 RNN layers for spatial and temporal resolution, for EEG & Facial Expression for emotion classification | Stacking 2 RNN layers for spatial and temporal resolution, for EEG & Facial Expression for emotion classification | Emotion Classification for short emotional films/clips (SEED dataset) | Leverage RNN for both spatial and temporal features | (NeuroScan) | Emotions | SEED | Public | SEED: 15 subjects Assumed: 15 subjects x 15 movies x 4min x 2 exp (9s windows, no overlap) | 12000 | 1800 | 15 | 62 | 1000 | None | No | No | No | DE descriptors (?) - Freq Bands (256-point FFT + Hanning Window (1s) for 5 F-Bands) | Frequency-domain | N/M | N/M | STRNN (Spatial-Temporal RNN) | RNN | Spatial & Temporal features representation with stacked RNNs | 1) To learn spatial dependencies, a quad-directional spatial RNN (SRNN) layer is first employed 2) Then, a bi-directional temporal RNN (TRNN) layer is further stacked on SRNN to capture long-term temporal dependencies | Yes | Not clear... (to be reviewed) | SRNN: 1 TRNN: 1 | 2 | ReLU Sigmoid | N/M | N/M | 3 (Softmax) | N/M | N/M | N/M | Back Propagation Through Time (BPTT) | N/M | N/M | N/M | N/M | N/M | N/M | Cross-entropy | Inter | No | No | Train: 9 Sessions Test: 6 Sessions | Accuracy | accuracy | N/M | N/M | Accuracy: 89.5% | None | None | No | No | No | A multidirection SRNN layer and a bi-direction TRNN layer are hierarchi- cally employed to learn spatial and temporal dependencies layer by layer. To adapt the multichannel EEG signals to the proposed STRNN framework, the spatial scanning order of electrodes are specified by spatial coordinates and tempo- ral variation information is involved by slicing a window on the extracted DE feature sequences. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Zhang2018 | |||||||||||
| 56 | Individual Recognition in Schizophrenia using Deep Learning Methods with Random Forest and Voting Classifiers: Insights from Resting State EEG Streams | 2018 | Chu, Qiu, Liu, Ling, Zhang & Wang | IEEE Transactions on Neural Systems and Rehabilitation Engineering | Yes | Journal | Big Data and AI Research Center of Shanghai Jiaotong University | China | 7 | Classification of EEG signals | Clinical | Schizophrenia | Detection | New Approach | Using Random Forest and Voting Classifiers with a CNN for Individual Recognition in Schizophrenia | Resting State, Eyes Open. (300s each) | Automatic feature extraction | (BrainProducts) | Raw EEG | Internal Recordings | Private | 120 Subjects x 300 seconds | 360000 | 600 | 120 | 64 | 1000 | Offline | 1) Occular Correction (with Brain Vision Analyzer's algos) 2) Re-Referenced to Common Average 3) Pass-Band Filter (IIR): 0.01 - 50Hz | Yes | Yes* | Yes | 1) Raw EEG 2) Freq Bands | Raw EEG | Divide by max | N/M | CNN, RNN, and MLP | CNN | 3 Conv Layers, ELU, 3 Dropout 0.5, 3 Max Pooling + Dropout 0.25, 3 FCs, 1 voting (RF, softmax or SVM) | N/M | Yes | Not clear | CNN: 6 MLP: 4 RNN: 2 | 6 | ELU | Dropout (0.5, 0.25) | Yes | 3 | High risk Schizophrenia Healthy | 3 Classes (replaced Softmax with Random Forrest) | N/M | Standard | Standard | N/M | N/M | N/M | N/M | N/M | N/M | No | N/M | Inter | Yes | Yes (no detail) | Train: 50% Test: 50% | Accuracy | accuracy | NVIDIA GeForce GTX 750 | N/M | FES: 96.7% CHR: 81.6% HC: 99.2% | ANNV, RNNV, CNNV, ANNV+mSVM, RNN+mSVM, CNN+mSVM, ANN+RF, RNN+RF, CNN+RF | DL | No | No | No | "In conclusion, we have shown that CNNV-RF performs better than softmax and CNNV-mSVM on a well-known dataset (mnist) and resting state EEG streams used in this paper. Switching from softmax or mSVM to RF is incredibly simple and appears ro be helpful for classification problems." | N/M | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | Chu2017 | ||||||||
| 57 | An EEG-based Image Annotation System | 2018 | Parekh, Subramanian, Roy & Jawahar | National Conference on Computer Vision, Pattern Recognition, Image Processing, and Graphics | Yes | Conference | IIIT Hyderabad, India University of Glasgow, Singapore National Brain Research Centre, Manesar, India | India | 11 | Classification of EEG signals | BCI | Reactive | RSVP | Novel Approach: Image classification based on subject's P300 | Using CNN (EEGNet) to classify images based on P300. RSVP with Oddball. | RSVP. Images from Caltech101 and VOC 2012. Oddball Paradigm for P300 | Not mentioned why DL... | EPOC (Emotiv) | RSVP P300 | Internal Recordings | Private | 5 subjects x 3 sessions x 2 (Test/Train) x 6 min 25 block of 100 image (2x test/train) x 100ms/image (1s windows) | 7500 | 125 | 5 | 14 | 128 | 1) Baseline power removal using the 0.5 second pre-stimulus samples 2) Band-Pass filter: 0.1 - 45 Hz 3) ICA to remove artifacts (eye-blinks, and eye and muscle movements) | Yes | Yes | Yes | P300 | Other | N/A | Braindecode | CNN (EEGNet) | CNN | They add a Outlier Removal "Feature". They used a pre-trained VGG-16 on predicted target image. (to reduce false-positive due to class imbalance) | see EEGNet & Braindecode | No | 1s Windows (Raw EEG) | 3 | 3 | ELU | N/M (see Braindecode / EEGNet) | N/M | 2 Target / Non-Target | N/M | N/M | N/M | Adam | Adam | N/M | N/M | N/M | N/M | N/M | Categorical Cross-Entropy | Inter | 5-Fold CV | k-fold | Train: 2500 images / subject Test: 2500 images / subject | F1-Score (Due to a heavy class imbalance between T/non-T, we use F1-score) | f1-score | NVIDIA GEFORCE GTX 1080 Ti | N/M | [DS: CT101] Before outliers removal: F1: 0.71 Precision: 0.66 Recall: 0.81 [DS: CT101] After outliers removal: F1: 0.68 Precision: 0.63 Recall: 0.72 [DS: VOC2012] Before outliers removal: F1: 0.88 Precision: 0.99 Recall: 0.81 [DS: VOC2012] After outliers removal: F1: 0.83 Precision: 0.97 Recall: 0.72 | None | None | No | No | No | Our annotation system exclusively relies on the P300 ERP signature, which is elicited upon the viewer detecting a pre-specified object class in the displayed image. A further outlier removal procedure based on binary feature-based clustering significantly improves annotation performance. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Parekh2018 | |||||||||||
| 58 | EEGNet: A Compact Convolutional Neural Network for EEG-based Brain-Computer Interfaces | 2018 | Lawhern, Solon, Waytowich, Gordon, Hung & Lance | Journal of Neural Engineering | Yes | Journal | U.S. Army Lab,DCS Corporation, Columbia University, Georgetown University Medical Center | USA | 17 | Classification of EEG signals | BCI | Active & Reactive | MI & ERP | Novel Approach: DN that can be used for different BCI paradigms | Compare EEGNet with SOTA ML for different BCI Paradigms | Visual P300 ERN Movement-related cortical potentials Sensory Motor Rhythms | Allows robust feature extraction | ActiveTwo (BioSemi), N/M, ActiveTwo (BioSemi), N/M | 1) P300 2) ERN 3) Movement-related cortical potentials 4) SMR | Internal Recordings; Kaggle: Inria BCI challenge; Internal Recordings; BCI Competition IV - IIa | Both | P300: 15 subject x 2000 trials [1s windows] ERN: 26 subjects x 340 trials [1.25s windows] MRCP: 13 subjects x 1100 trials [1.5s windows] SMR: 9 subjects x 288 trials [3s windows] | 30000; 8840; 14300; 2592 | 500; 184.2; 357.5; 129.6 | 15; 26; 13; 9 | 64; 56; 64; 22 | 512; 600; 1024; 250 | Offline | 1) Rereferencing (linked mastoids or earlobes) 2) Bandpass filter (1 - 40 Hz, 0.1-40 Hz or 4-40 Hz) 3) Downsampled to 128 Hz (** Different approaches! e.g. Used PREP Pipeline for #3) | Yes | No | No | Raw EEG | Raw EEG | Exponential moving average | Keras + Tensorflow | CNN | CNN | Layer 1: 1D Temporal Filters Layer 2: Depthwise 2D Conv Layer 3: Separable 2D Conv | 1D temp. conv. at L1 to learn frequency filters. Depthwise: Inspired in part by the Filter-Bank Common Spatial Pattern (FBCSP) algorithm. Separable: explicitly decoupling the relationship within and across feature maps by first learning a kernel summarizing each feature map individually, then optimally merging the outputs afterwards | Yes | Channels x Time | 3 | 3 | ELU | Dropout, weight decay | Yes | P300: 2 ERN: 2 MRCP: 2 SMR: 4 | P300: Target/Non-target ERN: Error/No error MRCP: Left/Right hand SMR: left hand/right hand/feet/tongue | (depends on the task) (softmax) | 1) 1,066 2) 1,082 3) 1,098 4) 796 | Within-Subject and Cross-Subject If classes are umbalanced, we apply class-weight to the loss function whenever the data is imbalanced | Standard | Adam | Adam | N/M | 64 | N/M | N/M | N/M | Categorical cross-entropy + Class weight if unbalanced | Both | Intrasubject: 4-Fold CV Intersubject: Leave some subjects out (different nb folds and ratio for each per task) | k-fold; Leave-N-Subjects-Out | [Intra] Train: 50% [Intra] Valid: 25% [Intra] Test: 25% [Inter] Different ratios of subjects for training, for validation and for test | Accuracy ROC AUC | accuracy, ROC AUC | NVidia Quadro M6000 | N/M | See paper for full breakdown. TLDR; Doesn't outperform or underperform anything by a lot. (however, it uses two orders of magnitude fewer parameters) | DeepConvNet (Schirrmeister, 2017) ShallowConvNet (Schirrmeister, 2017) Riemannian EEG (Barachant, 2015) FBCSP | DL & Trad. | Repeated-measures ANOVA | 1) Summarizing averaged outputs of hidden unit activations. 2) Visualizing the convolutional kernel weights. 3) Calculating single-trial feature relevance on the classification decision Also used DeepLIFT (Shrikumar 2017) | Analysis of activations, Analysis of weights, Ablation of filters, DeepLIFT | In this work we proposed EEGNet, a compact convolutional neural network for EEG-based BCIs that can generalize across different BCI paradigms (e.g. ERP and oscillatory-based) in the presence of limited data and can produce interpretable features. To the best of our knowledge, this represents the first work that has validated the use of a single network architecture across multiple BCI datasets, each with their own feature characteristics and data set sizes. Through the use of feature visualization and ablation analysis, we show that neurophysiologically interpretable features can be extracted from the EEGNet model | N/M | Yes | GitHub | No | Yannick Roy | Hubert Banville | Yes | Lawhern2018 | ||||||||
| 59 | A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series | 2018 | Chambon, Galtier, Arnal, Wainrib & Gramfort | IEEE Transations on Neural Systems and Rehabilitation Engineering | Yes | Journal | Telecom ParisTech, Inria, Université Paris-Saclay | France | 12 | Classification of EEG signals | Clinical | Sleep | Staging | Improve State-of-the-Art | Sleep | N/M | Sleep events | MASS | Public | MASS (61 out of 62) 61 nights (8h) each from diff subjects (30s windows, no overlap) | 58560 | 29280 | 61 | 20 | 128 | 1) Low-pass @30Hz | Yes | N/M | N/M | Raw EEG + EOG and raw EMG | Raw EEG | z-score | Keras + Tensorflow | ConvNet | CNN | 3 conv layers + dense (per modality) | Layer 1: spatial filter Layers 2, 3: temporal filters | Yes | Nb channels * 30 s | 4 | 4 | Linear, ReLU, Softmax | 25% (last layer) | Yes | 5 | <10^5 | 1) Training on a single 30-s epoch 2) Freezing net, and train last layer on multi-epochs | Pre-training | Adam | Adam | Random searches with the hyperopt Python packag | Yes | No | Categorical cross-entropy | Inter | Leave-p-subject-out 5 random permutations | Leave-N-Subjects-Out | Train: 41 records Valid: 10 records Test: 10 records | Balanced accuracy F1-score, Precision, Sensitivity, Specificity, Confusion matrix | balanced accuracy, f1-score, precision, sensitivity, specificity, confusion matrix | N/M* | ~250 s | Acc: ~80% Bal. acc.: ~80% Kappa: ~0.7 F1 score: ~0.71 | Gradient boosting on time domain and freq. domain features Univariatie ConvNets from Tsinalis et al. (2016) and Supratak et al. (2017) | DL & Trad. | No | Occlusion sensitivity | Occlusion of input | 1D convolution provided a speed-up vs. 2D convolutions Smaller number of parameters than other studies Temporal context helps for some classes, but not for others; recurrent architectures could help Size of dataset matters | No | N/A | Yes | Hubert Banville | TBR | TBC | Chambon2018 | ||||||||||||||||
| 60 | Converting Your Thoughts to Texts: Enabling Brain Typing via Deep Feature Learning of EEG Signals | 2018 | Zhang, Yao, Sheng, Kanhere, Gu, Zhang | IEEE International Conference on Pervasive Computing and Communications (PerCom) | Yes | Conference | University of New South Wales Macquarie University RMIT University | Australia | 10 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA | Joint CNN & LSTM + AE for Motor Imagery (5 classes) | Motor Imagery (5 classes) (see eegmmidb dataset) | EEG processing is time-consuming and depend on human expertise. SOTA models achieve 70-80% which is not enough. | N/M, EPOC (Emotiv) | Motor Imagery | eegmmidb; Internal Recordings | Public | 1) eegmmidb: 28,000 samples x 10 subjects (28000 points @ 160Hz = 175s/subject) 2) Internal (Emotiv): 34,560 samples x 7 subjects (28000 points @ 128Hz = 270s/subject) (window length = 1 point) | 28000; 34560 | 29.2; 31.5 | 10; 7 | 64; 14 | 160; 128 | Offline and Online | N/M | N/M | No | No | Raw EEG (None) | Raw EEG | N/M | N/M | CNN + LSTM + linear AE + XGB (classification) | Other | CNN & LSTM are parallel, then combined for the AE then XGB classifier | CNN for Spatial and RNN for Sequential info | Yes | 1 x 64 (sample x channels) | LSTM: 6 layers CNN: 2 Conv + 2 FC | 6 | ReLU, Sigmoid, tanh | L2 | Yes | 5 | eegmmidb: eye closed, left hand, right hand, both hands, both feet emotiv: up arrow, down arrow, left arrow, right arrow, eye closed | 5 | N/M | Standard optimization | Standard | LSTM & CNN: Adam AE: RMSProp | Adam | Full table on optim params | 7000 | N/M (they have tried many config, manually I suppose) | N/M | N/M | LSTM + CNN: Cross-Entropy AE: MSE | Intra | No | No | Train: 75% Test: 25% | accuracy, precision, recall, F1 score, ROC curve, and ROC AUC | accuracy, precision, recall, f1-score, ROC, ROC AUC | N/M | 2000 s | DS #1 - Accuracy: 0.955 DS #2 - Accuracy: 0.9427 | Baselines: KNN, SVM, RF, LDA, AdaBoost, RNN, CNN Externals: Almoari, Sun, Mohammad, Major, Shenoy, Tonic, Rashid, Ward, Sita, Pinheiro. (all different papers, see Table IV) | DL & Trad. | No | No | No | The classification accuracy of the public dataset (eegmmidb) is consistently higher than the local real-world dataset (emotiv). Our future work will focus on improving the accuracy in the person-independent scenario, wherein some subjects participate in the training and the rest of subjects involve in the testing. | N/M | Yes | GitHub | No | Yannick Roy | Hubert Banville | Yes | Zhang2017g | ||||||||
| 61 | MindID: Person Identification from Brain Waves through Attention-based Recurrent Neural Network | 2018 | Zhang, Yao, Kanhere, Liu, Gu & Chen | ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies | Yes | Conference | University of New South Wales, Australia Tsinghua University RMIT University, Australia | Australia | 20 | Classification of EEG signals | Personal trait/attribute | Person identification | Improve SOTA: EEG for Person Identification | Use RNN on EEG for Person Indentification | 3 Different Datasets. (they claim that Delta has the most personal info) | The DL motivation is not clear. They want to improve SOTA. | EPOC (Emotiv), N/M | Delta Band* | Internal Recordings; eegmmidb | Both | DS1 (EID-M): 21,000 Samples/Subject. Total: 168,000 DS2 (EID-S): 7,000 Samples/Subject. Total: 56,000 DS3 (EEG-S): 8 subjects x 7,000 samples (window length = 1 point) | 168000; 56000; 56000 | 21.9; 7.3; 5.8 | 8; 8; 8 | 14; 14; 64 | 128; 128; 160 | 1) Remove DC Offset (substract) 2) Band-Pass Filter: 0.5 - 4Hz (using only Delta) | Yes | No | No | Delta Band | Frequency-domain | z-score | Matlab | Attention-based Encoder-Decoder RNN + XGB Classifier | RNN | Encoder, Decoder, Attention Module + XGB Classifier | N/M | Yes | 1x14 Delta Bands / Channel (not clear about the dimensionality) | Encoder: 3 FC (164) + 1 LSTM (164) Decoder: 1 FC (164) | 4 | N/M | L2 | Yes | 8 One-Hot Label (ID - 8 Subjects) | N/M | N/M | N/M | Adam | Adam | LR: | 21,000 samples (?) | N/M | N/M | N/M | Cross-Entropy | Inter | No | No | Train / Test DS1: 147,000 / 21,000 DS2: 49,000 / 7,000 DS3: 49,000 / 7,000 | Precision Recall F1-Score | precision, recall, f1-score | Nvidia Titan X Pascal 768G memory 145 TB PCIe SSD | N/M | Precision | Recall | F1-Score DS #1: 0.982 | 0.982 | 0.982 DS #2: 0.988 | 0.988 | 0.988 DS #3: 0.999 | 0.999 | 0.999 | SVM, RF, KNN, AdaBoost, LDA, XGB, RNN | DL & Trad. | No | No | No | Moreover, the pre-trained model should be updated for a period of time since the user’s EEG data is gradually changed with the environmental factors such as age, mental state, and living style. One of our future work is to develop an online learning system which is enabled to automatically update the training dataset based on the testing data which is collected during the operating period. | Limited by the local experimental conditions, our study only gathered EEG data from 8 subjects with few trials. The dataset is only divided into two categories (Multi and Single), which is not enough to explore the change trend of the identification accuracy with the increase of data trials. | Yes | GDrive | Yes | Yannick Roy | TBR | Yes | Zhang2017e | ||||||||||||
| 62 | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 63 | A convolutional neural network for steady state visual evoked potential classification under ambulatory environment | 2017 | Kwak, Muller & Lee | PLOS One | No | Journal | Korea University, TU Berlin | South Korea | Classification of EEG signals | BCI | Reactive | SSVEP | Improve SOTA | Improve robustness of SSVEP BCIs for exoskeleton control in ambulatory conditions | SSVEP | - | MOVE (BrainProducts) | SSVEP | Internal Recordings | Private | 2 datasets (50x5s + 250x5s) x 7 subjects 5s trial into 300 x 2s trials (2s sliding window, 10ms shift size) | 630000 | 175 | 7 | 8 | 1000 | Offline | 1) Notch filter @60Hz 2) Band pass from 4-40 Hz | Yes | No | No | 120 FFT bins from 5-35 Hz | Frequency-domain | min-max | N/M | 1,2) CNN 3) MLP | CNN | 1) CNN (3 layers) 2) CNN (4 layers) 3) MLP (3 layers) | First conv layer: spatial filter Second conv layer: spectral filter | Yes | 120 x 8 Freq x channels | 1) 3 2) 4 3) 3 | 4 | Sigmoid | N/M | N/M | 5 | Walk Forward Turn Left Turn Right Stand Up Sit Down | 5 | N/M | Standard optimization | Standard | SGD | SGD | Learning rate: 0.1 | N/M | N/M | N/M | Sliding Window (Shifted by [10-60ms] over 2s win) | N/M | Intra | 10-Fold | k-fold | Train: 90% Test: 10% Chronological split | Accuracy | accuracy | N/M | N/M | Static condition: up to 99.28% Ambulatory condition: up to 94.03% | CCA, MSI, CCA + kNN | Traditional pipeline | Yes (not clear what method) | Visualization of activations | Analysis of activations | CNN-1 (3 layers) was the most robust. Since architecture is pretty simple, no regularization is used. | Artefacts in ambulatory settings | No | N/A | No | Hubert Banville | Yannick Roy | TBC | kwak2017 | |||||||||
| 64 | Mental Tasks Classification using EEG signal, Discrete Wavelet Transform and Neural Network | 2017 | Padmanabh, Shastri & Biradar | Discovery | No | Journal | Savitribai Phule Pune University | India | Classification of EEG signals | BCI | Active | Mental tasks | [Classification of 5 different mental tasks, via Wavelet & ANNs (PNN & MLP)] | 5 Mental Tasks (Baseline, Multiplication, Rotation, Counting, Letter composition) | 7P511 (Grass Instruments) | Keirn & Aunon (1989) | Public | 5 subjects x 5 tasks x 5 trials x 10s @250Hz (1s windows) | 1250 | 20.8 | 5 | 6 | 250 | 1) Band-Pass filter: 0.1-100Hz | Yes | N/M | N/M | Frequency-domain | MATLAB & NNtool | MLP PNN | FC | No | 200x1 | 2 (20; 15) | 2 | N/M | N/M | Standard | N/M | N/M | Learning Rate: 0.9 | N/M | N/M | No | MSE | N/M | N/M | No | N/M | accuracy | accuracy | N/M | N/M | MLP: 92% NPP: 100% | None | None | No | No | No | No | N/A | No | Yannick Roy | TBR | TBC | Padmanabh2017 | ||||||||||||||||||||||||||
| 65 | Cross-session classification of mental workload levels using EEG and an adaptive deep learning model | 2017 | Yin & Zhang | Biomedical Signal Processing and Control | No | Journal | University of Shanghai for Science and Technology East China University of Science and Technology | China | Classification of EEG signals | Monitoring | Cognitive | Mental workload | New approach | ACAMS (Automation-enhanced Cabin Air Management System) | (Nihon Kohden) | PSD | Internal Recordings | Private | 7 subjects x (5min + 6x15min + 5min) x 2 sessions (2s windows, no overlap) | 42000 | 1400 | 7 | 11 | 500 | 1) Low-Pass filter: 40Hz 2) ICA for EOG artifacts | Yes | Yes | Yes | PSD Avg. Power: T (5–7.5 Hz), A (8–13.5 Hz), B1 (14–20 Hz), B2 (20.5–30 Hz), G (30.5–40 Hz) | Frequency-domain | MATLAB | SDAE | AE | Adaptive Stacked Denoising AutoEncoder | Yes | 55x1 EEG PSD Features | 6 | 6 | N/M | N/M | 2 | N/M* | Other | N/M | N/M | Grid search | Yes | Gaussian noise on Freq features | N/M* | Intra | N/M | No | Train: 66% Test: 33% | accuracy, confusion matrix, sensitivity, specificity | accuracy, confusion matrix, sensitivity, specificity | N/M* | [Complicated ... read me again ...] SDAE > State of the art. | ANN, NB, kNN, SVMlin, SVMrbf, BSV, SDAE | DL & Trad. | Wilcoxon sign-rank test | 3D scatter plots of layer activations | Analysis of activations | It is evident that the proposed method is superior to those shallow and static classifiers when the comprehensive cortical information is adopted as the network inputs. | No | N/A | No | Yannick Roy | TBR | TBC | Yin2017a | ||||||||||||||||||||||
| 66 | Generative Adversarial Networks Conditioned by Brain Signals | 2017 | Palazzo, Spampinato, Kavasidis, Giordano & Shah | ICCV | No | Conference | University of Catania, University of Central Florida | Italy | 9 | Generation of data | Generating images conditioned on EEG | New approach: generating images conditioned on EEG | Generating images using GANs conditioned by EEG representation | Visual presentation of images | Allows image generation | BrainAmp (BrainProducts) | Raw EEG | Internal Recordings | Private | 6 Subjects x 50 images x 40 classes 1400s per subject (4 sessions of 350s) 2000 images x 6 subjects = 12000 minus exclusions = 11,466 valid samples | 11466 | 140 | 6 | 128 | 1000 | 1) Hardware notch filter: 49-51 Hz 2) Band-pass filter: 14-70 Hz 3) Non-uniform quantization of the voltage values | Yes | N/M | N/M | Raw EEG | Raw EEG | N/M | N/M* | 1) LSTM for EEG encoder 2) DCGAN for image generation | Other | Conditional DCGAN (conditioning G and D) | N/M | Yes | Nb channels * 0.5 s | 1) 2 2) 5 (generator), 6 (discriminator) | 6 | 1) ReLU 2) ReLU | N/M | N/M | 1) 40 2) 64 x 64 | N/M | 1) Train encoder to predict image category from raw EEG 2a) Train GAN on images without EEG features 2b) Train GAN condtioned on average (across subs) EEG representation learned by the encoder | Other | Adam (lr=0.001) | Adam | 1) N/M 2) Batch normalization | 1) 16 2) N/M | N/M | N/M | 1) Nothing on EEG 2) Resizing + Flipping on Images | 1) categorical cross-entropy. 2) non-saturating | Inter | N/M; [TBD] | No | Train: 80% Valid: 10% Test: 10% | 1) Accuracy 2) Inception score, Inception accuracy | accuracy, inception score, inception accuracy | 2 Titan X Pascal | N/M | Encoder: 83.9% GAN: IS: 4-6.5, acc: 43% | No | None | No | No | No | Conditioning vector (i.e. EEG representations) are noisy, which makes harder to learn how an appropriate conditioning vector. | Suffers from classes with high internal variability Dataset is small | No | N/A | No | Isabela Albuquerque | Hubert Banville | TBC | Palazzo2017 | |||||||||||||
| 67 | The effects of pre-filtering and individualizing components for electroencephalography neural network classification | 2017 | Major & Conrad | IEEE SoutheastCon | No | Conference | University of North Carolina (Charlotte) | USA | 6 | Classification of EEG signals | BCI | Active | Motor imagery | Improve State-of-the-Art: Exploring impact of ICA preprocessing | Analyze effectiveness of using ICA to enhance EEG that will be processed by a neural network | Motor imagery | "Since every brain computer interface (BCI) has to be tailored for each person it is advantageous to use a neural network" | N/M | Raw EEG | eegmmidb | Public | 109 subjects x 14 experiments (12x2min + 2x1min) Not clear how many samples they used... (??s windows) | N/M | 2834 | 109 | 64 | 160 | Offline | 1) Band pass filter: 8-30Hz | Yes | Yes | Yes | Raw EEG | Raw EEG | N/M | Matlab | MLP | FC | N/M | N/M | Yes | 16x ? | 10 | 10 | N/M | N/M | N/M | 2 | Left Grasp Right Grasp | 2 | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Inter | No | No | Train: 2/3 Test: 1/3 | Accuracy | accuracy | N/M | N/M | With ICA: 68% Without ICA: 56% | No | None | No | No | No | Applying ICA to raw data improves the neural network performance. | N/M | No | N/A | No | Isabela Albuquerque | Yannick Roy | TBC | Major2017 | ||||||||
| 68 | Convolutional neural network-based transfer learning and knowledge distillation using multi-subject data in motor imagery BCI | 2017 | Sakhavi & Guan | IEEE Conference on Neural Engineering | No | Conference | NUS & NTU | Singapore | 4 | Classification of EEG signals | BCI | Active | Motor imagery | Transfer learning (from one subject to another) | Reduce calibration time in a BCI using transfer learning | Motor imagery | Reduce BCI's calibration time | N/M | Raw EEG | BCI Competition IV - IIa | Public | BCI competition IV-2a dataset (4 out of 9 subjects) x 4 classes x 72 samples x 2 session x 4 seconds. After removing data: ∼ 1000 per class per session (4s windows, no overlap) | 8000 | 153.6 | 9 | 22 | 250 | 1) Bandpass between 0.5-100 Hz 2) Notch filter @50 Hz | Yes | N/M | N/M | FBCSP in 9 frequency bands, then extracting envelope | Frequency-domain | weird z-scoring | Torch7 | CNN + MLP | CNN | CNN: 5 layers MLP: 1 layer | - | Yes | CNN: 32x40 MLP: 32 | CNN: 4 conv, 1 FC MLP: 1 FC | 5 | ReLU | N/M | N/M | CNN: 128 MLP: 128 | N/M | 1) Pre-train CNN+MLP on N-1 subjects 2) Fine-tune pre-trained network on 1 subject | Pre-training | Adam | Adam | N/M | N/M | N/M | N/M | N/M | KL divergence | Inter | Leave-N-Samples-Out | Leave-N-Samples-Out | Train: 5, 10, 20 samples / class Test: Remaining | Test set accuracy | accuracy | N/M | N/M | Average acc: 69.71% | SVM | Traditional pipeline | Wilcoxon sign-rank test | No | No | Best results (average across subjects) show significant improvement with respect to SVM. However, there is high variability | Choosing hyperparameter lambda | No | N/A | No | Hubert Banville | Isabela Albuquerque | TBC | Sakhavi2017 | |||||||||||
| 69 | Single-trial EEG classification of motor imagery using deep convolutional neural networks | 2017 | Tang, Li & Sun | Optik - International Journal for Light and Electron Optics | No | Journal | Zhejiang University of Technology | China | 8 | Classification of EEG signals | BCI | Active | Motor imagery | New Approach | CNN for MI on Single Trial | Motor Imagery | Automated feature extraction | ActiveTwo (BioSemi) | SMR - ERD/ERS | Internal Recordings | Private | 2 subjects x 460 trials 3s epochs, and splitted in 50ms windows (50ms windows) | 55200 | 46 | 2 | 28 | 1000 | Offline | 1) [Hardware] Notch Filter: 50Hz 2) [Hardware] Band-Pass Filter: 0.5-100Hz 3) [Software] Band-Pass Filter: 8-30Hz | Yes | No | No | SMR - ERD/ERS | Frequency-domain | N/M | N/M | CNN | CNN | Activation Function: Hyperbolic Tangent | N/M | Yes | 28x60 Channels x Time Points | 2 Conv 1 FC | 3 | Tanh Sigmoid | N/M | N/M | 2 | Left hand Right hand | 2 | N/M | Standard | Standard | GD | SGD | N/M | N/M | N/M | N/M | No | N/M* | Intra | 10-Fold CV | k-fold | Train: 80% Test: 20% | confusion matrix, accuracy, ROC, precision, recall, f-score | confusion matrix, accuracy, ROC, precision, recall, f-score | N/M | N/M | Accuracy: 86.41% | Power+SVM CSP+SVM AR+SVM | Traditional pipeline | ANOVA | No | No | "The results demonstrate that CNN can further improve classification performance compared with other three conventional methods." | N/M | No | N/A | No | Yannick Roy | Isabela Albuquerque | TBC | Tang2017 | ||||||||
| 70 | Pattern Recognition of Momentary Mental Workload Based on Multi-Channel Electrophysiological Data and Ensemble Convolutional Neural Networks | 2017 | Zhang, Li & Wang | Frontiers in Neuroscience | No | Journal | East China University of Science and Technology | China | 16 | Classification of EEG signals | Monitoring | Cognitive | Mental workload | Improve State-of-the-Art | MWL classification with CNN & ECNN | ACAMS (Automation-enhanced Cabin Air Management System) | N/M | (Nihon Kohden) | PSD | Internal Recordings | Private | 6 subjects x 2 sessions x 10 tasks x 5 min (2s windows) | 18000 | 600 | 6 | 10 | 500 | Offline | 1) Low-Pass filter: 40Hz | Yes | No | No | PSD (STFT) Avg. Power: D (1-4Hz), T (5–8 Hz), A (9–13 Hz), B1 (14–16 Hz), B2 (17–30 Hz), G (31–40 Hz) | Frequency-domain | N/M | Python Matlab | CNN ECNN | CNN | Many architectures tested | N/M | Yes | 102x10 (not clear what x what) | [2, 10] (tested many) | 10 | ReLU | N/M | N/M | 4 and 7 | Low/Normal/High/ - Unloaded - Very Low/Low/Medium/High/Very High/Overloaded | 4 and 7 | N/M | N/M | N/M | Nesterov Momentum Adagrad Adadelta Adam | Adam | (see paper, they describe each optimizer params) | N/M | N/M | N/M | N/M | Cross-Entropy | Inter | 5-Fold CV | k-fold | Train: 50% Test: 50% | Accuracy Precision F-Measure G-Measure | accuracy, precision, f-measure, g-measure | Single Intel core i5 CPU, 4-GB memory, Windows | N/M | 93% | LDA NB SDA | Traditional pipeline | No | No | No | "It was found that the deeper CNN model with the small convolutional kernels leads to improved classification performance." [YR] --> Like in other fields... | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Zhang2017 | ||||||||
| 71 | Deep RNN learning for EEG based functional brain state inference | 2017 | Patnaik, Moharkar & Chaudhari | International Conference on Advances in Computing, Communication and Control (ICAC3) | No | Conference | Xavier Institute of Engineering, Mahim, Mumbai M G M Inst. of Health Sciences, Navi Mumbai | India | Classification of EEG signals | BCI | Active | Mental tasks | New Approach: Brain State Inference with RNN using Alpha Phase Coherence | 5 Tasks: Baseline, Multiplications, Rotations, Letter Composition, Visual Counting (not using baseline) | N/M | ERD/ERS (looking at Alpha Cross Coherence - Occipital/Center) | Keirn & Aunon (1989) | Public | DB: 7 subjects x 10 sessions x 5 tasks x 10s Then they say they used 65 instances for 4 activities. not clear... (Sliding window of 50 samples) | N/M | 33 | 7 | 6 | 250 | 1) Band-Pass Filter: 0.1-100Hz (Hardware) 2) ICA for EOG Artifacts 3) DWT to get Alpha Sub-Bands 4) Hilbert Transform (no-overlap) for Phase Coherence | Yes | Yes | Yes | Alpha Sub-Bands Phase Coherence | Frequency-domain | N/M* | Elman's RNN with Bottlenect | RNN | A 5-layer network with 53-400 - 50-200-20-T | Yes | [Shape Not Mentioned] | 5 | 5 | N/M | N/M | 4 Classes | N/M* | Standard | N/M* | N/M | N/M | N/M | No | MSE | Inter | No | No | Train: 40 instances (/65) Test: 25 instances (/65) | Accuracy | accuracy | N/M* | 90% for two tasks 82% for three tasks 77% for all the four tasks | No | None | No | No | No | "In this research, a RNN model is trained to identify the phase coherence patterns of EEG alpha-bands. Difference between EEG signals from central and occipital (C1-O1 & C2- O2) locations is considered to compute phase coherence patterns for various activities." | No | N/A | No | Yannick Roy | TBR | TBC | Patnaik2017 | ||||||||||||||||||||||
| 72 | Deep Convolutional Neural Networks for Interpretable Analysis of EEG Sleep Stage Scoring | 2017 | Vilamala, Madsen & Hansen | IEEE International Workshop on Machine Learning for Signal Processing | Yes | Conference | Technical University of Denmark Danish Research Centre for Magnetic Resonance | Denmark | Classification of EEG signals | Clinical | Sleep | Staging | New Approach: CNN for Sleep Stages | Sleep | N/M | PSD | Sleep EDF | Public | SleepEDF 2 whole nights x 20 subjects (2 x ~10h x 20) 30s windows | 48000 | 24000 | 20 | 1 | 100 | Multitaper Spectral Estimation | Yes | N/M | N/M | Spectrogram log values (from Multitaper Spectral Estim.) | Frequency-domain | N/M* | CNN | CNN | VGGNET Activation Function: ReLU & Softmax Xavier’s initialisation. | No | 224x224 (RGB Image) | 16 | 16 | Dropout | Yes | 5 (Sleep Stage) | N/M* | Pre-training | Adam | Adam | Learning Rate: 10^-5 Mini-batch: 250 Decay Rate 1st & 2nd moments 0.9 & 0.999 | N/M | N/M | No | Categorical cross-entropy | Inter | Leave-One-Subject-Out | Leave-One-Subject-Out | Train: 15 subjects Valid: 4 subjects Test: 1 subjects | Precision Sensitivity F1-score Accuracy | precision, sensitivity, f1-score, accuracy | N/M* | [VGG-FE] Precision: 91, Sensitivity: 73, F1-S: 81, Accuracy: 83 [VGG-FT] Precision: 93, Sensitivity: 78, F1-S: 84, Accuracy: 86 | SSAE, CNN | DL | No | Sensititvity maps | Saliency map | Further improvement of the method includes better hyperparameter optimisation when generating the spectral images | No | N/A | Yes | Yannick Roy | TBR | TBC | Vilamala2017 | |||||||||||||||||||||
| 73 | Deep long short-term memory structures model temporal dependencies improving cognitive workload estimation | 2017 | Hefron, Borghetti, Christensen & Kabban | Pattern Recognition Letters | No | Journal | Air Force Institute Air Force Research Laboratory | USA | 9 | Classification of EEG signals | Monitoring | Cognitive | Mental workload | Improve State-of-the-Art: MWL classification with RNNs (LSTM). | Multi-Attribute Task Battery (MATB) environment | Using deep RNNs to account for temporal dependence considerably improves day-to-day feature stationarity | N/M | PSD (Raw EEG) | Internal Recordings | Private | 6 of 8 subjects x 5 sessions x 6 of 9 trials x 5 min This process yielded 380 features for each second and approximately 9000 observations per individual for the five day period. (10s slidding windows, 9s overlap) | 54000 | 900 | 6 | 19 | 256 | The power spectral density was determined for 30 points spread out over a logspace from 3 Hz to 55 Hz by extracting power from complex Morlet wavelets [9] . Each wavelet was 2 s in length | Yes | N/M | N/M | Mean, Variance, Skewness, Kurtosis of PSD (delta (1–4), theta (4–8), alpha (8–14), beta (15–30), and gamma (30–55)) + all possible combinations of M, V, S, K. | Frequency-domain | Keras, Theano | LSTM | RNN | N/M* | Yes | 600 x 30 x F (batch size, temporal depth in seconds, and number of features) (F varies between 90 and 380 features) | 2 LSTM Layers (50 and 10 units | 2 | Dropout | Yes | 1 (low or high WL) | N/M* | Standard | Mini-batch gradient descent (600 obs. per batch) Adam, Dropout 20% | Adam | 600 | Random search | Yes | No | Binary Cross-Entropy | Intra | 4-Fold CV | k-fold | Train: 3 days Valid: 1 day Test: 1 day | Accuracy | accuracy | N/M* | 93% (using all measures: M/V/S/K) | linear SVM (SVM-L), Radial Basis Function (RBF) SVM (SVM-R), feedforward ANN (ANN), deeply stacked simple RNN (RNN-D), single LSTM (LSTM-S), and deeply stacked LSTM (LSTM-D) | DL & Trad. | ANOVA, Tukey HSD | No | No | There is an abundance of future work to be pursued in this area. Due to time constraints and computational complexity, only a select number of deep architectures were examined during this re- search. A thorough evaluation of different deep RNN architectures to include variations in the depth of hidden layer recurrent con- nections, stacking of different sized LSTM layers, and interleaving fully-connected feedforward layers between sequence-to-sequence recurrent layers may yield additional improvement. | No | N/A | No | Yannick Roy | TBR | TBC | Hefron2017 | |||||||||||||||||||
| 74 | The signature of robot action success in EEG signals of a human observer: Decoding and visualization using deep convolutional neural networks | 2017 | Behncke, Schirrmeister, Burgard & Ball | Arxiv | Yes | Preprint | Albert-Ludwigs-University Freiburg University Medical Center Freiburg | Germany | 6 | Classification of EEG signals | BCI | Reactive | ERP | Novel Approach: DL for Robot Error Detection | Comparing CNN to rLDA and FB-CSP (both state of the art) for error detection in human-robot interaction | Participant watching short videos of robots "performing naturalistic actions either in a correct or an erroneous manner" | Deep Learning has been tried for other EEG decoding tasks | N/M | Error Potential | Internal Recordings | Private | 5 subjects x 720 trials + 12 subjects x 800 trials x ~20s/trials (using ~2.8s/trials) | 13200 | 616 | 17 | 128 | N/M | Offline | 1) Re-reference to common average (CAR) 2) Downsampled to 250 Hz | Yes | No | No | Raw EEG | Raw EEG | Electrode-wise exponential moving standardization | Braindecode | CNN | CNN | Deep ConvNet from braindecode paper | Layer 1: temporal filtering, Layer 2: spatial filtering, with no non-linearity in-between (Braindecode) | No | Time x channels | 5 | 5 | ELU | Dropout Early stopping | Yes | 2 | Error No error | 2 | N/M | Standard optimization | Standard | Adam | Adam | N/M | N/M | N/M | N/M | N/M | Categorical cross-entropy | Intra | No | No | N/M | Accuracy | accuracy | N/M | N/M | KPO Error (2.5-5s): (78.2 ± 8.4) % KPO Error (3.3-7.5s): (71.9 ± 7.6) % RGO Error (4.8-6.3s): (59.6 ± 6.4) % RGO Error (4-7s): (64.6 ± 6.1) % | rLDA FB-CSP (CNN is better) | Traditional pipeline | Permutation test on individual decoding results Wilcoxon signed-rank tests | Correlation of changes in ConvNet predictions with perturbation changes in 1) input spectral amplitudes and 2) time domain signals to obtain information about what the deep ConvNets learned from the data | Input-perturbation network-prediction correlation maps | "Among other recent advances in the field of deep learning research, automatic hyperparameter optimization and architecture search, including recurrent and residual network architectures, data augmentation, using 3-D convolutions, or increasing the amount of training data all have the potential to further increase ConvNet performance." | N/M | No | N/A | No | Yannick Roy | Hubert Banville | TBC | Behncke2017 | ||||||||
| 75 | Deep learning with convolutional neural networks for EEG decoding and visualization | 2017 | Schirrmeister, Springenberg, Fiederer, Glasstetter, Eggensperger, Tangermann, Hutter, Burgard, Ball | Human Brain Mapping | Yes | Journal | University of Freiburg | Germany | 30 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA Feature visualization/interpretability | Find out best CNN architecture for EEG decoding | Motor imagery/execution | Can learn from raw data | N/M | None | BCI Competition IV - IIa; Internal Recordings; BCI Competition IV - IIb; Mixed Imagery Dataset | Both | DS #1 - BCI Comp IV - IIa: 9 * 2 * 288 = 5184 x 4s DS #2 - Internal Recordings: 14 * 1000 = 14000 x 4s DS #3 - BCI Competition IV - IIb: 9 * 720 = 6480 x 4s DS #4 - Mixed Imagery Dataset: 4009 trials / 37830 w (DS4: 2s window, 1.5s overlap) | 5184; 14000; 6480; 37830 | 345.6; 933.33; 432; 267 | 9; 14; 9; 4 | 22; 44; 3; 64 | 250; 250; 250; 250 | BCI Competition Datasets: 1) Lowpass @38 Hz | Yes | Yes (removed trials with at least one channel > 800 uV) | Yes | Raw EEG | Raw EEG | Electrode-wise exponential moving standardization | Lasagne | CNN | CNN | 1) Deep ConvNet 2) Shallow ConvNet 3) Hybrid of 1) and 2) with 2 dense layers 4) ResNet | 1) Layer 1: temporal filtering, Layer 2: spatial filtering, with no non-linearity in-between 2) Embedding FBCSP in a ConvNet 3) Combining 1 and 2 4) 2 layers like in 1) | Yes | 1) 5 2) 2 3) max(2, 5) + 2 = 7 4) 31 | 31 | 1) ELU 2) Square, log 3) ELU, square & log 4) ELU | Dropout (0.5) Early stopping | Yes | 2 or 4 | N/M | Standard optimization | Standard | Adam | Adam | Batch norm | N/M | N/M | N/M | Crops (sliding windows within 1 trial) | Categorical cross-entropy For cropped training: "Tied loss function" | Intra | No | No | 1) 288 - 288 2) 880 - 160 3) 400 - 320 4) Variable per subject | Accuracy Confusion matrices | accuracy, confusion matrix | Geforce GTX Titan Black Intel Xeon @2.60 GHz with 32 cores 128 GB RAM | N/M | Filter bank common spatial patterns | Traditional pipeline | Wilcoxon sign-rank test | Input-feature unit-output correlation maps (visualization of correlation between spectral bands and receptive fields) Input-perturbation network-prediction correlation map (perturbing the input and visualizing change in output of net) | Input-feature unit-output correlation maps, Input-perturbation network-prediction correlation maps | ConvNets reached FBCSP accuracies ConvNet design choices substantially affects decoding accuracies Recent DL advances substantially increases accuracies ResNet performed worse than deep ConvNet Cropped training strategy improves performance on higher frequencies And much more! | ConvNets can be too flexible, especially if there is a specific type of brain activity that a user should use | Yes | GitHub | Yes | Hubert Banville | TBR | TBC | Schirrmeister2017 | |||||||||||||
| 76 | Optimal Feature Selection and Deep Learning Ensembles Method for Emotion Recognition From Human Brain EEG Sensors | 2017 | Mehmood, Du & Lee | IEEE Access | No | Journal | Chonbuk National University, Nanjing University of Posts and Telecommunications | South Korea | 10 | Classification of EEG signals | Monitoring | Affective | Emotion | Improve SOTA | Ensemble Method with DL and others to improve SOTA in EEG emotion classification | Watching "Emotional" Images from IAPS database. | Using ensemble approach. | EPOC (Emotiv) | Emotions | Internal Recordings | Private | 21 subjects x 4 classes x 2 sessions x 45 trials 360 epochs, 368s / session (1.5s windows, no overlap) | 7560 | 189 | 21 | 14 | 128 | 1) Artifact Removal (cites: Gómez-Herrero et al., 2006) 2) Filtering (cites: Widmann et al., 2012) 3) Epoching | Yes | Yes | Yes | Hjorth parameters for different frequency ranges + ANOVA feature selection | Frequency-domain | N/M | EEGLAB Matlab WEKA | "Deep Learning" (they don't even specify) | N/M | Ensemble: LDA, KNN, SVM, Naive/Bayes-Net, DT, RF, Deep Learning They don't describe the DL model at all | N/M | Yes | 3 Hjorth params for each of the 5 frequencies (?) | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Pre-Training and Fine Tuning | Pre-training | N/M | N/M | N/M | N/M | N/M | N/M | No | N/M | Inter | 10-Fold CV | k-fold | Train: 90% Test: 10% | Accuracy | accuracy | SOTA Server 4 TITAN-X (Pascal) | N/M | Accuracy: 76.62% | Jirayucharoensak et al., 2014 (SAE): 46/50% Chanel et al., 2006 (FDA, Naive Bayes): 72% Khalili et al., 2008 (LDA, KNN): 61% Horlings et al., 2008 (SVM): 37/32% Jenke et al., 2014 (...): 45% Yin et al., 2017 (SAE, Ensemble): 84/83% Atkinson et al., 2016 (...): 73/73% | DL & Trad. | ANOVA | No | No | Comparatively, the proposed method performs better than existing emotion recognition methods. The proposed feature selection method OF obtained the best emotion recognition rates of 76.6% for Voting ensembles method. Based on our results, we conclude that optimal feature selection is a good choice for enhancing the performance of EEG-based emotion recognition. | To further improve emotion recognition performance, we need to explore additional feature combinations with more emotional classes in the arousal–valence domain. | No | N/A | No | Yannick Roy | TBR | TBC | Mehmood2017 | |||||||||||
| 77 | Emotion Recognition based on EEG using LSTM Recurrent Neural Network | 2017 | Alhagry, Fahmy & El-Khoribi | International Journal of Advanced Computer Science and Applications (IJACSA) | No | Journal | Cairo University | Egypt | Classification of EEG signals | Monitoring | Affective | Emotion | Improve SOTA: Using LSTM on raw EEG to classify emotions (arousal, valence, liking) | Emotion Classification on DEAP (Like or Dislike video) | N/M | Raw EEG | DEAP | Public | DEAP: 32 subjects x 40 x 1min 12x 5s windows per video (5s window, no overlap) | 15360 | 1280 | 32 | 32 | 512 | 1) Downsampled to 128Hz (in the dataset) 2) Re-reference to Common Average (in the dataset) 3) Eye Artifacts Removed (in the dataset) 4) High-Pass Filter [freq not mentioned] | Yes | Yes | Yes | Raw EEG (None) | Raw EEG | Keras, TensorFlow | LSTM | RNN | AF: ReLU and Sigmoid | Yes | 5s segments x 32 channels (672 x 32) | 2 LSTM Layers (64,32) + 1 Dropout (0.2) + 1 FC | 3 | Dropout | Yes | 3 Classes | 5534113 | Standard | RMSProp, LR:0.001 | Other | N/M | N/M | No | N/M* | Intra | 4-Fold CV | k-fold | Train: 75% Test: 25% | Average Accuracy | accuracy | N/M* | Arousal: 85.65% Valence: 85.45% Liking: 87.99% | Traditional pipelines Koelstra et al., [2]: 62 | 56 | 55 % Atkinson ... [3]: 73 | 73 | - % Yoon and Chung [6]: 70 | 70 | - % Naser and Saha [7]: 66 | 64 | 70 % proposed method: 86 | 85 | 88 % | Traditional pipeline | No | No | No | Results show that the proposed method is a very promising choice for emotion recognition, because of its powerful ability to learn features from raw data directly. It achieves high average accuracy over participants compared to the traditional feature extraction techniques. | No | N/A | No | Yannick Roy | TBR | TBC | Alhagry2017 | ||||||||||||||||||||||
| 78 | Intent Recognition in Smart Living Through Deep Recurrent Neural Networks | 2017 | Zhang, Yao, Huang, Sheng & Wang | International Conference on Neural Information Processing (ICONIP) | Yes | Conference | University of New South Wales, AU Macquarie University, AU Singapore Management University, Singapore | Australia | 11 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA | Using LSTM on multiclass BCI open dataset Use hyperparameter fine-tuning method | Motor Imagery (see eegmmidb dataset) | Explore multiclass as opposed to binary classification like many others. BCI at home will be multiclasses. | N/M | Intent / Motor Imagery | eegmmidb | Public | eegmmidb: 10 subjects x 28,000 samples (28000 points @ 160Hz = 175s/subject) (window length = 1 point) | 28000 | 29.2 | 10 | 64 | 160 | None | No | N/M | N/M | Raw EEG (None) | Raw EEG | N/A | N/M | LSTM | RNN | N/A | N/M | Yes | 1 x 64 (sample x channels) | 5 | 5 | Sigmoid | L2 | Yes | 5 | eegmmidb: eye closed, left hand, right hand, both hands, both feet emotiv: up arrow, down arrow, left arrow, right arrow, eye closed | 5 | N/M | N/M | N/M | Adam | Adam | LR: 0.004 Lambda: 0.005 | N/M | Orthogonal Array (OA) experiment method | Yes | N/M | Cross-Entropy | Inter | No | No | Train: 75% Test: 25% | Accuracy Recall F1 Score ROC | accuracy, recall, f1-score, ROC | N/M | N/M | Accuracy: 0.9545 Recall: 0.9228 F1: 0.9382 AUC: 0.9985 | Almoari [2] 0.7497, Sun [13] 0.65, Major [4] 0.68, Shenoy [12] 0.8206, Tolic [16] 0.6821, Ward [19] 0.8, Pinheiro [10] 0.8505 KNN (k=3) 0.8369, SVM 0.5082, RF 0.7739, LDA 0.5127, AdaBoost 0.3431, CNN 0.8409 | DL & Trad. | No | No | No | To achieve optimal recognition accuracy, we employ OA to op- timize the hyper-parameters. In this paper, we select five most common hyper-parameters including λ (the coefficient of L2 norm), lr (learning rate), Ki(the hid- den layer nodes size), I (the number of layers), and nb (the number of batches). | N/A | Yes | Website | No | Yannick Roy | TBR | TBC | Zhang2017d | |||||||||
| 79 | Deep Recurrent Neural Networks for seizure detection and early seizure detection systems | 2017 | Talathi | Arxiv | Yes | Preprint | Lawrence Livermore National Lab | USA | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA: Using RNN for early seizure dectection | Using GRU-RNN for early seizure detection | Resting State, Eyes Open, Eyes Closed, Seizures. | Using available data to test RNNs for seizure detection. | N/M | Seizures | Bonn University | Public | Bonn University 5 x 100 x 23.6s 173.61 x 23.6 = 4097 --> 51 segments x 80 (0.46s windows) | 25500 | 197 | 15 | 1 | 173.6 | None (see dataset preprocessing steps) | No | N/M | N/M | Raw EEG (None) | Raw EEG | N/A | Keras | GRU (RNN) | RNN | GRU -> FC -> GRU | GRU for RNN long-term dependencies, but control the vanishing gradient | Yes | 51 x 80 x 1 (51 EEG sub-segment x 80 values x 1 channel) | GRU: 2 FC: 1 | 3 | N/M | N/M | N/M | 3 (Logistic Regression with Softmax) | In the order of 100,000 | (1) We train the RNN in stateful-mode*. (2) Rescaling the learning rate by factor 0.1 at each 100th epoch | Standard | Adam | Adam | LR: 0.01 | N/M | N/M | N/M | N/M | N/M | Inter | No | No | Train: 50% Test: 50% | Accuracy | accuracy | N/M | N/M | 98% Accuracy within the first 5 sec (3 classes: Healthy vs Ictal vs InterIctal) | They mentioned (A. T. Tzallas et al., 2007) getting 98% accuracy (ANN). | Traditional pipeline | No | No | No | This findings offers a strong support to the utility of GRU-RNN model for use in early-seizure detection system that can be extremely useful for developing closed loop seizure control systems where timely intervention can be leveraged to abate seizure progression | - | No | N/A | No | Yannick Roy | TBR | TBC | Talathi2017 | ||||||||||||
| 80 | DeepSleepNet: a Model for Automatic Sleep Stage Scoring based on Raw Single-Channel EEG | 2017 | Supratak, Dong, Wu & Guo | Arxiv | Yes | Preprint | Imperial College London | UK | 11 | Classification of EEG signals | Clinical | Sleep | Staging | Improve SOTA: Using CNN+LSTM for Sleep Stage Scoring from Raw EEG | Combining CNN + LSTM for Raw EEG and testing it on 2 different existing datasets | Sleep | Using RNN (LSTM) to capture time depencies in sleep stages. | N/M | Sleep Stages | MASS; Sleep EDF | Public | MASS: Used SS3, PSG recordings from 62 subjects Sleep EDF: Used 20 subjects (30s windows, no overlap) | 58600; 41950 | 29300; 20975 | 62; 20 | 20; 2 | 256; 100 | 1) Notch filter: 60Hz 2) Band-pass filter: 0.30 - 100Hz | Yes | No | No | Raw EEG (None) | Raw EEG | N/M | TensorLayer eTRIKS | CNN + bi-LSTM | CNN+RNN | 1D Conv, Batch Norm, Max Pooling | First part is representation learning, which can be trained to learn filters to extract time-invariant features from each of raw single-channel EEG epochs. The second part is sequence residual learning, which can be trained to encode the temporal information. | Yes | 30s EEG Epoch (2 diff sampling freq) | 2 CNN 2 bi-LSTM | 4 | ReLU | L2 Dropout (50%) | Yes | 5 Sleep Stages (Softmax) | N/M | The two-step training algorithm (their technique) to prevent from suffering from class imbalance. The algorithm first pre-trains the representation learning part of the model and then fine-tunes the whole model using two different learning rates. | Pre-training | Adam | Adam | LR: 0.0001 b1: 0.9 b2: 0.999 | 100 | N/M | N/M | Oversampling to balance classes (duplicating minority sleep stages) | Cross-Entropy | Inter | DS 1 - MASS) 31-Fold DS 2 - Sleep EDF) 20-Fold | k-fold | DS 1 Train: 60 subjects DS 1 Valid: 2 subjects DS 2 Train: 30 subjects DS 2 Valid: 1 subject | Precision (PR) Recall (RE) F1-score (F1) macro-averaging F1-score (MF1) Accuracy (ACC) Cohen’s Kappa coefficient (κ) | precision, recall, f1-score, macro-averaging f1-score, accuracy, Cohen's kappa | NVIDIA GeForce GTX980 | The training time for each validation fold was approximately 3 hours on each node | Sleep EDF - Acc: 82.0 Sleep EDF - MF1: 76.9 Sleep EDF - k: 0.76 MASS - Acc: 86.2 MASS - MF1: 81.7 MASS - k: 0.80 | Traditional pipelines & DL Sleep EDF: Y.-L. Hsu et al., 2013 Sleep EDF: R. Sharma et al., 2017 Sleep EDF: A. R. Hassan et al., 2017 Sleep EDF: O. Tsinalis et al., 2016a Sleep EDF: O. Tsinalis et al., 2016b MASS: H. Dong et al., 2016 | DL & Trad. | No | Visualization of filter activations | Analysis of activations | It achieved similar overall accuracy and macro F1-score compared to the state-of-the-art hand-engineering methods on both the MASS and Sleep-EDF datasets, which have different properties such as sampling rate and scoring standards (AASM and R&K). | N/M | Yes | GitHub | No | Yannick Roy | TBR | TBC | Supratak2017 | |||||||||||
| 81 | Mixed Neural Network Approach for Temporal Sleep Stage Classification | 2017 | Dong, Supratak, Pan, Wu, Matthews & Guo | IEEE Transaction on Neural Systems and Rehabilitation Engineering | Yes | Journal | Imperial College London | UK | 11 | Classification of EEG signals | Clinical | Sleep | Staging | Improve SOTA: Using Mixed NN on 1 channel EEG for Sleep Stage Scoring | Combining MLP + LSTM on 1-Channel Raw EEG from an existing (open) dataset | Sleep | Using RNN (LSTM) to capture time depencies in sleep stages and using a single, frontal (skin) electrode. | N/M | Sleep Stages | MASS | Public | MASS: 62 subjects (~ 494h) (30s windows, no overlap) | 58600 | 29300 | 62 | 1 | 256 | N/M Seems to directly do SFTF for freq features | N/M | No | No | PSD Features | Frequency-domain | N/M | Theano | Mixed NN (MNN) MLP + LSTM | RNN | N/M | Our MNN is composed of a rectifier neural network which suitable for detecting naturally sparse patterns [18], and a long short-term memory (LSTM) for detection of temporally sequential patterns [19] | Yes | 30s EEG Epoch PSD | MLP: [2,5] LSTM: 1 (200-1000) | 6 | ReLU | Dropout | Yes | 5 Sleep Stages (Softmax) | N/M | N/M | N/M | SGD | SGD | LR: 0.01 Momentum: 0.9 no weight decay | 500 | Manual fine tune | Yes | Oversampling to balance classes | Cross-Entropy | Inter | 31-Fold CV | k-fold | Train: 60 subjects Valid: 2 subjects | Macro F1-score (MF1) Accuracy (ACC) Recall (RE) Precision (PR) | macro f1-score, accuracy, recall, precision | NVIDIA 630 | 2 days | MF1: 80.50 ACC: 85.92 | SVM: 75.01 | 79.70 (best with sequence 2) RF: 72.44 | 81.67 (best with sequence 3) MLP: 77.23 | 81.43 (best with sequence 4) | DL & Trad. | No | No | No | (1) In terms of convenience, wearing the F4 channel near the hair line is imperfect. Other frontal EEG channels such as Fp2 and Fpz are easier to wear, but these channels have lesser information about stage W, N1, N2 and N3. (2) In our experiment, we tried to add fully connected layers between LSTM and softmax, and vary their hidden sizes, but no improvement was found. | Less inofrmation in low frontal (skin) channels (They've identified 3 challenges) Challenge 1. Heterogeneity Challenge 2. Temporal Pattern Recognition Challenge 3. Comfort | No | N/A | No | Yannick Roy | TBR | TBC | dong2018mixed | |||||||||||
| 82 | SLEEPNET: Automated Sleep Staging System via Deep Learning | 2017 | Biswal, Kulas, Sun, Goparaju, Westover, Bianchi & Sun | Arxiv | Yes | Preprint | Georgia Institute of Technology Nanyang Technological University Massachusetts General Hospital | USA | 17 | Classification of EEG signals | Clinical | Sleep | Staging | Improve SOTA: Using CNN, RNN, CRNN for Sleep Stage Scoring | Trying CNN, LSTM, RCNN on 10,000 subjects on Raw EEG, Expert Feature Set and Freq Bands for Sleep Stage Scoring | Sleep | Leveraging huge dataset (3.2TB) of 10,000 subjects to apply deep learning | N/M | Sleep Stages | Internal Recordings | Private | 10,000 overnight PSGs x ~8h / patient 80000 hours 3.2TB of data! Each 8h ~ 950-1000 labels (avg at 975) (30s windows, no overlap) | 9750000 | 4800000 | 10000 | 6 | 200 | None | No | No | No | 3 Sets of Features: 1) Raw EEG 2) Experts Defined Features 3) Spectrogram | Combination | N/M | Tensorflow CUDA 8.0 | 1) CNN 2) RNN 3) RCNN | CNN+RNN | 1) CNN: 1D Conv for Raw EEG / 2D Conv for Freq Features 2) RNN: Look back steps in RNN : [3,5,10,20,30] | By combining a RNN with CNN, we can have a hybrid model, namely, Recurrent-Convolutional Neural Networks (RCNN), which is able to extract features present in a spectrogram and preserve the long-term temporal relationship present in the EEG data | No | 30s EEG Epoch (depending on feature set) | RNN: 5 LSTM (1000) | 5 | ReLU | Dropout | Yes | 5 Classes | N/M | N/M | N/M | N/M | N/M | LR: [0.01 - 0.00001] | N/M | We performed 50 iterations of random search over a set of parameter choices for hyper-parameter tuning | Yes | N/M | Categorical Cross-Entropy | Inter | 50 Iterations of random search for hyper-parameter tuning | Train-Valid-Test | Train: 8700 patients Valid: 300 patients Test: 1000 patients | Accuracy Cohen's Kappa | accuracy, Cohen's kappa | Intel Xeon E5-2640, 256GB RAM, four Nvidia Titan X | between 40 -100 min | [RNN] - Expert Defined Features: [Acc] 85.76 | 79.46 [k] [RNN] - Spectrogram Features: [Acc] 79.21 | 73.83 [k] [RNN] - Waveform Features: [Acc] 79.46 | 72.46 [k] --- [RCNN] - Expert Defined Features: [Acc] 81.67 | 76.38 [k] [RCNN] - Spectrogram Features: [Acc] 81.47 | 74.37 [k] [RCNN] - Waveform Features: [Acc] 79.81 | 73.52 [k] | Logistic Regression Tree Boosting MLP CNN RNN RCNN | DL & Trad. | No | No | No | On 1000 held-out testing patients, the best performing algorithm achieved an expert-algorithm level of inter-rater agreement of 85.76% with Kappa value 79.46%, exceeding previously reported levels of expert-expert inter-rater agreement for sleep EEG staging. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Biswal2017 | |||||||||||
| 83 | DeepKey: An EEG and Gait Based Dual-Authentication System | 2017 | Zhang, Yao, Chen, Wang, Sheng & Gu | Arxiv | Yes | Preprint | University of New SouthWales Macquarie University RMIT University | Australia | 20 | Classification of EEG signals | Personal trait/attribute | Person identification | Improve SOTA: EEG for Person Identification | Use AR+RNN+SVM on EEG+Gait for Person Identification | Motor Imagery (see eegmmidb dataset) + Gait (PAMAP2 dataset) | The DL motivation is not clear. They want to improve SOTA. | N/M | Raw EEG | eegmmidb | Public | eegmmidb: 8 subjects x 13,500 samples 13,500 samples / 90 per window = 150 examples per subjects, 1200 total (90 points windows, no overlap) | 1200 | 11.3 | 8 | 64 | 160 | None (AR) | No | No | No | Raw EEG | Raw EEG | N/M | N/M | AR + RNN + SVM | RNN | N/M | AR for pre-processing, RNN for feature extracting, and SVM for classification. Auto-regressive Coefficients (AR) is one of the most widely used pre-processing methods on EEG data | Yes | 150x13x64 (150 segments, 13 coefficients (AR), 64 features/nodes) | 5 RNN (64) | 5 | N/M | L2 | Yes | 8 One-Hot Label (ID - 8 Subjects) | N/M | N/M | N/M | Adam | Adam | lambda is set as 0.004 while learning rate is set as 0.005 | 8 mini-batch with the shape of [150, 13, 64] | Orthogonal Array Experiment Method | Yes | N/M | Log Loss Function | Inter | No | No | Train: 87.5% Test: 12.5% | Accuracy | accuracy | N/M | N/M | Highest Accuracy: 0.9841 Gait: 0.999 Combined: 0.983 | [45]: PSD + cross-correlation values, [8]: Customized Threshold, [17]: Low-pass filter+wavelets+ ANN, [3]: Bandpass FIR filter +ECOC + SVM, [44]: IAF + delta band EEG + Cross-correlation & mahalobonis, [22]: CSP +LDA, [23]: AR + SVM | Traditional pipeline | No | No | No | The Gait Identification Model adopts a 7-layer deep learning model to process gait data and classify subjects’ IDs, achieving an accuracy of 0.999. The EEG Identification Model combines three components (auto-regressive coefficients, the RNN structure, and an SVM classifier) and achieves the accuracy of 0.9841 on a public dataset. Overall, the DeepKey authentication system obtains a FAR of 0 and a FRR of 0.019. | N/M | Yes | GitHub | No | Yannick Roy | TBR | Yes | Zhang2017c | ||||||||||||
| 84 | Multi-Person Brain Activity Recognition via Comprehensive EEG Signal Analysis | 2017 | Zhang, Yao, Zhang, Wang, Sheng, Gu | EAI International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services | Yes | Conference | University of New South Wales, Australia Singapore Management University Macquarie University, Australia RMIT University, Australia | Australia | 10 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA | Use AE + XGB for BCI-MI 5 classes (eegmmidb + internal recordings) | Motor Imagery (see eegmmidb dataset) | Deep learning should be able to generalize better across subjects and across classes, instead of binary classif. | N/M | Motor Imagery | eegmmidb | Public | eegmidb: 20 subjects x 28000 samples Total: 560,000 EEG samples (window length = 1 point) | 560000 | 58.3 | 20 | 64 | 160 | N/M | N/M | No | No | Raw EEG (None) | Raw EEG | z-score | N/M | AE + XGB Classifier | AE | Encoder, Decoder + XGB Classifier | N/M | Yes | 64x?? Channels x Raw EEG time window | 1 (64) Input - Encoder - Decoder - Classifier (XGB) | 1 | N/M | L2 | Yes | 5 | N/M | N/M | N/M | RMSProp | Other | LR: 0.01 | There are 9 mini-batches and the batch size is 17,280. | N/M | N/M | N/M | MSE | Inter | No | No | Train: 532,000 Test: 28,000 | Accuracy Precision Recall F1-Score ROC ROC AUC | accuracy, precision, recall, f1-score, ROC, ROC AUC | Nvidia Titan X Pascal 768G memory 145 TB PCIe SSD | See charts | Accuracy: 0.794 Precision: 0.7991 Recall: 0.781 F1 score: 0.7883 AUC: 0.9456 | SVM, RNN, LDA, RNN+SVM, CNN, DT, AdaBoost, RF XGBoost, PCA+XGBoost, PCA+AE+XGBoost, EIG+AE+XGBoost, EIG+PCA+XGBoost, DWT+XGBoost, SAE+XGBoost, AE+XGBoost | DL & Trad. | No | No | No | As part of our future work, we will build multi-view model of multi-class EEG signals to improve the classification performance. In particular, we plan to establish multiple models with each single model dealing with a single class. Following this philosophy, the correlation between test sample and each model can be calculated in the test stage and the sample can be classified to the class with minimum correlation coefficient. | N/M | No | N/A | No | Yannick Roy | TBR | Yes | Zhang2017a | |||||||||||
| 85 | Neurology-as-a-Service for the Developing World | 2017 | Dharamsi, Das, Pedapati, Bramble, Muthusamy, Samulowitz, Varshney, Rajamanickam, Thomas & Dauwels | Arxiv | Yes | Preprint | IBM Research AI Nanyang Technological University | USA | 5 | Classification of EEG signals | BCI | Active | Motor imagery | Improve SOTA: Use DL on the Cloud | Use DL on the Cloud for developing countries. Starting with a BCI Tasks (MI) | MI: Feet and Hands, real / imagined. | To develop neurology-as-a-service to learn features automatically from the data. This would help developing countries | N/M | Motor Imagery | eegmmidb | Public | eegmidb: 103 out of 109 subjects, 12 out of 14 tasks Segments 0.8 sec and sliding window of 0.05 sec The prepared data consisted of 17,232 samples (window length = 0.8s ??? not clear) | 17232 | N/M | 103 | 64 | 160 | 1) Bandpass: 3 - 30Hz 2) Generate Spectrogram: Hanning window & NFFT (128) | Yes | No (mention it, but only filters) | No | Spectrograms | Frequency-domain | N/A | N/M (Cloud) | CNN | CNN | N/M | N/M | No | 3D (channels x freq x time) | [1-3 3D CNN] [0-2 FC] | 5 | N/M | Dropout | Yes | N/M | N/M | N/M | N/M | (hyperparameters are automatically fine-tuned using an optimizer) | N/M | LR: 0.001 | N/M | Random Optimizer | Yes | N/M | N/M | Inter | No | No | Train: 70% Test: 30% | Accuracy | accuracy | N/M (Cloud) | N/M | Best accuracy: 63.4% | PCA-SVM | Traditional pipeline | No | No | No | As part of our next steps, we plan to use this framework on a dataset aimed at classification of epileptic seizures and/or pathological/normal EEG. We would also like to see how the framework performs using other hyperparameter optimization techniques including Bayesian optimization. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Dharamsi2017 | |||||||||||
| 86 | Deep Architectures for Automated Seizure Detection in Scalp EEGs | 2017 | Golmohammadi, Ziyabari, Shah, de Diego, Obeid, Picone | Arxiv | Yes | Preprint | Neural Engineering Data Consortium, Temple University | USA | 8 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA: Comparing different deep architectures | Compare HMM+sAE, HMM+LSTM, IPCA+LSTM, CNN+MLP, CNN+LSTM | Ongoing EEG recording, with and without seizures. | With big EEG corpus now available we can explore deep learning. | (Natus), (Nihon Kohden) | Seizures | TUH; Duke Seizure Corpus | Both | TUHS & DUZS: 1,864,012s ~ 517.8h, 159 subjects Multiple models with different window sizes | N/M | 31067 | 159 | 22; -1 | 250; N/M | N/M | N/M | N/M | N/M | LFCCs + First & Second Derivative of LFCCs | Other | N/A | N/M | 1) HMM + sAE 2) HMM + LSTM 3) IPCA + LSTM 4) CNN + MLP 5) CNN + LSTM | CNN+RNN | 2D Conv Layers -> Flatten -> 1D Conv Layer -> LSTM (output 1s data) -> LSTM -> 2-way sigmoid | They tried different architectures trying to capture Spatio-Temporal information. They also use Time-Freq Features, not raw EEG as is. | Yes | 210 @ 22 x 26 x 1 (Frames @ Channels * Features * 1) (to be reviewed) | 3x 2D CNN + 1x 1D FC CNN + 2x Bi-LSTM (CNN + LSTM) (see paper for others) | 6 | ELU | Dropout | Yes | 2-way Sigmoid | N/M | Trained + Eval on TUSZ and only Eval on DUSZ | Standard | Adam | Adam | N/M | N/M | N/M | N/M | N/M | MSE | Inter | Train-Valid-Test | Train-Valid-Test | Train: 614,382 (sec) Valid: 647,948 (sec) Test: 601,682 (sec) | Sensitivity Specificity | sensitivity, specificity | N/M | N/M | CNN + LSTM gave the best results. TUSZ - Sensitivity: 30.83% | Specificity: 96.86% DUSZ - Sensitivity: 33.71% | Specificity: 70.72% | HMM + Gaussian mixture + AE They compared 7 Optimizer Methods. (e.g. Adam, SGD, etc.) They compared 6 Activation Functions. (e.g. Tanh, Sigmoid, etc.) CNN + LSTM, Adam, ELU is the best combinaison | DL & Trad. | No | No | No | This is a significant finding because the Duke corpus was collected with different instrumentation and at different hospitals. Our work shows that deep learning architectures that integrate spatial and temporal contexts are critical to achieving state of the art performance and will enable a new generation of clinically-acceptable technology. | Access to labeled data and $ to label the data and make it public. | No | N/A | Yes | Yannick Roy | TBR | TBC | Golmohammadi2017a | |||||||||||
| 87 | Neonatal Seizure Detection using Convolutional Neural Networks | 2017 | O'Shea, Lightbody, Boylan, Temko | IEEE 27th International Workshop on Machine Learning for Signal Processing | Yes | Conference | Irish Centre for Fetal and Neonatal Translational Research, University College Cork | Ireland | 6 | Classification of EEG signals | Clinical | Epilepsy | Detection | New Approach | CNN on (preprocessed) raw EEG for neonatal seizure detection | Ongoing EEG recording, with and without seizures. | CNN works well on audio signal, why not on EEG. | N/M | Seizure | Internal Recordings | Private | 835 hours with 1389 seizures from 18 subjects splitted into 8s, 50% overlap. (not clear 50% versus 7s overlap / 1s shift) (8s window, 7s overlap) | 3006000 | 50100 | 18 | 8 | 256 | Band-pass filter: 0.5 - 12.8Hz Down sampled: 32Hz EEG Split into 8s windows (12.5% overlap) | Yes | No | No | Raw EEG 8 sec windows (1 sec shift) | Raw EEG | N/A | Keras | 1D - CNN | CNN | Conv - Batch Norm. - Pooling Output Layer: GAP (not dense) | "...1D CNNs wide convolutional filters (1-4s, 32-128 samples) significantly improved the performance". Sample size filters were used. In contrast to larger filter lengths allow the learning the various filters in a hierarchical manner [21]. | Yes | 256x1 (8s x 1 channel) | 6 | 6 | RELU Softmax | Batch Norm | Yes | 2 (Seizure vs Non-seizure) | 16,930 | The network was trained for 100 epochs, after each epoch the validation AUC was calculated. | Standard | SGD | SGD | LR: 0.003 LR -= 10% every 20 iterations Nesterov Momentum: 0.9 | 2048 | N/M | N/M | Sliding Window (Shifted by 1s, 7/8 overlap) | Categorical Cross-Entropy | Inter | Leave-One-Subject-Out | Leave-One-Subject-Out | Train: 17 subjects Test: 1 subject | ROC AUC | ROC AUC | N/M | N/M | AUC: 97.1% AUC90: 82.9% | SVM | Traditional pipeline | No | No | No | "We have also tried max pooling, which led to slightly inferior results in our experiments." "Initially, the EEG was converted to time-frequency images (spectrograms) and 2D CNNs were utilized, adopted from the area of image processing [17] – this architecture proved unsuccessful in the seizure detection task." | N/M | No | N/A | No | Yannick Roy | TBR | Yes | OShea2017 | |||||||||||
| 88 | Improving classification accuracy of feedforward neural networks for spiking neuromorphic chips | 2017 | Yepes, Tang & Mashford | International Joint Conference on Artificial Intelligence | Yes | Conference | IBM Research, VIC, Australia | Australia | 7 | Improvement of processing tools | Hardware optimization | Neuromorphic chips | New Approach: Running DL on Neuromorphic Chips | Compare constrained network for a neuromorphic chip on 2 datasets, vs unconstrained version of NN | MNIST & EEG Data from Nurse et al., 2015 (BCI-MI) | Implement DL/DNNs on a chip. | N/M | Motor Imagery | Nurse et al. (2015) | Public | From [Nurse et al., 2015]: 1 subject ~ 30 min 480/468 examples for training, 66/95 for testing (window length N/M) | 1109 | 30 | 1 | -1 | 1000 | N/M | N/M | No | No | N/M | N/M | [0, 1] | Matlab | [Esser et al., 2015] | N/M | [Esser et al., 2015] | [Esser et al., 2015] | No | N/M | Small Network: 3 Large Network: 4 | 4 | N/M | N/M | N/M | 2 | N/M | N/M | N/M | N/M | N/M | LR: 0.1 | 25 | N/M | N/M | No | N/M | Intra | No | No | Train: 80% Test: 20% | Accuracy | accuracy | TrueNorth (IBM Chip) | (see paper) | EEG Data: 86% MNIST: 98-99% | No | None | No | No | No | Furthermore, analysis of the learnt parameters pro- vide insights that might complement hardware design, thus providing a more efficient deployment of the trained models. The trained models use a small portion of the TrueNorth chip (30 cores vs. 4096 avail- able in the current version of the chip), thus requiring a much less than 70mW to work, which makes these models suitable for portable autonomous devices with large autonomy. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Yepes2017 | ||||||||||||
| 89 | Automatic Analysis of EEGs Using Big Data and Hybrid Deep Learning Architectures | 2017 | Golmohammadi, Hossein, Torbati, Lopez De Diego, Obeid & Picone | Arxiv | Yes | Preprint | Temple University Jibo, Inc., Redwood City | USA | 20 | Classification of EEG signals | Clinical | Epilepsy | Detection | New Approach: Hybrid HMM & SdA for Epilepsy | Using an Hybrid 3 Passes Model, combining HMM & Stacked Denoising AutoEncoders for Epilepsy classification | Ongoing EEG recording, with and without seizures. (TUH Dataset) | Deep Learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction | (Natus) | Seizure | TUH | Public | TUH Corpus Training from 359 sessions, Evaluation from 159 sessions. 113453 events total. Splitted 10s windows -> 1s epoch -> 0.1s frames (0.1 window, 0.2s overlap) | 113453 | 18909 | 518 | 128 | 1024 | PCA | Yes | No | No | Cepstral coefficient-based feature extraction approach based on Linear Frequency Cepstral Coefficients (LFCCs) | Other | N/M | Theano | 3x Stacked denoising Autoencoders (SDAE) | AE | 3 Passes. (1) HMM -> (2) SDAEs -> (3) NLP (2) PCA -> Out of Sample -> 3 SDAEs in parallel -> Enhancer (combining 3 SDAEs) | Not your typical DL-EEG approach... | Yes | 3 [Nodes from 100-800] | 3 | N/M | N/M | N/M | 6 Classes | N/M | Training of these three SDAE networks is done in two steps: pre-training and fine-tuning. Denoising autoencoders are stacked to form a deep network. The unsupervised pre-training of such an architecture is done one layer at a time. | Pre-training | Minibatch Stochastic Gradient Descent | SGD | LR: [0.1-0.5] | [100-300] | N/M | N/M | Out-of-sample technique (van der Maaten, 2009) | Cross-Entropy | Inter | No | No | Train: 84,032 events Test: 29,421 events | Sensitivity Specificity | sensitivity, specificity | N/M | N/M | Pass: Sensitivity | Specificity 1 (HMM): 86.78 | 17.70 2 (SDAE): 78.93 | 4.40 3 (SLM): 90.10 | 4.88 | No | None | No | No | No | A summary of the results for different stages of processing is shown in Table 12. The overall performance of the multi-pass hybrid HMM/deep learning classification system is promising: more than 90% sensitivity and less than 5% specificity. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Golmohammadi2017 | ||||||||||||
| 90 | Multimodal deep learning approach for joint EEG-EMG data compression and classification | 2017 | Ben Said, Mohamed, Elfouly, Harras & Wang | IEEE Wireless Communications and Networking Conference | Yes | Conference | Qatar University Carnegie Melon University University of British Columbia | Qatar | 6 | Classification of EEG signals | Monitoring | Affective | Emotion | New Approach: Compressing joint EEG-EMG with an autoencoder | Compression & Classification of joint EMG + EEG on DEAP dataset with SAE | Watching music videos (DEAP Dataset) | Deep learning approach has emerged as one of the possible techniques to exploit the correlation of the data from multiple modalities. Compression for mobile health data. | N/M | Emotions | DEAP | Public | DEAP 32 subjects x 40 videos x 63s (6s windows) | 23040 | 1280 | 32 | -1 | 128 | 1) 6s Windows 2) Whitened 3) Normalized | Yes | N/M | N/M | Raw EEG + EMG (None) | Raw EEG | z-score | N/M | SAE | AE | N/M | Deep learning approach has emerged as one of the possible techniques to exploit the correlation ofthe data from multiple modalities | No | N/M | 2 SAE | 2 | Sigmoid | L2 | Yes | N/M 1) Compressed data 2) Classification | N/M | Greedy-layer wise | Pre-training | N/M | N/M | N/M | N/M | N/M | N/M | Duplicated multimodal data keeping values from 1 modality, setting the other modality to 0. And vice-versa. | Square Euclidean Distance | Inter | No | No | [Compress] Train: 50% [Compress] Test: 50% [Classif] Train: 75% [Classif] Test: 25% | 1) Compression: Distortion 2) Classification: Accuracy | distortion, accuracy | N/M | N/M | 1) [Compression] Distortion: EMG = 13.85% | EEG = 12% 2) [Classification] Accuracy: 78.1% | Discrete Wavelet Transform (DWT) [26] Compressed Sensing (CS) [27] (distortion: 22% | 17.21%) 2D compression approach which is based on SPIHT and FastICA [28] | Traditional pipeline | No | No | No | 1) Compression: Distortion 2) Classification | N/M | No | N/A | No | Yannick Roy | TBR | Yes | BenSaid2017a | |||||||||||
| 91 | Deep Learning for Fatigue Estimation on the Basis of Multimodal Human-Machine Interactions | 2017 | Gordienko, Stirenko, Kochura, Alienin, Novotarskiy & Gordienko | Arxiv | Yes | Preprint | National Technical University of Ukraine | Ukraine | 12 | Classification of EEG signals | Monitoring | Physical | Exercise | New Approach | Multi-modal fatigue (and activity) estimation | Different activities (sports) while having different sensors | Use Multimodal Models. to combine different modalities with a NN. | OpenBCI (OpenBCI) | Multimodal | Internal Recordings | Private | N/M Not much information on the EEG data | N/M | N/M | N/M | -1 | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | DNN | N/M | N/M | N/M | No | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Mean Residual Deviance (MRD) Mean Absolute Error (MAE) | N/M | N/M | No | N/M | Mean Residual Deviance (MRD) Mean Absolute Error (MAE) | mean residual deviance, mean absolute error | N/M | N/M | See Paper (not really relevant / meaningful for this paper) | N/M | None | No | No | No | The main achievement is the multimodal data measured can be used as a training dataset for measuring and recognizing the intensity and physical load on the person by means of the machine learning approaches. | N/M | No | N/A | No | Yannick Roy | TBR | Yes | Gordienko2017 | |||||||||||
| 92 | Towards Deep Modeling of Music Semantics using EEG Regularizers | 2017 | Raposo, Matos, Ribeiro, Tang & Yu | Arxiv | Yes | Preprint | Universidade de Lisboa | Portugal | 5 | Classification of EEG signals | Music semantics | Improve SOTA on music semantics | Modeling of music audio semantics | Listening to music | Previous success of CNNs in music audio modeling | OpenBCI (OpenBCI) | None | Internal Recordings | Private | 60 music segments + 2 noise + 2 songs x 18 subjects music duration = average of 15.13s samples approx: 60 * 15.13 / 1.5 * 18 = 10894 (1.5s windows) | 10894 | 272.3 | 18 | 16 | 250 | 1) Highpass 0.5Hz 2) Notch at 50Hz | Yes | Yes | Yes | Raw EEG + Audio embeddings | Raw EEG | Rescaled [-1, 1] | N/M | CNN | CNN | N/M | N/M | Yes | N/M | 5 | 5 | ReLU | No | N/M | 128 | N/M | 1) Train audio+lyrics embeddings model 2) Train audio embeddings+EEG embeddings model | Standard | N/M | N/M | N/M | 102 | N/M | N/M | N/M | CCA between embeddings | Inter | 5-Fold CV | k-fold | Train: 80% Test: 20% | Mean Reciprocal Rank (MRR) | mean reciprocal rank | GeForce GTX 1080 | 20 minutes | Outperformed Spotfiy by ~1%, but did not perform better than the SOTA (by a small margin) | Spotify embeddings and current SOTA (Choi) | DL & Trad. | No | No | No | Proposed approach did not outperformed SOTA but SOTA was trained on more than 2083 hours of music, whereas the proposed method needs less than 3 hours of both music and EEG | N/M | No | N/A | No | Isabela Albuquerque | TBR | TBC | Raposo2017 | |||||||||||||
| 93 | Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals | 2017 | Acharya, Oh, Hagiwara, Tan & Adeli | Computers in Biology and Medicine | No | Journal | Ngee Ann Polytechnic, Singapore SUSS University, Singapore University of Malaya, Malaysia The Ohio State University, US | Singapore | Classification of EEG signals | Clinical | Epilepsy | Detection | New Approach: CNN for Epilepsy (claiming its a new approach, but it's not...) | 13-Layers CNN for Epilepsy | Ongoing EEG recording, with and without seizures. | To develop a computer-aided diagnosis (CAD) to classify EEG | N/M | Seizures | Bonn University | Public | Bonn University: B,D,E 3 x 100 x 23.6s (23.6s windows) | 300 | 118 | 10 | 1 | 173.6 | None | No | N/M | N/M | Raw EEG | Raw EEG | z-score | N/M | CNN | CNN | 1D CNN Conv / Max Pooling | Yes | 4097x1 | 1D CNN: 10 FC: 2 | 12 | ReLU | L1 | Yes | 3 (Softmax with 3 classes) | N/M | A conventional backpropagation (BP) [32] with a batch size of 3 is employed in this work to train CNN. | Standard | Adam | Adam | Lambda: 0.7 LR: 1x10^-3 Momentum: 0.3 | 3 | Trial and Error | Yes | No | N/M | Inter | 10-Fold CV | k-fold | Train: 90% Valid: 30% of 90% Test: 10% | Accuracy Specificity Sensitivity | accuracy, specificity, sensitivity | Intel Xeon 2.40 GHz (E5620) 24 GB RAM | 150 epochs 12.8s / epochs = 32 min | Accuracy: 88.7% Sensitivity: 95% Specificity: 90% | Many other SOTA (check paper) They performed worse than most previous SOTA | Traditional pipeline | No | No | No | The advantage of the model presented in this paper, however, is separate steps of feature extraction and feature selection are not required in this work. Nevertheless, the main drawback of this work is the lack of huge EEG database | Amount of data | No | N/A | Yes | Yannick Roy | TBR | TBC | Acharya2017 | |||||||||||||
| 94 | Electroencephalogram-based decoding cognitive states using convolutional neural network and likelihood ratio based score fusion | 2017 | Zafar, Dass & Malik | Plos One | No | Journal | Universiti Teknologu PETRONAS | Malaysia | 23 | Classification of EEG signals | BCI | Reactive | RSVP | Improve SOTA | Decode seen images by extracting features with a CNN | Watching natural images from 5 classes | Features learned automatically can be more efficient | (EGI) | None | Internal Recordings | Private | 26 subjects x 21min (1s windows) | 13520 | 546 | 26 | 128 | 250 | [Hardware: Bandpass from 0.1 to 100 Hz] 1) Bandpass from 0.3 to 30 Hz 2) Removal of eye artefacts | Yes | Yes | Yes | Raw EEG | Raw EEG | N/M | N/M | CNN | CNN | Modified LeNet CNN is *just* for feature extraction (feature selection and classification is done separately) | Temporal 1D conv in first layer | Yes | 128 x 250 | 2 | 2 | Sigmoid, tanh | N/M | N/M | 128 x 11 x 100 | N/M | ??? | Standard | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Monte-Carlo 100-Fold CV | Train-Valid-Test | Train: 90% Test: 10% | Accuracy Specificity Sensitivity | accuracy, specificity sensitivity | N/M | N/M | Accuracy (across participants, 5-class): 40% | Discrete Wavelet Transform + SVM | Traditional pipeline | two-sample t-test, ANOVA | No | No | Amount of data | No | N/A | No | Hubert Banville | TBR | TBC | Zafar2017 | ||||||||||||
| 95 | Deep Convolutional Neural Network for Emotion Recognition Using EEG and Peripheral Physiological Signal | 2017 | Lin, Li & Sun | International Conference on Image and Graphics | No | Conference | College of Computer Science of Zhejiang University, Hangzhou, China | China | Classification of EEG signals | Monitoring | Affective | Emotion | Improve SOTA: AlexNet on DEAP | AlexNet on Images (Raw EEG + Freq Bands) + other physiological sensors | Watching videos (check out DEAP details) | Using AlexNet on DEAP | N/M | Emotions | DEAP | Public | DEAP 32 subjects x 1min x 40 videos (6s windows) | 12800 | 1280 | 32 | 32 | 512 | 1) Downsampling to 128Hz 2) Band-Pass Filter: 4.0 - 45Hz 3) Average to Common Reference* (?) | Yes | No | No | EEG -> 6 gray images (Raw EEG + Freq Bands) + 81 features from other physiological sensors | Frequency-domain | min-max | N/M | CNN | CNN | AlexNet | AlexNet is great for images, frequency bands can be converted to images... | Yes | 6 Gray Images (2D) | 5 CNN 1 FC (81+500) | 6 | N/M | N/M | N/M | 2 Softmax | N/M | Fine-tuning AlexNet | Pre-training | SGD | SGD | LR: 0.001 (decreases every 500 iterations) | 200 | Empirically | Yes | N/M | N/M | Inter | 10-Fold CV | k-fold | Train: 90% Test: 10% | Accuracy F1-Score | accuracy, f1-score | N/M | N/M | Arousal - Accuracy: 87.30% Arousal - F1-Score: 78.24% Valence - Accuracy: 85.50% Valence - F1-Score: 80.06% | Many other SOTA (check paper) They outperform all others. | Traditional pipeline | No | No | No | To achieve better performances, data preprocessing of the original signal was also adopted. The provided experimental results prove the effectiveness and validate the proposed contributions of our method by achieving superior performance over the existing methods on DEAP Dataset. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Lin2017 | ||||||||||||
| 96 | Cross-subject recognition of operator functional states via EEG and switching deep belief networks with adaptive weights | 2017 | Yin & Zhang | Neurocomputing | No | Journal | University of Shanghai | China | 18 | Classification of EEG signals | Monitoring | Cognitive | Mental workload & fatigue | Improve SOTA on cross-subject operator functional state recogntion | Exploit "new" improvements in deep learning | Cabin air management simulation (AutoCAMS) | Using switching deep belief network with adaptive weights | (Nihon Kohden) | None | Internal Recordings | Private | 8 subjects x 1080 EEG Segments per subject (2s windows) | 8640 | 288 | 8 | 11 | 500 | Offline | 1) Adaptive exponential smooth (to remove outliers) | Yes | Yes | Yes | Centroid frequency, log-energy entropy, mean, five power components, Shannon entropy, sum of energy, variance, zero-crossing rate of each channel and power differences between channel pairs | Frequency-domain | z-score | Matlab 2011b | Switching DBN | DBN | One DBN per subject | The member DBN is switched at different time instants to fit the non-stationarity of the EEG features recorded from a novel testing subjects. | Yes | 152x1 | 4 | 4 | Sigmoid | N/M | N/M | 3 | Low MW Medium MW High MW | 3 | N/M | Unsupervised pre-training of DBNs to learn representation of features for each subject (layer by layer). Supervised fine-tuning of the complete model. | Pre-training | N/M | N/M | Pre-training: 0.1 Fine-tuning: 1 | 10 | N/M | N/M | Gaussian noise to feature vector | N/M | Inter | Leave-One-Subject-Out | Leave-One-Subject-Out | Train: 7 subjects Test: 1 subject | Accuracy True positive True negative False positive False negative | accuracy, true positives, true negatives, false positives, false negatives | AMD4CPU 1.9GHz, 8G RAM | N/M | Mental workload: 77% Mental fatigure: 68% MW+MF: 54% | KNN, Naive Bayes, Logistic Regression, LSSVM, SAE, DBN (all with and without PCA) | DL & Trad. | Two-tailed Wilcoxon sign-rank test | No | No | Results of the proposed method outperform all baselines. When the number of subjects increases, the performance gap between SDBN and baselines increases, suggesting that the number of subjects plays a fundamental role. | Number of subjects is crucial to obtain a good performance | No | N/A | No | Isabela Albuquerque | Yannick Roy | TBC | Yin2017 | ||||||||
| 97 | Vowel classification from imagined speech using sub-band EEG frequencies and deep belief networks | 2017 | Sree & Kavita | IEEE International Conference on Signal Processing, Communications and Networking | No | Conference | SSN College of Engineering | India | 4 | Classification of EEG signals | BCI | Active | Speech decoding | Improve SOTA on vowel classification | Use DBNs to extract EEG features | Speech imagery | N/M | Super Spec (RMS) | None | Internal Recordings | Private | 5 subjects x 75s experiment x ?? trials Between 15-20 min per subject not clear... | N/M | 87.5 | 5 | 32 | 128 | 1) Band-pass 1-60Hz | Yes | No | No | Energy features of Wavelet transform: Root Mean Square, Mean Absolute Value, Integrated EEG, Simple Square Integral, Variance of EEG, Average Amplitude Change | Frequency-domain | z-score | N/M | DBN | DBN | N/M | N/M | No | N/M | 7 | 7 | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | LR: 0.002 | N/M | N/M | N/M | N/M | Log-likelihood | Inter | No | No | Train: 80% Test: 20% | Accuracy | accuracy | N/M | N/M | ~87.5% (I believe this the average value for all vowels and EEG bands) | No | None | No | No | No | Vowels were more accurately classified in the theta and gamma bands | N/M | No | N/A | No | Isabela Albuquerque | TBR | TBC | Sree2017 | |||||||||||
| 98 | Bullying incidences identification within an immersive environment using HD EEG-based analysis: A Swarm Decomposition and Deep Learning approach | 2017 | Baltatzis, Bintsi, Apostolidis & Hadjileontiadis | Nature Scientific Reports | No | Journal | Aristotle University of Thessaloniki, Khalifa University of Science and Technology | Greece | 8 | Classification of EEG signals | Monitoring | Affective | Bullying incidents | New task: classifying bullying stimuli | Classifying bullying stimuli in 2D or VR presentation | Watching stimuli (2D or in VR) of bullying situations | N/M | (EGI) | None | Internal Recordings | Upon request | T1: 256 × 256 × 14 × 17 (x3 SWD) (channels × samples × trials × subject) T2: 256 × 192 × 16 × 17 (x3 SWD) (channels × samples × trials × subject) not clear... | 1530 | N/M | 17 | 256 | 250 | 1) Bandpass 0.3-30 Hz 2) Artefact detection, bad channel replacement, baseline correction 3) Channel-wise normalization (- mean, / max) 4) Highpass @7Hz 5) Downsample to 128 Hz | Yes | Yes | Yes | 1) Swarm decomposition to get oscillatory modes 2) k-means clustering to re-order channels based on the respective distance to each other | Other | N/M | N/M | CNN | CNN | N/M | N/M | Yes | 256 x 128 | 2 | 2 | ReLU | N/M | N/M | 2 or 4 | N/M | Standard optimization | Standard | N/M | N/M | N/M | N/M | N/M | N/M | N/M | "Softmax" | Inter | 10-Fold CV | k-fold | Train: 75% Valid: 10% of 75% Test: 25% | Accuracy Precision Recall ROC AUC | accuracy, precision, recall, ROC AUC | N/M | N/M | 2-class: Accuracy, precision, recall, AUC (test): 0.937, 0.9403, 0.9395, 0.9869 4-class: Accuracy, precision, recall, AUC (test): 0.8858, 0.8775. 0.87475, 0.975 | No Swarm decomposition or clustering Just clustering Just Swarm decomposition | Traditional pipeline | No | No | No | Swarm Decomposition was an important step in getting high accuracy. Withouth k-means clustering the network was overfitting. | Larger nets take more resources | No | N/A | No | Hubert Banville | TBR | Yes | Baltatzis2017 | |||||||||||
| 99 | Classification and discrimination of focal and non-focal EEG signals based on deep neural network | 2017 | Taqi, Al-Azzo, Mariofanna & Al-Saadi | International Conference on Current Research in Computer Science and Information Technology (ICCIT) | No | Conference | University of Arkansas at Little Rock Asiacell Company for Telecommunication, Iraq | USA | 7 | Classification of EEG signals | Clinical | Epilepsy | Detection | Improve SOTA | Detecting Focal vs Non-Focal Seizures with existing Deep Nets: AlexNet, LeNet, GoogleNet | Seizures (Bern-Barcelona Dataset) | deep neural network (DNN) is a high-res model that get sophisticated hierarchical features. (e.g. AlexNet, LeNet, GoogleNet) | N/M | Seizures | Bern-Barcelona EEG DB | Public | Bern-Barcelona EEG DB (600 out of 3750) 600 signal pairs: 300/300 x 40s (40s windows ??) | 600 | 400 | 5 | -1 | 256 | None | No | No | No | None (Raw EEG) | Raw EEG | N/M | Caffe | N/M (pre-trained models) | N/M | N/M (pre-trained models) | Using SOTA Networks in Vision/Images for EEG. (AlexNet, LeNet, GoogleNet) | No | 256x256 (images) | N/M (pre-trained models) | N/M | N/M (pre-trained models) | N/M (pre-trained models) | N/M | 2 | N/M | pre-trained models (AlexNet, LeNet, GoogleNet) | Pre-training | N/M (pre-trained models) | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Inter | No | No | Train: 75% Test: 25% | Accuracy | accuracy | NVidia GPUs | N/M | LeNet, AlexNet, GoogleNet 100% (with different numbers of TEs) LeNet is the best compromise | Anindya et al., 2016 : 89.4% (EMD-DWT domain, K-nearest neighbor classifier) R. Sharma et al.,2015 : 84% (DWT domain, KNN, PNN, fuzzy and LS-SVM) R. Sharma et al.,2014 : 85% (EMD domain, LS-SVM classifier) | Traditional pipeline | No | No | No | As a future task, we are looking forward to investigating approaches for EEG signals classification of other diseases, drunk people, or ECG signals classification | N/M | No | N/A | No | Yannick Roy | TBR | Yes | Taqi2017 | |||||||||||
| 100 | Deep Transfer Learning for Cross-subject and Cross-experiment Prediction of Image Rapid Serial Visual Presentation Events from EEG Data | 2017 | Hajinoroozi, Mao & Lin | International Conference on Augmented Cognition | No | Conference | University of Texas at San Antonio National Sun Yat-sen University, Tawain | USA | 11 | Classification of EEG signals | BCI | Reactive | RSVP | Novel Approach: Transfer Learning | Transfer learning on RSVP task with CNN on Raw EEG: (1) Cross-Suject (2) Cross-Experiment | RSVP (3 datasets from 1990, 1999, 2013) | Transfer learning has a lot of potential for BCI training. | ActiveTwo (BioSemi) | RSVP | USA DoD (1999); USA Army (1990); Touryan et al. (2013) | Private | DS #1 - CT2WS: 15 subjects x 15min DS #2 - Static: 16 subjects x 15min DS #3 - Expertise: 10 subjects x 5 sessions x 60min (1s windows, no overlap) | 65831; 62553; 21680 | 1097.2; 1042.6; 361.3 | 15; 16; 10 | 64; 64; 256 | 512; 512; 512 | 1) Bandpass filter: 0.1 - 55 Hz 2) Downsampled to 128 Hz 3) Epoching: 1s window | Yes | No | No | None (Raw EEG) | Raw EEG | N/M | N/M | STCNN (Spatial-Temporal CNN) | CNN | Pretty much a CNN with a fancy name. 2 Conv Layers + 3 FC with dropout | Trying to capture Spatial and Temporal information from Raw EEG | Yes | 64x128 | CNN: 2 FC: 3 | 5 | ReLU | Dropout | Yes | 2 Target / Non-Target (softmax) | N/M | The paper is about transfer learning. Training on 1 dataset, then fine-tuning (or not) on the other. | Pre-training | N/M | N/M | N/M | N/M | N/M | N/M | N/M | N/M | Inter | 10-Fold CV | k-fold | Train: 90% Valid: 10% | ROC AUC | ROC AUC | N/M | N/M | Ranging from 73-77%, depending on source / target datasets and transfer type (Cross-Subject or Cross-Experiment) | Bagging, XLDA, LDA | Traditional pipeline | No | All Layers: Subject Specific CNN Layers: Mostly Subj. Specific All Layers: General Info | Analysis of performance with transferred layers | This study represents the first comprehensive investigation of CNN transferability for EEG based classification and our results provide important information that will guide the design of more sophisticated deep transfer learning algorithms for EEG based classifications in BCI applications. | N/M | No | N/A | No | Yannick Roy | TBR | TBC | Hajinoroozi2017 |