a b/README.md 

  1
  
# Annotation (or segmentation) of the electrocardiogram (ECG) with a long short-term memory neural network. 





  
  

  2
  
Here, I experimented with annotating peaks of the ECG using a recurrent neural network in tensorflow's Keras.





  
  

  3
  
In the beginning I struggled a bit to get the input/output right, which had to do with the way I tried to format ECG-peaks (as a sparse vector containing peaks (1) vs no peaks (0)). Aproaching it as a semantic segmentation problem (e.g. Seq2Seq) solved it for me. 





  
  

  4
  
<br>It seems to work well on the QT database of physionet, but there are some cases that it has never seen where it fails; I haven't played with augmenting the ecgs yet.





  
  

  5
  
 





  
  

  6
  
[5oct 2019]





  
  

  7
  
Since posting this 3 years ago, I noticed several publications using the (exact) same principle:





  
  

  8
  
- Matlab: https://www.mathworks.com/help/signal/examples/waveform-segmentation-using-deep-learning.html





  
  

  9
  
- IEEE: https://ieeexplore.ieee.org/document/8333406





  
  

  10
  





  
  

  11
  





  
  

  12
  
## Model





  
  

  13
  





  
  

  14
  
```





  
  

  15
  
model = Sequential()





  
  

  16
  
model.add(Dense(32,W_regularizer=regularizers.l2(l=0.01), input_shape=(seqlength, features)))





  
  

  17
  
model.add(Bidirectional(LSTM(32, return_sequences=True)))#, input_shape=(seqlength, features)) ) ### bidirectional ---><---





  
  

  18
  
model.add(Dropout(0.2))





  
  

  19
  
model.add(BatchNormalization())





  
  

  20
  
model.add(Dense(64, activation='relu',W_regularizer=regularizers.l2(l=0.01)))





  
  

  21
  
model.add(Dropout(0.2))





  
  

  22
  
model.add(BatchNormalization())





  
  

  23
  
model.add(Dense(dimout, activation='softmax'))





  
  

  24
  
adam = optimizers.adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)





  
  

  25
  
model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy']) 





  
  

  26
  
```





  
  

  27
  





  
  

  28
  
## Getting Started





  
  

  29
  
- Download ECG data using something like <code>wget -r -l1 --no-parent https://physionet.org/physiobank/database/qtdb/</code>





  
  

  30
  
- Run, with as the first argument the directory where the ECG data is stored; or set <code>qtdbpath</code>.





  
  

  31
  





  
  

  32
  
## Output





  
  

  33
  
A 2 lead ECG, the colors indicate P-Pend(yellow),Pend-Q(green), Q-R(red),R-S(purple),S-Tend(brown),Tend-P(pink). Training took about an hour on 6 cores. 





  
  

  34
  
![example output](https://github.com/niekverw/Deep-Learning-Based-ECG-Annotator/blob/master/images/example.png?raw=true)





  
  

  35
  
- Colors at the top show true annotations





  
  

  36
  
- Colors at the bottom show predicted annotations





  
  

  37
  





  
  

  38
  
## Dependencies





  
  

  39
  
I haven't got a list of all dependencies; but use this:  





  
  

  40
  
- wfdb 1.3.4 ( not the newest >2.0, thanks BrettMontague); pip install wfdb==1.3.4