
# MetaPred
The repo is code (baseline and the proposed MetaPred) for paper MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records by [Xi Sheryl Zhang](https://www.xi-zhang.net), [Fengyi Tang](https://github.com/af1tang), [Hiroko H. Dodge](https://medicine.umich.edu/dept/neurology/hiroko-dodge-phd), [Jiayu Zhou](https://jiayuzhou.github.io), and [Fei Wang](https://sites.google.com/site/cornellwanglab/home).  

## Overview

MetaPred is a meta-learning framework for Clinical Risk Prediction using limited patient Electronic Health Records (EHRs). We given an example in the following figure:

<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/task-design.png?raw=true" alt=" Illustration of the proposed learning procedure" width="500"></p>

Suppose we have multiple domains, our goal is to predict Alzheimer’s disease with few labeled patients, which give rise to a low-resource classification. The idea is to employ labeled patients from high-resource domains and design a learning to transfer framework with sources and a simulated target in meta-learning. There are four steps: (1) constructing episodes by sampling from the source domains and the simulated target domain; (2) learn the parameters of predictors in an episode-by-episode manner; (3) fine-tuning the model parameters on the genuine target domain; (4) predicting the target clinical risk. We respectively implemented Convolutional Neural Network (CNN) and Long-Shot Term Memory (LSTM) Network as base predictors. The model overview (meta-training procedure) is shown as follows:

<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/MetaPred.png?raw=true" alt="MetaPred framework overview" width="750"></p>

The entire learning procedure can be viewed as: iteratively transfer the parameter Θ learned from source domains through utilizing it as the initialization of the parameter that needs to be updated in the target domain.

## Results

The learned representations of patients in five disease domains are visualized by t-SNE. In detail, AD, PD, DM, AM, MCI are abbreviations of Alzheimer's Disease, Parkinson's Disease, Dementia, Amnesia and Mild Cognitive Impairment, respectively. As a patient might suffer multiple diseases, there is supposed to be some overlaps among the given domains.

<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/patient_vis_metapred.png?raw=true" alt="Visualization of patient representation learned by MetaPred" width="500"></p>

To demonstrate the effectiveness of the proposed MetaPred in the context of domain adaptation, we compare it with the state-of-the-art meta-learning algorithm ``Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks`` (MAML). The results on Alzheimer's Disease domain are presented in terms of AUC and F1-Score.

<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/vs_maml_ad_cnn.png?raw=true" alt="Performance comparison of MetaPred and MAML on the top of Alzheimer's Disease" width="500"></p>


## Requirements
This package has the following requirements:
* `Python 3.x`
* [TensorFlow 1.5](https://github.com/tensorflow/tensorflow)
* [Progress Bar](https://progressbar-2.readthedocs.io/en/latest/index.html)


## Usage
### Baseline in Sequential Data Modeling 
The [baseline implementation](https://github.com/sheryl-ai/MetaPred/tree/master/baselines) includes:
* Logistic Regression
* K-Nearest Neighbors 
* XGBoost 
* SVM
* Random Forest
* MLP
* LSTM
* CNN

which can be used in any sort of sequence modeling, especially for EHRs data, directly.

### How to Run
To run MetaPred on EHR data, you need to revise the learning settings in main.py and the network hyperparameters in model.py. Then run the shell script metapred.sh.
```bash
bash metapred.sh
```
Our settings of learning procedures are:
```bash
python main.py --method='cnn' --metatrain_iterations=10000 --meta_batch_size=32 --update_batch_size=4 --meta_lr=0.001 --update_lr=1e-5 --num_updates=4 --n_total_batches=500000
```
or
```bash
python main.py --method='rnn' --metatrain_iterations=10000 --meta_batch_size=32 --update_batch_size=4 --meta_lr=0.001 --update_lr=1e-5 --num_updates=4 --n_total_batches=500000
```

### Additional Material
There is implementations used in:

Chelsea Finn, Pieter Abbeel, Sergey Levine, [Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks](https://arxiv.org/abs/1703.03400), International Conference on Machine Learning (ICML), 2017.

## References 
If you happen to use our work, please consider citing our paper: 
```
@inproceedings{Zhang:2019:MMC:3292500.3330779,
 author = {Zhang, Xi Sheryl and Tang, Fengyi and Dodge, Hiroko H. and Zhou, Jiayu and Wang, Fei},
 title = {MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records},
 booktitle = {Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
 series = {KDD '19},
 year = {2019},
 location = {Anchorage, AK, USA},
 pages = {2487--2495},
} 
```
This paper can be accessed on : [MetaPred] (https://dl.acm.org/citation.cfm?id=3330779)


	a/README.md		b/README.md
1	# MetaPred	1	# MetaPred
2	The repo is code (baseline and the proposed MetaPred) for paper MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records by [Xi Sheryl Zhang](https://www.xi-zhang.net), [Fengyi Tang](https://github.com/af1tang), [Hiroko H. Dodge](https://medicine.umich.edu/dept/neurology/hiroko-dodge-phd), [Jiayu Zhou](https://jiayuzhou.github.io), and [Fei Wang](https://sites.google.com/site/cornellwanglab/home).	2	The repo is code (baseline and the proposed MetaPred) for paper MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records by [Xi Sheryl Zhang](https://www.xi-zhang.net), [Fengyi Tang](https://github.com/af1tang), [Hiroko H. Dodge](https://medicine.umich.edu/dept/neurology/hiroko-dodge-phd), [Jiayu Zhou](https://jiayuzhou.github.io), and [Fei Wang](https://sites.google.com/site/cornellwanglab/home).
3		3
4	## Overview	4	## Overview
5		5
6	MetaPred is a meta-learning framework for Clinical Risk Prediction using limited patient Electronic Health Records (EHRs). We given an example in the following figure:	6	MetaPred is a meta-learning framework for Clinical Risk Prediction using limited patient Electronic Health Records (EHRs). We given an example in the following figure:
7		7
8	<p align="center"><img src="figures/task-design.png" alt=" Illustration of the proposed learning procedure" width="500"></p>	8	<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/task-design.png?raw=true" alt=" Illustration of the proposed learning procedure" width="500"></p>
9		9
10	Suppose we have multiple domains, our goal is to predict Alzheimer’s disease with few labeled patients, which give rise to a low-resource classification. The idea is to employ labeled patients from high-resource domains and design a learning to transfer framework with sources and a simulated target in meta-learning. There are four steps: (1) constructing episodes by sampling from the source domains and the simulated target domain; (2) learn the parameters of predictors in an episode-by-episode manner; (3) fine-tuning the model parameters on the genuine target domain; (4) predicting the target clinical risk. We respectively implemented Convolutional Neural Network (CNN) and Long-Shot Term Memory (LSTM) Network as base predictors. The model overview (meta-training procedure) is shown as follows:	10	Suppose we have multiple domains, our goal is to predict Alzheimer’s disease with few labeled patients, which give rise to a low-resource classification. The idea is to employ labeled patients from high-resource domains and design a learning to transfer framework with sources and a simulated target in meta-learning. There are four steps: (1) constructing episodes by sampling from the source domains and the simulated target domain; (2) learn the parameters of predictors in an episode-by-episode manner; (3) fine-tuning the model parameters on the genuine target domain; (4) predicting the target clinical risk. We respectively implemented Convolutional Neural Network (CNN) and Long-Shot Term Memory (LSTM) Network as base predictors. The model overview (meta-training procedure) is shown as follows:
11		11
12	<p align="center"><img src="figures/MetaPred.png" alt="MetaPred framework overview" width="750"></p>	12	<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/MetaPred.png?raw=true" alt="MetaPred framework overview" width="750"></p>
13		13
14	The entire learning procedure can be viewed as: iteratively transfer the parameter Θ learned from source domains through utilizing it as the initialization of the parameter that needs to be updated in the target domain.	14	The entire learning procedure can be viewed as: iteratively transfer the parameter Θ learned from source domains through utilizing it as the initialization of the parameter that needs to be updated in the target domain.
15		15
16	## Results	16	## Results
17		17
18	The learned representations of patients in five disease domains are visualized by t-SNE. In detail, AD, PD, DM, AM, MCI are abbreviations of Alzheimer's Disease, Parkinson's Disease, Dementia, Amnesia and Mild Cognitive Impairment, respectively. As a patient might suffer multiple diseases, there is supposed to be some overlaps among the given domains.	18	The learned representations of patients in five disease domains are visualized by t-SNE. In detail, AD, PD, DM, AM, MCI are abbreviations of Alzheimer's Disease, Parkinson's Disease, Dementia, Amnesia and Mild Cognitive Impairment, respectively. As a patient might suffer multiple diseases, there is supposed to be some overlaps among the given domains.
19		19
20	<p align="center"><img src="figures/patient_vis_metapred.png" alt="Visualization of patient representation learned by MetaPred" width="500"></p>	20	<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/patient_vis_metapred.png?raw=true" alt="Visualization of patient representation learned by MetaPred" width="500"></p>
21		21
22	To demonstrate the effectiveness of the proposed MetaPred in the context of domain adaptation, we compare it with the state-of-the-art meta-learning algorithm ``Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks`` (MAML). The results on Alzheimer's Disease domain are presented in terms of AUC and F1-Score.	22	To demonstrate the effectiveness of the proposed MetaPred in the context of domain adaptation, we compare it with the state-of-the-art meta-learning algorithm ``Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks`` (MAML). The results on Alzheimer's Disease domain are presented in terms of AUC and F1-Score.
23		23
24	<p align="center"><img src="figures/vs_maml_ad_cnn.png" alt="Performance comparison of MetaPred and MAML on the top of Alzheimer's Disease" width="500"></p>	24	<p align="center"><img src="https://github.com/sheryl-ai/MetaPred/blob/master/figures/vs_maml_ad_cnn.png?raw=true" alt="Performance comparison of MetaPred and MAML on the top of Alzheimer's Disease" width="500"></p>
25		25
26		26
27	## Requirements	27	## Requirements
28	This package has the following requirements:	28	This package has the following requirements:
29	* `Python 3.x`	29	* `Python 3.x`
30	* [TensorFlow 1.5](https://github.com/tensorflow/tensorflow)	30	* [TensorFlow 1.5](https://github.com/tensorflow/tensorflow)
31	* [Progress Bar](https://progressbar-2.readthedocs.io/en/latest/index.html)	31	* [Progress Bar](https://progressbar-2.readthedocs.io/en/latest/index.html)
32		32
33		33
34	## Usage	34	## Usage
35	### Baseline in Sequential Data Modeling	35	### Baseline in Sequential Data Modeling
36	The [baseline implementation](https://github.com/sheryl-ai/MetaPred/tree/master/baselines) includes:	36	The [baseline implementation](https://github.com/sheryl-ai/MetaPred/tree/master/baselines) includes:
37	* Logistic Regression	37	* Logistic Regression
38	* K-Nearest Neighbors	38	* K-Nearest Neighbors
39	* XGBoost	39	* XGBoost
40	* SVM	40	* SVM
41	* Random Forest	41	* Random Forest
42	* MLP	42	* MLP
43	* LSTM	43	* LSTM
44	* CNN	44	* CNN
45		45
46	which can be used in any sort of sequence modeling, especially for EHRs data, directly.	46	which can be used in any sort of sequence modeling, especially for EHRs data, directly.
47		47
48	### How to Run	48	### How to Run
49	To run MetaPred on EHR data, you need to revise the learning settings in main.py and the network hyperparameters in model.py. Then run the shell script metapred.sh.	49	To run MetaPred on EHR data, you need to revise the learning settings in main.py and the network hyperparameters in model.py. Then run the shell script metapred.sh.
50	```bash	50	```bash
51	bash metapred.sh	51	bash metapred.sh
52	```	52	```
53	Our settings of learning procedures are:	53	Our settings of learning procedures are:
54	```bash	54	```bash
55	python main.py --method='cnn' --metatrain_iterations=10000 --meta_batch_size=32 --update_batch_size=4 --meta_lr=0.001 --update_lr=1e-5 --num_updates=4 --n_total_batches=500000	55	python main.py --method='cnn' --metatrain_iterations=10000 --meta_batch_size=32 --update_batch_size=4 --meta_lr=0.001 --update_lr=1e-5 --num_updates=4 --n_total_batches=500000
56	```	56	```
57	or	57	or
58	```bash	58	```bash
59	python main.py --method='rnn' --metatrain_iterations=10000 --meta_batch_size=32 --update_batch_size=4 --meta_lr=0.001 --update_lr=1e-5 --num_updates=4 --n_total_batches=500000	59	python main.py --method='rnn' --metatrain_iterations=10000 --meta_batch_size=32 --update_batch_size=4 --meta_lr=0.001 --update_lr=1e-5 --num_updates=4 --n_total_batches=500000
60	```	60	```
61		61
62	### Additional Material	62	### Additional Material
63	There is implementations used in:	63	There is implementations used in:
64		64
65	Chelsea Finn, Pieter Abbeel, Sergey Levine, [Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks](https://arxiv.org/abs/1703.03400), International Conference on Machine Learning (ICML), 2017.	65	Chelsea Finn, Pieter Abbeel, Sergey Levine, [Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks](https://arxiv.org/abs/1703.03400), International Conference on Machine Learning (ICML), 2017.
66		66
67	## References	67	## References
68	If you happen to use our work, please consider citing our paper:	68	If you happen to use our work, please consider citing our paper:
69	```	69	```
70	@inproceedings{Zhang:2019:MMC:3292500.3330779,	70	@inproceedings{Zhang:2019:MMC:3292500.3330779,
71	author = {Zhang, Xi Sheryl and Tang, Fengyi and Dodge, Hiroko H. and Zhou, Jiayu and Wang, Fei},	71	author = {Zhang, Xi Sheryl and Tang, Fengyi and Dodge, Hiroko H. and Zhou, Jiayu and Wang, Fei},
72	title = {MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records},	72	title = {MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records},
73	booktitle = {Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},	73	booktitle = {Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
74	series = {KDD '19},	74	series = {KDD '19},
75	year = {2019},	75	year = {2019},
76	location = {Anchorage, AK, USA},	76	location = {Anchorage, AK, USA},
77	pages = {2487--2495},	77	pages = {2487--2495},
78	}	78	}
79	```	79	```
80	This paper can be accessed on : [MetaPred] (https://dl.acm.org/citation.cfm?id=3330779)	80	This paper can be accessed on : [MetaPred] (https://dl.acm.org/citation.cfm?id=3330779)
81		81