This repository is the official implementation of Clinical Risk Prediction with Temporal Probabilistic Asymmetric Multi-Task Learning (AAAI 2021). https://arxiv.org/abs/2006.12777
To install requirements:
pip install -r requirements.txt
To train the models in the paper, run this command:
python run.py
Also for P-AMTL experiment on MNIST-variant dataset, you can run the same command inside MNIST directory. Please see MNIST Toy experiment section below for details.
Please look up and modify config.py for selecting base model and datasets. For example, if you want to run tp_amtl model on physionet2012 dataset, you can modify config.py as follows:
config.mtl_model = "tp_amtl"
config.task_code = "physionet2012"
Please put model name on config.mtl_model and dataset on config.task_code (both in lower-case). For tp_amtl (epistemic) and tp_amtl (aleatoric), config.mtl_model input can be written as below:
config.mtl_model = "tp_amtl_1" #tp_amtl (epistemic)
config.mtl_model = "tp_amtl_2" #tp_amtl (aleatoric)
For single-task version of UA,LSTM and RETAIN model, you can simply select config.tasks = [i] where i is the task you want to run (0,1,2,etc.). Also, for RETAIN-Kendall model, you need to change config.tasks [i for i in range(len(tasks))] where i is task id (0,1,2, etc.). Below is example for physionet2012 (also can be applied to MIMIC-III Heart Failure).
config.tasks = [0,1,2,3]
Requirement:
python3.6.9
torch
tqdm
matplotlib
sklearn
Training: Please download mnist_rotation_back_image_new.zip (https://sites.google.com/a/lisa.iro.umontreal.ca/public_static_twiki/variations-on-the-mnist-digits) into a folder named mnist_rotation_back inside the MNIST directory and unzip. Then run this command:
python run.py
We experiment on four datasets that we compile for clinical risk prediction from two open-source electronic health records (EHR) datasets.
MIMIC-III Infection: code for train/test/valid separation
Physionet2012: code for train/test/valid separation
MIMIC-III Heart Failure: code for train/test/valid separation
MIMIC-III Respiratory Failure: code for train/test/valid separation
For MIMIC-III Respiratory Failure dataset, you can set as below on task_codes/mimic_respiratory_failure.py to experiment on dataset with full 37,818 instances.
path = 'mimic_respiratory_failure/30000/'
Our model achieves the following performance on four datasets:
