This is the source code for VLDB 2024 submission.
Thanks for your interest in our work.
- Install python 3.6, Pytorch 1.10.2.
- If you want to use the GPU, please install the CUDA accordingly.
As for the MIMIC3 dataset, you must submit the application for data access from https://mimic.physionet.org/. After downloading the CSVs, you first need to build the benchmark dataset according to the mimic3-benchmarks, and filter features according to the MIMIC-Extract.
Before training the CohortNet, please update the configuration first.
- update the root path for dataset.py
- update the log file path for main.py and train.ipynb.
You first need to train the CohortNet in the following three steps:
- Process a dataset for training the CohortNet.
- Train the Multi-Channel Feature Learning Module for fine-grained patient representation learning with the jupyter file train.ipynb.
- Train the Cohort Discovery Module, Cohort Learning Module, and Cohort Exploitation Module for the automatic cohort discovery, comprehensive cohort learning, and personalized cohort exploitation with the jupyter file train_cohort.ipynb. (Please update the model path before the training)
If you want to use your own dataset, please design the SetLoader, and open the mode for the forward imputation (i.e. impute the missing data with the last observation) and standardization before training the CohortNet.