Final Project, CS598 DLH Deep Learning for Healthcare, UIUC
This repository is a Pytorch implementation of HiTANet: Hierarchical Time-Aware Attention Networks for Risk Prediction on Electronic Health Records. Part of the code is adapted from the official implementation.
To train and test the HiTANet model, you just need to download and run the CS598DL4H_FP_HiTANet.ipynb file with coresponding datasets (COPD, HF, or HF-sample).
To extract the dataset from MIMIC-III, you can download and run the CS598_extract_data.ipynb file with the DIAGNOSES_ICD and ADMISSIONS tables in MIMIC-III.
Our model achieves the following performance on the sample of the heart failure dataset:
| Model name | Accuracy | Precision | Recall | F1-score | AUC |
|---|---|---|---|---|---|
| Our Implementation on Sample Dataset | 0.772 | 0.823 | 0.897 | 0.857 | 0.630 |
| Reported Result on Full Dataset | 0.823 | 0.724 | 0.587 | 0.647 | 0.564 |
More experiements on HiTANET-C1 and HiTANet-C2 is available here.