|
--- |
|
license: unlicense |
|
task_categories: |
|
- reinforcement-learning |
|
language: |
|
- en |
|
tags: |
|
- medical |
|
pretty_name: epicare |
|
size_categories: |
|
- 100K<n<1M |
|
--- |
|
|
|
# EpiCare Dataset README |
|
|
|
## Overview |
|
|
|
This repository contains a dataset of 64 `.hdf5` files following the D4RL format standard, designed for use in testing offline reinforcement learning (RL) and off-policy evaluation (OPE) methods. The dataset is split into train and test files for two behavior policies (SMART and SoC) on the EpiCare environment. |
|
|
|
The EpiCare environment, described in detail in our [project repository](https://github.com/masonhargrave/EpiCare), is a benchmark designed to mimic the challenges associated with applying RL to longitudinal healthcare settings. This dataset is useful for evaluating the performance of RL models in these complex scenarios. |
|
|
|
## Directory Structure |
|
|
|
The dataset is organized into the following directory structure: |
|
|
|
``` |
|
smart/ |
|
├── test_seed_1.hdf5 |
|
├── test_seed_2.hdf5 |
|
├── test_seed_3.hdf5 |
|
├── test_seed_4.hdf5 |
|
├── test_seed_5.hdf5 |
|
├── test_seed_6.hdf5 |
|
├── test_seed_7.hdf5 |
|
├── test_seed_8.hdf5 |
|
├── train_seed_1.hdf5 |
|
├── train_seed_2.hdf5 |
|
├── train_seed_3.hdf5 |
|
├── train_seed_4.hdf5 |
|
├── train_seed_5.hdf5 |
|
├── train_seed_6.hdf5 |
|
├── train_seed_7.hdf5 |
|
├── train_seed_8.hdf5 |
|
soc/ |
|
├── test_seed_1.hdf5 |
|
├── test_seed_2.hdf5 |
|
├── test_seed_3.hdf5 |
|
├── test_seed_4.hdf5 |
|
├── test_seed_5.hdf5 |
|
├── test_seed_6.hdf5 |
|
├── test_seed_7.hdf5 |
|
├── test_seed_8.hdf5 |
|
├── train_seed_1.hdf5 |
|
├── train_seed_2.hdf5 |
|
├── train_seed_3.hdf5 |
|
├── train_seed_4.hdf5 |
|
├── train_seed_5.hdf5 |
|
├── train_seed_6.hdf5 |
|
├── train_seed_7.hdf5 |
|
├── train_seed_8.hdf5 |
|
``` |
|
|
|
## Behavior Policies |
|
|
|
### SMART Policy |
|
|
|
The Sequential Multiple Assignment Randomized Trial (SMART) policy models treatment selection for a simulated clinical trial. This policy adheres to a weighted random selection process where each treatment's likelihood of selection is based on its expected reward. The SMART policy is widely used in clinical trials to balance exploration and exploitation, providing synthetic clinical trial data for training RL models. |
|
|
|
### SoC Policy |
|
|
|
The Standard of Care (SoC) policy aims to simulate the performance of a hypothetical clinician following best practices without access to latent disease states. This policy models a conservative approach to treatment selection, avoiding actions that would exacerbate any current symptoms beyond a safe threshold. The SoC policy serves as a baseline to compare the performance and safety of RL algorithms. |
|
|
|
## Seeds and Environment Generation |
|
|
|
The `seed` number in each file name refers to the random seed used to generate the EpiCare environment. Each seed creates a distinct EpiCare environment, which can be thought of as representing a completely different disease population. This variability allows researchers to evaluate the generalizability of their models across diverse simulated patient cohorts. |
|
|
|
## Dataset Details |
|
|
|
Each `.hdf5` file contains 131,072 episodes of data, effectively representing that many patients. Each episode consists of a maximum of 8 time steps. |
|
|
|
## Usage |
|
|
|
This dataset can be used to test and benchmark offline RL and OPE methods. The `.hdf5` files are compatible with the D4RL format, making them easy to integrate with existing RL frameworks and libraries. Researchers can use this dataset to train and evaluate their models, ensuring reproducibility and comparability of results. |
|
|
|
## Reference |
|
|
|
For more details on the EpiCare environment and the design considerations behind this dataset, please refer to the accompanying paper and visit our [project repository](https://github.com/masonhargrave/EpiCare). |
|
|
|
## Contact |
|
|
|
For questions or support, please contact: |
|
|
|
- Mason Hargrave ([email protected]) |
|
- Alex Spaeth ([email protected]) |
|
- Logan Grosenick ([email protected]) |
