RL-RRT-ICU-AKI

the reinforcement learning in RRT operation optimation of ICU-AKI patients (in MIMIC and eICU)

1. Abstract

2. Pipeline

1. Patient Trajectories Select and Construction.
2. MDP Process Recognition,
3. RL model (DQN) Training
4. Validation and Utilities.

2.1 how to use

1. Do data preparation work.
2. Set the parameter in parameter.py
3. Run the start.py

3. Code Introduction

3.1 Data Preparation

3.1.1 SQL from database

v6_reward_function.sql get the aki_patients and RRT treatment(action), eGFR(reward).

20240421-Factor4discontinuation.sql get the chart record for patients states related to AKI treatment. (supported by Minqi Xiong)

3.1.2 Python process code

dataset_con.ipynb: original data generated by strategy step (specific time period) and Missing value completion dataset_con.ipynb: different trajectory data combine. dataset_desc.ipynb: data description and stat.

3.2 Data process

trainmodel.StateConstructor

utils.Dataloader

3.3 Model Training

trainmodel.StateConstructor

trainmodel.RL

exp.DQNexperiment

3.4 Validation and Utility

test.TestClass

Reference

https://github.com/microsoft/med-deadend/tree/main

https://github.com/MIT-LCP/mimic-code/tree/v2.1.0/mimic-iv/concepts