AI Clinician: reinforcement learning in intensive care

Code for a reinforcement learning model applied to the management of intravenous fluids and vasopressors in patients with sepsis in intensive care.

Author: Dr Matthieu Komorowski, Imperial College London, 2015-2019 - [email protected]

The 2 datasets used in the research are:

MIMIC-III : https://mimic.physionet.org/
eICU-RI: not publicly available in full, subset available here: http://eicu-crd.mit.edu/

Cohort definition: all adult patients fulfilling the sepsis-3 definition: http://jamanetwork.com/journals/jama/fullarticle/2492881

This repository contains:

I. the Jupyter notebook to perform data extraction in MIMIC-III (AIClinician_Data_extract_MIMIC3_140219.ipynb)

II. the Matlab code to identify the cohort of patients with sepsis in MIMIC-III (AIClinician_sepsis3_def_160219.m)

III. the Matlab code to re-create the MIMIC-III dataset (AIClinician_MIMIC3_dataset_160219.m)

IV. the Matlab code (AIClinician_core_160219.m) to:

build 500 different discrete state and action MDP models from the MIMIC-III training dataset;
select the best policy from off-policy evaluation on the MIMIC-III validation set;
test this optimal policy on the eICU-RI dataset;
compute the main results and key figures.

V. Additional files:

The unique identifiers for these patients in both datasets are provided (patientIDs_MIMIC3.csv and patientIDs_eRI.csv). Note: you'll need to add 200,000 to all the patient identifiers in patientIDs_MIMIC3 to match the numbering found in the initiail database.
A detailed desciption of the datasets (Dataset description Komorowski 111118.xlsx).
deloutabove.m and deloutbelow.m : custom functions to delete outliers.
fixgaps.m : custom function for linear interpolation.
reference_matrices.mat contains the mapping of item identifiers and concepts, and data for sample-and-hold technique.
SAH.m : custom function to perform sample-and-hold.

External files and toolboxes used:

2lambda123 / ai_clinician Goto Github PK