This case study examines the fairness and generalizability of a MIMIC-trained model published in the peer-reviewed article on Multitask learning and benchmarking with clinical time series data by Harutyunyan et al., with the source code available here. The case study is based on the following fairness and generalizability assessment framework:
The source code to create an in-hospital mortality benchmark cohort and train a channel-wise LSTM model on it has been updated, extended to a new data source (i.e. STARR) and augmented by a comprehensive analysis framework to examine generalizability and fairness of the risk prediction model. All code necessary to reconstruct the same analyses can be found in this repository.
This repository contains two main directories:
benchmarks: Pipeline to construct in-hospital mortality (IHM) cohorts- mimic: Using MIMIC-III v1.4 relational database
- starr: Using STARR_DE relational database
- Shared scripts
models: Training, testing and evaluation protocols
Detailed usage instructions for how to create benchmark cohorts, use the model architectures and analyze the output can be found in the README's of the corresponding folders and subfolders.