Continual Learning of Longitudinal Health Records

Repo for reproducing the experiments in Continual Learning of Longitudinal Health Records (2021). Release v0.1 of the project corresponds to published results.

Experiments evaluate various continual learning strategies on standard ICU predictive tasks exhibiting covariate shift. Task outcomes are binary, and input data are multi-modal time-series from patient ICU admissions.

Setup

1. Clone this repo to your local machine.

2. Request access to MIMIC-III and eICU-CRD.¹

3. Download the preprocessed datasets to the /data subfolder.

4. Create and activate a virtual environment:

   conda env create -f environment.yml
   conda activate env-continual

Results

To reproduce main results:

python3 main.py --train

Figures will be saved to /results/figs. Instructions to reproduce supplementary experiments can be found here. Bespoke experiments can be specified with appropriate flags e.g:

python3 main.py --domain_shift hospital --outcome mortality_48h --models CNN --strategies EWC Replay --validate --train

A complete list of available options can be found here or with python3 main.py --help.

Citation

If you use any of this code in your work, please reference us:

@misc{armstrong2021continual,
      title={Continual learning of longitudinal health records}, 
      author={J. Armstrong and D. Clifton},
      year={2021},
      eprint={2112.11944},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

---

Stack

For standardisation of ICU predictive task definitions, feature pre-processing, and Continual Learning method implementations, we use the following tools:

Tool	Source
ICU Data	MIMIC-III eICU-CRD
Data preprocessing / task definition	FIDDLE
Continual Learning strategies	Avalanche

Note

Temporal Domain Incremental learning experiments require linkage with original MIMIC-III dataset. Requires downloading ADMISSIONS.csv from MIMIC-III to the /data/mimic3/ folder.