Machine Learning for Healthcare Research Reproducibility Challenge

OMSCS CSE6250 Spring 2023

Contributors

Jessica Buzzelli, Xi Lu

Credit

Code sourced from the Survival-MDN and SODEN repos.

More information on the models and how to use the subrepositories can be found in the corresponding READMEs: Survival-MDN, SODEN.

Usage

Experiments from this course project were run from the Survival MDN repository with minor edits to accomodate unclear package versions.

To reproduce our experiments (assumes Windows machine with a dedicated GPU):

1. pip install -r requirements.txt (primary dependencies: numpy, pandas, lifelines, torch)
2. cd Survival-MDN
3. python generate_model_runner.py to populate run_models.ps1
4. Run run_models.ps1 to generate model config and run 100 experiments on each of the 3 datasets' 10 folds

Per generate_model_runner.py, model training and evaluation output are saved to Survival-MDN/output/{dataset}/{split}.

To generate tables featured in the final report, run helpers/parse_outputs.ipynb and helpers/label_distribution.ipynb; outputs will be saved to the helpers/ directory as appendix_A.txt, results.txt (model evaluation statistics), and train_durations.txt (time-to-train statistics by dataset).