Experiments for the paper "Towards symbolic regression for interpretable clinical decision scores"
Abstract: Symbolic Regression (SR), a form of supervised learning, attempts to solve the NP-hard problem of searching both function form and its free parameters. When applied to healthcare, it should have a clear decision process, often found in decision trees. We introduce Brush, a multi-objective SR algorithm with flexible split-wise operations compatible with non-linear optimization, designed for both regression and classification problems. Brush achieves Pareto-optimal performance on SRBench, and was applied to design clinical decision support using the CART and MEWS scoring systems, achieving high accuracy and interpretable models. Compared to decision trees, random forests, and other SR methods, Brush achieves comparable or superior predictive performance while producing simpler models.
MIMIC is a public dataset, but to access it you must first do the CITI data training and upload your conclusion certificate to them, and also accept a use agreement term.
The steps are described on physionet website: https://physionet.org/content/mimic-iv-ed/2.2/.
We recomend to create a clean python virtual environment with python=3.8, the latest one compatible with PSTree.
Then, install the requirements with pip install -r requirements.txt.
Medcodes was taken from https://github.com/topspinj/medcodes/tree/master/medcodes and it is not available through PyPi. YOu can just clone it inside data_prep and it should be visible to medcode_utils.py.
If PSTree or Brush fails to install, you can go to their GitHub pages and follow the detailed specific instructions.
Your experiments should be ready to go.
The script submit_jobs.py is encharged of searching for previously finished experiments, as well as submit subprocesses (or SLURM jobs) for each individual experiment.
We provide a file named run.sh with the experiment configurations we used for the paper experiments.