A flexible symbolic regression method for constructing interpretable clinical prediction models | JoVE Visualize

Area of Science:

* Artificial Intelligence in Healthcare
* Clinical Informatics
* Machine Learning for Medical Applications

Background:

* Machine learning (ML) models for clinical decision support (CDS) often lack interpretability, hindering clinical adoption and patient safety.
* Scaling ML-driven CDS requires models that are both accurate and intuitively understandable by clinicians.
* Electronic health records (EHR) contain high-dimensional data suitable for complex predictive modeling.

Purpose of the Study:

* To adapt a symbolic regression method, the Feature Engineering Automation Tool (FEAT), for training interpretable ML models from EHR data.
* To evaluate FEAT's performance in classifying specific hypertension phenotypes and its generalizability across diverse clinical tasks.
* To demonstrate that FEAT can generate accurate and clinically intuitive predictive models for CDS applications.

Main Methods:

* Applied the Feature Engineering Automation Tool (FEAT), a symbolic regression approach, to high-dimensional EHR data.
* Trained and evaluated FEAT models for classifying hypertension, hypertension with unexplained hypokalemia, and apparent treatment-resistant hypertension (aTRH) in 1200 subjects.
* Assessed FEAT's generalizability on 25 benchmark clinical phenotyping tasks using the MIMIC-III critical care database, comparing against penalized linear models.

Main Results:

* FEAT models achieved equivalent or superior discriminative performance compared to other interpretable models for hypertension phenotypes (p < 0.001) and were significantly smaller (p < 1x10^-6).
* A six-feature FEAT model for aTRH demonstrated high discriminative ability (PPV=0.70, sensitivity=0.62) and clinical intuitiveness.
* Across 25 MIMIC-III tasks, FEAT models outperformed penalized linear models in AUC under comparable dimensionality constraints (p < 6x10^-6).

Conclusions:

* FEAT successfully trains machine learning models from EHR data that are both highly accurate and intuitively interpretable for clinicians.
* The developed models facilitate the safe and effective expansion of ML-triggered clinical decision support across various healthcare settings.
* FEAT offers a promising approach for bridging the gap between complex ML algorithms and practical clinical implementation.