Recursive Feature Elimination with Cross-Validation (RFECV) for Optimal Feature Subset Selection in Cardiovascular Disease Prediction

Authors

  • RAPHAEL OZIGHOR ENIHE Baze University, Abuja

Keywords:

Cardiovascular Disease, Feature Selection, Recursive Feature Elimination , SMOTE-Tomek, Extra Trees, Machine Learning

Abstract

High-dimensional clinical datasets frequently contain redundant or weakly informative features that add noise, increase computational cost, and can obscure the clinically relevant predictors of cardiovascular disease (CVD). This study applies Recursive Feature Elimination with Cross-Validation (RFECV), using an Extra Trees Regressor as the base estimator and time-series-split cross-validation, to a SMOTE-TOMEK-balanced cardiovascular disease dataset of 1,000 patient records (1,696 after balancing) collected at the University of Abuja Teaching Hospital (UATH), Nigeria. RFECV reduced the original fourteen predictors to nine optimal features which are: Age, Smoke, Diabetes, High Cholesterol, BMI, Family History of CVD, Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), and Proteinuria, ranking Body Mass Index (importance = 0.409) and Systolic Blood Pressure (importance = 0.139) as the two most influential predictors. Ten machine learning classifiers were evaluated before and after feature selection using precision, recall, F1-score, accuracy, ROC-AUC and confusion-matrix analysis. Feature selection improved or maintained performance for most models,  most notably LightGBM (accuracy from 0.93 to 0.95) and Extra Trees (accuracy from 0.92 to 0.95); while reducing false negatives substantially for LightGBM (from 18 to 2) and modestly for Random Forest and Extra Trees. The results demonstrate that RFECV can simultaneously reduce clinical data-collection burden, improve model interpretability, and preserve or enhance predictive performance in CVD risk modelling.

DOI: https://doi.org/10.5281/zenodo.23410758 

Downloads

Published

2026-10-16