An Explainable Approach for Heart Disease Prediction Using an Evolutionary Rule-Based Learning Classifier System on Real-World ICU Data

Abstract

Cardiovascular disease is the leading cause of death globally, yet few machine learning models for cardiac risk prediction are clinically interpretable. This paper develops an Explainable Artificial Intelligence (XAI) approach for heart disease risk classification on the MIMIC-IV-Ext Cardiac Disease dataset (v1.0.0, PhysioNet), comprising 4,748 de-identified ICU admissions. A multimodal set of 35 features including 14 laboratory biomarkers and NLP-derived clinical narrative indicators is constructed, with binary risk labels generated by keyword-majority voting over four clinical free-text fields. Class imbalance (78.1% High Risk vs. 21.9% Low Risk) is addressed by applying SMOTE only to the training set. The proposed classifier, ExSTraCS, a Michigan-style Learning Classifier System with kappa-based evolutionary fitness, is evaluated against five baselines, achieving 84.53% held-out test accuracy (84.29% ± 0.60% cross-validated), 90.19% F1-score, and 0.8757 AUROC (with 89.30% precision, 91.11% recall and 61.06% specificity), close to the strongest baseline; McNemar and DeLong tests assess statistical significance. ExSTraCS is a hybrid approach in which a gradient-boosting engine performs prediction, while interpretability is delivered through an evolved IF-THEN rule layer and post hoc SHAP and LIME explanations, whose faithfulness is measured using feature-perturbation metrics. Because several text-derived features share keywords with the labelling rule, a leakage-controlled experiment using only laboratory biomarkers bounds the impact of feature–label correlation. All three explanation layers agree that Troponin T, ECG abnormality, and HPI-derived symptom features are the most reliable predictors, and a cost-sensitive threshold analysis identifies the optimal operating point given the high clinical cost of false negatives. The results show that models offering both competitive predictive accuracy and clinical interpretability are achievable.