A Cross-Validated LBX Ensemble Framework for Reliable and Clinically Interpretable Heart Disease Prediction

This study presents a novel and clinically interpretable machine learning framework for early prediction of heart disease, addressing key limitations of previous models through enhanced data preprocessing, effective feature selection, and a hybrid ensemble approach. The methodology includes rigorous data cleaning, outlier treatment using the interquartile range (IQR), and class balancing via Synthetic Minority Over-sampling Technique (SMOTE) to ensure data quality and equal representation. A multi-stage feature selection strategy, CCRF—combining correlation analysis, chi-square testing, and recursive feature elimination—reduced the feature set to eight, improving computational efficiency by 27% without compromising predictive power. At the core of the framework is the LBX Ensemble, a hybrid model that merges the interpretability of Logistic Regression with the non-linear learning strength of XGBoost. This approach achieved outstanding results: 94.2% accuracy, 93.5% precision, 95.6% recall, 94.5 F1-score, and an AUC of 0.965. The model’s robustness was validated through stratified 10-fold cross-validation with a low standard deviation of 1.4%. These findings highlight the LBX Ensemble’s potential as a reliable, accurate, and explainable tool for clinical decision support in cardiovascular risk assessment.