Research on the Influencing Factors of Heart Disease based on Logistic Regression and XGBoost

Abstract

This study leverages the UCI Cleveland Heart Disease dataset (297 complete records, 14 features) to present a two-stage pipeline that couples rigorous feature engineering with hybrid modeling and compares Logistic Regression (LR) against Gradient Boosting Decision Trees (GBDT/XGBoost) for cardiac risk prediction. Mutual information and recursive feature elimination first isolate the most informative variables-chest-pain type, maximum heart rate, exercise-induced ST depression, ST slope, and number of major vessels-whose clinical relevance is well established. GBDT then markedly outperforms LR across all evaluated metrics: accuracy 89.2% vs 83.7%, recall 86.7% vs 80.0%, and AUC 0.925 vs 0.874, demonstrating the value of capturing non-linear interactions among risk factors. Feature-importance analysis corroborates these predictors’ medical interpretability. The authors acknowledge limitations arising from the modest sample size and the ongoing need for transparent models; they recommend expanding the data, integrating SHAP explanations, and incorporating real-time monitoring. Overall, explainable GBDT-based tools can support clinicians in early identification of high-risk individuals, enabling personalized interventions and improved patient outcomes.