Heart failure risk prediction based on machine learning and interpretability analysis.

Area of Science:

• Machine Learning in Cardiovascular Disease
• Clinical Risk Stratification
• Explainable Artificial Intelligence (XAI)

Background:

• Heart failure (HF) poses a global health challenge, necessitating accurate early risk stratification.
• Existing machine learning (ML) models for HF prediction lack standardized benchmarking and robust explainability.
• Class imbalance issues and reliance on single explainable AI (XAI) methods limit current ML applications in HF.

Purpose of the Study:

• To conduct a comprehensive benchmark of 10 ML algorithms for HF risk stratification.
• To establish and validate a dual-XAI framework (SHAP + LIME) for interpreting ML models in HF prediction.
• To address limitations in current ML applications, including algorithm comparison and explainability.

Main Methods:

• Analyzed a public dataset of 2,169 HF patients with 15 clinical features.
• Compared 10 ML algorithms (5 traditional, 5 ensemble) under standardized conditions with stratified sampling.
• Evaluated performance using ROC-AUC, accuracy, precision, recall, F1-score, and PR-AUC; applied SHAP and LIME for model interpretation.

Main Results:

• Logistic Regression demonstrated optimal performance (ROC-AUC = 0.9451, accuracy = 88.25%), surpassing ensemble methods.
• Left Ventricular Ejection Fraction (LVEF) was the dominant predictor, confirmed by SHAP and LIME analysis.
• The dual-XAI framework showed 100% concordance in top-3 predictor rankings, highlighting LVEF, diabetes, and age.

Conclusions:

• A systematic ML algorithm benchmark identified Logistic Regression as a top performer for HF risk stratification.
• A validated dual-XAI framework provides reliable interpretation, addressing single-method biases.
• Simple, interpretable models are effective for HF prediction in moderate-sized datasets, offering practical clinical guidance.