Enhanced heart failure mortality prediction through model-independent hybrid feature selection and explainable...

Area of Science:

• Cardiology and Artificial Intelligence
• Biomedical Informatics and Machine Learning

Background:

• Heart failure (HF) presents a significant global health burden with high mortality rates.
• Accurate prediction of HF mortality is crucial for effective clinical management and improved patient outcomes.
• Existing machine learning (ML) models for HF mortality prediction are limited by feature selection dependency and lack of generalizability.

Purpose of the Study:

• To introduce and validate a novel, model-independent hybrid feature selection methodology for enhanced 1-year all-cause mortality prediction in HF patients.
• To identify a robust, compact, and interpretable subset of features for HF mortality prediction using echocardiographic and demographic data.

Main Methods:

• A hybrid feature selection approach combining Extremely Randomized Trees (Extra-Trees) and non-linear correlation measures was developed.
• Optimal feature subset identification was achieved through loss graph inspection, resulting in seven key predictive features.
• Seven ML models were trained and evaluated on both full and selected feature sets, with performance assessed using SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

• The proposed hybrid feature selection method identified a highly informative subset of seven features, reducing feature count by 80% without compromising predictive performance.
• Most ML models maintained or improved their 1-year mortality prediction accuracy with the reduced feature set compared to using the full dataset.
• The selected feature subset demonstrated superior predictive accuracy for HF mortality compared to conventional feature selection techniques across all evaluated ML models.

Conclusions:

• The novel hybrid feature selection methodology offers a robust, generalizable, and explainable approach for predicting HF mortality.
• The identified compact feature subset facilitates personalized management strategies for HF patients.
• This study highlights the potential of advanced feature selection techniques to improve ML-driven clinical decision support in cardiology.