Explainable machine learning predictions to support personalized cardiology strategies.

Area of Science:

• Cardiology
• Gerontology
• Machine Learning
• Computational Biology

Background:

• Physical activity is crucial for cardiac health in older adults, but its effects can vary significantly among individuals.
• Current machine learning (ML) models for predictive health lack patient-specific rationalization, limiting personalized interventions.
• Understanding the precise impact of physical activity on cardiac structure is essential for tailored health strategies.

Purpose of the Study:

• To apply SHapley Additive exPlanations (SHAP), a novel explainable ML methodology, to understand the heterogeneous effects of physical activity on cardiac structure in older adults.
• To demonstrate the capability of SHAP to provide patient-level insights into how physical activity features influence left ventricular (LV) structure.
• To explore the integration of explainable ML into personalized cardiology.

Main Methods:

• A dataset of 86 older adults (mean age 72 ± 4 years) was analyzed, correlating physical activity levels with changes in left ventricular (LV) structure.
• SHapley Additive exPlanations (SHAP) was employed to visualize the magnitude and direction of impact of physical activity features on LV structure for individual patients.
• A Random Forest Regressor model was trained and validated using k-cross-validation, achieving optimal performance metrics (low mean squared error, mean absolute error) on the test set.

Main Results:

• SHAP analysis provided intelligible visualizations, illustrating the specific impact of physical activity features on individual patients' LV structure.
• The Random Forest Regressor model demonstrated robust predictive performance, validated on an independent test set.
• Force plots indicated the magnitude of influence of different physical activity parameters (red for higher impact, blue for lower impact) on predicted LV structure.

Conclusions:

• Explainable ML, specifically SHAP, can identify patient-specific physical activity features that predict cardiac structure.
• These findings support the incorporation of explainable ML into personalized cardiology, enabling tailored physical activity interventions.
• The study highlights a promising approach for enhancing the precision and individualization of cardiac health management strategies in older adults.