Explainable machine learning for stroke risk prediction: a comparative study with SHAP-based interpretation.

Area of Science:

• Biomedical Informatics
• Computational Medicine
• Public Health

Background:

• Stroke is a leading global cause of death and disability.
• Early screening and risk prediction are critical for stroke management.
• Traditional methods face challenges with complex data relationships and interpretability.

Purpose of the Study:

• To evaluate and compare the performance of various machine learning models for stroke risk prediction.
• To identify key predictors of stroke using interpretability techniques.
• To assess the clinical utility and resource implications of different predictive models.

Main Methods:

• Developed and compared Logistic Regression, Random Forest, XGBoost, CatBoost, Multi-layer Perceptron (MLP) neural network, and ensemble models.
• Utilized SHapley Additive exPlanations (SHAP) for feature contribution analysis.
• Employed confusion matrices and Precision-Recall curves for performance evaluation and compared training times.

Main Results:

• Ensemble and neural network models showed superior predictive performance over traditional algorithms.
• The MLP model demonstrated high recall for identifying stroke patients.
• Hypertension, average blood glucose level, and age were identified as significant risk factors via SHAP analysis.

Conclusions:

• Machine learning offers significant advantages for accurate stroke risk prediction.
• Integrating model interpretability enhances clinical trust and utility.
• Findings provide a methodological reference for stroke risk stratification and early warning systems.