Risk-Aware Explainable Artificial Intelligence for Early Stroke Screening and Clinical Prioritization

Stroke screening requires models that identify high-risk individuals while remaining calibrated, interpretable, and sensitive to the consequences of missed cases. This study evaluated a risk-aware explainable artificial intelligence framework for early stroke screening and clinical prioritization using the Kaggle Stroke Prediction Competition dataset. The analysis included 4,088 labeled observations with a binary stroke outcome and demographic, cardiovascular, metabolic, and behavioral predictors. Class-weighted logistic regression, random forest, and gradient boosting were developed under a stratified 70% development, 15% calibration, and 15% locked-test design, with probability calibration and F2-based threshold selection. On the locked test set, gradient boosting achieved the highest ROC-AUC and PR-AUC and the lowest Brier score among the evaluated classifiers, with ROC-AUC 0.894, PR-AUC 0.365, and Brier score 0.0400. Permutation importance identified age, body mass index, and average glucose level as the leading predictors. Risk stratification placed 80.0% of test stroke cases in the high-priority group. Adult-only robustness analysis yielded lower ROC-AUC 0.825 and PR-AUC 0.206, indicating sensitivity of performance to population definition within this evaluated study cohort. These findings demonstrate that calibrated, class-sensitive, and explainable modeling can support transparent stroke screening and prioritization while complementing, rather than replacing, clinical judgment.

Hali tarjima qilinmagan