A Hybrid Kernel SHAP and Enhanced ALE Framework for Explainable Stroke Risk Assessment

Stroke remains a leading cause of death and disability worldwide, yet machine learning (ML) models capable of predicting risk are rarely trusted clinically due to their "black-box" nature.This study introduces an explainability-centered hybrid framework, ALE-SHAP, evaluated on 5,110 patient records.A class-weighted Random Forest classifier addressed severe class imbalance (𝟒.𝟗% stroke prevalence), achieving 𝟖𝟔.𝟖𝟗%accuracy and an AUC-ROC of 𝟎.𝟕𝟖𝟑.To open the black box, we paired Kernel SHAP which identified age and average glucose level as dominant risk drivers with a novel Enhanced Accumulated Local Effects (ALE) algorithm featuring adaptive binning, spline smoothing, and bootstrapped confidence intervals.Unlike traditional Partial Dependence Plots, which introduced extrapolation errors due to feature multi-collinearity (𝒓 = 𝟎.𝟑𝟑 between age and BMI), our Enhanced ALE revealed clinically valid, non-linear risk thresholds: a sharp inflection point at age 50, an inverted U-shaped BMI curve peaking at 𝟐𝟖.𝟒 kg/m 𝟐 (mirroring the clinical "obesity paradox"), and a dual-peak glucose risk profile.The framework achieved high explanation fidelity (𝑫 = 𝟎.𝟗𝟐𝟓) and stability (𝑺 = 𝟎.𝟖𝟖𝟒), offering a reliable, clinically aligned tool for early stroke prevention.