A Shapley Value-Based Framework for Transparent Machine Learning in ROI-Based Lesion-Symptom Mapping of Aphasia

This study benchmarks predictive models in ROI-based lesion–symptom mapping (LSM) for aphasia and integrates SHapley Additive exPlanations (SHAP) to deliver transparent neurodiagnostic predictions. Lesion–ROI overlaps were calculated for nine language-related regions in 70 patients with aphasia. Seven machine learning models (XGBoost, SVR-Linear, SVR-RBF, Random Forest, Ridge, Lasso, and MLP) were trained to predict standardized behavioral scores. Model performance was assessed using R², RMSE, MAE, AUC, F1 score, sensitivity, and specificity. SHAP provided global and local feature attributions, which were statistically validated against Pearson’s r and two-sample t-statistics via Spearman’s ρ and 5,000-permutation testing. SVR-RBF achieved the highest predictive accuracy (R² = 0.927; RMSE = 1.227; AUC = 0.938), with SHAP consistently highlighting the Left Arcuate Fasciculus and Broca’s area as top contributors. Spearman correlations between mean |SHAP| and univariate metrics were ρ = 0.75 (p = 0.020; perm p = 0.025) for Pearson’s r and ρ = 0.72 (p = 0.030; perm p = 0.019) for t-statistics. Our SHAP-augmented ROI-based LSM framework predicts the severity of aphasia and provides explanations at a specific level, thereby increasing transparency and understanding. By elucidating how the AI arrives at its predictions, this approach enhances transparency, fosters clinician trust in AI-driven diagnostics, and guides personalized rehabilitation strategies.