Toward Interpretable Machine Learning Models for Grading Prostate Cancer: A Multicenter Development and Validation...

To develop and validate an interpretable machine learning model for distinguishing prostate cancer (PCa) with Gleason Score (GS) ≤ 3+4 from those with GS ≥ 4+3, and to explore the predictive value and biological significance of core radiomic features.

This study enrolled a retrospective multicenter cohort, including 225 PCa patients. Two radiologists manually segmented intratumoral and peritumoral regions of interest on T2-weighted imaging and extracted radiomic features. We built machine learning models by integrating clinical variables with intratumoral and peritumoral radiomic features. To enhance model interpretability, we used SHapley Additive exPlanations (SHAP), correlation analysis, and mediation analysis. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, balanced accuracy, F1 score, and other relevant metrics.

The combined model integrating clinical features and IntraPeri_1mm radiomic features achieved the best overall performance. It demonstrated robust discriminative performance, with AUCs of 0.975 (95% CI, 0.945-1.000) in the training set, 0.854 (95% CI, 0.717-0.991) in the internal test set, and 0.795 (95% CI, 0.694-0.895) in the external validation set. SHAP analysis identified the core predictive feature. This feature correlated positively with high GS and showed significant yet weak correlations with aggressive clinical-radiological indicators. Mediation analysis revealed that this feature partially mediated the relationship between PI-RADS score and GS (indirect effect = 0.309, 95% CI, 0.061-0.875).

The interpretable machine learning model demonstrates excellent performance in predicting PCa grading.