Multimodal Fusion of mpMRI Radiomics, Clinical Features, and Hematological Biomarkers Enhances Machine Learning-Based...

To develop and validate an interpretable machine-learning model for early prediction of biochemical recurrence (BCR) by fusing multiparametric MRI radiomics, clinical variables, and hematological biomarkers. A total of 172 patients (115 BCR-positive, 57 BCR-negative), who underwent RP with a median follow-up of 37 months, from Taizhou People's Hospital were retrospectively enrolled. Preoperative ADC and T2-weighted images were manually segmented to extract radiomic features; laboratory and clinical data included PSA, routine blood counts, lipid profile, and coagulation indices. After LASSO selection, seven machine-learning algorithms were trained and internally validated. The best-performing model was interpreted with SHAP analysis to quantify feature contributions. PSA emerged as the dominant predictor, exhibiting the highest mean absolute SHAP value and thereby exerting the greatest influence on model output. Among coagulation parameters, D-dimer and APTT were significantly associated with BCR (p = 0.007 and p = 0.036, respectively). Integration of ADC + T2 radiomics with clinical/hematological variables via a SVM classifier yielded the highest performance: AUC 0.916 (95% CI: 0.891-0.941) in the training set and 0.820 (95% CI: 0.762-0.878) in the validation set. SHAP revealed that ADC/T2 radiomic features constituted 60% of the top-ten predictors, while D-dimer exhibited a paradoxical negative SHAP value, suggesting a potential protective association. Multimodal fusion of mpMRI radiomics, clinical feature, and hematological biomarkers substantially improves BCR prediction. The interpretable model highlights PSA, and coagulation dysfunction, as key recurrence drivers, offering a clinically actionable tool for personalized management.