Dosiomics-based prediction of late genitourinary toxicity after Cyberknife for prostate cancer: a SHAP-explainable machine learning approach

Abstract

Abstract Objective To evaluate whether organ-specific dosiomics extracted from CyberKnife® prostate SBRT plans can predict late genitourinary (GU) toxicity and to provide interpretable risk drivers using SHapley Additive exPlanations (SHAP) analysis. Methods Eighty-seven patients with localized prostate adenocarcinoma treated with CyberKnife® SBRT (35–36.25 Gy in five fractions) were retrospectively analyzed. Late GU toxicity was graded using the RTOG scale; the endpoint was grade ≥ 1 toxicity occurring ≥ 6 months after treatment (follow-up up to 5 years). DICOM RTDOSE/RTSTRUCT data were used to extract PyRadiomics dosiomics features from prostate, PTV, bladder, rectum, and penile bulb. After patient-level train/test splitting (80/20), feature filtering and LASSO selection were applied, a classifier was trained with group-aware cross-validation and class weighting. Model performance was assessed on an independent hold-out test set with bootstrap confidence intervals. SHAP analysis was used for explainability. Results Late GU toxicity occurred in 17.24% of patients. ROC-AUC was 0.85 ± 0.10 (95% CI 0.60–1.00) and accuracy was 0.88 ± 0.04 (95% CI 0.83–0.94). SHAP analysiss identified a compact set of plausible predictors reflecting both organ geometry and dose-pattern heterogeneity across prostate and bladder. Conclusions An explainable dosiomics-driven approach showed encouraging discrimination for late GU toxicity prediction after CyberKnife® prostate SBRT and provided clinically interpretable risk drivers. Multicenter validation is warranted to confirm generalizability.