Dosiomics-based prediction of late genitourinary side-effects after CyberKnife for prostate cancer: a SHAP-explainable machine learning approach

To evaluate whether organ-specific dosiomics extracted from CyberKnife ® prostate SBRT plans can predict late genitourinary (GU) adverse events, and to provide clinically interpretable risk drivers using SHapley Additive exPlanations (SHAP) analysis. Eighty-seven patients with localized prostate adenocarcinoma treated with CyberKnife ® SBRT (35–36.25 Gy in five fractions) were retrospectively analyzed. Late GU adverse events were graded using the RTOG scale; the endpoint was grade 1 or higher occurring beyond 6 months (follow-up up to 5 years). PyRadiomics dosiomic features were extracted from prostate, PTV, bladder, rectum, penile bulb and a surrogate urethra model (SUM). Four models were developed on the same patients and partition, using clinical and patient-reported variables, conventional dose–volume parameters, dosiomic features, or their combination. After patient-level train/test splitting (80/20), feature filtering and LASSO selection were applied and a classifier was trained with group-aware cross-validation and class weighting. Performance was assessed on an independent hold-out test set with bootstrap confidence intervals, feature selection stability across independent training sets, and SHAP contributions related to the conventional dose–volume parameters of each structure. Late GU adverse events occurred in 17.2% of patients. The dosiomic model achieved the highest discrimination (ROC-AUC 0.92, 95% CI 0.69–1.00; PR-AUC 0.84, 95% CI 0.53–1.00, against a prevalence reference of 0.167), followed by the combined (ROC-AUC 0.90, 95% CI 0.73–1.00; PR-AUC 0.68, 95% CI 0.37–1.00), the DVH (ROC-AUC 0.86, 95% CI 0.68–1.00; PR-AUC 0.65, 95% CI 0.35–1.00) and the clinical models (ROC-AUC 0.46, 95% CI 0.22–0.70; PR-AUC 0.23, 95% CI 0.16–0.37). Adding the clinical and dose–volume blocks to the dosiomic features did not improve discrimination. Balanced accuracy was also computed (reference value 0.50): 0.80 for the dosiomic, 0.70 for the DVH and 0.43 for the clinical model. SHAP analysis identified a compact set of plausible predictors reflecting both organ geometry and dose-pattern spatial heterogeneity across bladder, urethra and prostate. An explainable dosiomics-driven approach showed encouraging discrimination for late GU adverse event prediction after CyberKnife ® prostate SBRT and provided clinically interpretable risk drivers. Multicenter validation is warranted to confirm generalizability.

Authors

Institutions

Publication Details

Journal
BMC Cancer
Published
2026-11-17
DOI
https://doi.org/10.1186/s12885-026-17279-2
Primary Topic
Radiomics and Machine Learning in Medical Imaging
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Dosiomics-based prediction of late genitourinary side-effects after CyberKnife for prostate cancer: a SHAP-explainable machine learning approach

Savino Cilla, Cecilia Arrichiello, Donato Pezzulla, Sisto Perdonà et al.
3 links
BMC Cancer
Radiomics and Machine Learning in Medical Imaging
article

Dosiomics-based prediction of late genitourinary side-effects after CyberKnife for prostate cancer: a SHAP-explainable machine learning approach

Savino Cilla, Cecilia Arrichiello, Donato Pezzulla, Sisto Perdonà, Esmeralda Scipilliti, Francesca Buonanno, Rossella Di Franco, Alessandro Izzo, Sabrina Rossetti, Gianluca Ametrano, Valentina d'Alesio, Dario Franzese, Vincenzo Ravo, Francesco Passaro, Sandro Pignata, Valentina Borzillo, Simona Mercogliano, Rocco Mottareale, Marcello Serra
article en

Abstract

To evaluate whether organ-specific dosiomics extracted from CyberKnife ® prostate SBRT plans can predict late genitourinary (GU) adverse events, and to provide clinically interpretable risk drivers using SHapley Additive exPlanations (SHAP) analysis. Eighty-seven patients with localized prostate adenocarcinoma treated with CyberKnife ® SBRT (35–36.25 Gy in five fractions) were retrospectively analyzed. Late GU adverse events were graded using the RTOG scale; the endpoint was grade 1 or higher occurring beyond 6 months (follow-up up to 5 years). PyRadiomics dosiomic features were extracted from prostate, PTV, bladder, rectum, penile bulb and a surrogate urethra model (SUM). Four models were developed on the same patients and partition, using clinical and patient-reported variables, conventional dose–volume parameters, dosiomic features, or their combination. After patient-level train/test splitting (80/20), feature filtering and LASSO selection were applied and a classifier was trained with group-aware cross-validation and class weighting. Performance was assessed on an independent hold-out test set with bootstrap confidence intervals, feature selection stability across independent training sets, and SHAP contributions related to the conventional dose–volume parameters of each structure. Late GU adverse events occurred in 17.2% of patients. The dosiomic model achieved the highest discrimination (ROC-AUC 0.92, 95% CI 0.69–1.00; PR-AUC 0.84, 95% CI 0.53–1.00, against a prevalence reference of 0.167), followed by the combined (ROC-AUC 0.90, 95% CI 0.73–1.00; PR-AUC 0.68, 95% CI 0.37–1.00), the DVH (ROC-AUC 0.86, 95% CI 0.68–1.00; PR-AUC 0.65, 95% CI 0.35–1.00) and the clinical models (ROC-AUC 0.46, 95% CI 0.22–0.70; PR-AUC 0.23, 95% CI 0.16–0.37). Adding the clinical and dose–volume blocks to the dosiomic features did not improve discrimination. Balanced accuracy was also computed (reference value 0.50): 0.80 for the dosiomic, 0.70 for the DVH and 0.43 for the clinical model. SHAP analysis identified a compact set of plausible predictors reflecting both organ geometry and dose-pattern spatial heterogeneity across bladder, urethra and prostate. An explainable dosiomics-driven approach showed encouraging discrimination for late GU adverse event prediction after CyberKnife ® prostate SBRT and provided clinically interpretable risk drivers. Multicenter validation is warranted to confirm generalizability.

BMC Cancer
Istituto Nazionale Tumori IRCCS "Fondazione G. Pascale" (IT)
Ministero della Salute
Good health and well-being
Radiomics and Machine Learning in Medical Imaging
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon