Development and evaluation of a combined machine learning model for the prediction of acute GU toxicity in prostate cancer radiotherapy.
Abstract
Purpose
To develop and internally evaluate a compact multimodal machine learning model for predicting acute genitourinary (GU) toxicity after prostate cancer radiotherapy and to assess the effect of validation design on estimated performance.
Methods
A retrospective cohort of 92 patients was analyzed, including 84 training patients and 8 hold-out test patients. Acute GU toxicity was modeled as a binary endpoint (RTOG Grade ≥ 1 versus Grade 0). From 495 clinical, DVH, radiomic, and dosiomic predictors, compact feature sets (k = 10 and k = 15) were derived using L1-based stability selection. Primary internal performance was estimated using repeated stratified nested cross-validation, with preprocessing, feature selection, and model fitting confined to each outer training fold. Results were compared with non-nested cross-validation using global feature selection before fold partitioning.
Results
Acute GU toxicity occurred in 33 patients, including 26 Grade 1 and 7 Grade 2 events. In nested cross-validation, the clinical-only baseline achieved the highest performance (ROC-AUC = 0.714 ± 0.108; PR-AUC = 0.626 ± 0.136), exceeding the compact k = 10 multimodal model (ROC-AUC = 0.611 ± 0.139; PR-AUC = 0.578 ± 0.159). For the same k = 10 model, non-nested cross-validation yielded higher estimates (ROC-AUC = 0.794 ± 0.116; PR-AUC = 0.793 ± 0.117), consistent with optimistic performance inflation from global feature preselection.
Conclusions
Clinical variables remained the strongest predictors in this small cohort. Compact multimodal models were feasible and interpretable but did not improve leakage-free internal performance. These findings support nested cross-validation in small high-dimensional radiomics and dosiomics studies.