Considerations on external validation, potential data leakage, and model selection in a PET/CT-based radiomics model predicting residual disease in ovarian cancer

Ye et al. recently reported that a PET/CT-based radiomics model incorporating clinical variables and using a support vector machine classifier outperformed conventional scoring systems in predicting complete cytoreduction in ovarian cancer (J Ovarian Res 2026). We appreciate this valuable contribution and wish to raise three points for consideration. First, although the validation cohort was held out from model development, it shares institution and scanner platform with part of the training set. The Siemens platform, which contributed 103 of 153 training patients, was not represented in the validation cohort. The results may therefore primarily reflect reproducibility on the GE platform, and broader cross-center or cross-scanner generalizability may warrant further evaluation. Second, SMOTE oversampling and mRMR feature selection are described as applied to the entire training cohort before nested cross-validation; if this reflects the actual workflow, fold-level data leakage could be a consideration. Third, selecting the SVM from seven models using identical cross-validation estimates and reporting the maximum could, in principle, introduce some optimistic selection bias. These points are offered in the hope of further strengthening the evidence base for broader multicenter use.