Answers are grounded in this paper and cite their sources.
Repeated biopsies from the same patient may enter both the training and test sets of an ovarian cancer transcriptomic model, leading to overly optimistic performance estimates. We compared specimen-stratified with patient-grouped threefold cross-validation in three public cohorts (107 within-cohort patients; 246 specimens or biopsies). We used three fixed classifiers and 20 repeated partitions, and aggregated out-of-fold scores at the patient level. The mean differences in area under the receiver operating characteristic curve (AUC) were 0.126, 0.339 and 0.137, with conditional patient-bootstrap intervals of 0.084–0.171, 0.241–0.450 and 0.087–0.197. After labels were randomized consistently within patients, specimen-split support vector machines still achieved mean AUCs of 0.652–0.830. Two additional sources (36 patients; 117 measurements) showed similar predictability of artificial labels. Across all five sources, patient-grouped within-fold null AUCs were near 0.5. Independently permuting specimen labels obscured this effect. In contrast, a spatial subset restricted to pretreatment biopsies retained an AUC difference of 0.131. These findings show how patient identity can inflate apparent performance in repeated-biopsy transcriptomic analyses. We provide an auditable ovarian cancer benchmark using established validation controls, not a new algorithm or a clinically validated predictor.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: