Evaluation effects can rival reported concordance gains in cross-cancer survival prediction

Selecting the best test step in cross-cancer survival prediction can yield concordance gains of 0.03 to 0.05, rivaling reported 0.020–0.045 gains across 13 TCGA cancers.

• Why it matters: Reported cross-cancer learning improvements may be inflated or obscured by evaluation effects, including how models are selected and how cancers are weighted, yet this comparison was unclear.
• What they did: Across 13 TCGA cancers (5,761 patients; 1,526 disease-specific deaths), the authors used five repeats of five-fold cross-validation with paired bootstrap intervals, testing equal-weight and random-effects pooling, plus test-split fusion-weight tuning and VAECox reimplementation.
• The result: When selection avoided test data, borrowing improved concordance by 0.009 (0.001 to 0.016) under equal weighting, but test-step selection raised apparent gains to 0.017 and 0.03–0.05, while single-run estimates varied by at least 0.02 in 11%–42% of runs.

AI summary of the abstract · may contain errors · not medical advice · read the paper.