Preprocessing leakage inflates radiomics AUC by 0.09 and quadruples model disagreement

Across 50 radiomics datasets, SMOTE leakage inflated AUC by 0.09 (95% CI 0.07-0.11), quadrupled the Rashomon set, and 79% of the effect was mediated by performance ceiling inflation. Near-optimal pipelines disagreed on 20-30% of patients.

• SMOTE leakage inflated pooled AUC by 0.09 (95% CI 0.07–0.11), exceeding 0.05 in 64% of datasets.
• Leakage quadrupled the Rashomon set (from 0.06 to 0.23), with 79% of the increase mediated by best-AUC inflation.
• Among near-optimal pipelines (ε=5%), patient-level classification agreement was only moderate (Fleiss κ=0.58), with <30% feature overlap.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.