Target Leakage Reduces Cardiovascular Screening Model AUROC by 0.05

An audit of ten classifiers across 442,067 respondents shows that removing two post-diagnostic features reduces AUROC by 0.049-0.051. The study demonstrates that reported high accuracy in survey-based cardiovascular models is driven by feature-tier target leakage rather than model architecture.