A study criticized for data leakage used anti-hypertensive-drug use as a feature to predict hypertension, even though that information would not be available when predicting the outcome for a new patient.
Notes on verification
Confirmed by the original 2021 JMIR comment (Chiavegatto Filho et al.) critiquing Ye et al. 2018, plus independent citation in a 2024 arXiv preprocessing pitfalls paper and the Kapoor & Narayanan review. Multiple independent sources corroborate the specific mechanism of leakage.
Sources
- Leakage and the reproducibility crisis in machine-learning-based science (cell.com)
- https://pmc.ncbi.nlm.nih.gov/articles/PMC7880048/ (pmc.ncbi.nlm.nih.gov)
- https://www.jmir.org/2021/2/e10969 (jmir.org)
- https://arxiv.org/pdf/2412.03491 (arxiv.org)
- https://www.jmir.org/2018/1/e22/ (jmir.org)