AI Cautionary Guide: Pitfalls and Strategies for the Use of Machine Learning in Medical Research

Machine learning (ML) is increasingly utilized in medical research. However, underrecognized methodological pitfalls can result in misleading conclusions, which if acted upon can result in patient mismanagement and harm. As a result, we have provided a compilation of common pitfalls and practical safeguard considerations for ML-enabled medical research. Examples emphasize the importance of integrating clinical/biologic knowledge and statistical/epidemiologic principles, illustrated with key case examples. Topics include (1) Data drift, which can create artificial trends when coding systems or clinical practices change over time; (2) Feature construction that should be clinically grounded, as models may otherwise learn operational proxies rather than true biology; (3) Accidental data leakage, which can inflate apparent model performance and collapse on real-world deployment; (4) Competing risks and differential time-at-risk that can generate spurious associations in safety analyses if not modeled explicitly; (5) Confounding and collider bias, which can distort findings, particularly in EHR-based studies; (6) For rare outcomes, inappropriate evaluation measures can overstate performance and mislead explainability analyses; (7) Finally, pooled multi-trial analyses require careful handling of the intrinsic heterogeneity to avoid extracting study artifacts rather than generalizable signals. By outlining many of the common potential issues, with lessons learned, we aim to provide the medical and research community with practical insights to help foster trustworthy ML-enabled medical research.