Medical Research

AI has changed medical research in two directions at once, and they pull against each other.

On the useful side, literature screening and data extraction that consumed weeks of a systematic review can now be done in hours, imaging analysis pipelines that required a specialist can be run by a research fellow, and multi-omic datasets too large for manual analysis have become tractable.

On the other side, the same tools have made it trivially easy to produce a plausible-looking paper with a fatal methodological flaw. A large share of published medical AI research reports only internal validation, and models that report excellent internal and no external validation should be assumed not to generalize.

The reproducibility problem is specific

Medical AI reproducibility fails for reasons that are not mysterious. Training data is frequently unavailable for privacy reasons. Preprocessing steps go undocumented. Model weights are not shared. The result is a literature where a published performance figure often cannot be checked by anyone.

Shortcut learning is the recurring embarrassment: models that identified the hospital from image artifacts and used its base rate, or keyed on a laterality marker rather than pathology. These score well internally and collapse externally, and nothing in the training process signals that the wrong thing was learned.

Reading a medical AI paper

Ask four things. Was it validated externally. Was it validated prospectively. What was the reference standard and who produced it. And what was the disease prevalence in the evaluation set compared with the intended deployment population.