Library / research-methodology-guide
Understanding biomedical AI
Audience: patients-public, students-learners ยท Review state: published
What AI in medicine means
Artificial intelligence in medicine is software that learns patterns from health data rather than following hand-written rules. Machine learning models are trained on examples, such as labelled scans or records, and then make predictions on new cases.
Reviews of the field describe applications across imaging, diagnosis support, drug discovery, hospital operations and personalized treatment planning, with imaging among the most mature uses.
Where it is actually used
Radiology and pathology tools flag suspicious findings for a specialist to review. Prediction models estimate risks, such as deterioration on a ward. Natural language systems extract structured facts from clinical notes, and drug-discovery models screen candidate molecules far faster than bench experiments can.
In nearly all deployed systems the model supports a clinician's decision rather than replacing it, because accountability for care remains with people.
How to judge an AI claim
The questions are the same as for any medical evidence: what data trained the model, does that data resemble the patients it will be used on, how was performance measured, and was it tested on data it never saw during development.
A model that excels on its training hospital's data can fail at another hospital with different equipment, coding habits and population. External validation, testing at sites uninvolved in development, is the check that separates promise from product.
The known failure modes
Bias is the central risk: models trained on unrepresentative data can perform worse for the very groups already underserved by medicine. Dataset shift, where real-world inputs drift from training data, silently degrades performance over time.
Opacity compounds both: many models cannot explain individual predictions, so errors need monitoring systems to catch them. Regulators now treat high-risk medical AI as software requiring evidence, surveillance and periodic review.
Limitations
Educational overview of AI in medicine, not an evaluation of any specific product and not medical advice.