AI in the Clinic: A Builder's Guide to AI in Healthcare (AI Engineering in Practice)

The model is the easy part. Medical AI has the strongest published results of any applied AI domain and the hardest path to deployment. Between a good AUC and a safe product sit: dataset bias that harms identifiable groups, a European regulatory stack most engineers have never read, clinicians who have good reasons not to trust you, and a base rate that quietly destroys the positive predictive value your test set promised. AI in the Clinic is a builder's field guide to what actually happens between a research result and a product a clinician will use on a real patient. Written developer-to-developer, for people who can already train a model and now need to understand why that was never the hard part. Inside: Why "beats a radiologist" rarely means what the headline implies, and the four questions that separate a real result from a laboratory artifact The base-rate trap that can turn a 94%-sensitive test into a coin flip at the bedside Knowledge graphs for drug discovery, and why a ranked hypothesis list is not a drug The documented, named mechanisms by which clinical AI harms identifiable patient groups — and what remediation actually exists Why a model that works at one hospital fails at the next, and what federated learning does and does not fix Calibration, decision curves, and why explainability does not deliver what it is sold as delivering EU MDR classification, the AI Act's medical-device intersection, and what ISO 42001 buys you on top of ISO 13485 and IEC 62304 Turning a research result into a regulated product without becoming a permanent pilot Behavioural health AI: the widest gap between deployed scale and evidentiary support in the whole field Every performance number in this book carries a named source, a study design, and a limitation. Retrospective, prospective and deployed are held apart throughout. Where the honest answer is "nobody knows yet," that is what the book says. This is a book for people who build software, not medical or legal advice.