Book

Every research on medical artificial intelligence opens with a number — an accuracy, an area under the curve, a claim to match or exceed the specialists. Yet hospitals remain cautious, and for good reason: a model can be highly accurate and still be untrustworthy. The distance between the two is where the real work of medical AI lies, and it is the subject of this book. More Than Accuracy argues that the most consequential biases in clinical AI are introduced not by the data but by the method — by the under-examined choices about which features a model may use, how its inputs are transformed, and how its decisions are explained. These biases are dangerous precisely because they survive good data and hide behind the very accuracy we use to certify success; often they even inflate it. Across four case studies — thalassemia diagnosis from blood data, lung-cancer classification from CT imaging, thyroid-nodule assessment through structural modelling, and a deliberately non-clinical mirror in materials engineering — the book shows that respecting the data need not cost performance. A method that refuses to normalise, rebalance, or extract, and that grounds interpretability in the model rather than in an unreliable post-hoc tool, can match and even exceed the heavily engineered pipelines it declines to imitate. Along the way, the book dismantles the field’s faith in post-hoc explanation, showing through impossibility results and real experiments why a plausible attribution is not a trustworthy one, and offers a disciplined alternative built on model-independent evidence and triangulation. Rigorous yet readable, and honest about its own limits, this is a book for practitioners, clinicians, and anyone who suspects the field’s confidence has outrun its evidence. “In high-stakes AI, restraint is a discipline, not a deficiency — and trust is earned not by the size of a number, but by the integrity of the path taken to reach it.”