The COACH theory formalizes clinical reasoning in the development and validation of medical artificial intelligence systems

Abstract

The development and deployment of medical artificial intelligence (AI) systems are limited by two persistent gaps: clinician knowledge is often introduced late in development as labels or annotations, and deployed systems often provide outputs that are difficult to reconcile with clinical reasoning. We propose COACH, a clinician-led, logic-guided framework for designing, training, validating, and monitoring medical AI systems. In COACH, clinicians define the intended use, decompose the clinical decision into named clinical features and reasoning steps, and specify acceptable evidence, uncertainty, and error modes; engineers translate these requirements into data specifications, feature extractors, model architectures, interfaces, and validation protocols; AI models learn to detect, quantify, and integrate these features. We define logical anthropomorphic design as aligning the model’s decision pathway with an explicit clinician-defined reasoning chain, and behavioral anthropomorphic design as aligning data acquisition, uncertainty communication, evidence presentation, and escalation behavior with clinical workflow. COACH is not proposed as a replacement for deep learning or foundation models, but as a way to constrain and audit them through clinical logic. This Perspective delineates an implementation pipeline, feature-dictionary method, data flow, evaluation strategy, and an endoscopy case exemplar to position COACH relative to post-hoc explainable AI, rule-based expert systems, and human-in-the-loop approaches. We argue that COACH may improve auditability, clinician trust, and data efficiency in selected high-stakes tasks, while prospective validation is required before claims of clinical benefit can be made.