The representational bottleneck in rehabilitation AI: from human-cognitive proxies to pathway-based clinical representations

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Artificial intelligence (AI) in healthcare is often framed as a problem of model performance. This view overlooks a prior representational problem: many clinical data are human-cognitive proxies built for communication, documentation, classification, or administration rather than high-fidelity computational access to clinical phenomena. In rehabilitation medicine, movement quality, compensatory strategy, sensorimotor integration, fatigue-related deterioration, and postural instability are frequently compressed into proxy measures such as TUG, BBS, FMA, ECOG performance status, diagnosis codes, or clinical notes. We propose Clinical Reality Translation-Construction (CRTC), a representational design framework that specifies how clinically meaningful phenomena can be constructed as AI-readable entities before model development. CRTC argues that the bottleneck in rehabilitation AI is often representational before it is algorithmic. Its central task is not merely to automate interpretation of existing human-readable proxies, but to construct pathway-based clinical representations: AI-readable entities that model how clinical problems emerge through bodily impairment, temporal adaptation, environmental coupling, habitual compensation, symptom perception, institutional classification, and lived disability. Drawing on digital phenotyping, digital biomarkers, ambient AI documentation, imaging AI, digital twins, wearable monitoring, and rehabilitation kinematics, this Perspective outlines CRTC and criteria for evaluating whether richer representations improve prediction, explanation, actionability, and equitable implementation. CRTC therefore shifts the primary design question from how AI should learn to what should become available for learning.