Cliniflow: Effects of AI, Human, and Hybrid Diagnostic Advice (and Advice Timing) on Physicians’ Diagnostic Decisions

This preregistration covers a randomized experiment on the Human Diagnosis Project (Human Dx) platform. Physician participants diagnose internal medicine case vignettes and receive diagnostic advice from either (i) a 4-LLM ensemble, (ii) a 4-physician ensemble, (iii) a hybrid collective (2 physicians + 2 LLMs), or (iv) no advice (control). Advice is presented either early (together with case details and before a diagnosis is made) or late (after all clinical findings are shown and participants made their initial diagnosis). Outcomes include diagnostic accuracy (top-1/top-3/top-5 and mean reciprocal rank (MRR), of the correct diagnosis), advice uptake (similarity to advice and change after advice), and diagnostic confidence. Analyses will use hierarchical Bayesian models with varying intercepts by case.