Optimizing Human-AI Collaboration in Medical Diagnostics

Abstract

While AI increasingly outperforms human experts in medical diagnostics, fundamental questions remain about optimizing human-AI collaboration, particularly regarding workflow design and integration with clinical practice. We examine these challenges through a randomized controlled trial involving 30 physiatrists using deep learning models for video-based medical diagnostics across varying task difficulties. The study contrasts two workflow designs: a two-step workflow, where human experts make initial decisions before consulting AI, and a one-step workflow, where decisions are made concurrently. Results reveal that the two-step workflow significantly enhances accuracy in medium-difficulty tasks without extending decision time, but this advantage diminishes for highly complex cases or with more experienced physiatrists. The findings highlight how task difficulty and workflow design critically influence the success of human-AI collaboration, offering practical insights for implementing AI in clinical settings while preserving the value of human expertise.

Ähnliche Arbeiten