MPD²-Router: Multi-Expert Learning-to-Defer Router Accounts for Expert Availability in Medical AI

arXiv·low signal

Introduces a learning-to-defer router for glaucoma screening that accounts for real-world constraints: heterogeneous expert availability, varying reliability across case types, and partial expert coverage. Standard L2D assumes a single always-available human expert; this handles the realistic scenario of multiple specialists with different schedules and competencies. Advances the human-AI collaboration pattern for safety-critical deployments.