CRC-Router Adds Risk-Constrained Routing to Medical AI Agents

Summary

Medical agentic AI can propagate autonomous errors into downstream clinical decisions, making reliable escalation essential. CRC-Router is a risk-constrained, uncertainty-aware routing module designed for both conventional medical prediction models and agentic systems. It combines multiple uncertainty signals with predictive scores to create a per-finding routing vector, estimates wrong-accept risk with a lightweight per-finding model, and uses Conformal Risk Control to calibrate acceptance thresholds against a user-specified risk target. In chest X-ray multi-finding triage on the NIH ChestX-ray14 dataset, it achieved the strongest empirical risk-coverage trade-off among the evaluated baselines. The result held both when CRC-Router was used as a standalone routing layer and when it was integrated into the MedRAX medical agent. The authors present the module as model-agnostic and compatible with existing predictive and agentic pipelines. Code is publicly available.