Researchers have proposed CRC-Router, a lightweight layer that decides whether a medical AI system should accept a case or send it for further review. The goal is selective automation: handle routine work without allowing uncertain errors to flow into clinical decisions unchecked.
The router combines uncertainty measures with a model’s prediction score, estimates the risk of a wrong acceptance for each finding and calibrates thresholds to a user-chosen risk limit. Conformal risk control, a statistical calibration method, keeps decisions aligned with that target.
In experiments on multi-finding chest X-ray triage using the NIH ChestX-ray14 dataset, CRC-Router produced the strongest risk-coverage trade-off among evaluated baselines. It worked alone and when attached to the MedRAX medical agent. This is research evidence, not clinical clearance: performance on one dataset does not establish safety across hospitals, populations or imaging systems. The code is publicly available for further testing.