Researchers in Singapore have built a safety net for one of primary care’s quietest failure modes: the referral a doctor intended to make and never sent.
Writing in npj Digital Medicine, a team from National University Polyclinics and the National University Health System describes MR-AI, a system that reads consultation notes and searches for referral intent that has no matching order attached. It was evaluated in a quasi-experimental study across seven primary care clinics, scanning about 25,000 notes each week.
Seven times more catches
Weekly detection of missed referrals rose roughly sevenfold, a 4.7 percentage point increase. Of 428 flagged cases, clinicians verified 225. A further 31 arrived through patient reports, bringing total detections to 256. Specificity was 99.9% and sensitivity 48.8%.
In a blinded one-week comparison, MR-AI was measured against seven physician reviewers on 83 flagged cases. The model’s positive predictive value was 61.8%, while humans confirmed 51.8%. Running the system was projected to need about 41 fewer staff hours.
Why it matters beyond the numbers
Missed referrals carry real clinical and medico-legal exposure, and until now they have been largely invisible to health systems. A tool built on notes rather than referral databases means the safety net works before anything is lodged.
The authors are careful about what the result does not show. They note that sustained improvement will need complementary changes to clinic workflow, and that cost-effectiveness still requires formal assessment. The study received no specific funding.
