AI referral checks still need clinician review

Wayne Han Lee and colleagues evaluated MR-AI across seven Singapore primary-care clinics. Their October 30 study in npj Digital Medicine is peer-reviewed accepted research, shared before final editing.

The system combines rules and a language model to find documented referral intentions without matching orders. Clinicians verify its flags.

The study supports additional surveillance, but the system missed about half the reference-standard cases. Accuracy came from one week; human adjudication was imperfect. This observational evaluation does not establish improved patient outcomes or performance in other health systems.