Copenhagen study: ML lifts stroke call recognition to 63%

On 2021 emergency calls in Copenhagen, human call-takers recognised 52.7% of strokes; a machine learning pipeline trained on six prior years of calls recognised 63.0%, with positive predictive value climbing from 17.1% to 24.9%, at p < 0.0001.

That is the headline result in a retrospective study published in npj Digital Medicine by researchers at Copenhagen University Hospital, Copenhagen Emergency Medical Services, and Corti. Training calls ran from 1 January 2015 to 31 December 2020; 2021 was held back as the test set. Each call is first transcribed by an automatic speech recognition model, then classified as stroke or non-stroke by a text model.

The authors are careful about scope. "This study does not imply that a machine-learning model can replace medical call-takers," they write, framing the pipeline instead as one that "may become a supportive tool for call-takers, aiding in early and accurate stroke recognition."

A PPV of 24.9% still means roughly three in four model stroke flags are false positives. The abstract reports no live deployment, no patient outcomes, and no time-to-treatment numbers; those would need a separate prospective study.