Copenhagen ML model spots strokes on helpline calls at 63%

A machine-learning pipeline reviewing audio from Copenhagen's non-emergency medical helpline recognised stroke in 63.0% of cases, against 52.7% for human call-takers, in a retrospective study published in npj Digital Medicine. Positive predictive value also rose, from 17.1% to 24.9%, with the differences significant at p < 0.0001.

The system, from researchers at Copenhagen Emergency Medical Services, Corti and the Technical University of Denmark, chains a speech-recognition network to a text classifier trained on MH-1813 helpline calls from 2015 through 2020 and evaluated on 2021 calls, with the Danish Stroke Registry as ground truth. Stroke accounted for only 0.250% of MH-1813 calls in 2021, a rarity the authors cite as part of why individual call-takers struggle.

Performance was uneven across patient sex. "The model performed significantly better on men than on women," the authors write, adding that "women may experience different and more challenging-to-identify symptoms than men." They also report that "the model exhibited less bias than call-takers did."

The paper is careful about scope. "This study does not imply that a machine-learning model can replace medical call-takers," the authors write; rather, "the model should be used as a supportive tool for call-takers in the decision-making process."