ML beats Copenhagen call-takers at spotting stroke calls

Copenhagen call-takers correctly flagged just 52.7% of stroke cases on incoming helpline calls. A machine learning pipeline trained on six years of the same recordings reached 63.0% on the 2021 test set, according to a retrospective study published in npj Digital Medicine.

Positive predictive value rose from 17.1% (95% CI 15.5–18.6%) for human call-takers to 24.9% (24.3–25.5%) for the model, with the paper reporting p<0.0001. The system runs automatic speech recognition over the call and passes the transcript to an ensemble of five multi-layer perceptrons; ground truth came from the Danish Stroke Registry.

The authors are blunt about the limit. "This study does not imply that a machine learning model can replace medical call-takers...the model should be used as a supportive tool," they write, concluding that the framework "may become a supportive tool for call-takers."

The training data came from a single service — Copenhagen Emergency Medical Services, covering roughly 1.9 million residents across the 1-1-2 emergency and 1813 medical helpline lines — and the analysis is retrospective, not a live trial.