ML beats call-takers at spotting stroke in Copenhagen helpline calls

Call-takers at Copenhagen's non-emergency medical helpline MH-1813 catch 52.7% of strokes they are told about, at a positive predictive value of 17.1%. A machine learning pipeline trained on six years of those same calls caught 63.0%, at a PPV of 24.9%.

Those numbers come from a retrospective study in npj Digital Medicine by researchers at Copenhagen University Hospital, Corti, and the Technical University of Denmark, published 19 December 2023. The team transcribed calls with an automatic speech recognition model and ran the transcripts through an ensemble of five multi-layer perceptrons; ground truth came from linking patients to the Danish Stroke Registry.

The authors are explicit about why any of this matters: "Advanced stroke treatment is time-dependent and, therefore, relies on recognition by call-takers at prehospital telehealth services to ensure fast hospitalisation," the abstract states. Their framing is supportive rather than replacement. The pipeline "may become a supportive tool for call-takers, aiding in early and accurate stroke recognition."

Training used calls "from 1 January 2015 to 31 December 2020 in Copenhagen," with 2021 held out as the test set. Across the matched corpus, 2,361,178 audio files were linked to records, with another 2,730,199 contacts lacking matching audio. Stroke prevalence in the 2021 test calls was 0.250%.

The gap between model and human is statistically significant at p < 0.0001. Two caveats sit in the setup itself: the paper is retrospective and run on one city's helpline, and two of the three equally-contributing lead authors list Corti as an affiliation. The abstract reports no prospective deployment.