ML tops Copenhagen helpline call-takers at spotting stroke

A machine-learning system reading transcripts of Copenhagen emergency calls recognised stroke in 63.0% of confirmed cases, against 52.7% for the human call-takers who answered the same lines, according to a retrospective study published in npj Digital Medicine. Positive predictive value also rose, from 17.1% to 24.9%, with the difference significant at p < 0.0001.

The pipeline pairs automatic speech recognition with an ensemble of five multi-layer perceptron classifiers. It was trained on Copenhagen Emergency Medical Services calls from 2015 to 2020 and tested on 2021 calls, 2,361,178 of which were matched to audio files. Stroke prevalence in the 2021 test set was 0.250%. Ground truth came from linking calls to the Danish Stroke Registry.

The authors, led by Jonathan Wenstrup with co-authors including Helle Collatz Christensen and Michael R. Sayre, describe the result as evidence the framework "significantly outperformed human call-takers in stroke recognition," noting that "the framework can thus be tuned to a sensitivity of around 73%, while still having a higher positive predictive value than the human call-taker."

The study is retrospective and runs on recorded calls rather than live audio, and the training corpus is Danish-language and Copenhagen-only. The paper frames the work as "the largest collection of audio calls from patients with stroke to date."