ML model beats Copenhagen dispatchers at spotting stroke calls

A machine learning pipeline applied to Copenhagen medical helpline recordings flagged stroke calls at 63.0% sensitivity, against 52.7% for the human call-takers handling the same line, according to a study published in npj Digital Medicine. Positive predictive value also rose, from 17.1% to 24.9%, with the authors reporting p < 0.0001.

The pipeline transcribes each call using automatic speech recognition and then classifies the text as stroke or non-stroke. Training used calls from 1 January 2015 to 31 December 2020 in Copenhagen; calls from 2021 were held out for testing. The team also tried heavier architectures and reported that 'We also tested convolutional, recurrent, and self-attention (i.e., Transformer) architectures. However, this did not improve performance.'

The authors frame the aim modestly. 'The model should be used as a supportive tool for call-takers in the decision-making process, contributing to a higher recognition,' they write, and flag the baseline problem it is trying to solve: 'approximately half of all patients with stroke do not receive the correct triage for their condition from call-takers.'

The paper is retrospective. It reports classification performance against recorded ground-truth diagnoses, not prospective outcomes on patients whose triage was actually changed by the model.