ML model beats Copenhagen call-takers at spotting stroke

Call-takers at Copenhagen Emergency Medical Services correctly flagged stroke in 52.7% of true stroke cases on the 2021 test set. A machine-learning pipeline, run retrospectively over the same calls, reached 63.0%. Both figures come from a study in npj Digital Medicine covering 2,361,178 matched call recordings from 2015 through 2021.

The pipeline pairs automatic speech recognition with an ensemble of five multi-layer perceptrons for text classification, trained on calls to Denmark's 1-1-2 emergency line and the MH-1813 medical helpline from 2015 to 2020 and evaluated on 2021 calls. Positive predictive value rose from 17.1% for call-takers to 24.9% for the model, with p<0.0001 across the reported metrics. Stroke represented 0.250% of all MH-1813 calls in the test year.

The authors frame it as decision support rather than automation: "the model should be used as a supportive tool for call-takers in the decision-making process, contributing to higher recognition of patients with stroke." Co-authors include researchers from Corti, the Copenhagen conversational-AI vendor that sells dispatch tooling to EMS agencies, alongside clinicians from Copenhagen University Hospital and the Danish Stroke Registry.

The study is retrospective. The model has not been evaluated inside a live call, and even at 24.9% PPV, three in four patients it would flag are still not having a stroke.