Copenhagen ML model beats call-takers at helpline stroke ID

Call-takers on Copenhagen's medical helpline recognised stroke in 52.7% of cases. A machine-learning pipeline run on the same audio reached 63.0%.

The study in npj Digital Medicine trained on calls recorded between 1 January 2015 and 31 December 2020 and tested on calls from 2021. Positive predictive value climbed from 17.1% for call-takers to 24.9% for the model, with p < 0.0001. The pipeline first transcribes audio with an automatic speech recognition model, then classifies the transcript as stroke or non-stroke.

The authors frame the result carefully: the system "may become a supportive tool for call-takers, aiding in early and accurate stroke recognition."

Stroke appeared in roughly 0.250% of MH-1813 helpline calls in 2021, which sets the base rate both sides are fighting. Model rights and source code belong to Corti, the AI company whose staff co-authored the paper; two of the authors held stock warrants.

Some caveats sit inside the methods. Calls where the call-taker had not entered a diagnostic category were excluded from validation and test sets, and the model performed worse on those. The paper reports no breakdown of word error rates by language, accent or dialect, and the authors flag that prior work has found higher transcription error rates for non-native speakers.