Copenhagen study: ML beats call-takers at phone stroke triage

On Copenhagen's non-emergency medical helpline, call-takers caught about half the strokes coming in by phone. A machine learning pipeline, run on the same calls after the fact, caught more.

The retrospective study in npj Digital Medicine matched 2,361,178 audio calls to Danish Stroke Registry records from 2015 through 2021, training on 2015-2020 and testing on 2021. Call-takers achieved 52.7% sensitivity (95% CI: 49.2-56.4%) and 17.1% positive predictive value. The model reached 63.0% sensitivity and 24.9% PPV, and could be tuned to around 73% sensitivity while still beating the humans on PPV.

The pipeline is two stages: automatic speech recognition converting Danish audio to text, then an ensemble of five multi-layer perceptrons classifying bag-of-words features. The authors cast it narrowly: "The framework should be used as a supportive tool for call-takers in the decision-making process, contributing to higher recognition of patients with stroke."

Not every subgroup came out even. The paper reports the model "performed better in males and patients aged 65+ than younger females," and performance degraded on 2021 calls that lacked diagnostic category codes. The two helplines involved are 1-1-2 and 1813, serving a 1.9-million-resident catchment. Corti, the audio-AI firm whose researchers co-authored the paper, has staff listed on affiliations alongside Copenhagen University Hospital and CEMS.