ML model tops Copenhagen call-takers at stroke recognition

In a retrospective study of calls to Copenhagen's emergency and medical helpline numbers, a machine-learning classifier spotted stroke more reliably than the humans answering the phone.

The paper, published in npj Digital Medicine, reports that call-takers reached a sensitivity of 52.7% with a positive predictive value of 17.1%. The model reached 63.0% and 24.9%. The gap was significant at p < 0.0001.

The pipeline is two stages. Calls are first transcribed by an automatic speech recognition model, then categorised as stroke or non-stroke by a text classifier. Training ran on calls from 1 January 2015 to 31 December 2020; calls from 2021 were held out for testing. In total, 2,361,178 contacts were successfully matched to electronic records.

Performance was not even across patients. The paper reports the model "significantly outperformed call-takers on all metrics in both age groups," but also "exhibited significantly higher PPV and F1 score in men than in women." The authors do not resolve why.

The limitations are stated flatly. "The mapping of call recordings to electronic records was incomplete due to technical limitations," and the stroke labels relied on patient-reported onset times. The authors frame the system as "a supportive tool for call-takers, aiding in early and accurate stroke recognition," not a replacement.