ML model tops Copenhagen call-takers on stroke recognition

A machine learning system trained on Copenhagen emergency calls recognized strokes in 63.0% of cases, against 52.7% for the human call-takers it was benchmarked against, according to a retrospective study published in npj Digital Medicine. Positive predictive value rose from 17.1% to 24.9%. The gap was statistically significant at p < 0.0001.

The training corpus is large for this setting: 2,361,178 calls with matched audio from 2015 through 2020, with 2021 held out as the test set. The data spans both the 1-1-2 emergency line and the non-emergency medical helpline 1813, serving a 1.9 million-person region. The pipeline is a speech recognition model built on convolutional and bidirectional LSTM layers, feeding an ensemble of five text classifiers.

The authors write that the framework "may become a supportive tool for call-takers, aiding in early and accurate stroke recognition," and stop short of endorsing deployment. They report the model performed better in men than women and better in patients aged 65 and up than in the 18-64 group, while noting it showed less bias along those axes than the call-takers did.

This is retrospective, single-city work, and the paper recommends further investigation before clinical use.