Copenhagen study: ML tops call-takers on stroke helpline calls

A machine-learning pipeline applied to Copenhagen emergency calls flagged suspected strokes with 63.0% sensitivity and 24.9% positive predictive value, against 52.7% and 17.1% for the human call-takers handling the same lines. The difference was statistically significant at p < 0.0001.

The retrospective study, published in npj Digital Medicine, used 2,361,178 recordings from Copenhagen's 1-1-2 emergency line and MH-1813 medical helpline between 2015 and 2020, validated against the Danish Stroke Registry, with 2021 calls held out as a test set. Calls were first transcribed by an automatic speech recognition model built from convolutional and bidirectional LSTM layers, then classified as stroke or non-stroke by an ensemble of five multi-layer perceptrons running on bag-of-words features.

The authors position the system as assistive rather than autonomous: a "supportive tool for call-takers, aiding in early and accurate stroke recognition." They also flag a gap the paper itself cannot close. "We did not have data to estimate potential bias related to race, ethnicity, language, accent, or dialects," they write, and note the training data is "legally restricted by Danish patient privacy and secrecy laws," which complicates external replication.