DenmarkHealthcare and life sciences2018–2019

Copenhagen Emergency Medical Services

Machine-learning alert for cardiac arrest during 112 calls

A Copenhagen trial of a cardiac-arrest alert on 112 calls found no significant gain in dispatcher recognition, though the model beat dispatchers on sensitivity

In progressIndependent study

Problem

Dispatchers miss roughly 25% of out-of-hospital cardiac arrests during emergency calls, which delays CPR.

Approach

A model listened to calls to Denmark's emergency number 112 in Copenhagen and flagged suspected arrests. Flagged calls were randomly assigned to show the alert to dispatchers or not. The double-masked trial ran from 1 September 2018 to 31 December 2019. The model processed 169,049 calls and flagged 5,242.

Results

90.5% vs 93.1%dispatcher recognition of confirmed cardiac arrest without vs with the alert (P = .15, not significant; researchers also reported 93.7% for the alert arm)Independent study[1]

85.0% vs 77.5%sensitivity of the model vs dispatchers aloneIndependent study[1]

17.8% vs 55.8%positive predictive value of the model vs dispatchersIndependent study[1]

Takeaway

A model that outperforms humans on sensitivity does not by itself change what dispatchers do. The authors named low specificity and the human-computer interaction as the next problems.

Sources

Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial
JAMA Network Open6 Jan 2021Research paper or report

Source summaries

[1]JAMA Network Open6 Jan 2021

Peer-reviewed randomized trial by Blomberg and colleagues. Of 5,242 flagged calls, 654 were confirmed cardiac arrests. Dispatcher recognition was 90.5% without the alert and 93.1% with it, a difference that was not significant, while the model had higher sensitivity but lower specificity and positive predictive value than dispatchers.

Full details are in the original sources.

Added 25 Oct 2026 · Updated 25 Oct 2026