AI Improves Stroke Recognition in Emergency Calls

May 26, 2023

An artificial intelligence (AI) tool outperformed human call handlers in recognizing patients with stroke from calls to the emergency services, a new study shows.

The AI model correctly identified more patients who were actually having a stroke than the human call operators and also had a higher positive predictive value, a measure of the proportion of predicted positive cases that are actually positive.

"The model did a better job in both measures. It flagged fewer patients in total with a suspected stroke but correctly identified more patients who were actually having a stroke than the human dispatcher," study co-author, Jonathan Wenstrup, MD, Copenhagen University Hospital–Herlev and Gentofte & Copenhagen Emergency Medical Services, Denmark, told | Medscape Cardiology.

Wenstrup presented the study on May 24 at the European Stroke Organisation Conference (ESOC) 2023 in Munich, Germany.

For the study, the researchers linked the Danish stroke registry — which contains information on every patient admitted to hospital with a stroke, including the time of onset of symptoms — with the emergency call registry, which has recordings of all phone calls to the medical helpline at Copenhagen emergency services. The calls were labeled as those from patients who subsequently turned out to be having a stroke and those who were determined not to be having a stroke.

The AI model was trained to transcribe the audio recordings of the emergency calls as text and look for differences between the stroke calls and the nonstroke calls.

The model was trained using data on 1.5 million calls to the emergency services between 2015 and 2020, of which 7370 turned out to be actual stroke cases. It was then tested on 2021 data on 344,000 calls of which 750 were stroke cases.

Results showed that the AI model correctly identified 63% of patients who were having a stroke, a better result than the human emergency call dispatchers who recognized just 52.7% of stroke cases.

The model also had a better positive predictive value — 24.9% vs 17.1% for the human dispatcher.

Combining the two measures together gives an overall F1 score (an overall measure of a test's accuracy) of 35.7 for the AI model compared with 25.8 for the human call handlers.

"The AI model recognized stroke better and has a lower rate of false positives than the actual emergency services dispatchers," Wenstrup commented.

He explained that stroke was a difficult condition to identify from calls to the emergency services. "Many cases go undetected at this stage, leading to delays in treatment that can have potentially life-threatening consequences for patients."

A limitation of this study is that it had a retrospective design. The model has not yet been tested in a live setting. "We need to do a study to see how it performs when implemented in real life," Wenstrup said.

He believes the AI tool could become an aid to help emergency telephone operators recognize patients who are having stroke.

"When they are talking to all sorts of different people calling in, this model could be running in the background and would flag up a warning that a particular patient has a high probability of having a stroke and should be prioritized for urgent care."

He added: "If the model performs similarly well in a real-life setting, then it could improve stroke recognition by the emergency call handlers, enabling more stroke patients to get the rapid advanced treatment that improves outcomes."

Wenstrup noted that further improvements to the model could expand its capabilities.

"In future, it may be possible to train the framework directly from the call audio, bypassing the transcription step, as well as incorporating nonword audio — such as a slurred voice — into the training data. However, given the promising results of this study, it is already clear that technologies like this have the capability to completely transform stroke diagnosis and care," he said.

The study was funded by Trygfonden (a nonprofit foundation in Denmark), as well as the University of Copenhagen, University Hospital Copenhagen–Herlev, and Gentofte and Innovation Fund Denmark. Wenstrup has disclosed no relevant financial relationships. The Machine Learning Framework was created by Corti, a private company, and other co-authors of the study are researchers at Corti.

European Stroke Organisation Conference (ESOC) 2023. Presented May 24, 2023.

For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.