Characteristics of out-of-hospital cardiac arrest calls unrecognised by machine learning versus medical dispatchers: a nationwide observational study

When a person's heart stops beating outside a hospital, every second counts. The path to survival begins with a single phone call to emergency services, where a dispatcher must instantly recognize that the caller is describing a cardiac arrest. This recognition triggers the immediate dispatch of an ambulance and, crucially, instructions for the caller to start chest compressions. However, human dispatchers are not perfect; they can miss the signs, often because the person in distress is making strange, gasping breaths that sound like snoring or because the caller is too panicked to speak clearly. In recent years, computer programs designed to listen to these calls have emerged as a potential safety net, capable of spotting patterns in speech that humans might overlook. The question facing emergency services is not just whether these machines work, but where they succeed and where they fail compared to the human experts sitting at the other end of the line.

Researchers in Sweden set out to investigate this exact dynamic by listening to a specific set of emergency calls where the human and the machine disagreed. They focused on ninety calls from 2018 where a real cardiac arrest occurred, but either the human dispatcher or the computer algorithm failed to identify it. By carefully reviewing the audio recordings of these calls, the team sought to understand the specific circumstances that led to these mistakes. Their goal was not to declare one side the winner, but to map the unique blind spots of each, hoping to find a way for them to work together more effectively.

The study revealed a clear split in how the two groups recognized the emergency. In more than half of these ninety cases, the computer model correctly identified the cardiac arrest while the human dispatcher missed it. In the remaining cases, the human dispatcher caught the emergency while the computer model failed to do so. This division was not random; it depended heavily on the situation and the people involved. The computer was significantly better at recognizing the emergency when the caller was alone, when the caller was a regular person rather than a medical professional, and when the call came from a private home. The machine also performed better when the caller knew the patient personally and when the call happened in the evening.

Conversely, the human dispatchers were more likely to spot the emergency when the caller was not alone and when someone was already performing chest compressions at the time of the call. The researchers found that when a caller was alone, the computer was far more likely to succeed than the human. This suggests that the quiet, focused nature of a single voice speaking in a home environment provides the computer with the clear audio signals it needs to work. In contrast, human dispatchers seemed to rely more on the chaotic context of a scene where others were present or where CPR was already underway, perhaps picking up on background noises or the urgency of multiple voices that confused the computer's listening software.

The study also highlighted how the background of the caller influenced the outcome. The computer was much more successful when the caller was a layperson, but it struggled when the caller was a healthcare professional. This suggests that medical professionals may describe the situation using different words or tones that the computer's training did not expect, whereas laypeople tend to use more standard descriptions of collapse and breathing that the machine recognizes easily. Similarly, the computer was more likely to succeed when the caller knew the patient, likely because familiar callers could provide more specific details about how the person's condition had changed, creating clearer linguistic clues for the software to follow.

Despite these differences, the researchers emphasized that neither side is perfect on its own. The computer missed calls where the scene was noisy or chaotic, while humans missed calls where the signs were subtle or the caller was isolated. The authors suggest that the most promising path forward is not to replace human dispatchers with machines, but to use the computer as a supportive tool. In this scenario, the machine would act as a second pair of ears, alerting the human dispatcher when it detects a cardiac arrest in situations where the human might be distracted or where the caller is alone and the machine performs best.

The findings come with a note of caution. The study looked at a relatively small group of calls where the two sides disagreed, and the researchers noted that some of the statistical patterns were not entirely precise due to this limited number. They also pointed out that the computer's performance might change as the technology evolves or as the types of calls change. However, the core observation remains robust: humans and machines make different kinds of mistakes. By understanding these specific differences, emergency services can design systems where the computer handles the quiet, isolated calls with high sensitivity, while the human dispatcher manages the complex, noisy scenes, creating a partnership that could save more lives than either could achieve alone.