Machine learning as a supportive tool to recognize cardiac arrest in emergency calls

Area of Science:

• Artificial Intelligence in Emergency Medicine
• Machine Learning for Healthcare Applications
• Cardiovascular Emergency Response Systems

Background:

• Emergency medical dispatchers miss approximately 25% of out-of-hospital cardiac arrest (OHCA) cases.
• Missed OHCA cases result in delayed cardiopulmonary resuscitation (CPR) instructions.
• Developing automated systems to aid dispatchers is crucial for improving survival rates.

Purpose of the Study:

• To evaluate a machine learning (ML) framework's ability to detect OHCA from emergency call audio.
• To compare the ML framework's performance against human emergency medical dispatchers.

Main Methods:

• Retrieved 108,607 emergency calls from Emergency Medical Dispatch Center Copenhagen (2014).
• Trained an ML framework to recognize OHCA from recorded emergency calls.
• Calculated sensitivity, specificity, and positive predictive value; compared ML to dispatcher performance and time-to-recognition.

Main Results:

• The ML framework achieved higher sensitivity (84.1%) compared to dispatchers (72.5%) for OHCA detection (p < 0.001).
• ML framework demonstrated lower specificity (97.3%) versus dispatchers (98.8%) (p < 0.001).
• Time-to-recognition was significantly faster with the ML framework (median 44s) than dispatchers (median 54s) (p < 0.001).

Conclusions:

• The ML framework significantly outperformed human dispatchers in identifying OHCA from emergency calls.
• Machine learning shows promise as a decision support tool for emergency medical dispatchers.
• Implementing ML can enhance the accuracy and speed of OHCA recognition in emergency dispatch.