Can a smartwatch save your life? Google researchers develop smartwatch algorithm to detect cardiac arrest

A machine learning algorithm running on a smartwatch demonstrated the ability to detect sudden loss of pulse with high specificity (99.99%) and moderate sensitivity (67.23%), according to a study led by Google Research. Designed to identify cardiac arrest events, the system can automatically place an emergency call when it senses an event has occurred, even if the user is unresponsive.

Out-of-hospital cardiac arrest (OHCA) is a significant cause of sudden cardiac death, with survival rates highly dependent on immediate recognition and intervention. Approximately 50–75% of OHCA cases are unwitnessed, reducing the likelihood of immediate medical response and resuscitation. Researchers evaluated whether a smartwatch could autonomously detect pulseless events and contact emergency services while minimizing false alarms.

In the study, “Automated Loss of Pulse Detection on a Consumer Smartwatch,” published in Nature, researchers trained an algorithm using photoplethysmography (PPG) and motion data. The team then validated the system across six distinct cohorts, including controlled clinical settings and free-living real-world environments.

In an electrophysiology lab, 100 patients undergoing defibrillator testing experienced induced ventricular fibrillation, providing data on pulselessness. Another 99 participants experienced pulselessness through a tourniquet-induced arterial occlusion model. A larger free-living cohort of 948 users provided additional data without pulseless events.

220 participants wore the smartwatch passively in daily life to see how often false positives occurred. 135 participants were studied both in free-living conditions (for specificity) and in a controlled setting where their pulse was intentionally stopped via tourniquet-induced arterial occlusion (for sensitivity evaluation).

21 professionally trained stunt persons simulated out-of-hospital cardiac arrest collapses to evaluate the algorithm’s detection accuracy during high-motion events.

No statistical difference was found between PPG signals from ventricular fibrillation and arterial occlusion-induced pulselessness. Sensitivity for motionless pulseless events was 72%, while sensitivity for simulated collapse events was 53%. Specificity reached 99.99%, with one false emergency call per 21.67 user-years. The system identified pulselessness within 57 seconds, followed by a 20-second user response check before initiating a call.

Wearable devices that detect cardiac arrest could profoundly improve survival rates, particularly for unwitnessed events. To prevent unnecessary emergency calls, ensuring a lower false positive rate is a critical next step.

The algorithm was trained using controlled pulseless events and may not perfectly match real-world events. Ongoing real-world smartwatch data collection may enable further refinements to improve the algorithm’s accuracy and reliability in diverse conditions.

© 2025 Science X Network