The model was pre-trained without diagnostic labels on 492,347 hours of photoplethysmography and accelerometer data from 2,109 participants, learning which short sensor segments came from the same person [1][3]. The accelerometer also picks up the ballistocardiogram, the small body movements produced by each heartbeat. The system averages its readings over several days into a probability of hypertension [3].
As a confirmatory test in 48 people whose first office reading was elevated, the watch reached 78.4% sensitivity and 90.9% specificity, against 67.6% and 63.6% for a return office visit; multi-day home cuff monitoring scored 89.2% and 81.8% [1]. Google says more than half of adults with hypertension do not know they have it [1].
The authors call the system early feasibility research, not a replacement for arm cuffs or a standalone diagnosis [1]. Google says the findings helped inform the blood pressure trends feature in its Google Health app [1].