No cuff calibration needed: AI-PPG-ACC-HTN flags hypertension in 196 people with 90.0% specificity
Synopsis
Google Research, with Beth Israel Deaconess Medical Center and Harvard Medical School, used the photoplethysmography and accelerometer signals of a consumer smartwatch to passively identify a persistent hypertensive state without cuff calibration. Rather than estimating systolic and diastolic pressure at a moment in time, the system averages 15-second signal features over multiple days and combines them with height in a regularized logistic regression that outputs a probability. Against 24-hour ambulatory blood pressure monitoring in 196 participants free of known cardiovascular disease and not taking blood pressure medication, initial screening reached 65.8% sensitivity and 90.0% specificity, versus 55.3% sensitivity for concurrent office screening.
Interpretation
A consumer smartwatch can passively judge whether a person is in a persistent hypertensive state without cuff calibration, instead of only tracking relative change. Earlier commercial cuffless devices usually required periodic calibration with an arm cuff, so they tracked change from a baseline reading rather than identifying absolute hypertensive states. In a prospective, multi-center clinical study across five US sites, 196 participants free of known cardiovascular disease and not taking blood pressure medication underwent both smartwatch monitoring and 24-hour ambulatory blood pressure monitoring; smartwatch initial screening reached 65.8% sensitivity, 90.0% specificity and 80.6% positive predictive value, while concurrent office screening reached 55.3% sensitivity and 77.8% positive predictive value.
The system reframes the task from estimating a blood pressure value at one moment to judging an individual's persistent hypertensive state accumulated over multiple days. Point-in-time estimates of systolic and diastolic pressure are highly susceptible to transient environmental noise, so passive physiological signatures are aggregated over multiple days instead. Each user's 15-second feature blocks are mean-aggregated across multiple days of wear to wash out ephemeral noise, then combined with height and fed into a regularized logistic regression that outputs a continuous hypertension probability score.
The model learns features of vascular and mechanical cardiac function directly from raw waveforms, without manual feature engineering or medical labels. The features come from self-supervised contrastive learning rather than labeled data or handcrafted feature engineering. A 1-D deep convolutional neural network was trained on 492,347 hours of unlabeled sensor data from 2,109 consenting participants, pulling overlapping 15-second photoplethysmography and accelerometer waveforms from the same individual close together in a latent space.
Perspective
Validation covers 196 participants free of known cardiovascular disease and not taking blood pressure medication across five US sites, benchmarked against 24-hour ambulatory blood pressure monitoring; the authors state this is early feasibility research, that the methods are purely exploratory, and that they do not indicate any existing or upcoming commercial product feature or device capability.
The authors state that smartwatch screening is not intended to replace standard arm cuffs or provide a definitive standalone diagnosis; initial-screening sensitivity of 65.8% is moderate; confirmatory results come from a subset of 48 participants with elevated initial office blood pressure, and the authors call the approach competitive with multi-day home monitoring, which reached 89.2% sensitivity.