Back to news
researchGoogle2026-11-07

Google published research on November 7 detecting hypertension with consumer smartwatches without a cuff, in Nature

Google's multimodal AI-PPG-ACC-HTN system trained a 1D-CNN on over 490,000 hours of unlabeled data, achieving 90.0 percent specificity, 80.6 percent positive predictive value and 65.8 percent sensitivity in a 196-subject prospective trial.

On November 7, it was reported that on November 5, Google published a blog post introducing its latest research published in Nature. Google introduced a multimodal AI system called AI-PPG-ACC-HTN that detects signs of hypertension passively using ordinary consumer smartwatches without cuff calibration. The company said that unlike conventional approaches that try to estimate a single instantaneous systolic and diastolic reading, which is easily disturbed by transient environmental noise, the new system takes a different approach by aggregating passively collected physiological features across multiple consecutive days to identify latent and persistent hypertensive state in an individual.

Methodologically, the system first trains a 1D-CNN on more than 490,000 hours of unlabeled data, using contrastive learning so the model discovers complex cardiovascular physiological features on its own, then averages multi-day 15-second feature vectors to smooth transient noise and combines height as a proxy for arterial path length, feeding the result through a logistic regression model to output a probability score for sustained hypertension.

To evaluate performance, the team ran a prospective multicenter clinical trial at five clinical sites across the United States, enrolling 196 subjects with no cardiovascular history and not taking antihypertensive medication, using 24-hour ambulatory blood pressure monitoring as the clinical gold standard. In initial screening, specificity reached 90.0 percent, positive predictive value 80.6 percent and sensitivity 65.8 percent, ahead of the 55.3 percent sensitivity and 77.8 percent positive predictive value of a single outpatient screening. Confirmatory testing on those with elevated screening results reached a positive predictive value of 96.7 percent, specificity of 90.9 percent and sensitivity of 78.4 percent.

谷歌高血压检测智能手表自然杂志AI-PPG-ACC-HTN