Google Research model flags hypertension from smartwatch PPG and motion sensors

A cuff-free hypertension screen on the wrist

On November 5, 2026, Google Research published a blog post alongside a paper in Nature describing a system that can flag signs of hypertension using nothing more than a consumer smartwatch. The Nature paper, titled "Opportunistically Detecting Signs of Hypertension on a Consumer Smartwatch," presents a system called AI-PPG-ACC-HTN. It relies on two sensors that ship in ordinary smartwatches: the photoplethysmography (PPG) optical sensor, which tracks pulse-related changes at the wrist, and the accelerometer.

Crucially, the system is not a blood pressure meter. According to Google, it does not estimate momentary systolic and diastolic values. Instead, it estimates the probability that the wearer is in a sustained hypertensive state by pooling information across several days of passive wear.

How the model is built

The pipeline runs in two stages. In the first, a one-dimensional convolutional network learns general-purpose representations from unlabeled wrist data using a self-supervised contrastive objective. The training corpus is large: 492,347 hours of PPG and accelerometer signals from 2,109 participants, segmented into overlapping 15-second windows.

In the second stage, the learned features are averaged across a participant's wear days. A regularized logistic regression head, fed the aggregated representation together with the wearer's height, then outputs the probability of hypertension.

The validation study

The model was assessed in a prospective, multicenter study of 196 participants across five centers in the United States. None of the participants had known cardiovascular disease, and none were taking antihypertensive medication. Ground truth came from 24-hour ambulatory blood pressure monitoring, compared against seven days of real-world smartwatch wear.

For primary screening, the system achieved 65.8 percent sensitivity, 90 percent specificity, and a positive predictive value of 80.6 percent. An initial office blood pressure measurement in the same cohort recorded lower sensitivity, at 55.3 percent, and a lower positive predictive value of 77.8 percent.

On a subset of 48 participants whose baseline office readings were already elevated, a confirmatory application of the model performed markedly better, reaching 78.4 percent sensitivity, 90.9 percent specificity, and 96.7 percent positive predictive value.

What Google will and will not ship

Google frames the work as early-stage research. The company states that the output is not a diagnosis, is not a substitute for cuff-based measurement, and does not correspond to any existing or planned commercial device feature. That said, the study did shape the Health Guardian blood pressure tracking feature in the Google Health app.

Why it matters

Two things stand out in this work. The first is the reframing of the problem. Rather than chasing cuffless, beat-by-beat blood pressure estimation, a goal that has long resisted wearable devices, the system treats hypertension as a durable physiological state that can be inferred from days of noisy, passively collected sensor data. Averaging across wear days trades immediacy for robustness, and the reported numbers suggest the trade-off is good enough to outperform a single office measurement as a first-line screen.

The second is that passive screening on mass-market wearable hardware is turning into a concrete engineering discipline rather than a research curiosity, complete with large-scale self-supervised pretraining, prospective clinical protocols and explicit error metrics. For teams building on watch sensors, the hard questions are increasingly downstream of the model itself: choosing screening thresholds, validating across diverse populations, and handling the false negatives that any 65.8-percent-sensitivity screen will inevitably produce.