Google's AI-PPG-ACC-HTN Catches More Hypertension Cases Than Doctor Visits

Google researchers have developed a model that uses a consumer smartwatch’s optical pulse sensor, accelerometer, and several days of ordinary wear to flag likely sustained hypertension without prior cuff calibration. In a prospective clinical study, the system recorded higher screening sensitivity than office measurements while maintaining 90% specificity, according to the Google Research report.

The system, called AI-PPG-ACC-HTN, produces a hypertension probability score rather than a systolic or diastolic reading. The Nature paper validates that score against 24-hour ambulatory blood pressure monitoring, which measures blood pressure repeatedly during normal daytime activity and sleep. Google says the broader findings informed the Blood Pressure Trends component of Health Guardian in the Google Health app, although the paper does not state that the published model powers that feature.

Why annual cuffs miss cases

Hypertension affects more than one billion people worldwide, and over half of affected adults do not know they have it. Symptoms are uncommon, so detection often depends on an incidental cuff reading. Delayed diagnosis gives sustained high blood pressure more time to damage blood vessels and increase the risk of stroke, heart failure, and kidney disease.

Most commercial cuffless systems require periodic calibration against an arm cuff. They can then estimate changes from that baseline, but they cannot independently screen someone who has never measured their blood pressure. Google’s model targets that gap by estimating an underlying hypertensive state from repeated sensor observations.

A state estimate from ordinary wear

Blood pressure at a single moment changes with stress, caffeine, posture, exercise, temperature, and sleep. The model reduces that variability by combining short signal segments collected over multiple days, then classifying whether the resulting pattern is consistent with persistent hypertension.

Modern smartwatches already collect the two signal streams used by the model, although developer access to raw waveforms depends on the device and platform:

• Photoplethysmography (PPG): An optical sensor measures reflected light to track changes in blood volume at the wrist. The resulting waveform captures the shape and timing of each pulse.
• Accelerometry (ACC): A motion sensor identifies periods of stillness and captures faint wrist vibrations associated with the heart ejecting blood into the aorta. These vibrations form a ballistocardiogram.

Two stages produce one risk score

Training proceeds in two stages, using self-supervision to learn from large volumes of unlabeled sensor data before fitting a simpler hypertension classifier:

1. Waveform encoding: A one-dimensional convolutional neural network was pretrained with contrastive learning on 492,347 hours of data from 2,109 consenting participants. It learned to place time-aligned, 15-second PPG and accelerometer segments from the same person near one another in a latent representation while separating samples from different people. This process captures recurring person-level signal patterns without requiring a blood pressure label for every segment.
2. Person-level aggregation: The system averages encoded 15-second segments across multiple days, reducing the influence of motion, temperature, and other transient conditions. It combines those features with the wearer’s height, used as a proxy for arterial path length, and passes them to a regularized logistic regression classifier that returns a continuous hypertension probability.

Height matters because pulse waves travel along the arterial tree, and body size affects the path length and timing relationships visible in PPG and ballistocardiogram signals. The published approach therefore depends on physiological signals plus one demographic measurement, not sensor data alone.

Results against 24-hour monitoring

A prospective study enrolled 196 consenting participants across five US sites. Participants had no known cardiovascular disease and were not taking blood pressure medication. Each person wore the smartwatch and completed 24-hour ambulatory monitoring, the reference method used to determine hypertension status.

Sensitivity measures how often a test catches participants who have hypertension. Specificity measures how often it clears those who do not, while positive predictive value, or PPV, measures how often a positive result is correct. PPV varies with the prevalence of hypertension in the tested population.

For initial screening, the smartwatch model caught a larger share of hypertension cases than office measurement in this cohort while correctly excluding 90% of participants without the condition. The reported point estimates favor the model, although point estimates alone do not establish whether each difference was statistically significant.

Confirmatory analysis covered 48 participants whose initial office reading was elevated. The model reached 78.4% sensitivity, 90.9% specificity, and 96.7% PPV. It also performed competitively with multiday home cuff monitoring, a common out-of-clinic option for confirming hypertension.

Physiology inside the score

Because a contrastively trained convolutional network does not directly explain its predictions, the researchers averaged waveforms from participants at the highest and lowest ends of the predicted-risk distribution. High-risk participants showed three recurring patterns:

1. Narrower pulse width, which is consistent with arteries that expand less during each heartbeat.
2. A slower systolic upstroke and lower d/a ratio in the second derivative of the PPG waveform, reflecting changes in forward and reflected pressure waves.
3. Deeper K waves and wider IJK complexes in the accelerometer-derived ballistocardiogram, indicating differences in the mechanical signal generated by cardiac ejection.

These waveform differences align with established signs of vascular stiffening and support the model’s physiological plausibility. The analysis remains post hoc: averaged waveforms cannot prove which features drove an individual prediction or rule out demographic and behavioral confounders.

Where the evidence stops

The study supports passive risk screening in a defined research cohort. Several constraints limit broader conclusions:

• No direct blood pressure reading: The model returns a probability score and cannot report current systolic or diastolic pressure.
• Clinical confirmation remains necessary: A positive result requires validated cuff measurements and assessment by a clinician.
• Narrow study population: Validation excluded people with known cardiovascular disease and those taking blood pressure medication.
• Limited external validation: The study involved 196 participants at five US sites, leaving performance across larger populations, devices, skin characteristics, and health conditions unresolved.
• No public implementation: Google has not released the training code, model weights, decision threshold, or a device integration package.
• Exploratory status: The authors describe the methods as research and do not present the paper as an announcement of a commercial diagnostic product.

From risk score to clinical follow-up

A practical deployment could monitor eligible users over several days, surface an elevated-risk result, and direct them toward validated cuff testing. That workflow could reach people who rarely receive office screening, especially younger adults whose hypertension might otherwise remain undetected for years.

Production use would require regulatory clearance, transparent eligibility rules, privacy protections for continuous sensor data, and validation on each supported device. Developers would also need safeguards for missing data, poor sensor contact, motion artifacts, firmware changes, and population drift. The research shows that ordinary smartwatch signals can support sustained-hypertension screening, while diagnosis and treatment decisions remain grounded in validated blood pressure measurements.