A change in approach compared with instantaneous measurement
Google has presented an artificial intelligence system called AI-PPG-ACC-HTN that, according to the company, can detect signals compatible with persistent hypertension using an ordinary consumer smartwatch, without the need for calibration with a cuff. The research was published in the journal Nature and described in a company post on November 5, as reported by IT之家.
The starting point is a critique of traditional methods: instead of attempting to estimate a single systolic and diastolic pressure reading, exposed to momentary environmental noise, the system aggregates physiological features collected passively over multiple days. The stated goal is not to provide a point reading, but to estimate the probability that a person has an ongoing hypertensive state over time.
From a technical standpoint, the model relies on self-supervised pretraining of a convolutional neural network with more than 490,000 hours of unlabeled data, followed by a multi-day aggregation stage: features extracted from 15-second windows are averaged to attenuate noise, combined with the person's height, used as an indicator of arterial path length, and passed to a logistic regression model that returns a probability score.
The trial numbers
The evaluation was conducted in a prospective, multicenter study carried out at five clinical centers in the United States, involving 196 participants with no cardiovascular history and not treated with antihypertensive drugs. The comparison reference was 24-hour ambulatory monitoring, designated as the clinical reference standard.
In the initial screening phase, the system showed a specificity of 90.0%, a positive predictive value of 80.6%, and a sensitivity of 65.8%. According to the source, these results exceed those of a single outpatient visit, which would have a sensitivity of 55.3% and a positive predictive value of 77.8%.
In the confirmation phase, applied to people with high screening scores, the positive predictive value rose to 96.7%, with specificity of 90.9% and sensitivity of 78.4%. Google maintains that these data are superior to repeating the measurement in an outpatient setting and comparable to continuous home cuff monitoring, which, however, requires several days of use.
What remains to be clarified
The research points to a promising direction for passive screening, but the sample of 196 people and the controlled setting of the study do not allow conclusions to be drawn about the system's validity in larger populations or under everyday-use conditions. The material does not specify which smartwatches were used or whether the system was tested on devices from different manufacturers.
Another element to assess is the transition from research to product: according to IT之家, the technology already underpins the blood pressure trend monitoring feature in Google's Health app. However, it is not indicated whether and when this feature will be available to users, in which countries, with which compatible devices, and with what regulatory approvals, if any.
No data are reported on false negatives in specific subgroups, on variability across ethnic populations, or on how the system performs in the presence of conditions that alter the optical signal or the motion detected by the accelerometer. The sensitivity and specificity percentages remain statements by the study and the company, to be verified in clinical practice.
