AI
Google studies passive detection of hypertension signs using consumer smartwatches
Google Research published a multimodal AI system called AI-PPG-ACC-HTN that uses PPG and accelerometer signals from ordinary smartwatches to aggregate data over multiple days and identify sustained hypertension status without cuff calibration. The study says its accuracy, benchmarked against 24-hour ambulatory blood pressure monitoring, approaches that of traditional cuff-based procedures.
Google Research researchers Paolo Di Achille and Ming-Zher Poh Poh described an AI system that uses standard consumer smartwatch sensors to passively detect signs of hypertension. Hypertension affects more than 1000000000 people worldwide, but more than half of adults with the condition are unaware of it; missed diagnosis among people aged 18 to 44 can also delay early intervention and increase long-term cardiovascular risks such as stroke and heart failure.
Existing cuffless wearable devices generally require periodic calibration with a traditional arm cuff and can only track changes relative to baseline readings, making it difficult to independently identify absolute hypertension status. The new system, AI-PPG-ACC-HTN, does not estimate precise systolic and diastolic blood pressure at a given moment. Instead, it aggregates passive physiological signals over multiple days to determine whether an individual has a sustained hypertensive state.
Model training had two stages. First, self-supervised contrastive learning was used to train a one-dimensional deep convolutional neural network on a total of 492,347 hours of unlabeled sensor data from 2,109 consenting participants, allowing the model to learn representations of overlapping 15-second PPG and accelerometer waveforms from the same person and discover features of vascular and cardiac mechanical function. The system then averages and aggregates feature blocks generated from multiple days of wear by a user to reduce the effects of short-term noise such as sudden movement or temperature changes, and combines height as a proxy for arterial path length before a regularized logistic regression classifier outputs a continuous hypertension probability.
The study has been published in Nature and was validated against 24-hour ambulatory blood pressure monitoring, the clinical gold standard. Google said the system’s accuracy in initial screening and confirmatory screening is comparable to traditional cuff-based diagnostic procedures; the findings were also used for the blood-pressure trends feature of Health Guardian in the Google Health app. Available materials did not fully list the sample size or final performance figures for the multicenter clinical study.