Google unveils watch-based AI system for non-invasive hypertension screening in nature study

Google has published research in Nature detailing a novel, multi-modal AI system named AI-PPG-ACC-HTN. This technology allows for the passive detection of hypertension signs using standard consumer smartwatches, eliminating the need for traditional cuff calibration. Unlike conventional methods that rely on single, easily perturbed readings, the new system aggregates physiological data collected over multiple days. It utilizes self-supervised pre-training on over 490,000 hours of unlabeled data via a 1D-CNN, followed by averaging 15-second feature vectors across days to smooth transient noise. The system then outputs a sustained hypertension probability score using logistic regression, incorporating height to account for arterial path length. Clinical trials involving 196 subjects without prior cardiovascular history or antihypertensive medication, benchmarked against 24-hour dynamic blood pressure monitoring, yielded strong results. Initial screening achieved a sensitivity of 65.8% and a specificity of 90.0%, surpassing the 55.3% sensitivity and 77.8% PPV of standard office screenings. For confirmation in elevated individuals, the system reached a PPV of 96.7% and a sensitivity of 78.4%, comparable to multi-day home cuff monitoring. This technology is slated to underpin the health monitoring features within Google Health applications.