正文 · AI 翻译
IT Home reported on November 7 that Google published a blog post on November 5 local time introducing its latest research findings published in Nature.
Google introduced a multimodal AI system called AI-PPG-ACC-HTN. Without cuff calibration, the system can passively detect signs of hypertension using an ordinary consumer smartwatch.
According to Google, conventional approaches try to estimate individual, momentary systolic and diastolic blood pressure readings, which are highly susceptible to transient environmental noise. The new system takes a different approach: by aggregating passive physiological features collected continuously over multiple days, it focuses on identifying an individual's underlying, persistent hypertension.
Self-supervised pretraining: A 1D-CNN was trained on more than 490,000 hours of unlabeled data, using contrastive learning to enable the model to discover complex cardiovascular physiological features on its own.
Multi-day aggregation and classification: Feature vectors from 15-second intervals over multiple days are averaged to smooth transient noise and combined with height, which represents arterial path length. A logistic regression model then outputs a probability score for persistent hypertension.
To assess its performance, the research team conducted a prospective, multicenter clinical trial at 5 clinical centers across the United States. The trial enrolled 196 participants with no history of cardiovascular disease who were not taking blood pressure medication, and used the clinical gold standard, 24-hour ambulatory blood pressure monitoring, as its reference. The results showed:
Initial screening: Specificity was 90.0%, positive predictive value (PPV) was 80.6%, and sensitivity was 65.8% (better than the 55.3% sensitivity and 77.8% PPV of a single initial in-office screening).
Confirmatory testing (for people whose initial screening results were elevated): PPV reached 96.7%, specificity reached 90.9%, and sensitivity reached 78.4%, outperforming repeat in-office measurements and comparable to continuous home cuff monitoring that requires several days of wear.
IT Home has learned that these research findings have provided core technical support for the blood pressure trends “Health Guardian” feature in Google's health app.