Empirical Health Trains Wearable Foundation Model with 87% Accuracy at Detecting High Blood Pressure

Our health is determined by the sum of our daily habits, but measuring the impact of any specific action is difficult. Blood tests are the gold standard in medicine but are only measured every few months; consumer wearables measure daily health signals, but are not medically actionable.

What if we could bridge the gap by predicting medical ground truth from realtime wearable signals?

Today, Empirical Health is announcing a research study on a new foundation model that predicts blood test results and diagnoses from wearable data. The research was accepted at the Timeseries for Health workshop at NeurIPS, the top artificial intelligence conference.

Empirical’s model, JETS, was able to detect high blood pressure with 87% accuracy, as well atrial flutter (70% accuracy), ME/CFS (81% accuracy), and sick sinus syndrome (87% accuracy). JETS (joint embedding for timeseries) was trained on on 3 million person-days of wearable data (including Apple Watch, Fitbit, Pixel Watch, and Samsung Galaxy Watch) with 63 independent timeseries. It uses the JEPA architecture proposed by Yann LeCun, formerly Meta’s Chief Scientist.

This touches several themes:

• Blood testing meets wearables. Many wearables recently launched lab testing, but don’t attempt to link your wearable and blood test data. This is one of the first published research studies that shows how to link the two using artificial intelligence.
• AI beyond LLMs. Large language models have been wildly successful, but we’re out of text to train on–there is no second internet. The next frontier in AI is physiological ground truth, like those generated by wearables. This study shows one path toward using physiological ground truth to create health superintelligence.
• Extracting meaningful signal from wearables. The system trains twin encoders where one sees the full sequence and the other sees only ~30%, learning to align their latent representations without reconstructing raw signals. This encourages learning meaning, not surface details of wearable data.