Bold claim: Wearable data can reveal real medical truths, not just daily vibes. But the real story is how close we are to bridging everyday signals with medical ground truth.
Empirical Health unveils a new foundation model that predicts blood test results and medical diagnoses from wearable data, a development showcased at the Timeseries for Health workshop during NeurIPS 2025—the premier AI conference.
Conventional wisdom holds that blood tests are the gold standard and wearables offer only everyday health hints. This research pushes beyond that divide by translating real-time wearable signals into medically meaningful predictions.
Key details from the study:
- The model, named JETS (joint embedding for timeseries), achieved 87% accuracy in detecting high blood pressure. It also identified atrial flutter (70% accuracy), myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) (81% accuracy), and sick sinus syndrome (87% accuracy).
- JETS was trained on an extraordinary 3 million person-days of wearable data from diverse devices, including Apple Watch, Fitbit, Pixel Watch, and Samsung Galaxy Watch, spanning 63 independent timeseries. The architecture draws on JEPA, a design proposed by Yann LeCun.
This work touches several important themes:
- Bridging lab tests and wearables: While many wearables now offer lab-like features, few link wearable signals directly to blood test results. This study is among the first to demonstrate such a connection using AI.
- AI beyond large language models: LLMs have shown impressive progress in text, but there’s a growing belief that the next leap will hinge on physiological ground truth from wearables. This research illustrates one route toward building health-focused intelligence grounded in real-world physiology.
- Extracting meaningful signals: The model uses a twin-encoder setup where one stream sees the full sequence and the other sees roughly 30% of it. The system learns to align their latent representations without reconstructing raw signals, prioritizing the extraction of meaningful health signals over incidental surface details.
About Empirical Health
Empirical Health aims to prevent one million heart attacks. The program starts with 100+ biomarkers, models an individual’s risk for heart disease, and then crafts a personalized prevention plan.
The company was founded by a physician with experience at Kaiser and a former Google ML lead, and it completed Y Combinator S23. To date, more than 100,000 people have used Empirical Health’s services.
Questions to consider: If wearable-derived predictions can approach or even rival traditional testing for certain conditions, how should clinicians integrate these tools into everyday care? What safeguards are needed to manage false positives and ensure patient trust? Do you believe this represents a practical path toward medical superintelligence, or is it a promising but early step with limits to scale and generalizability?
