AI that learns health from everyday life
Most health data is collected in a clinic, on the day someone already feels sick. Our models learn from the signals people generate between visits: wearables, short check-ins, and, with permission, health records. Life sciences teams use the same consented, longitudinal data for their own research and AI.
Inside our models
Our research team builds health models the way large language models are built: pretrain once on lots of unlabeled everyday data, then fine-tune for each condition.
Foundation model
Pretrained on everyday wearable data
- Learns from 28-day windows of daily data by filling in hidden values, using ~80 million rows in proof-of-concept runs
- Beats training from scratch at 10,000 and 20,000 labels, where most real conditions sit
- Encoding each data type separately, then fusing, lifted hypertension average precision from 44% to ~60% on wearables and demographics alone
Data value
Measuring what each source adds
- We add demographics, environment, wearables, surveys, and EHR one layer at a time to see where each helps
- Eight short questions improved prediction more than wearables or demographics on their own
- EHR helps most for chronic conditions: hypertension average precision rose from ~60% to ~83%.
Smart onboarding
Eight questions, 96% of a full survey
- A greedy algorithm picks the question that adds the most new coverage across all target conditions
- Mental-health questions carry the most signal; skin-condition questions the least
- When two wordings perform the same, we choose the one that's easiest to answer
Condition spotlights
Some conditions leave a much clearer trace in everyday data than others. Here is where our models stand today.
Autoimmune
0.86
AUC for rheumatoid arthritis flares
People in the top 5% of predicted risk are about 10 times more likely to have the condition than the general population.
Respiratory
0.75
AUC for COPD
The model reliably detects that something has changed in the wearable data; rating severity near the clinical threshold remains the hard part, and we say so.
Cardiometabolic & under-diagnosed
0.85+
AUC for hypertension
Using wearables, demographics, EHR, and 10 targeted questions. For under-diagnosed conditions like sleep apnea, moderate precision can still find many people who would benefit from screening.
What our models learn from
Wearables
Steps, heart rate, and sleep.
Check-ins & PROs
Short surveys and patient-reported outcomes.
Demographics
Age, sex, and ethnicity only; social attributes are excluded.
Health records
Electronic health records, with permission.
In development
Environment and weather, labs, richer EHR history, and JEPA-style pretraining across sensor and clinical data.
How we keep results honest
Health AI can easily produce numbers that look impressive but don't hold up. Much of our method is designed to prevent that.
Label leakage
No answers hidden in the inputs
- A hypertension model that sees ICD-10 I10 is reading the answer. We remove whole families of codes.
- We look for less obvious proxies too, such as a diabetes complication standing in for the diabetes label.
- Leakage rules are set at pretraining, so they carry into every condition model.
Metrics
Precision read against prevalence
- A heart-attack model with an AUC of 0.86 had about 3% precision: five times the baseline, but far less than the AUC suggests
- We compare models on average precision, always on the same held-out test set
- We report real decision thresholds: hypertension precision is 83% at a 95% threshold and 31% at 50%
Validation
Tests that match real use
- User splits test people the model has never seen; temporal splits test future episodes in people it knows
- Split design often show that some early “signals” are artifacts, not real predictors
- We report against the label source: self-reports and medical records agree only about two-thirds of the time
What life sciences teams build with it
From cohort selection to model validation, the same data and models support three kinds of work.
Clinical development
Find and enrich cohorts
- Risk scoring across ~90 conditions from wearable, survey, and demographic data
- Recontact flagged participants to confirm eligibility before a study starts
R&D and digital measures
Develop digital measures
- Detect change in daily wearable signals between clinic visits
- Pair passive signals with patient-reported anchors on the same daily timeline
Data science & AI
Train and validate models
- Fine-tune on small labeled sets, where pretraining on everyday data helps most
- Check predictions prospectively by asking participants what happened
What makes the data AI-ready
The questions data science teams ask about any dataset, answered up front.
Data science in action
Peer-reviewed research and sponsor programs built on everyday health data.
Eli Lilly · Neurology
Digital signals of cognitive impairment
Wearables, apps, and sleep sensors used to develop new endpoints and find digital signals of cognitive impairment.
Research · COVID-19
An early warning algorithm for COVID-19
Smartphone and wearable data showed potential for individual-level early warnings of infection.
Johnson & Johnson · Cardiology
Earlier AFib detection to reduce stroke risk
Support for the multi-year, observational Heartline Study of iPhone and Apple Watch features.
Longitudinal cohorts across key conditions
Engaged, recontactable cohorts with time-aligned wearable, survey, and clinical data across 100+ conditions.
VACCINES & INFECTIOUS DISEASEFlu and respiratory illness, symptom follow-up, and vaccination journeys
Discuss an AI-ready dataset
Tell us the model, measure, or cohort you're working on. Our data scientists will show you which signals, labels, and study designs fit, and how we'd validate the results.