Google Research's SensorWM Turns Fitbit Data Into 12-Week Health Coaching Plans

Google Research has published an investigational machine-learning framework that uses wearable data to simulate how changes in sleep, activity, and exercise could affect a person’s health over 12 weeks. The system combines an interpretable health-age score, a model of hypothetical future trajectories, an optimizer, and a language model that converts the selected trajectory into a coaching plan.

The project appears in a coaching paper and a companion CKM age paper. Both studies use retrospective cohorts and offline evaluations. The framework remains research infrastructure, with no released clinical product or evidence yet from prospective trials.

From wearable logs to forecasts

Most wearable products summarize past behavior through daily totals, trends, and alerts. Google’s framework estimates counterfactual trajectories: modeled outcomes under hypothetical routines such as taking more steps, sleeping longer, or completing more cardiovascular exercise. That approach gives a coaching system structured estimates of how each proposed routine could change measured signals over time.

An age score designed to change

Cardiovascular-kidney-metabolic syndrome, or CKM, describes the connected progression of cardiovascular disease, kidney dysfunction, obesity, insulin resistance, and type 2 diabetes. These conditions can develop for years before symptoms appear, while sleep, movement, and cardiovascular fitness influence several of the associated risk factors.

Many biological-age models are trained to predict chronological age, which can make their scores depend heavily on stable traits and obscure the effect of a specific behavior. Coachable Health Age, or CHA, instead expresses a modeled, behavior-linked score in years using signals that can change over time.

The core model uses resting heart rate, daily steps, sleep duration, and weekly Active Zone Minutes, Fitbit’s intensity-weighted measure of physical activity. It can also incorporate sleep-schedule regularity, estimated VO2 max for aerobic fitness, and insulin-resistance measurements when available.

To reduce double counting, the model adjusts for correlations among inputs such as resting heart rate, exercise, and sleep. Its main output is the health-age gap, calculated as CHA minus chronological age. A positive gap indicates that the modeled health profile appears older than the person’s calendar age; a negative gap indicates a younger modeled profile.

Three cohorts test different claims

The validation program examines whether CHA predicts future disease, responds to sustained behavior changes, and aligns with established clinical scores:

• Future risk: In the All of Us program, 22,748 participants contributed more than 68,000 person-years of Fitbit and electronic health-record follow-up. Participants who were healthy at baseline and had a health age above their chronological age were more likely to develop hypertension, cardiovascular disease, or type 2 diabetes during the following eight years.
• Incremental prediction: Adding CHA to age, sex, and body mass index improved disease-risk prediction in the reported analyses, with the largest gains among adults younger than 50.
• Responsiveness: In a longitudinal cohort of 1,464 Fitbit users followed for two years, sustained habit changes coincided with lower CHA scores. This result shows that the metric responds to its behavioral inputs; it does not establish improved clinical outcomes.
• Clinical alignment: In the WEAR-ME study, CHA aligned more closely with a modified American Heart Association Life’s Essential 8 score and a metabolic-syndrome severity score than wearable clocks trained to predict chronological age.

SensorWM maps routines to trajectories

SensorWM was pretrained on 2.28 million person-weeks of de-identified data from about 100,000 consenting Fitbit users. Given two weeks of baseline wearable data and a proposed routine, it forecasts 20 physiological and behavioral markers across the next 12 weeks. The paper reports stronger forecasting performance than general-purpose time-series models such as TimesFM, while retaining patterns associated with diabetes, hypertension, and insulin resistance.

An action-conditioned world model learns how a current state may change after a proposed action. In this system, the state consists of wearable-derived measurements, while the actions specify targets for steps, sleep, and cardiovascular activity. Its forecasts estimate patterns in observational data and cannot establish that a prescribed behavior will cause the predicted result.

The planning pipeline uses those forecasts in three stages:

1. It simulates 125 candidate combinations of steps, sleep, and cardiovascular activity for each person, along with a status-quo trajectory.
2. It ranks the candidates by projected health-age reduction, the person’s historical ability to change each habit, and predefined clinical safety bounds.
3. It gives the selected trajectory to a language model, which writes a progressive 12-week plan.

The optimization step adjusts ambition to historical flexibility. The paper gives examples of a projected 2.3-year reduction for a person with greater flexibility and a 1.3-year reduction for someone whose past data suggests tighter constraints. These values are model outputs, not guaranteed reductions in disease risk or biological aging.

Can simulated futures improve an LLM plan?

The evaluation held the underlying language model constant and changed the context supplied to it. One version received raw wearable history, while the other received SensorWM’s counterfactual trajectories. Five certified lifestyle coaches and five board-certified physicians reviewed 49 participant profiles spanning five behavioral personas, producing 490 paired reviews. Automated evaluations covered the full WEAR-ME cohort of 2,780 participants.

• Lifestyle coaches selected the simulation-grounded plan as best overall in 71.4% of evaluations, with p < 0.001.
• Coaches rated the simulation-grounded plan more likely to support long-term adherence in 71.0% of evaluations.
• Physicians judged 99.2% of the counterfactual plans to be within safe exercise-intensity limits and 97.6% of their physiological claims to be accurate.
• Reviewers associated the counterfactual plans with a stronger sense of agency in 97% of ratings, compared with 59% for plans based on raw histories.

Holding the language model constant isolates the value of the modeled context: plans generated from structured simulations received higher expert ratings than plans generated from raw time-series data. The study measured reviewer preferences and automated criteria, leaving actual adherence, safety, and health outcomes for prospective evaluation.

Blood markers separate the CKM domains

The companion study extends the wearable model with routine blood measurements from 819 WEAR-ME participants. It combines resting heart rate, steps, Active Zone Minutes, and sleep with three clinical markers: cholesterol-to-HDL ratio for cardiovascular risk, estimated glomerular filtration rate for kidney function, and HOMA-IR for insulin resistance.

Kidney decline sometimes appeared independently of metabolic and cardiovascular strain, supporting separate estimates for the three CKM domains. The resulting CKM health-age gap correlated with the modified Life’s Essential 8 score at r = -0.58 and with metabolic-syndrome severity at r = 0.65. The negative first correlation reflects the scoring direction: higher Life’s Essential 8 scores indicate better health.

A reusable pattern with deployment gaps

For engineering teams, the framework offers a modular design: an inspectable metric defines the objective, a time-series model generates candidate futures, an optimizer applies feasibility and safety constraints, and a language model handles explanation and planning. Separating those responsibilities makes each stage easier to test than an agent that receives raw logs and produces recommendations in one step.

The published evidence leaves several deployment questions open:

• Prospective efficacy: A trial must test whether people follow the plans and achieve the predicted outcomes.
• Causal validity: Observational forecasts require validation before they can support claims about the effects of an intervention.
• Generalization: Performance needs measurement across devices, populations, missing-data patterns, and changing sensor firmware.
• Safety: Model-based bounds and physician ratings cannot replace clinical monitoring, escalation rules, and contraindication checks.
• Operations: A deployed system would need calibration monitoring, drift detection, privacy controls, audit trails, and appropriate regulatory review.

The research provides concrete evidence that simulated, structured context can improve expert assessments of LLM-generated health plans. Clinical benefit remains an open question until prospective studies test the complete loop from recommendation to behavior change and measured outcome.