An investigational research framework describes how consumer wearables, an interpretable coachable health-age metric (CHA), and an action-conditioned sensor world model (SensorWM) can be combined to simulate personalized 12-week “what-if” habit trajectories and produce clinician-vetted coaching plans. The system was trained and validated on millions of person-weeks of Fitbit data and can be extended, when routine blood tests are available, to profile multi-organ cardiovascular–kidney–metabolic (CKM) health.
Why this matters
CKM syndrome—interconnected cardiovascular, kidney and metabolic disease—often develops silently over years and is influenced by daily behaviors such as movement, sleep and cardiovascular load. Moving from passive wearable tracking to actionable, personalized guidance requires metrics and models that (1) depend on modifiable behaviors, (2) forecast how physiology adapts over weeks, and (3) translate simulations into feasible, safe coaching plans.
The coachable health age (CHA)
CHA is an interpretable health-age metric built solely from modifiable, passively measured daily behaviors and physiology. The core signals are resting heart rate, daily step count, sleep duration and weekly Active Zone Minutes; optional inputs include VO2max, sleep regularity and insulin resistance when available. CHA computes a health age delta—the difference between CHA and chronological age—where a negative delta indicates a physiologically younger profile and a positive delta indicates accelerated aging.
Validation highlights:
- All of Us research program (n = 22,748): across more than 68,000 person-years of consenting Fitbit and electronic health record follow-up, participants whose CHA exceeded their chronological age at baseline were significantly more likely to develop hypertension, cardiovascular disease or type 2 diabetes over the next eight years. Adding CHA to standard clinical predictors (age, sex, BMI) significantly improved disease risk prediction, notably for adults under 50.
- Longitudinal Fitbit cohort (n = 1,464): tracking users across consecutive 3-month windows over two years showed that users with sustained habit changes experienced decreases in CHA.
- WEAR-ME study: unlike wearable clocks trained to predict chronological age, CHA aligned better with the modified American Heart Association Life's Essential 8 cardiovascular score, the metabolic syndrome severity score, and participants’ self-reported health.
Sensor World Model (SensorWM) and counterfactual trajectories
CHA indicates current standing; converting that into personalized plans requires forecasting how physiology will change if habits change. SensorWM is an action-conditioned world model pre-trained on 2.28 million person-weeks of de-identified Fitbit data from roughly 100,000 consenting users. Given two weeks of baseline wearable data, SensorWM forecasts how 20 physiological markers will evolve over 12 weeks under different step, sleep and cardio routines, outperforming general-purpose time-series foundation models such as TimesFM while preserving clinical patterns linked to diabetes, hypertension and insulin resistance.
The system pipeline:
- Simulate: SensorWM rolls out 125 candidate 12-week habit combinations per person and compares them to a status-quo projection.
- Filter for realism and safety: a feasibility-constrained optimizer balances projected health-age reductions against each person’s historical habit flexibility and clinical safety bounds, selecting more ambitious targets for those with high flexibility (e.g., −2.3 years) and gentler ramps for more constrained individuals (e.g., −1.3 years).
- Coach: a language model agent converts the top-ranked counterfactual trajectories into supportive, progressive 12-week coaching plans.
Expert evaluation with coaches and clinicians
To test whether grounding a language model in world-model simulations improves guidance, the team ran a blinded head-to-head study with five certified lifestyle coaches and five board-certified primary care and lifestyle medicine physicians reviewing 49 participant profiles (490 paired reviews total), plus automated evaluations across the full WEAR-ME cohort (n = 2,780).
- Strong expert preference: lifestyle coaches selected the counterfactual-grounded plan as best overall in 71.4% of evaluations (p < 0.001) and rated it more likely to drive long-term adherence in 71.0% of evaluations (p < 0.001).
- Clinician-verified safety: physicians rated 99.2% of counterfactual plans within safe exercise intensity limits and judged 97.6% of their physiological claims to be accurate. Coaches and clinicians reported that showing a concrete contrast between status-quo trajectories and achievable “what-if” options provided clear motivation and enhanced perceived agency.
Extending to CKM health using routine blood biomarkers
Because the framework is modular, it can incorporate routine lab tests. In a companion study evaluated in 819 WEAR-ME participants, the multimodal CKM health age model combined continuous wearable signals (resting heart rate, steps, active zone minutes, sleep) with three routine blood biomarkers: cholesterol-to-HDL ratio (cardiovascular), eGFR (kidney filtration) and HOMA-IR (metabolic insulin resistance). By computing normalized deviance scores for each marker relative to healthy clinical thresholds, the CKM model showed that multi-organ aging does not follow a single template: metabolic and cardiovascular strain often correlate, but early kidney decline frequently appears independently. The CKM health age gap correlated closely with objective clinical scores (modified Life's Essential 8 r = −0.58; metabolic syndrome severity r = 0.65) and with participants’ day-to-day sense of health.
Limitations and next steps
These evaluations were retrospective and offline; the models are investigational research prototypes, not clinical diagnostic tools. Prospective trials are needed to determine whether interactive counterfactual coaching improves long-term adherence and clinical outcomes.
Contributors
This work is joint research across Google Research and partner teams. Contributors named by the authors include Zeinab Esmaeilpour, Simon A. Lee, Menglian Zhou, Marius Guerard, Sohrab Sidhu, Alexandru Solot, Hamid Palangi, Pramod Rudrapatna, Ming-Zher Poh, Max A. Xu, Daniel McDuff, Dimitrios Antos, Yuzhe Yang, Jeffrey N. Chiang (from UCLA), Anthony Z. Faranesh, Conor Heneghan, Lindsey Sunden, Mark Malhotra, Cathy Speed, Javier L. Prieto, Shwetak Patel, A. Ali Heydari, and Ahmed A. Metwally.
Conclusion
The proposed combination of CHA, SensorWM and a language-model coach demonstrates a practical route from passive wearable monitoring to personalized, safety-constrained habit plans and short-term forecasts. When augmented with routine lab biomarkers, the framework can profile organ-specific CKM strain. Further prospective research is required to confirm clinical benefit and long-term behavior change.
