Google Research scientists created a machine learning framework to estimate a metric called ‘coachable health age’ from wearable signals. The team built an action-conditioned world model named SensorWM using Fitbit data from millions of person-weeks. The system simulates 12-week ‘what-if’ habit trajectories to build personalized health coaching plans.
The researchers trained SensorWM on 2.28 million person-weeks of Fitbit data and tested the framework on 18,423 participants. The study found that individuals with a younger health age had higher rates of disease-free survival. Google researchers expanded the model to profile cardiovascular-kidney-metabolic syndrome using blood biomarkers alongside wearable signals.
Blinded evaluations by coaches and clinicians showed that plans grounded in counterfactual simulations were preferred nearly three-to-one. The research uses retrospective cohorts and describes the models as investigational prototypes rather than clinical diagnostics. This work offers a new way for developers to build health interventions that predict long-term outcomes from daily activity.
