A smartwatch can tell you how much you slept and walked yesterday, but not what an extra half hour of sleep would do for your heart over the next three months. Google Research has published an investigational system that tries to answer that question. It turns everyday wearable data into a health age you can change, then simulates how different habits would move it over 12 weeks.
The first piece is what Google calls a coachable health age. Rather than training a model to guess a person's birth year, it is built from four signals a consumer wearable already measures, resting heart rate, daily steps, sleep duration and weekly Active Zone Minutes, plus optional extras such as VO2max. The difference between that figure and your real age is the number to coach. In the All of Us research programme, covering 22,748 people and more than 68,000 person-years of Fitbit and health-record data, healthy participants whose health age was older than their calendar age were significantly more likely to develop hypertension, heart disease or type 2 diabetes over eight years.
The second piece is a forecasting model called SensorWM, trained on 2.28 million person-weeks of de-identified data from roughly 100,000 Fitbit users. Given two weeks of data, it predicts how 20 physiological markers will change over 12 weeks under different routines, and Google says it beats general time-series models such as TimesFM, which we review. For each person it tries 125 candidate habit combinations, filters out ones that are unrealistic or unsafe, and a language model turns the best into a coaching plan. In a blinded study with five lifestyle coaches and five physicians reviewing 490 paired plans, coaches picked the simulation-based plan as best in 71.4% of evaluations, and physicians rated 99.2% within safe exercise limits.
For anyone already wearing a Fitbit or Pixel Watch, this hints at where Google's health coaching is going: from a dashboard of yesterday's numbers to concrete "if you do this, expect that" advice. A companion study adds blood tests, such as cholesterol ratio and kidney filtration, to profile heart, kidney and metabolic strain separately, and found early kidney decline often appears on its own.
The limits are important. These are research prototypes, not diagnostic tools, and all evaluations were run offline on past data rather than in a trial where people actually followed the plans. Expert preference for a plan is not evidence that it improves health. The cohorts are drawn from Fitbit users who chose to share data, who may not represent everyone.