Personalized health coaching via coachable health age and counterfactual wearable trajectories
Organizations: Google Research · University of California, Los Angeles
Abstract
Personalized health coaching requires moving beyond retrospective wearable summaries toward quantitative, actionable behavioral prescriptions. We introduce Coachable Health Age (CHA), an interpretable health-age metric built from modifiable physiological and behavioral signals passively captured by consumer wearables. Across the WEAR-ME (N=2,780), All of Us (N=22,748), and longitudinal Fitbit (N=1,464) cohorts, CHA aligns with established cardiometabolic health constructs and improves prognostic discrimination for incident cardiovascular and metabolic disease beyond age, sex, and body mass index. To enable prospective behavioral optimization, we developed Sensor World Model (SensorWM), an action-conditioned wearable world model pretrained on over 2.2 million person-weeks of longitudinal sensor data from approximately 100,000 participants. SensorWM accurately forecasts multi-week physiological responses to candidate lifestyle changes (), outperforming generic time-series foundation models while preserving downstream clinical signal for diabetes, hypertension, and other health outcomes. Finally, we integrate CHA and SensorWM into an agentic counterfactual coaching system to synthesize personalized 12-week coaching plans. In blinded paired reviews, certified lifestyle coaches preferred counterfactual-grounded plans over plans grounded in wearable history alone for overall quality (71.4%, P < 0.001) and anticipated adherence (71.0%, P < 0.001), and board-certified clinicians judged 99.2% of counterfactual-grounded plans to meet clinical safety criteria (vs. 97.6% of comparison plans). Automated evaluations across the full cohort (N=2,780) also favored counterfactual-grounded plans. By closing the loop between longitudinal tracking, trajectory prediction, and counterfactual coaching, our framework provides a scalable path to systematically improve cardiometabolic health.
Figures & tables
| Model & Covariate | Log-HR ( ) | HR (95% CI) | -stat | -value |
| A. Lagged Time-Varying Cox Model ( ; intervals; events) | ||||
| Lagged Health Age Delta ( , +1 yr) | 1.04 (1.02–1.05) | |||
| Baseline chronological age (+1 yr) | 1.03 (1.02–1.03) | |||
| Baseline BMI (+1 ) | 1.03 (1.03–1.04) | |||
| Sex: Male (vs. Female) | 1.09 (0.94–1.27) | |||
| Sex: Other / Prefer not to answer / Skipped | 0.96 (0.59–1.58) | |||
| Cohort ( ; Events) | Baseline Clinical | + Coachable | C-index [95% CI] | -value |
| (Age + Sex + BMI) | Health Age | |||
| Full Cohort ( ; ) | 0.6485 [0.6340–0.6632] | 0.6541 [0.6390–0.6685] | +0.0056 [+0.0018, +0.0096] | |
| 5-Fold CV ( ; ) | 0.6475 [0.6330–0.6615] | 0.6529 [0.6390–0.6668] | +0.0053 [+0.0012, +0.0098] | |
| Adults yr ( ; ) | 0.6648 [0.6420–0.6910] | 0.6762 [0.6525–0.7010] | +0.0114 [+0.0048, +0.0181] |
| Evaluation Level | Guiding Question | Assessment Dimensions |
| L1. Reaction | Will the user engage? | 1a. Clarity and readability |
| 1b. Tone and motivation | ||
| 1c. Personal relevance | ||
| L2. Learning | Does the user understand? | 2a. Data comprehension |
| 2b. Causal understanding | ||
| 2c. Insight and stakes (“why now”) |
| Dim | Rubric dimension | CF | Generic |
| 1a | Clarity & readability | 5.00 | 5.00 |
| 1b | Tone & motivation | 5.00 | 4.96 |
| 1c | Personal relevance | 4.99 | 4.40 |
| 2a | Data comprehension | 4.99 | 4.22 |
| 2b | Causal understanding | 4.98 | 3.76 |
| 2c | Insight / stakes (“why now”) | 4.98 | 2.86 |
| Contrast | W / T / L | Preferred | |||
| Counterfactual Generic | 1.86 | 2,570 / 3 / 40 | 98.2% |
Appendix figures & tables36 assets
Supplementary material from the paper’s appendix.
Appendix
| Core | Knot / Slope | Reference | Unadjusted | WEAR-ME | Adjusted |
| Domain | Parameter | ( ) | Weight ( ) | ||
| Resting heart rate | Knot at | ||||
| Knot at | |||||
| Daily steps | Per (cap ) | ||||
| Sleep duration | Per hour | ||||
| Per hour |
| Modeling Type | Predictable from Wearables | Explainable | Feature Flexibility | Engageable | Coachable | Movable | Grounded |
| Chronological Age | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Predicted Chronological Age (SensorFM or PPG-FM) | ✓ | ✓ | ✗ | ||||
| PhenoAge | ✗ | ✓ | ✓ | ✓ | |||
| LLM-generated Health Age | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Coachable Health Age (wearable features) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Input Configuration | Temperature ( ) | Average SD (years) | Single-Run Consistency ( ) | Average Consistency ( ) |
| Identical Metrics as Health Age | 0.1 | 2.72 | 0.54 | 0.78 |
| 0.5 | 3.54 | 0.09 | 0.23 | |
| 1.0 | 3.20 | 0.11 | 0.28 | |
| All Raw Metrics | 0.1 | 1.35 | 0.54 | 0.78 |
| 0.5 | 2.06 | 0.30 | 0.57 | |
| 1.0 | 3.37 | 0.20 | 0.42 |
| Comparison | Raw Signals (adj. ) | Curated Metrics (adj. ) |
| vs | 0.0751 | 0.0039 |
| vs | 0.0689 | 0.0024 |
| vs | 1.0000 |
| Input Configuration | Temp ( ) | Modified LE8 ( ) | MSSS ( ) |
| All Available Signals | 0.1 | -0.110 | 0.063 |
| 0.5 | -0.103 | 0.061 | |
| 1.0 | -0.105 | 0.053 | |
| Curated Metrics as Health Age | 0.1 | -0.214 | 0.122 |
| 0.5 | -0.189 | 0.099 | |
| 1.0 | -0.192 | 0.123 |
| Feature | Group | Description | Map |
|---|---|---|---|
| DEM Demographics | |||
| Age | DEM | Chronological age in years. | State |
| Gender | DEM | Self-reported sex or gender. | State |
| BMI | DEM | Body mass index. | State |
| CRD Cardiovascular | |||
| Heart Rate Variability | CRD | Beat-to-beat variability in heart rhythm. | State |
| Persona ID & Clinical Archetype | Selection Criteria (Steps, Sleep, BMI, HA) | Full WEAR-ME Sub-Cohort ( , % of 2,780) | Sub-Cohort Means (Age, Steps, Sleep, BMI, RHR, HA) | Evaluated Sample ( Valid Cases; PIDs) | Evaluated Sample Baseline (Mean SD) | Primary Counterfactual Coaching Test |
|---|---|---|---|---|---|---|
| P1: Optimal Baseline “The High-Vitality Longevity Pacer” | Steps: k/day Sleep: 7.0–9.0 h BMI: 18.5–24.9 HA: yr | (2.81%) | Age: 53.0 yrs Steps: 13,775 Sleep: 7.42 h BMI: 22.9 kg/m 2 RHR: 60.1 bpm HA: yrs | 12611 , 72404 , 58662 , 55814 , 56084 , 55957 , 71670 , 72398 , 71845 , 56638 | Age: yrs Female: 40.0% (4/10) BMI: kg/m 2 Steps: Sleep: h RHR: bpm HA: yrs HA: yrs | Ceiling Effects & Maintenance: Preventing overtraining and framing vitality preservation against age-related drift. |
| P2: Multi-Risk Challenger “The Sedentary Fatigued Metabolic Seeker” | Steps: k/day Sleep: h BMI: HA: yr | (9.96%) | Age: 49.8 yrs Steps: 3,341 Sleep: 6.03 h BMI: 36.3 kg/m 2 RHR: 65.4 bpm HA: yrs | 186449 , 56440 , 13533 , 70268 , 70494 , 56314 , 57778 , 58098 , 72185 , 56458 | Age: yrs Female: 70.0% (7/10) BMI: kg/m 2 Steps: Sleep: h RHR: bpm HA: yrs HA: yrs | Multi-Lever Prioritization: Avoiding cognitive overwhelm by offering low-friction entry pathways (Sleep vs. Steps) in high-risk patients. |
| P3: Recovery Bottleneck “The Active Striver at Burnout Risk” | Steps: k/day Sleep: h BMI: Healthy/OW HA: Youthful/Sync | (12.73%) | Age: 49.1 yrs Steps: 12,738 Sleep: 6.23 h BMI: 26.9 kg/m 2 RHR: 62.6 bpm HA: yrs | 70621 , 71746 , 71100 , 58093 , 70552 , 13851 , 55974 , 70488 , 71651 , 58419 | Age: yrs Female: 10.0% (1/10) BMI: kg/m 2 Steps: Sleep: h RHR: bpm HA: yrs HA: yrs | Lever Discrimination: Recognizing saturated activity volume and isolating sleep extension and recovery rather than adding steps. |
| P4: Activity Bottleneck “The Sedentary Age-Accelerated Individual” | Steps: k/day Sleep: 7.0–9.0 h BMI: OW/Healthy HA: yr | (8.17%) | Age: 48.2 yrs Steps: 3,190 Sleep: 7.67 h BMI: 32.0 kg/m 2 RHR: 64.3 bpm HA: yrs | 12803 , 13530 , 13656 , 56193 , 72532 , 72585 , 72223 , 71774 , 57671 , 58005 | Age: yrs Female: 100.0% (10/10) BMI: kg/m 2 Steps: Sleep: h RHR: bpm HA: yrs HA: yrs | Targeted Activity Escalation: Isolating progressive step accumulation while validating and preserving healthy sleep habits. |
| P5: Balanced Maintainer “The Balanced Health Seeker” | Steps: 5k–10k/day Sleep: 7.0–9.0 h BMI: Healthy/OW HA: to | (5.14%) | Age: 46.6 yrs Steps: 5,539 Sleep: 7.48 h BMI: 29.9 kg/m 2 RHR: 67.1 bpm HA: yrs | valid † 55807 , 56085 , 56144 , 58421 , 70909 , 57709 , 57065 , 58023 , 57497 | Age: yrs Female: 66.7% (6/9) BMI: kg/m 2 Steps: Sleep: h RHR: bpm HA: yrs HA: yrs | Fine-Tuning & Progression: Nudging median compliant users from baseline maintenance into cardiovascular fitness gains. |
| Total / Overall | 5 Stratified Personas | 1,079 / 2,780 (38.81%) | Weighted Sub-Cohort | valid (50 sampled) | Age: yrs Female: 57.1% (28/49) BMI: kg/m 2 Steps: Sleep: h RHR: bpm HA: yrs HA: yrs | Comprehensive Evaluation Across the Risk Continuum |
| Evaluation Criterion / Statistic | Counterfactual Coaching | Generic Coaching |
|---|---|---|
| SxS: Best Overall Plan | ||
| Scale | Forced-choice win | Forced-choice win |
| Full export ( ) | 71.4% ( ) | 28.6% ( ) |
| 95% CI | 65.5–76.7% | 23.3–34.5% |
| Net difference | +42.9 percentage points favoring Counterfactual Coaching. | |
| Exact -value | Two-sided exact binomial . | |
| Evaluation Criterion / Statistic | Generic Coaching | Counterfactual Coaching |
|---|---|---|
| Safe Intensity Limits (Primary Safety Gate) | ||
| Scale | Binary (Yes/No) | Binary (Yes/No) |
| Fig. 1 & 5 bar values | 97.6% Yes (panel 1 e) 98% Yes (panel 5 b) | 99.2% Yes (panel 1 e) 99% Yes (panel 5 b) |
| Full export: Yes | 239/245 (97.6%) | 243/245 (99.2%) |
| 95% CI | 94.8–98.9% | 97.1–99.8% |
| Full distribution: No | 6/245 (2.4%) | 2/245 (0.8%) |
| Shared Dimension / Statistic | Lifestyle Coaches ( ) | Clinicians ( ) |
|---|---|---|
| Contrasts Benefits vs. Status Quo | ||
| Scale | Binary (% Yes) | Binary (% Yes) |
| Counterfactual | 73.5% | 92.2% |
| Generic | 5.7% | 7.8% |
| Net difference (CF Gen) | +67.8 percentage points | +84.5 percentage points |
| Exact -value | ||
| Dimension / Endpoint | Coachable Persona | Result |
|---|---|---|
| Coach SxS: Best Overall Plan (Counterfactual vs. Generic; exact binomial ) | ||
| P1: Optimal Baseline k steps/day; 7–9 h sleep | 72.0% vs. 28.0% 36/50 | |
| P2: Multi-Risk Challenger k steps/day; h sleep; high BMI | 80.0% vs. 20.0% 40/50 | |
| P3: Recovery Bottleneck k steps/day; h sleep | 60.0% vs. 40.0% 30/50 | |
| P4: Activity Bottleneck k steps/day; 7–9 h sleep | 74.0% vs. 26.0% 37/50 | |
| P5: Balanced Maintainer 5–10k steps/day; 7–9 h sleep ; 9 cases | 71.1% vs. 28.9% 32/45 | |
| Expert Track & Rater ID | Counterfactual Win Rate / Key Rubric Highlights | Representative Verbatim Qualitative Feedback (Unblinded Post-Analysis) |
|---|---|---|
| Lifestyle Coach 1 Valid cases: | Best Overall: 89.8% (44/49) Adherence: 98.0% (48/49) Contrasts Benefits: 98% CF vs. 2% Gen | “Although both plans are helpful, supportive, and use data and behavior change techniques, [the Counterfactual plan] provides much more detail and dives deeper into personalizing the plan and providing options for the user to choose, giving the user the autonomy to select the best option for them which will be stronger in long-term success.” (PIDs 186449 , 55807 ) |
| Lifestyle Coach 2 Valid cases: | Best Overall: 53.1% (26/49) Adherence: 44.9% (22/49) Trajectory Control: 92% CF vs. 67% Gen | “[The Counterfactual plan] is by far more effective with the greatest likelihood to impact both short-term and long-term behavior. While both positively reinforce behavior at the outset, [the Generic plan] fails to contrast short- and long-term impacts nor does it explain the science-backed ‘why’. In P3 (PID 70621 ), [Counterfactual] provides the best opportunity for user agency, whereas [Generic] misses sleep as the key factor for someone already at 10,000+ steps/day.” |
| Lifestyle Coach 3 Valid cases: | Best Overall: 98.0% (48/49) Adherence: 98.0% (48/49) Compelling Reason: 98% CF vs. 4% Gen | “I loved how thorough the coach was with detailing what would happen to the user by choosing the plan and following it to their long-term benefits… [The Counterfactual plan] would give me the motivation to start the plan I want to choose and know what benefits could come from doing so. [The Generic plan] does not do that at all.” Constructive critique on PID 186449 : “Needs to clear up the phrasing of adding +1,000 steps per day so the user does not confuse a daily target with adding another 1,000 every single day.” |
| Lifestyle Coach 4 Valid cases: | Best Overall: 57.1% (28/49) Adherence: 53.1% (26/49) Behavior Strategies: 37% CF vs. 39% Gen | “In [the Generic plan], there is not an active contrast here for the user, and there could have been more tangible points to connect the dots on where their health is heading if they choose to adopt this plan. [The Counterfactual plan] connects the metrics directly to future outcomes and gives the user tangible options.” |
| Lifestyle Coach 5 Valid cases: | Best Overall: 59.2% (29/49) Adherence: 61.2% (30/49) Trajectory Control: 100% CF vs. 98% Gen | “[The Counterfactual plan] is more flexible with multiple pathways, while [the Generic plan] is more rigid and increasingly takes more time for the user to complete, which can be a barrier to success. In [the Generic plan], each metric is touched upon separately with nothing that addresses the bigger ‘why’ for making these changes.” |
| Clinician 1 Valid cases: | Safety Gate: 100% CF vs. 94% Gen Contrasts Benefits: 98% CF vs. 2% Gen Personalization: 88% CF vs. 82% Gen | “[The Counterfactual plan] is much more thorough in explaining the physiological consequences of staying the same versus making changes, and gives options of paths to follow that are flexible and the user’s choice.” |
Healthcare
Healthcare AI develops and evaluates models and systems for clinical care, medical research, public health, and healthcare operations.
See all 635 papers