OpenMHC: Accelerating the Science of Wearable Foundation Models
Organizations: Stanford University · Imperial College London · Charité – Universitätsmedizin Berlin · University of Washington ·
Abstract
Mobile and wearable devices offer an unprecedented opportunity for continuous, passive health monitoring and active health coaching. However, the largest wearable datasets are proprietary, and leading wearable foundation models trained on them are rarely open weight or include reproducible training code. To accelerate open science in wearable health, we introduce OpenMyHeartCounts (OpenMHC), the largest and most comprehensive wearable health dataset broadly accessible to qualified researchers, alongside open-source implementations and weights for recent wearable foundation models. Derived from over a decade of data collected through the My Heart Counts study app, OpenMHC comprises over 67M participant-hours of wearable recordings across up to 19 sensor channels (e.g., step count, heart rate, sleep, workouts) and up to 170 health, lifestyle, mood, and behavioral variables from 11,894 participants. Furthermore, we establish a benchmark for standardized evaluation across three tasks: (1) health and behavioral outcome prediction, (2) multivariate data imputation, and (3) time-series forecasting. Benchmarking classical methods alongside recent wearable and time-series foundation models, we find that while foundation models perform strongly, carefully tuned classical baselines like XGBoost remain highly competitive. By uniting broad-access wearable data with open code and model weights, OpenMHC provides a framework needed to drive reproducible advances in wearable health AI.
Figures & tables
| Dataset | N | Hours | Source | Duration | Modalities | Access |
| China Kadoorie [ 16 ] | 22k | 3M | Axivity | 7 days | ACC | |
| NAKO Germany [ 80 ] | 74k | 10M | ActiGraph | 7 days | ACC | |
| UK Biobank [ 23 ] | 100k | 17M | Axivity | 7 days | ACC | |
| NHANES [ 4 ] | 4k | <868k | ActiGraph | 9 days | ACC | |
| All of Us [ 73 , 28 ] | 30k | 264M | Fitbit | Multi-year | ST,HR,SL | |
| Fitbit (LSM-1) [ 58 ] | 165k | 40M | Fitbit | 18 mo. | 26 features* |
| Method | Demo- graphics | Medical Conditions & Risk | Vitals & Blood Biomarkers | Mental Well-Being | Sleep & Lifestyle | |||
| Statistical Models | ||||||||
| MultiRocket [ 75 ] | ||||||||
| XGBoost [ 15 ] | ||||||||
| Linear (reference) | ||||||||
| Supervised Neural Models | ||||||||
| GRU-D [ 14 ] | ||||||||
| Method | Activity | Physio. | Sleep | Workout | Semantic | |||
| Single-day imputation | ||||||||
| Statistical Models | ||||||||
| Linear | ||||||||
| LOCF (reference) | ||||||||
| Temporal mode | ||||||||
| Mode | ||||||||
Appendix figures & tables54 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Training data | Hardware | Time |
| WBM | 83,889 eligible weeks | 1 80-GB GPU | 3 GPU-h / 20 epochs |
| LSM-2 | 666,175 daily samples | 1 A100 40-GB | 45 GPU-h / 25 epochs |
| Component | Hardware | Time |
| WBM inference | 1 L40S | 10 GPU-min |
| LSM-2 inference | 1 A100 40-GB | 6 GPU-h |
| Forecasting ( Chronos-2 ) | 1 A100 80-GB | 1 GPU-h |
| Forecasting ( Toto ) | 1 A100 80-GB | 33 GPU-h |
| Forecasting ( MixLinear / DLinear / SegRNN ) | 1 A100 80-GB | 6 GPU-min |
| Imputation ( LSM-2 ) | 1 RTX 3090 | 5 GPU-h |
| Method | All | Age 40 | Age 40 | Female | Male | ||
| Statistical Models | |||||||
| MultiRocket | |||||||
| XGBoost | |||||||
| Linear | |||||||
| Method | All | Age 40 | Age 40 | Female | Male | ||
| Single-day imputation | |||||||
| Statistical Models | |||||||
| Linear | |||||||
| LOCF | |||||||
| Method | All | Age 40 | Age 40 | Female | Male | ||
| Statistical Methods | |||||||
| Seasonal Naive | |||||||
| AutoARIMA | |||||||
| AutoETS | |||||||
| Category | Channel | Unit | n | Avg. Days | Span |
| Phone | StepCount | steps/min | 11,631 | 234.4 | 4,730 |
| DistanceWalkingRunning | meter/min | 11,630 | 234.6 | 4,730 | |
| FlightsClimbed | count/min | 2,752 | 430.6 | 4,730 | |
| Watch | StepCount | steps/min | 7,010 | 202.2 | 4,307 |
| DistanceWalkingRunning | meter/min | 7,006 | 200.6 | 4,307 | |
| HeartRate | count/s | 7,020 | 202.0 | 4,307 |
| Characteristic | Train ( ) | Validation ( ) | Test ( ) | Overall ( ) |
| Age (years) | ||||
| Coverage | 7,136 (100%) | 1,189 (100%) | 3,568 (100%) | 11,893 (100%) |
| Mean SD | 42.2 15.5 | 41.1 14.7 | 41.6 15.4 | 41.9 15.4 |
| Median [IQR] | 39 [30–53] | 39 [30–50] | 39 [30–52] | 39 [30–52] |
| Biological Sex | ||||
| Coverage | 6,395 (89.6%) | 1,064 (89.5%) | 3,170 (88.8%) | 10,629 (89.4%) |
| Year | Male (%) | Sex available, | Age (%) | Age available, |
|---|---|---|---|---|
| 2015 | 84.9 | 4,873 | 55.2 | 5,944 |
| 2016 | 73.9 | 2,015 | 47.5 | 2,043 |
| 2017 | 78.8 | 1,189 | 44.2 | 1,194 |
| 2018 | 65.0 | 1,229 | 45.3 | 1,233 |
| 2019 | 66.9 | 534 | 51.5 | 540 |
| 2020 | 62.9 | 464 | 42.3 | 468 |
| Channel (%) | 2015 | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| iPhone step count | 98.6 | 98.0 | 97.8 | 98.3 | 99.2 | 99.7 | 99.8 | 99.9 | 99.9 | 99.8 | 99.7 |
| iPhone walking/running distance | 98.6 | 98.0 | 97.8 | 98.3 | 99.2 | 99.7 | 99.8 | 99.9 | 99.9 | 99.8 | 99.7 |
| iPhone flights climbed | 75.4 | 80.3 | 85.5 | 86.8 | 89.8 | 91.3 | 92.0 | 92.8 | 92.8 | 92.9 | 92.1 |
| Watch step count | 46.5 | 60.0 | 76.8 | 78.5 | 77.7 | 78.4 | 81.9 | 81.2 | 80.6 | 80.4 | 81.9 |
| Watch walking/running distance | 46.6 | 60.2 | 77.0 | 78.7 | 77.8 | 78.6 | 82.0 | 81.2 | 80.6 | 80.4 | 81.9 |
| Watch heart rate | 46.5 | 60.1 | 76.9 | 78.7 | 77.8 | 78.6 | 82.0 | 81.2 | 80.6 | 80.4 | 81.9 |
| HealthKit Type | Filter | Threshold |
| StepCount | Max rate | 5 steps/s |
| DistanceWalkingRunning | Max rate | 10 m/s |
| ActiveEnergyBurned | Max rate | 250 cal/s |
| DistanceCycling | Max rate | 25 m/s |
| AppleStandTime | Max rate | 1 min/min |
| HeartRate | Valid range | 40–200 bpm |
| Channel Group | Presence Criterion | Zero | NaN |
| Continuous HK (0–6) | 10% of days | No activity | Stream unavailable |
| Sleep (7–8) | Any occurrence | Not asleep | No sleep data |
| Workouts (9–18) | Any occurrence | Not exercising | Never logged |
| Metrics | Variable Names | Considerations |
| Blood Pressure | SystolicBloodPressure, blood_pressure_categories | Swapped if Systolic BP was lower than Diastolic BP; both values required. Diastolic BP is not released; it is used only for this swap and the categories |
| BMI | BMI_values, BMI_categories | Height 1.4–2.1m, weight 40kg, and BMI 60 |
| Body weight | WeightKilograms | Entries of exactly 30kg or 300kg removed as sentinel values |
| Wake-up time & bedtime | WakeUpTime_categories, GoSleepTime_categories | Profile times converted to decimal hours, then binned |
| Psychological | feel_worthwhile1–4, satisfiedwith_life, happiness_categories | Thresholds based on distributions |
| Happiness (long) | happiness | Filtered to 3 responses per participant |
| Stage | Train | Validation | Test |
| Day-level filtering (Appendix D.3.1 ) | |||
| Released cohort (all days) | 7,136 / 1,682,478 | 1,189 / 285,540 | 3,569 / 854,482 |
| Wear-time filter ( 12 h wear) | 3,602 / 681,251 | 630 / 124,327 | 1,799 / 352,570 |
| Retained days (+ low variance) | 3,542 / 666,175 | 622 / 122,107 | 1,777 / 348,580 |
| Track 1: Health and Behavior Prediction | |||
| Largest outcome (Age) | 3,429 / 656,086 | 606 / 118,941 | 1,730 / 342,145 |
| Outcome | Train | Validation | Test |
| Demographics | |||
| Age | 3,429 / 656,086 | 606 / 118,941 | 1,730 / 342,145 |
| Biological Sex | 3,341 / 602,448 | 588 / 111,903 | 1,674 / 317,426 |
| Medical Conditions & Risk | |||
| Cardiovascular conditions (8) | 2,197 / 446,973 | 368 / 69,925 | 1,091 / 242,231 |
| Diabetes, Hypertension | 1,048 / 291,237 | 174 / 47,287 | 525 / 173,133 |
| Train | Validation | Test | ||||
| Channel | Participants | Hours | Participants | Hours | Participants | Hours |
| Phone | ||||||
| StepCount | 3,437 | 15,345k | 603 | 2,726k | 1,719 | 8,057k |
| DistanceWalkingRunning | 3,437 | 15,349k | 602 | 2,727k | 1,719 | 8,059k |
| FlightsClimbed | 1,217 | 15,204k | 220 | 2,696k | 647 | 7,708k |
| Watch | ||||||
| No filter | ||||
| h wear | h wear | h wear | (all days) | |
| Participants Retained | 2,920 | 6,031 | 9,661 | 11,894 |
| Age (mean SD) | ||||
| BMI (mean SD) | ||||
| Hypertension % | 26.8 | 26.2 | 25.9 | |
| Diabetes % | 7.2 | 6.0 | 5.8 | 5.8 |
| Outcome | Self-report question/source | Type | Construction details |
|---|---|---|---|
| Demographics (2 outcomes) | |||
| Age | enrollment_info.json birthdate and last_labels.json last-survey timestamp | Continuous | Whole years from birthdate to last-survey timestamp; participants with computed age outside are dropped. |
| Biological Sex | “Gender”, recorded by Apple’s HealthKit as one of Female / Male / Other (HealthKit field HKBiologicalSex ) | Binary | Male is encoded as and Female as ; "Other" option absent in observed data. |
| Medical Conditions & Risk (12 outcomes) | |||
| Atrial Fibrillation | heart_disease : “Have you been diagnosed with any of the below diseases?” | Binary | if the participant selected “Atrial fibrillation (Afib)”. |
| Cardiovascular Disease | heart_disease : “Have you been diagnosed with any of the below diseases?”; vascular : “Which vascular disease diagnosis have you received?” | Binary | if the participant selected any non-“None of the above” option in either multi-select question. |
| Outcome | Source | Outcome Type | Included Participants | Label Summary |
|---|---|---|---|---|
| Demographics | ||||
| Age ( 1095 days) | Enrollment | Continuous | 5,765 (97.0%) | 43.0 15.1 years |
| Biological Sex | HealthKit | Binary | 5,603 (94.3%) | Male: 79.0% |
| Medical Conditions & Risk ( 365 days) | ||||
| Atrial Fibrillation | Survey | Binary | 3,656 (61.5%) | Prevalence 2.0% |
| Cardiovascular Disease | Survey | Binary | 3,656 (61.5%) | Prevalence 11.2% |
| Item | Mean gap | Agreement | ||
| Biological sex | 4,014 | 19.4 wk | 99.7% | 0.99 |
| Diabetes | 963 | 16.6 wk | 98.5% | 0.85 |
| Hypertension | 963 | 16.6 wk | 93.1% | 0.83 |
| Sleep disorder | 2,040 | 35.1 wk | 92.2% | 0.70 |
| Sense of purpose (time-varying) | 3,115 | 15.4 wk | 31.0% | — |
| Parameter | Value |
| Architecture | |
| Encoder type | Bidirectional Mamba2 |
| Encoder layers | 4 |
| Embedding dimension ( ) | 256 |
| Tokenizer hidden dimension ( ) | 64 |
| FFN expansion factor | 4 |
| Hyperparameter | Search space | Selected value |
| Token drop probability ( drop_prob ) | Uniform | |
| KoLeo regularization weight ( lambda_koleo ) | Uniform | |
| Learning rate ( lr ) | Log-uniform |
| Domain | Task | Metric | Linear | MultiRocket | XGBoost | GRU-D | LSM-2 | WBM | Toto | Chronos-2 |
| Demographics | Age | Pearson | ||||||||
| Biological Sex | AUPRC | |||||||||
| Domain Average Rank | ||||||||||
| Domain Skill Score (%) | ||||||||||
| Medical conditions & risk | Atrial Fibrillation | AUPRC | ||||||||
| Cardiovascular Disease | AUPRC |
| Method | Demo- graphics | Medical Conditions & Risk | Vitals & Blood Biomarkers | Mental Well-Being | Sleep & Lifestyle | ||
| LSM-2 | |||||||
| XGBoost | |||||||
| WBM | |||||||
| Linear (reference) |
| Method | Demo- graphics | Medical Conditions | Vitals & Blood Biomarkers | Mental Well-Being | Sleep & Lifestyle | ||
| LSM-2 Full Cohort (min) | |||||||
| LSM-2 WBM Cohort (min) | |||||||
| WBM | |||||||
| LSM-2 WBM Cohort (hour) | |||||||
| Linear (reference) |
| Domain | Task | Metric | Linear | XGBoost | LSM-2 | Gem-2.5 St | Gem-3.1 St | Gem-2.5 Vis | Gem-3.1 Vis | PHA-DSA |
| Demographics | Age | Pearson | ||||||||
| Biological Sex | AUPRC | |||||||||
| Domain Average Rank | ||||||||||
| Domain Skill Score (%) | ||||||||||
| Medical conditions & risk | Atrial Fibrillation | AUPRC | ||||||||
| Cardiovascular Disease | AUPRC | |||||||||
| Approach | Real-world failure mode | Tier | Channels masked | Key parameters |
| Random noise | Sporadic transmission errors | Structural | C + B | , |
| Temporal slice | Device removal or power-off | Structural | C + B | , |
| Signal slice | Complete sensor/modality failure | Structural | C + B | |
| Sleep gap | Device removed during sleep | Semantic | C + B | data-driven |
| Workout gap | Motion-artifact dropout in exercise | Semantic | C only | data-driven |
| Intensity failure | Sensor saturation at high intensity | Semantic | C only | BPM, min |
| Method | Random noise | Temporal slice | Signal slice | Sleep gap | Workout gap | Intensity failure | ||
| Single-day imputation | ||||||||
| Statistical Models | ||||||||
| Linear | ||||||||
| LOCF (reference) | ||||||||
| Temporal mode | ||||||||
| Mode | ||||||||
| Method | Random MAE | Random AUC | Temporal MAE | Temporal AUC | Signal MAE | Signal AUC | Sleep MAE | Workout MAE | Intensity MAE |
| Statistical Models | |||||||||
| Linear | |||||||||
| Temporal mean | |||||||||
| LOCF (reference) | |||||||||
| Temporal mode | |||||||||
| Mode | |||||||||
| Method | Random MAE | Random AUC | Temporal MAE | Temporal AUC | Signal MAE | Signal AUC | Sleep MAE | Workout MAE | Intensity MAE |
| Statistical Models | |||||||||
| Pers. temp. mean | |||||||||
| Pers. mean | |||||||||
| Pers. mode | |||||||||
| Neural Models | |||||||||
| LSM-2 (7-day) | |||||||||
| Method | Default / reference | Random 10-min patches | Structural masks |
| DLinear [ 87 ] | |||
| BRITS [ 13 ] | |||
| LSM-2 [ 83 ] | – | – |
| Variant | Layer type | Complexity | Tokens / layer | Multi-day |
| LSM-2 (Daily) | Full | No | ||
| LSM-2-Weekly | Full | Yes | ||
| LSM-2-Sparse | Day-local | — | ||
| LSM-2-Sparse | Cross-day | Yes |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Weight decay ( wd ) | log-U | |
| Batch size |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Weight decay ( wd ) | log-U | |
| Decoder depth ( decoder_depth ) | ||
| Mask ratio ( mask_ratio ) |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| Moving avg. window size | ||
| d_model | fixed | |
| ORT weight | log-U | |
| MIT weight | log-U |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| RNN hidden size |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| Moving avg. window size | ||
| n_layers | fixed | |
| d_model | fixed | |
| d_ffn | fixed |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| Dropout | U | |
| n_layers | fixed | |
| d_model | fixed | |
| d_ffn | fixed |
| Model | Hyperparameter | ||
| DLinear | Batch size | Learning rate | Window size of moving average |
| Search Range | |||
| MixLinear | Batch size | Learning rate | Segment length |
| Search Range | |||
| SegRNN | Batch size | Learning rate | Segment length |
| Search Range | |||
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| Moving avg. window size |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| Segment length ( period_len ) |
| Hyperparameter | Search space | Selected value |
| Learning rate ( lr ) | log-U | |
| Batch size | ||
| Segment length ( seg_len ) |
| Parameter | Value |
| Base model | Datadog/Toto-Open-Base-1.0 |
| Max context length | 2,048 |
| Number of channels ( ) | 19 |
| Forecast horizon | 24 hours |
| Best checkpoint | epoch 24, val_loss |
| Fine-tuning method | LoRA ( , , dropout ) |
| Parameter | Value |
| Base model | amazon/chronos-2 |
| Max context length | 8,192 |
| Number of channels ( ) | 19 |
| Input steps ( n_steps ) | 336 |
| Prediction steps ( n_pred_steps ) | 24 |
| Context length (minimum past) | 168 |
| Hyperparameter | Search space | Selected |
| Learning rate | Log-uniform | |
| LoRA rank ( ) | 16 | |
| LoRA alpha ( ) | 16 | |
| LoRA dropout | Uniform | 0.0 |
Wearable Health Monitoring
Wearable health monitoring uses body-worn sensors and AI to track physiological and behavioral signals for continuous, personalized health assessment.
See all 119 papers