WearableBridge: Task-Guided Latent Transport for Free-Living Activity Recognition
Organizations: Herbert Wertheim School of Public Health and Human Longevity Science, University of California San Diego · Department of Statistics, University of Michigan, Ann Arbor
Abstract
Wrist-worn accelerometers make activity monitoring feasible at scale, but incidental arm motion, device orientation, and shifting context blur the signal that separates activities. Calibration studies supply higher-quality references such as concurrent thigh recordings or supervised laboratory wrist sessions, but they are expensive, cover few participants, and are rarely available at deployment. How can limited reference data improve recognition for participants who wear only a wrist device? We propose WearableBridge, which holds the reference encoder and classifier fixed and learns a task-guided transport map, the bridge, that carries each wrist feature into the region the classifier associates with its activity. Reference features or the classifier's predictions indicate the destination, and a classification loss penalizes wrist features that land in the region of another activity. We instantiate WearableBridge with three bridges: Mapper regresses onto reference features, Distillation matches the classifier's class probabilities, and MeanFlow learns a one-step flow from wrist toward reference features. On 40 held-out free-living PAAWS participants, all three bridges improve macro-F1 in every training run, and the strongest, MeanFlow, raises mean macro-F1 from 35.1% to 56.0% with concurrent thigh references (Thigh2Wrist), and from 54.0% to 63.5% with group-matched laboratory wrist references (Lab2Freeliving). Fitted only on PAAWS, MeanFlow also raises macro-F1 on the CAPTURE-24 cohort from 42.8% to 54.7%. The gains replicate on selfBACK, hold across four reference encoders, and grow with temporal context. A study can therefore collect reference recordings for a calibration subset and improve wrist-only recognition for every other participant, training only a bridge of 0.13 to 2.44 million parameters while the reference encoder and classifier stay unchanged.
Figures & tables
| Overall | Class F1 | ||||||
|---|---|---|---|---|---|---|---|
| Method | Macro-F1 | Bal. acc. | Accuracy | Cycling | Sitting | Standing | Walking |
| Unadapted reference (baseline) | 35.07 0.21 | 42.45 0.13 | 51.80 0.30 | 2.77 0.03 | 70.27 0.53 | 18.13 0.24 | 49.12 1.11 |
| Mapper | 54.13 0.54 | 67.00 1.91 | 73.83 1.93 | 17.14 3.81 | 83.82 1.72 | 35.69 0.48 | 79.87 0.40 |
| MeanFlow | 55.98 0.51 | 69.87 0.73 | 76.11 1.64 | 19.01 2.56 | 85.66 1.41 | 38.07 0.50 | 81.17 0.11 |
| Distillation | 55.06 0.32 | 66.35 0.89 | 75.93 1.85 | 17.84 2.43 | 85.52 1.55 | 36.78 0.16 | 80.12 0.48 |
| Measured thigh (reference sensor) | 82.50 0.33 | 89.62 0.10 | 94.97 0.01 | 61.81 1.45 | 97.97 0.02 | 82.87 0.12 | 87.37 0.08 |
| Baseline | WearableBridge | Comparators | Reference sensor | ||||
|---|---|---|---|---|---|---|---|
| Metric | Unadapted | Mapper | MeanFlow | Distillation | Activity-only | Direct wrist | Measured thigh |
| Macro-F1 | 25.73 0.19 | 65.80 1.04 | 66.74 0.38 | 65.43 1.06 | 67.76 1.34 | 63.59 0.07 | 64.42 0.37 |
| Bal. acc. | 32.96 0.09 | 67.44 0.41 | 68.11 0.19 | 67.18 0.71 | 68.93 1.58 | 64.62 0.14 | 65.51 0.21 |
| Accuracy | 35.31 0.10 | 70.61 1.28 | 71.37 0.42 | 70.54 1.17 | 71.89 1.53 | 68.78 0.60 | 67.88 0.22 |
| Overall | Class F1 | ||||||
|---|---|---|---|---|---|---|---|
| Method | Macro-F1 | Bal. acc. | Accuracy | Cycling | Sitting | Standing | Walking |
| Direct wrist classifier | 54.60 0.29 | 69.61 0.15 | 76.56 0.47 | 15.11 1.07 | 86.31 0.38 | 37.94 0.47 | 79.04 0.24 |
| Activity-only bridge | 53.21 1.08 | 64.81 0.65 | 73.29 1.99 | 17.07 2.36 | 83.46 1.63 | 34.12 1.25 | 78.20 0.27 |
| MeanFlow | 55.98 0.51 | 69.87 0.73 | 76.11 1.64 | 19.01 2.56 | 85.66 1.41 | 38.07 0.50 | 81.17 0.11 |
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting | MoCA projection/classifier | Other encoders | Bridges |
|---|---|---|---|
| Learning rate | |||
| Weight decay | |||
| Epochs | 30 | 30 | 5–30 |
| Precision | Mixed | Mixed | Full |
| Activity | Thigh2Wrist | Lab2Freeliving |
|---|---|---|
| Cycling | 9,552 | 9,552 |
| Sitting | 1,261,789 | 1,267,245 |
| Standing | 176,233 | 177,148 |
| Walking | 140,795 | 141,826 |
| Total | 1,588,369 | 1,595,771 |
| Data and design | Comparator | Difference | 95% interval |
|---|---|---|---|
| PAAWS Thigh2Wrist | Unadapted reference | 20.91 | 17.31 to 24.61 |
| PAAWS Lab2Freeliving | Unadapted reference | 9.54 | 1.74 to 13.52 |
| selfBACK Thigh2Wrist | Unadapted reference | 41.01 | 31.70 to 50.59 |
| CAPTURE-24 transfer | Unadapted reference | 11.98 | 8.33 to 14.60 |
| selfBACK Thigh2Wrist | Direct wrist classifier | 3.16 | to 7.89 |
| CAPTURE-24 transfer | Direct wrist classifier | to 2.31 |
| Overall | Class F1 | ||||||
|---|---|---|---|---|---|---|---|
| Method | Macro-F1 | Bal. acc. | Accuracy | Cycling | Sitting | Standing | Walking |
| Unadapted reference (baseline) | 42.75 0.01 | 47.09 0.04 | 68.78 0.06 | 5.38 0.14 | 84.12 0.05 | 26.23 0.04 | 55.27 0.05 |
| MeanFlow | 54.73 1.58 | 57.50 1.18 | 69.40 0.94 | 50.64 4.94 | 83.16 0.84 | 22.57 0.45 | 62.56 0.63 |
| Direct wrist classifier | 55.58 0.42 | 65.17 0.54 | 72.17 0.48 | 50.30 2.63 | 85.81 0.39 | 25.15 0.14 | 61.08 0.83 |
| Baseline | WearableBridge | |||
|---|---|---|---|---|
| Reference encoder | Unadapted reference | Mapper | MeanFlow | Distillation |
| DeepConvLSTM | 33.73 4.30 | 43.93 0.93 | 43.36 1.22 | 44.20 0.92 |
| InceptionTime | 32.49 5.47 | 52.04 1.71 | 52.37 2.66 | 52.63 1.38 |
| HART | 31.67 2.78 | 48.21 0.73 | 49.20 0.72 | 47.90 0.59 |
| Thigh2Wrist | Lab2Freeliving | |||
|---|---|---|---|---|
| Method | Macro-F1 (%) | macro-F1 (pp) | Macro-F1 (%) | macro-F1 (pp) |
| Unadapted reference (baseline) | 35.53 0.25 | +0.46 | 58.16 1.70 | +4.16 |
| Mapper | 60.28 0.78 | +6.15 | 66.49 0.24 | +5.59 |
| MeanFlow | 61.70 0.75 | +5.72 | 68.90 1.42 | +5.36 |
| Distillation | 60.93 0.43 | +5.87 | 67.55 1.32 | +5.26 |
| Method | Macro-F1 | Bal. acc. | Accuracy |
|---|---|---|---|
| MeanFlow (frozen encoder) | 55.98 0.51 | 69.87 0.73 | 76.11 1.64 |
| Joint fine-tuning | 62.56 0.59 | 72.82 1.88 | 79.62 0.86 |
| Method | Macro-F1 |
|---|---|
| Unadapted reference | 14.46 0.12 |
| Direct wrist classifier | 40.55 0.34 |
| MeanFlow | 44.04 0.69 |
| MeanFlow minus Direct (pp) | +3.49 (2.58 to 4.34) |
Human Activity Recognition
Human activity recognition classifies actions and behaviors from wearable sensors, video, radio signals, and other measurements.
See all 239 papers