Research map / Label-efficient and robust learning / Recognition and long-tail
Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a wearable health behavior foundation model (WBM) pretrained with contrastive SSL on 2.5B hours of hourly behavioral wearable metrics from 162K people, selecting a TST tokenizer with bidirectional Mamba-2, and shows it complements a PPG foundation model on 57 health tasks.
Problem
Foundation models for wearables have mostly modeled low-level sensor signals, while higher-level behavioral metrics (irregularly sampled, heavily missing) are underexplored for health detection tasks.
Approach
Pretrain an encoder-only foundation model on weekly hourly-aggregated behavioral wearable data using a dense (TST-style, mean-imputed with missingness masks) tokenization and bidirectional Mamba-2 backbone with a subject-level contrastive InfoNCE + KoLeo objective, then combine its embeddings with an existing PPG model.
Key results
- WBM+PPG reaches age MAE 2.46 vs 2.89 for PPG alone and 7.89 for the summary-statistics baseline.
- WBM outperforms the baseline in 39 of 47 medical history/medication outcomes and in all 8 time-varying tasks, with WBM+PPG best in 42 of 47 outcomes.
- On sleep duration regression WBM achieves R^2 0.590 versus 0.110 for PPG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WBM embeddings instead of baseline summary statistics | Health detection tasks · 47 baseline disease and medication outcomes | AUROC: median AUROC improvement of 0.017 | Helped |
| Combine WBM and PPG embeddings by concatenation | Health detection tasks · 47 baseline disease and medication outcomes | AUROC: median AUROC improvement of 0.009 | Helped |
| Combine WBM with PPG embeddings | Time-varying health detection · Diabetes | AUROC: 0.829 -> 0.828 | Hurt |
| Use WBM rather than PPG embeddings | Time-varying health detection · Sleep duration, sleep efficiency, and REM sleep duration | R²: 0.110 -> 0.590; 0.182 -> 0.424; 0.230 -> 0.326 | Helped |
| Use TST tokenization with Mamba-2 rather than other tokenizer and backbone combinations (ablation) | WBM pretraining · Age prediction | MAE | Helped |
| Replace TST with Tuple or mTAN tokenization (ablation) | WBM pretraining · Age prediction | MAE | Hurt |
| Replace Mamba-2 with a Transformer backbone (ablation) | WBM pretraining · Age prediction | MAE | Hurt |
| Use global-mean imputation rather than subject-specific or week-specific mean imputation (ablation) | TST tokenizer · Downstream health detection tasks | Performance | Helped |
| Replace contrastive pretraining with a masked autoencoder (ablation) | Wearable behavior foundation model · Downstream health detection tasks | Performance | Hurt |
Surprising findings
- The dense TST tokenizer with global-mean imputation outperformed more sophisticated irregular-time-series tokenizers (Tuple, mTAN).
- A bidirectional Mamba-2 backbone outperformed Transformer variants, and masked autoencoder pretraining performed poorly in initial experiments.
Limitations
- Trained and evaluated only on Apple Watch/iPhone-derived data, so generalization to other devices is untested.
- Contrastive SSL requires defining positive pairs, which may limit downstream task types.
- Many labels come from self-reported surveys and may be inaccurate.
Experimental setup
Datasets: Apple Heart and Movement Study (AHMS) behavioral wearable data AHMS self-reported medical history and medication surveys
Benchmarks: age prediction biological sex prediction baseline medical history and medication outcomes (47) time-varying health detection (diabetes, pregnancy, infection, injury) sleep regression (duration, efficiency, deep and REM sleep)
Metrics: MAE AUROC R^2
Baselines: Summary statistics baseline (mean and std of 27 variables plus demographics) PPG foundation model
Builds on: Mamba-2 InfoNCE KoLeo regularization Time Series Transformer
Closest papers by meaning
- Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.91)
- OpenMHC: Accelerating the Science of Wearable Foundation Models (NeurIPS 2026 · similarity 0.90)
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.89)
- Large-scale Training of Foundation Models for Wearable Biosignals (ICLR 2024 · similarity 0.88)
- FEEL: Quantifying Heterogeneity in Physiological Signals for Generalizable Emotion Recognition (NeurIPS 2025 · similarity 0.87)
- UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.