SSL-Wearables (HARNet)

SSL-Wearables, whose pretrained networks are distributed as HARNet, is a family of self-supervised deep learning models for human activity recognition (HAR) from wrist-worn accelerometer data. Developed by the Oxford Wearables Group (OxWearables) at the University of Oxford and published in npj Digital Medicine in 2024, the work addresses a persistent bottleneck in digital health: labeled accelerometer data is scarce and expensive to collect, while raw sensor data from population-scale studies is abundant but unlabeled. By pretraining on enormous quantities of unlabeled signal, the authors produce feature extractors that transfer to small labeled datasets and substantially outperform models trained from scratch.

The central contribution is leveraging the UK Biobank accelerometer study — approximately 700,000 person-days of wrist-worn triaxial accelerometer recordings — as an unlabeled corpus for self-supervised pretraining. This is among the largest self-supervised efforts applied to wearable time series, and it demonstrates that the representation-learning paradigm that transformed vision and language can be applied to consumer-grade movement sensors.

Critically, the resulting models generalize beyond their pretraining cohort, maintaining strong performance across different sensor devices, study populations, and free-living versus laboratory environments — addressing a long-standing concern that HAR models overfit to the specific hardware and protocol used to collect their training data.

#Key Features

• Massive self-supervised pretraining: The feature extractor is pretrained on roughly 700,000 person-days of unlabeled UK Biobank accelerometer data, removing the dependence on costly manual activity annotation.
• Multi-task pretext objectives: Pretraining combines three self-supervised tasks — arrow-of-time (temporal direction) prediction, channel permutation detection, and time-warping detection — which jointly teach the network temporally and structurally meaningful motion representations.
• Three time-scale variants: Released as HARNet5, HARNet10, and HARNet30 for 5-, 10-, and 30-second windows (150, 300, and 900 samples at 30 Hz), letting downstream users match the model to their labeling granularity.
• Cross-device and cross-cohort generalization: Models transfer across external datasets, different participant populations, living environments, and sensor hardware rather than being tied to UK Biobank's specific device.
• Drop-in PyTorch Hub access: Pretrained weights load in one line via torch.hub.load('OxWearables/ssl-wearables', 'harnet5', pretrained=True), providing a ready feature extractor with an untrained classification head for fine-tuning.

#Technical Details

The architecture is a ResNet-style one-dimensional convolutional neural network that ingests 3-channel (x, y, z) accelerometer signals sampled at 30 Hz. The model is split into a pretrained convolutional feature extractor and a task-specific classifier head that users fine-tune on their own labeled data. Pretraining uses the multi-task self-supervised objectives described above on the UK Biobank corpus; downstream adaptation requires only the modest labeled set for the target task. Evaluated on eight HAR benchmark datasets spanning different devices and populations, the self-supervised models consistently outperform strong supervised baselines, with reported F1 relative improvements of 2.5–130.9% (median 24.4%). Gains are largest on small benchmarks such as the ADL dataset and smallest on already data-rich benchmarks like Capture-24.

#Applications

SSL-Wearables targets researchers in physical activity epidemiology, digital health, and behavioral science who need to classify activities (e.g., sleep, sedentary behavior, walking, exercise) from wrist accelerometers but have limited labeled data. Because the pretrained feature extractor transfers across devices and cohorts, groups running clinical trials, cohort studies, or consumer-wearable analyses can fine-tune a strong model on a few hundred labeled examples rather than collecting and annotating data at scale. The one-line PyTorch Hub interface makes it straightforward to integrate into existing accelerometer-processing pipelines.

#Impact

The work established self-supervised learning as a practical paradigm for wearable sensor data, showing that population-scale unlabeled cohorts like UK Biobank can be mined to build broadly transferable activity-recognition models. The openly released HARNet weights have become a widely used starting point for accelerometer-based HAR, lowering the barrier for labeled-data-poor studies and informing downstream digital biomarker research. A key limitation is licensing: the code and weights are released for academic use only, with commercial use requiring a license from Oxford University Innovation, and the models are specialized to wrist-worn triaxial accelerometry rather than arbitrary biosignals.