Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Wearable accelerometers enable large-scale health monitoring, yet learning robust human-activity representations has been constrained by scarce labeled data. While self-supervised learning offers a remedy, existing methods treat sensor streams as unstructured time series, overlooking the underlying biological structure of human movement, a factor we argue is critical for effective Human Activity Recognition (HAR). We introduce a novel tokenization strategy grounded in the *submovement theory* of motor control, which posits that continuous wrist motion is composed of elementary basis functions called submovements. We define our token as the *movement segment*, a computationally tractable unit of motion composed of a finite sequence of submovements. By pretraining a Transformer encoder via masked reconstruction of these tokens, we shift the learning focus from local waveform morphology to high-level structural and temporal organization. Pretrained on the NHANES corpus ($\approx$ 28k hours; $\approx$ 11k participants), our representations outperform strong wearable SSL baselines across six subject-disjoint HAR benchmarks. Code and pretrained weights are available at [https://prithvitarale.github.io/biopm-site/](https://prithvitarale.github.io/biopm-site/).

Smartwatches record wrist movements throughout the day. These signals can help AI models recognize activities such as walking, eating, or climbing stairs, and may one day help track progress in motor rehabilitation or detect symptoms of conditions such as Parkinson’s disease, stroke, or ataxia. The problem is that teaching models to recognize each activity usually requires large amounts of human-labeled data, which is expensive to collect. One way around this is to let a model first learn from unlabeled movement recordings, before asking it to solve specific tasks with labels. Many current methods do this by hiding random parts of the signal and training the model to fill in the blanks. But human movement is not random: it is composed of distinct, meaningful actions, much like sentences are built from words. These actions are especially important because research has shown that their execution reveals changes linked to motor impairments. This motivated our central idea: a model may learn the language of human movement better if it learns to fill in the missing movement “words,” rather than the missing random signal fragments. Our method automatically finds these movement “words” in smartwatch sensor data, hides some of them, and trains the model to fill in the blanks using the surrounding movement context. After learning from about 28,000 hours of public wearable data, our model recognized daily activities more accurately than comparable models across six datasets. This approach could help future wearable devices better understand everyday movement, supporting applications in fitness tracking, elder care, rehabilitation, and motor health assessment.