Wearable Accelerometer Foundation Models for Health via Knowledge Distillation

Summary

Apple Research demonstrates that a generalist accelerometry foundation model can be trained via cross-modal knowledge distillation from a PPG foundation model — without any annotations. The core insight: PPG encodes rich health information (HR, HRV, cardiovascular state) that accelerometry alone cannot capture, but PPG and accelerometry are simultaneously recorded on the Apple Watch. This temporal co-occurrence enables an unsupervised KD framework.