Edge-Based Sleep-Stage Classification with Change Point Detection 🛌

Wearable devices can monitor sleep stages, but running sophisticated sleep stage classification models on-device is challenging due to limited resources. In this work, we develop a lightweight model for real-time sleep stage classification on the edge (e.g., Apple Watch) with a novel focus on wake-up timing. We incorporate change point detection (CPD) to identify transitions between sleep stages and evaluate performance using a timing-based metric, since traditional accuracy is limited by noisy sensor data and imperfect labels. Experiments on the Sleep-Accel dataset demonstrate that while overall classification accuracy is relatively low (around 40%–50%), the inclusion of CPD often improves the prediction of optimal wake-up times. Our approach achieves only modest F1 and accuracy scores, but can trigger a smart alarm within mere minutes of the ideal wake stage, highlighting the importance of timing-based evaluation for smart alarm applications.

Shani Kagan Micheletti, Diego Cerretti, Thomas Adler

• Report

• Presentation