This project assesses whether PSG-derived sleep stages can be recovered from wearable data. PSG is the gold standard for sleep staging because it measures signals such as brain waves and eye movements that directly characterize sleep stages. However, PSG is typically conducted in a lab, making it costly, time-consuming, and difficult to scale. Wearable signals are indirect proxies for sleep stage, but their accessibility makes them valuable if they can support reliable staging.

Using DREAMT wearable data, I developed a leakage-aware deep learning pipeline to classify sleep epochs as Wake, Non-REM, or REM. The key finding is that feature fusion, temporal context, and biologically informed modeling improve performance. Limitations include REM detection, transition errors, and participant-level variability. Signal ablation suggests future work should emphasize cardiovascular signals—BVP, IBI, and HR—to simplify dataset merging and improve performance through larger sample sizes.