Wearable sleep tracking can't replace EEG, study finds 0.304 kappa gap

A new study from Yi Wang (arXiv, July 2026) systematically decomposes the performance of wearable-compatible sleep staging by introducing a four-layer controlled framework covering signal source, physiological representation, temporal prior, and decision layers. Using a compact Mamba2 model across a signal-quality ladder—including Apple Watch Sleep-Accel (N=31), Sleep Heart Health Study (N=195, with ECG, respiratory, and SpO2), and Sleep-EDF-20 as an EEG+EOG reference—the research quantifies exactly what information is lost when moving from lab-grade EEG to consumer wearables.

Laboratory cardiorespiratory signals reach a Cohen's kappa of 0.492, while consumer heart rate and accelerometry from wearables achieve only 0.255—less than half the reference performance of 0.796 from EEG+EOG. The residual gap of 0.304 stems from missing cortical information, not temporal modeling alone. A confidence-based abstention mechanism provides a calibrated operating mode: removing the 20% lowest-confidence epochs boosts kappa from 0.452 to 0.512, while a label-shuffled control collapses to -0.003. These results confirm that non-EEG sleep staging is best suited for coarse, confidence-aware sleep-structure monitoring rather than EEG-equivalent five-class clinical staging.