Minimal-Sensor Sleep Monitoring for Consumer Devices via Robust EOG Encoding and Adaptive Calibration

Single-channel electrooculogram (EOG) signals are attractive for wearable sleep monitoring because they can be easily integrated into headbands and consumer devices. However, their high susceptibility to noise and motion artifacts demands robust algorithms for reliable sleep staging. We propose a lightweight and physiologically inspired framework that combines spectrum derivative features with the Flexible Variational Information Bottleneck (FVIB). Spectrum derivatives enhance time–frequency representations of ocular dynamics, while FVIB provides strong regularization and post-hoc calibration, ensuring robustness under noisy conditions. Experimental results demonstrate stable and balanced performance across sleep stages, highlighting improved discrimination of both transitional and stable sleep states under noisy conditions. By avoiding Electroencephalography (EEG) and targeting low-power, embedded deployment, our method enables continuous and personalized sleep analysis via personal or home monitoring systems, paving the way for accessible consumer health technologies.