A Temporally Causal Two-Layer Framework for Wearable EEG Sleep Staging and Short-Horizon Sleep-Transition Forecasting

Reliable online sleep monitoring requires sleep-stage estimates and forecasts based only on information available at the time of prediction. This study developed a temporally causal two-layer framework for wearable electroencephalography (EEG) sleep staging and short-horizon sleep-transition forecasting. Using 128 Bitbrain Open Access Sleep (BOAS) recordings from 100 participants, the first layer (M0) performed five-stage sleep classification with causal signal processing and participant-isolated nested validation. The second layer (M1) used strictly cross-fitted histories of M0 stage probabilities to predict transitions from wake or N1 (W/N1) to N2, N3, or REM sleep (N2+) within 60, 90, or 120 s after each observed 30-s epoch. M1 eligibility was determined from M0-predicted stages during training and evaluation, without requiring polysomnography (PSG) stage labels at inference. M0 achieved 0.830 accuracy and 0.704 macro-F1. Under M0-predicted eligibility, M1 achieved AUROC values of 0.801, 0.791, and 0.786 at 60, 90, and 120 s, respectively; complementary PSG-conditioned evaluation yielded 0.814, 0.803, and 0.801. Prediction was stronger during wakefulness than N1, and the benefits of probability history were specific to the future-sleep task. Future-perturbation tests confirmed unchanged pre-boundary outputs under the tested conditions. A frozen M0 replay on three Muse S Gen2 nights further identified stage-specific limitations in cross-device transfer. These findings establish an auditable framework for temporally causal wearable EEG analysis and demonstrate the potential of short-horizon sleep-transition forecasting while identifying its current generalization boundaries.