A Lightweight Convolutional–Recurrent Network for Automatic Sleep Staging from Single-Channel EEG: Evaluation in Healthy and Insomnia-Related Sleep

This is the authors' abstract. We don't add key points for this paper.

Objective, repeatable assessment of sleep architecture is valuable in insomnia, but attended polysomnography is costly and poorly suited to repeated home use. Single-channel electroencephalography (EEG) of the kind recorded by wearable devices offers an alternative, yet most automatic staging models are developed on healthy sleep and are too large for embedded hardware. We describe a convolutional–recurrent network with 0.33 million parameters that stages raw single-channel EEG (Fpz–Cz). Two depthwise-separable convolutional branches encode each 30-s epoch, a bidirectional gated recurrent unit models inter-epoch context and a class-balanced loss counters stage imbalance. Under 10-fold subject-independent cross-validation on the Sleep-EDF Expanded database, the model reached 85.2% accuracy (95% CI 84.6–85.8), macro-F1 0.81 and Cohen's κ 0.80 in healthy sleepers; no significant difference was detected from the best-performing baseline, which has about four times as many parameters (McNemar p=0.21). In participants with mild difficulty falling asleep, agreement was lower (accuracy 82.7%, 95% CI 81.5–83.9; κ 0.77), with N1 the weakest stage (F1 0.50). A compact model can approach the agreement of larger models, but performance in diagnosed insomnia and on wearable hardware remains to be established.