Optimal Electroencephalogram and Electrooculogram Signal Combination for Deep Learning-Based Sleep Staging

Objective: The traditional sleep staging involves manual scoring of electroencephalogram (EEG), electrooculogram (EOG), and electromyogram signals during polysomnography (PSG), which is laborious and susceptible to human errors. Previous studies have explored several automated sleep staging methods utilizing fewer EEG signals, for which deep-learning methods have shown promising results. Despite the availability of various signals and signal combinations in PSGs, the performance of different signal combinations for accurate sleep staging is not fully explored. We hypothesize that various EEG signal combinations will yield comparable performances, thus accurate automatic sleep staging could be achieved with a simplified measurement setup. We aim to identify the optimal EEG and EOG signal combinations for deep learning-based automatic sleep staging. Methods: Four EEG signals (F4-M1, C4-M1, O2-M1, and C3-M2) and one EOG signal (E1-M2) from 876 suspected obstructive sleep apnea subjects were studied. A total of 31 deep-learning models were trained utilizing different EEG and EOG signal combinations as input. The classification performance of automatic sleep staging against manual sleep staging was evaluated across five sleep stages. Results: The differences in classification performance among various EEG signal combinations were negligible, with accuracies ranging from 81% to 85% (Cohen's kappa, κ = 0.73–0.78). Incorporating the EOG signal into single EEG configurations improved accuracies by 1–2 percentage points, while the improvements were smaller when combining EOG with multiple EEG signals. Conclusions: The comparable classification performances among various signal combinations suggest that automatic sleep staging can be achieved with a simplified EEG and EOG measurement setup without compromising performance.