The discrete nature of traditional sleep staging limits the objective quantification of sleep quality, posing a significant barrier to applications such as optimizing sensory stimulation interventions. To address this, we propose a two-stage methodology to derive a continuous sleep depth score. First, an interpretable, deterministic Hybrid Rule Model generates physiologically-grounded sleep depth labels by integrating expert-annotated AASM stages with EEG spectral power. Subsequently, a 1D-Convolutional Neural Network (1D-CNN) is trained on these labels using a 38-dimensional feature vector to establish a practical, end-to-end estimation model. The rule-based labels demonstrated high physiological validity, correlating significantly with delta (r = 0.400) and alpha (r=-0.355) power. The trained 1D-CNN accurately replicated these labels on the SHHS dataset (R2 = 0.791, r = 0.894) and exhibited robust inter-subject generalization. This teacher-student framework provides a validated method for continuous sleep depth estimation, enabling more granular and objective sleep analysis than is possible with conventional staging.
Otsuka et al. (Wed,) studied this question.