CSPT-SleepNet: An Automatic Sleep Staging Framework Integrating Center-Driven Sequence Supervision and Multimodal Pyramid Token Encoding

Background: Automatic sleep staging is fundamental to sleep architecture analysis and the diagnosis of sleep disorders. Sleep staging requires jointly modeling complementary information across modalities, multi-scale physiological features, and temporal relationships across consecutive epochs. In target epoch prediction from consecutive epochs, non-target epochs primarily provide contextual information, but their ground-truth labels remain underused as auxiliary supervision. Methods: We propose CSPT-SleepNet, an automatic sleep staging framework integrating center-driven sequence supervision with multimodal pyramid token encoding. A Multimodal Pyramid Token Encoder extracts hierarchical multi-scale features from electroencephalogram, electrooculogram, and electromyogram signals, followed by intra-modal, cross-modal, and inter-epoch modeling for center-epoch classification. During training, the center epoch receives primary classification supervision, while non-center epochs receive weaker auxiliary supervision based on their ground-truth labels, providing stage-informed contextual support for the prediction target. Results: In subject-independent cross-validation on Sleep-EDF-20, Sleep-EDF-78, and ISRUC-S3, CSPT-SleepNet achieved accuracy (ACC)/macro F1 (MF1) scores of 87.93%/83.53%, 85.81%/80.87%, and 84.06%/82.86%, respectively. In a separate controlled comparison on Sleep-EDF-20, center-driven sequence supervision improved ACC and MF1 by 0.60 and 0.91 percentage points, respectively, over equal-weight sequence supervision. The full model has 3.418 million trainable parameters and requires 397.10 million floating-point operations for an input sequence of nine 30-s epochs. Conclusions: CSPT-SleepNet provides an effective framework for center-epoch sleep staging from consecutive multimodal inputs. Nevertheless, N1 recognition remains challenging, with F1 scores ranging from 55.3% to 62.7% across the three datasets.