Multimodal Sleep Stage Classification Method Based on Multi-channel Convolution and Neural Ordinary Differential Equations

Sleep is a critical state characterized by reduced mental and physical activity, with distinct brain wave patterns observed via electroencephalogram (EEG) and electrooculogram (EOG). Adequate sleep is essential for physiological and mental health, while poor sleep quality can impair attention, immunity, and increase chronic disease risk. Accurate sleep staging is vital for assessing sleep quality, but traditional manual methods are time-consuming and subjective. Existing automatic sleep staging approaches struggle to capture temporal dynamics, multi-scale signal characteristics, and multimodal data correlations, often yielding low accuracy, particularly with imbalanced sleep stage distributions. This study proposes a novel hybrid architecture integrating a multi-scale convolutional feature extractor with a Neural Ordinary Differential Equation (Neural ODE) network to enhance classification accuracy by capturing multi-scale temporal and dynamic features of EEG and EOG signals. A Transformer-based multimodal feature fusion mechanism leverages self-attention to establish cross-modal associations, improving feature integration. Additionally, a focal loss function is introduced to address category imbalance by dynamically adjusting weights, enhancing recognition of minority sleep stages. The proposed method offers significant improvements in automatic sleep staging, with potential for clinical and research applications.