MDAGCN: A multimodal dynamic adaptive graph convolutional network for sleep staging

Existing deep-learning-based sleep staging frameworks frequently rely on shared multimodal feature extractors, which may overlook modality-specific discriminative traits and insufficiently capture dynamic cross-modal dependencies. This study aims to improve automated sleep stage classification by proposing the Multimodal Dynamic Adaptive Graph Convolutional Network (MDAGCN) for heterogeneous polysomnography (PSG) analysis. MDAGCN employs specialized branches to capture the spatio-temporal characteristics of EEG, EOG, EMG, and ECG signals, and integrates multi-scale convolutions, lead-wise attention, and adaptive graph learning to model complementary multimodal interactions within a unified framework. Extensive evaluations on ISRUC-S3 and Sleep-EDF-78 demonstrate competitive performance. On ISRUC-S3, MDAGCN achieves an accuracy of 0.841, an F1-score of 0.825, and a Cohen's kappa of 0.795, matching the best reported accuracy and kappa and obtaining the highest F1-scores for the W, N2, and N3 stages. On Sleep-EDF-78, MDAGCN achieves the highest accuracy of 0.828 and Cohen's kappa of 0.766, improving over the strongest baseline by 0.3 and 0.6 percentage points, respectively. These results demonstrate the effectiveness of MDAGCN for multimodal sleep staging and highlight the importance of jointly modeling modality-specific characteristics and cross-modal physiological dependencies.