SleepMambaNet: a lightweight multimodal framework for sleep staging and sleep disorder screening with multicenter validation

Abstract Objective. Manual sleep staging is labor-intensive, whereas many automated systems remain computationally demanding and are validated on a single dataset. This study aimed to develop a lightweight multimodal framework for automated sleep staging and exploratory sleep-disorder screening from polysomnographic (PSG) signals. Approach. SleepMambaNet combines efficient channel attention to integrate complementary information across PSG channels with bidirectional Mamba modules to model long-range temporal dependencies between sleep epochs. The framework was evaluated on the public ISRUC-S1 and ISRUC-S3 datasets and a multicenter clinical cohort of 185 PSG recordings. Main results. In subject-independent experiments, SleepMambaNet achieved accuracy, macro- F 1, and Cohen’s kappa of 0.852, 0.824, and 0.803 on ISRUC-S3; 0.830, 0.801, and 0.773 on ISRUC-S1; and 0.854, 0.819, and 0.806 on the combined public and clinical data, respectively. The external cross-dataset evaluation yielded accuracy, macro- F 1, and kappa of 0.828, 0.812, and 0.779. For exploratory sleep-disorder screening, the model achieved accuracy of 0.892, F 1 score of 0.887, sensitivity of 0.956, specificity of 0.842, and precision-recall area under the curve of 0.831. Significance. These results show that cross-channel physiological information and inter-epoch temporal context can be modeled effectively with only 0.47 million trainable parameters across heterogeneous public and clinical PSG data. The combination of computational efficiency and cross-dataset performance provides a practical basis for prospective evaluation of automated PSG decision support, while the binary screening output should not be interpreted as a disease-specific or standalone clinical diagnosis.