A novel algorithm integrating multimodal information for improving OSA-related sleep staging.

PURPOSE: Obstructive Sleep Apnea (OSA) has severely disrupted the sleep structure of OSA patients, resulting in a decline in sleep staging performance for these patients. The majority of the automatic sleep staging methods have not addressed this issue. Therefore, this article proposes a neural network to enhance the sleep staging result for OSA patients. METHODS: By leveraging the easily extensible nature of Broad Learning System (BLS), we construct a dual-branch, multi-context and multimodal integration framework for sleep staging. The core of the proposed framework lies in its disease-aware design: it explicitly embeds SpO2-derived OSA-related physiological information into the staging pipeline. This targeted optimization effectively improves the sleep staging accuracy for OSA patients. RESULTS: The model was tested on more than 1.45 million samples across three heterogeneous public datasets. The proposed model demonstrated improvement in sleep staging performance for OSA patients, particularly those with severe OSA. Compared to the model lacking OSA-related enhancement, this model achieved increments of 0.18% in ACC, 1.14% in MF1, and 0.006 in Kappa for severe OSA patients in all three datasets. In addition, the sleep staging performance of our model achieved the effective and competitive performance compared with peer studies. CONCLUSION: The results validate our proposed model as an effective and competitive approach for enhancing sleep staging in OSA patients.