MIASS: A multi-interactive attention model for sleep staging via EEG and EOG signals

Abstract

Sleep staging is essential for sleep analysis. Recent studies have attempted to integrate multi-modal signals such as electroencephalogram (EEG) and electrooculogram (EOG) to enhance model sensitivity. However, these attempts still face limitations in effectively fusing multi-modal signals, particularly in capturing both global and fine-grained interaction information in sleep epochs simultaneously. To address this, we propose a multi-interactive model (MIASS) that integrates two core modules, the global information interaction (GII) module and the fine-grained information interaction (FII) module. The GII module can effectively capture the global correlation paradigm in EEG and EOG at the epoch level by combining the global channel and spatial attentions with a residual network. The FII module explores the fine-grained correlation paradigm between small EEG and EOG segments within epochs using the cross-attention mechanism to achieve more fine-grained interaction information. The combination of these modules increased the accuracy of the model up to 89.2%, 86.6% and 89.7% on the SleepEDF-20, SleepEDF-78 and SHHS datasets, respectively, which outperforms the comparison models by 0.2–5.7%. The ablation study confirmed the benefits of integrating global and fine-grained correlation paradigms to enhance sleep staging performance, and the model input study demonstrated that MIASS maintains good performance under various input conditions.