基于多模态多域特征融合的Res-P3D睡眠分期网络OA

Sleep staging is a fundamental basis for evaluating sleep quality and diagnosing sleep disorders.Most existing studies were based on single-modal and single-domain features,which made it difficult to comprehensively represent the complex physiological characteristics of sleep stages.Therefore,a multimodal and multi-domain feature fusion-based automatic sleep staging method was proposed.Firstly,electroencephalogram(EEG),electrooculogram(EOG),and electromyogram(EMG)signals were filtered into multiple frequency bands,and time-domain,time-frequency-domain,and frequency-domain features were extracted respectively to construct multi-dimensional representations.Then,residual pseudo-three-dimensional convolution was employed to extract joint temporal-spatial-spectral features from the time and time-frequency-domains,and a global context(GC)module was utilized to enhance the global context modeling capability of frequency-domain features.Thereafter,the outputs of the three branches were concatenated along the temporal axis and subsequently input into a GRU network for sleep stage classification.Experimental results show that the proposed model achieves favorable staging performance on both the ISRUC-S3 and Sleep-EDFx datasets,with accuracies of 83.6%and 78.1%respectively,verifying the effectiveness and robustness of the model.