Multimodal Polysomnography Based Automatic Sleep Stage Classification via Multiview Fusion Network

• Published in: IEEE transactions on instrumentation and measurement 2024-01, Vol.73, p.1-1
• Main Authors: Lin, Yiwei, Wang, Mengying, Hu, Fengdan, Cheng, Xu, Xu, Jinshan
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Classification; Deep Learning; Diagnosis; Effectiveness; Electroencephalography; Feature extraction; Feature maps; Frequency domain analysis; Machine learning; Modules; multi-modal physiological signals; Physiology; Quality assessment; Redundancy; Representations; Signal resolution; Sleep; sleep stage classification; Source code; Task analysis; Time-frequency analysis
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2023.3343781

Sleep staging is a standard diagnostic method for evaluating sleep quality, which would enable early diagnosis of sleep disorders as well as mental diseases. Polysomnography (PSG), a set of physiological signals recorded externally, is a standard media for sleep staging. Developing automatic algorithms to analyze the PSG signals with the purpose of better sleep staging is demanding, as manual assessment is tedious and time-consuming. However, it is challenged by the noisy nature of the PSG signals, i.e., features contributing to sleep staging are embedded in different types and scales of signals in both time and frequency domains. In this paper, we propose a hybrid deep learning architecture that uses multi-modal PSG signals, specifically electroencephalography (EEG) and electrooculogram (EOG), and their frequency representations as inputs, to accomplish sleep stage classification tasks. To this end, we design the Multi-Scale Local Feature Extractor (MSLFE) with a multi-branch CNN of different convolutional kernel sizes and the Global Relationship Modeling(GRM) module to extract features in both time and frequency domains effectively. A Cross Lined Feature fusing (CLF) module is further introduced to enable an effective fusion of multi-modal and multi-attribute features while avoiding bidirectional representation redundancy for high-quality feature maps. We carried out a set of experiments on the SleepEDF-ST and SleepEDF-SC to validate the effectiveness of the proposed method, where classification performance in terms of precision, recall, and F1 score higher than 84% are obtained on most of the sleep stages. Comparisons with state-of-the-art methods confirm the effectiveness of the proposed method in improving sleep quality evaluation and diagnosis. Source code is available at: https://github.com/ZJUT-CBS/MMNet.