Automatic sleep scoring: A deep learning architecture for multi-modality time series

• Published in: Journal of neuroscience methods Vol. 348; p. 108971
• Main Authors: Yan, Rui, Li, Fan, Zhou, Dong Dong, Ristaniemi, Tapani, Cong, Fengyu
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 15.01.2021
• Subjects: Automatic sleep scoring; Deep learning; Multi-modality analysis; Polysomnography; Multi-modality analysis; Deep learning; Automatic sleep scoring; Polysomnography
• ISSN: 0165-0270; 1872-678X; 1872-678X
• DOI: 10.1016/j.jneumeth.2020.108971

•A deep learning architecture is proposed to automate sleep scoring using multi-modality PSG signals.•A linear activation function is adopted in the first CNN layer to accommodate different numbers of input channels, which helps to addresschannel mismatches.•One LSTM module and two CNN moduleswith different kernels sizesare employedtocapture information across temporal and spatial scales.•The proposed model achieves good performanceon three disparate datasets withdifferent subject attributions, thereby demonstrating model generalizabilityon different disease populations.•Model transferability is demonstrated across three datasetswith different input channels and signal modalities. Sleep scoring is an essential but time-consuming process, and therefore automatic sleep scoring is crucial and urgent to help address the growing unmet needs for sleep research. This paper aims to develop a versatile deep-learning architecture to automate sleep scoring using raw polysomnography recordings. The model adopts a linear function to address different numbers of inputs, thereby extending model applications. Two-dimensional convolution neural networks are used to learn features from multi-modality polysomnographic signals, a “squeeze and excitation” block to recalibrate channel-wise features, together with a long short-term memory module to exploit long-range contextual relation. The learnt features are finally fed to the decision layer to generate predictions for sleep stages. Model performance is evaluated on three public datasets. For all tasks with different available channels, our model achieves outstanding performance not only on healthy subjects but even on patients with sleep disorders (SHHS: Acc-0.87, K-0.81; ISRUC: Acc-0.86, K-0.82; Sleep-EDF: Acc-0.86, K-0.81). The highest classification accuracy is achieved by a fusion of multiple polysomnographic signals. Compared to state-of-the-art methods that use the same dataset, the proposed model achieves a comparable or better performance, and exhibits low computational cost. The model demonstrates its transferability among different datasets, without changing model architecture or hyper-parameters across tasks. Good model transferability promotes the application of transfer learning on small group studies with mismatched channels. Due to demonstrated availability and versatility, the proposed method can be integrated with diverse polysomnography systems, thereby facilitating sleep monitoring in clinical or routine care.