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

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