A Self-attention-based Ensemble Convolution Neural Network Approach for Sleep Stage Classification with Merged Spectrogram

Scoring sleep data is a subjective and time-consuming. It takes more than one hour to score a whole night's PSG data. The automatic sleep staging method is needed to reduce clinical manpower. In this paper, an attention-based ensemble convolution neural network approach for sleep stage classification with merged spectrogram was proposed. All-night sleep physiological signals from 19 healthy individuals and 90 insomnia patients were used. First, the all-night polysomnography (PSG) signals including electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) were segmented into 30-sec segments. Subsequently, each segment was transformed into spectrograms by continuous wavelet transform and a simple merged processing was applied to generate the merged spectrograms with different viewpoint. Next, the three merged spectrogram groups with different viewpoint were utilized as an input of our proposed CNN with self-attention, named merged spectrogram Net (MS-Net). The three trained MS-Net models were used to form an ensemble MS-Net. The experimental results showed that the accuracy, kappa coefficient, and F1 score of the proposed method were 89.83%, 84.82%, and 85.09%, respectively. The results proved that the proposed deep learning approach, ensemble MS-Net, had highly accuracy for sleep PSG spectrogram classiflcation.
¹