Abstract:
Recently, there have been many trials to classify the sleep stage using a deep learning technique. In most previous studies, 1-D signal data have been used as input data ...Show MoreMetadata
Abstract:
Recently, there have been many trials to classify the sleep stage using a deep learning technique. In most previous studies, 1-D signal data have been used as input data and some studies have used spectrogram images. In this paper, we build an image-based dataset with the same style images as the images used by sleep scorers in order to enable the uses of state-of-the-art deep learning models and to utilize various augmentation techniques for image classification. Because our dataset is time-domain signal image data, it is easy for human experts to understand the results, unlike the spectrogram images which show frequency-domain data. Furthermore, through experiments, we find that there are differences in judging criteria for classifying sleep stages between institutions and it can cause inconsistency in classification results. To solve the inconsistency problem, the image-based deep learning model is extended further to consider the hidden information on the institution extracted from dataset. This leads to higher accuracy performance by allowing the model to consider the features of institutions. Experimental results show that the proposed method achieve the highest accuracy performance when compared to an exiting state-of-the-art model.
Published in: 2021 International Conference on Information and Communication Technology Convergence (ICTC)
Date of Conference: 20-22 October 2021
Date Added to IEEE Xplore: 07 December 2021
ISBN Information:
Print on Demand(PoD) ISSN: 2162-1233