Introduction of sub-band augmentation with machine learning to develop an insomnia classification model using single-channel EEG signals.

Abstract

Objective. Biological signals can be used to record sleep activities and can be used to identify sleep disorders. Insomnia is a sleep disorder that can be detected using supervised learning models developed using biological signal analysis. The baseline insomnia detection models segmented input signals based on various sleep stages, in which an imbalance in classes of the different subsets was visible.Approach. Leaning on sleep annotations for training data generation can overcome using electroencephalogram (EEG) augmentation, which trains the machine learning model based on the diverse nature of input EEG. The proposed work aims to generate a heterogeneity in the decomposed frequencies of EEG data using sub-band augmentation. The presented approach imposes the characteristics of various EEG frequencies when developing new data.Results. An excellent classification accuracy of 0.91, 0.90, and 0.866 can be visible in sub-band augmentation using signal scaling followed by noise addition and sliding window, respectively. An ensemble-bagged decision tree (EBDT) classifier was employed in developing the identification model incorporating all the sub-band augmentations with a significant accuracy of 0.986, a sensitivity of 1.0, and a specificity of 0.97. The proposed model also examines the features from smaller time segments of EEG in developing the training data for EBDT and shows an accuracy, sensitivity, and specificity corresponding to 0.97, 0.95, and 1.0.Significance. The presented model is simple, independent of supplementary data like sleep annotations describing sleep stages, and more suitable for disease detection bearing small datasets in training-data enhancement for classification.