An Insomnia Detection Model using Augmented Two-Class Multichannel EEG Frequencies

Clinical diagnosis of Insomnia relies on sleep-stage information from multiple electrodes for an accurate prediction. Existing insomnia detection models achieved satisfactory performance with single-channel training samples based on sleep-stage epochs, but the performance of classifiers using multiple-channel information has not been investigated. Also, differences in sleep stage distribution might lead to non-uniformity across training subsets. This study develops a model that does not rely on sleep-stage analysis, instead using traditional signal augmentation to efficiently manage multichannel data through time-frequency analysis of mixed-frequency characteristics across various Electroencephalogram (EEG) channels. A two-class Full-Sleep Ensemble Learning (FS-EL) model is implemented using augmented dual-frequency sub-bands that characterize deep sleep across EEG frequencies. It builds new sub-band representations using frequency cropping and superposes the components in subsequent iterations, reformulating them into a two-class model. The features extracted from newly generated composite signals have been evaluated using an Ensemble Bagged Decision Tree (EBDT), Random Forests (RF), and Gradient Boosting (GB). It achieves superior performance with a composite two-class FS-EL model trained across various datasets on Out-Of-Bag (OOB) samples. FS-EL model attained classification accuracy and Area Under The Curve (AUC) of 0.99 using k-fold cross-validation on EBDT, and sensitivity of 1.0 compared to composite five-class and non-composite single-class models. The proposed model improves performance by generating new composite signals that exploit multichannel information using a subset of the raw data, achieving perfect distinction of data points and obtaining efficient sample classification.