The Random Forest with Hybrid Principal Component Analysis for detecting Healthy and Insomnia individuals in three subtypes of Cyclic Alternating Pattern Sleep Study

Sleep disorders such as insomnia, sleep bruxism, narcolepsy, and sleep-disordered breathing have different sleep quality parameters such as cyclic alternating pattern phases and A subtype phases. This study aims to analyze and predict the A phase subtypes from electroencephalography (EEG) data. The EEG dataset for 27 (8 control, 2 bruxism, 8 insomnia, 5 narcolepsy, and 4 sleep-disordered breathing) people were downloaded and processed. The re-reference method is a common average reference (CAR), and the filtering method is a basic FIR method with lower and higher edge of frequency pass bands of 0.5 and 40 Hz, respectively. The sinusoidal artifacts (line noise) were removed with a CleanLine plug-in. The remaining noises were execrated visually and then Independent Component Analysis (ICA) with ICLabel plug-in. The Power Spectral Density (PSD) Parameters such as 0.5 to 2 Hz for Delta, 4 to 6 Hz for Theta, 8 to 12 for Alpha, 18 to 22 Hz for Beta and 30 to 60 for Gamma brainwaves, alpha peak frequency, and alpha asymmetry were calculated with eegstats plugin. The hybrid principal component analysis (HPCA) was used to estimate the functional, longitudinal, and location effects, and functional principal components analysis (FPCA) was estimated to reach at least 90% fraction of variance (FVE). Finally, the random forest with HPCA and FPCA were compared. The CAP time, NREM time, and CAP Rate were calculated. The PSD difference between groups is not statistically significant for most of the frequency bands, meaning they did not show any differences between different groups. The HPCA is estimated for each dimension with 90% FVE. The result of random forest on training and testing dataset with first to nine HPCA and FPCA eigenfunctions of time domain showed that in all subtypes, the HPCA performs better than FPCA according to the accuracy, sensitivity, and specificity. The random forest with HPCA had the best performance while the random forest without FPCA had an accuracy of about 40% in training and testing.