UMAP for Dimensionality Reduction in Sleep Stage Classification Using EEG Data

Sleep is vital for wellness, and electroencephalography (EEG) serves as an instrumental tool in the study of sleep. Sleep is classified into four stages: stages N1-N3, and rapid eye movement (REM). To acquire effective and robust EEG features for sleep detection and analysis, we explore the dimensionality reduction effects of Uniform Manifold Approximation and Projection (UMAP) on various features of the EEG signals. Compared with traditional band power analysis, UMAP demonstrated higher accuracy for sleep stage classification and better reliability. Using UMAP, we observed an average of 11% increase in accuracy and an average of 20% increase in Macro-F1 Score on the same dataset. Particularly, in the wakefulness stage, Macro-F1 Score increased by 23%. Moreover, the 2D visual analysis revealed the outstanding ability of UMAP to cluster EEG signals after significant dimensionality reduction of the data.