Samenvatting

Sleeping is an important part of human life as multiple sleep restriction studies have shown. However, its function is not yet completely understood even though researchers have been performing sleep experiments for years. These experiments require sleep stage scoring, which as of this moment is done visually by trained experts. However, this method has multiple drawbacks. In order to attempt to resolve these challenges, an automatic sleep scoring procedure is proposed in this document. Using Random Forest and K-nearest neighbor classifier algorithms, 4 separate models, for a 3 class (Wake, NREM, and REM) and a 6 class (Wake, S1, S2, S3, S4, REM) each, were trained. These models were trained on several feature sets: time domain features, Autoregressive (AR) model coefficient features, and Continuous Wavelet Transform (CWT) spectral band features. Furthermore, a Hidden Markov Model (HMM) was trained on a combined feature set where each sleep stage represented a hidden state. This allowed for a comparison between unsupervised and supervised algorithms based on the same feature set. Finally, an overall comparison of the methods concludes that the 3 class RF model based on the CWT feature set had the highest accuracy of 88.8% under the Cz-Fpz EEG channel.