EEG Sleep Stage Classification and Prediction using Entropy Feature Extraction

Over the past 2 decades, EEG research has rapidly advanced due to the discov eries of active electrodes, machine learning algorithms and more interest in the field of neuroscience. This has led to EEG research being much cheaper and accessible, but due to its stigma of previously being an extremely high-end research field and the sensitivity of the data required for the research, it has resulted in even more complex EEG experiments and applications. This thesis proposes a method of predicting sleep stages using a single EEG channel input BCI. The experiments were performed on Sleep-EDF’s Sleep cassette and Sleep telemetry datasets. We used entropy feature extraction to identify different sleep stages which were then fed into a classifier enabling us to then predict subsequent sleep stages. Entropy has been proven to be able to numerically depict non-linear time series data and it has been used in many other EEG BCIs. Three classifiers were compared to ascertain their various performances. SVM showed the most stable results at 75 percent prediction accuracy. While KNN has a higher maximum prediction accuracy than SVM at 88 percent accuracy, it has proven to be unstable and varies depending on the iterations and number of inputs. The final classifier used was the NN classifier which should result in the best accuracies com pared to the SVM and KNN at 87 percent or higher prediction accuracies. The results show that it is indeed possible to predict upcoming sleep stages with a single EEG channel with accuracies greater than 75 percent depending on the methods used.