Deep learning approaches to EEG and fMRI data: a comparative study for sleep stage classification

Abstract

In this thesis, for classification of sleep stages, we use deep learning techniques with the help of data from fMRI and EEG. The ConvLSTM models are applied for the fMRI data. The data for the EEG is worked on with the LSTM and Bidirectional LSTM. This can hence be seen as a work of optimizing the accuracy, the precision, and the generalizability of all these models with one another. The baselines for all these different types of data are built up using initial models. The LSTM baseline model has given an accuracy of 78.69% on testing for sleep staging with W (Wake), NREM-1, NREM-2, and NREM-3 using EEG data, which is highly effective with such data resolution in time. Meanwhile, the Bidirectional LSTM model performs better preprocessing for the temporal aspect and hence yields, on average, 80.60% accuracy for general classification on the same stages. This would make it a model that can capture the dynamic nature of the EEG data across these particular stages. In contrast, processed fMRI data starts with a 76.82% testing accuracy, while performance is readjusted based on the feature extraction spatial-temporal settings adopted in their ConvLSTM configurations to classify sleep stages W, NREM-1, NREM-2, and NREM-3, with special attention to the role of model configuration. These results show that functionalities of tailored deep learning play the most basic role in the high complexity domain of sleep stage classification. These findings are prerequisite for the future development of this area.