Application of Deep Neural Networks in Electroencephalography (EEG): Classification of User Intention

A primary topic of focus for Brain-Computer Interfaces (BCI) research is to accurately classify EEG signals associated with various motor imagery (MI) activities. Difficulties such as intra- and cross-subject variability, and restricted open-access data availability make establishing firm findings challenging. Deep learning approaches have gained widespread use in various application fields, including natural language processing and computer vision. We are interested in improving the performance of traditional classification methods in EEG-MI and the reliability and robustness of EEG-MI classification utilising deep networks. This thesis presents two frameworks based on deep Convolutional Neural Networks (dCNN). The first is a spectrogram and dCNN-based method that provides the network with information about the EEG data's power spectrum. The second is a topographical map and dCNN-based method capable of retaining the spatial, temporal, and spectral information contained in the data fed to the dCNN. The spectrogram based method with a 2-class MI classification problem performed with an average model accuracy of 91.81%, across individual participants. Results for the topographical maps based method demonstrate that it is capable of robustly and accurately identifying MI for two and three-class datasets and provides a feasible method for BCI applications employing EEG-MI.

We concluded in this investigation that for two-class MI datasets the proposed topographic and dCNN framework was able to generalise across individuals in a dataset, with its best performance at 94.12%. For the three-class MI datasets the best performance was 98.48%.