Imagined Speech Classification from Electroencephalography with a Features-Guided Capsule Neural Network

Imagined speech commands can be decoded from imagined speech modulated electroencephalography (EEG) with a range of different Deep Learning (DL) methods. Among these methods, Convolution Neural Networks (CNNs) achieved an attraction due to intuitive feature extraction and classification. However, recently Capsule Neural Network (CapsNet), as the state-of-the-art DL method, has been proposed to address the shortcomings in CNN-based models. In CapsNet architecture, instead of mean/max pooling in CNNs, an alternative Dynamic Routing (DR) is employed. To elaborate, after several convolution layers, a vector representation of neurons called capsule in the first layer is grouped in DR to form the next level capsule called parent capsule so that the part-whole hierarchical relationships are modelled. In this study, we validated prior approaches used in this field and for the first time, proposed a CapsNet-based architecture called Features-Guided CapsNet (FGCapsNet) to decode speech recordings. In FGCapsNet, we modified CapsNet architecture using multi-level feature maps. The performance of the proposed method is compared with the baselines: a feature-engineered method (regularised Linear Discriminant Analysis (rLDA) using filter bank common spatial patterns (FBCSP)) and CNN-based methods. Hyperparameter optimisation is performed using nested cross-validation (Nested-CV). Furthermore, two different scenarios for datasets are considered to evaluate the performance of the models: A different pairing of imagined-words and multiclass classification of imagined-vowels. The average accuracy results corresponding to the pairing of imagined-words demonstrated that FGCapsNet significantly outperformed Deep-CNN (>7.08%), Shallow-CNN (>8.4%) and FBCSP (>11.42%) with the chance accuracy of 50% (p