Explainable AI-Enhanced Interpretation for Inner Speech Classification Using EEG Signal

Brain-Computer Interfaces (BCIs) using Electroencephalogram (EEG) for inner speech classification have gained attention for their potential to offer a communication channel for individuals with speech disabilities. However, existing studies in this context fall short in achieving acceptable accuracy. To achieve this objective, this work introduces a methodology that enhances the accurate interpretation of EEG signal quality through a multi-step process. First, a band-pass filter combined with a notch filter is applied to attenuate noise and eliminate power line interference. Next, Independent Component Analysis (ICA) is utilized to automatically remove common artifacts. For the feature extraction process, temporal statistical features are utilized. The extracted features are classified using various machine learning and deep learning techniques for benchmarking purposes, such as Linear Discriminant Analysis (LDA), k-nearest neighbors (kNN), support vector machine (SVM), and deep convolutional neural networks (CNN). Finally, an explainable AI (XAI) algorithm is applied to further interpret the output of the classification model. This approach was applied to an existing dataset, containing four classes (up, down, right, left). The proposed approach demonstrated a notable improvement in accuracy, achieving approximately 63% for two-word and 31.5% for four-word combinations. In contrast, previous studies in the literature reported accuracy rates of just 58% and 29.67%, respectively. This enhancement highlights the effectiveness of the new methodology in processing and for inner speech classification.