Universal Semantic Feature Extraction from EEG Signals: A Task-Independent Framework

Extracting universal, task-independent semantic features from electroencephalography (EEG) signals remains a challenging and underexplored frontier. This contribution presents a novel unsupervised framework that integrates Convolutional Neural Networks(CNNs), Autoencoders, and Transformers to capture both low-level spatial-temporal patterns and high-level semantic representations. By leveraging self-attention and extensive unsupervised training, our proposed model is robust against inter-subject variability and generalizes effectively across diverse EEG paradigms: motor imagery (MI), steady-state visually evoked potentials (SSVEP), and event-related potentials (ERP, specifically P300). Experimental results highlight the state-of-the-art performance, with average accuracies of 83.50% and 84.84% on the MI datasets (BCICIV 2a and BCICIV 2b), 98.41% and 99.66% on SSVEP datasets (Lee2019-SSVEPand Nakanishi2015), and an average AUC of 91.80% across eight ERP datasets. Clustering and correlation analyses confirm the interpretability and structural consistency of the extracted features, while ablation studies suggest the near-optimality of the chosen architecture. Though computational demands and the data-intensive nature of Transformers pose challenges, our work sets a new standard in EEG analysis, establishing a foundation for universal semantic feature extraction across tasks and contexts.