EEG imagined speech neuro-signal preprocessing and deep learning classification

Abstract

This study presents an advanced approach for classifying imagined speech from Electroencephalography (EEG) signals, leveraging deep learning architectures and tailored preprocessing techniques. Five Convolutional Neural Network (CNN)–Long Short-Term Memory (LSTM) hybrid architectures are proposed and investigated, to extract spatial and temporal features in EEG signals, in conjunction with a proposed six-phase preprocessing pipeline combining Independent Component Analysis (ICA) for artifact attenuation with zero-phase Frequency-Domain Filtering (FD-F) and adaptive normalization. The proposed approach is evaluated across single- and multi-category classification and across multiple cross-validation strategies including random splits, GroupKFold and Leave-One-Subject-Out (LOSO) using weighted metrics, per-class, and per-subject analysis. Experiment results demonstrate the superior performance achieved by FD-F, and that by integrating the most effective proposed bidirectional temporal modeling architecture CNN-2-Bi-LSTM, with the proposed preprocessing pipeline, the approach achieves higher accuracy (exceeding 99%) for 30-class classification maintaining cross-subject generalization against state-of-the-art.