Efficient EEG Imagined Speech Decoding Using Channel Selection and Residual Attention-Based Lightweight Model
TL;DR
A novel decoding framework, ERP-guided dynamic wavelet phase-statistical residual network (ERP-DWPS-ResNet), is introduced that combines ERP-guided channel selection, dynamic wavelet (DW)-based signal augmentation, phase-statistical feature fusion, and a lightweight residual attention-based deep learning model for classification.
Abstract
Imagined speech is the internal articulation of words without vocalization. It offers a path for developing silent communication systems through noninvasive brain–computer interfaces (BCIs). The electroencephalographic (EEG) signals are high-dimensional, noisy, and nonstationary, making classification challenging. The primary issues arise from processing data from multiple EEG channels, many of which may contain irrelevant information. Various channel selection strategies have been investigated. Event-related potential (ERP)-based, statistical, entropy-based, spectral, and connectivity-driven approaches reduce computational overhead and enhance signal relevance. A hybrid strategy combining insights from multiple domains was also explored to select the most discriminative channels. Complementing this, a comprehensive investigation of multiple phase-based feature combinations was studied, such as those integrating phase with spectral power, entropy, Hjorth parameters, statistical descriptors, coherence, and time–frequency representations. These features capture dynamic phase interactions and neural synchrony patterns essential for imagined speech. To unify these contributions, a novel decoding framework, ERP-guided dynamic wavelet phase-statistical residual network (ERP-DWPS-ResNet), is introduced. This architecture combines ERP-guided channel selection, dynamic wavelet (DW)-based signal augmentation, phase-statistical feature fusion, and a lightweight residual attention-based deep learning model for classification. For classification, a novel lightweight deep learning model named residual hierarchically converging spatial and channel attention network (Res-HCSATNet) is introduced. The proposed framework achieves classification accuracies of 85.0%, 83.0%, 84.5%, and 94.1% for long words, short words, vowels, and short–long word pairs, respectively. Compared with transformer and BiLSTM baselines, the proposed Res-HCSATNet achieves comparable or higher accuracy with only 161 191 trainable parameters and an inference latency of 0.78–0.84 ms.