Selective regularized spatial features representation learning for motor imagery EEG based on alternating cascaded model.

Feature representation plays a pivotal role in the decoding of motor imagery electroencephalograph (MI-EEG) signals. Conventional spatial representations are often hindered by the operational time-frequency and noise interferences of MI-EEG. In response to this challenge, this paper proposed a novel method to S elective extract the R egularized S patial features and represent by the A lternating C ascaded M odel (SRS-ACM) for MI-EEG decoding. Initially, regularized common spatial patterns are derived from MI-EEG samples, capturing the relevance and redundancy between spatial features and labels. Subsequently, a global optimization framework is devised to select discriminative and comprehensive dimensions with the spatial features. Finally, an alternating cascade model, integrating sparse and collaborative representations, is developed to capture the intrinsic patterns with the selected spatial features, thereby facilitating MI-EEG decoding. Experimental evaluations were conducted across three publicly available MI-EEG datasets: BCI-III dataset 4a, BCI-IV dataset 1, and BCI-IV dataset 2a. The SRS-ACM model demonstrated notable performance improvements of 2.60 %, 6.15 %, and 1.14 % on these datasets, respectively. Ablation studies underscored the necessity and significance of the components within the SRS-ACM method, while parameter sensitivity analyses validated the robustness of the SRS-ACM approach in the development of practical MI-based brain-computer interfaces. • The selected regularized CSP representations based on supervised relevance and redundancy is applied for MI-EEG signals. • A sparse and collaborative alternating cascade model is utilized to classify selected spatial representations. • The proposed method is efficient for constructing MI-BCIs. [ABSTRACT FROM AUTHOR]