Neural correlating-guided decoding via static-prototype contrastive Mamba and spatiotemporal learning

Although substantial deep neural network models have been proposed to decode the electroencephalography (EEG) signals in brain-computer interface (BCI), their performance and explainability remain constrained by the inability to explicitly capture task-relevant neural patterns. In traditional BCI research, neural correlating analysis is a critical preliminary step to characterize shared neural signatures embedded in EEG signals. Motivated by it, we explore to propose an intuitive neural correlating-guided approach to endow each EEG sample with explicit task-aligned neural correlating patterns. To this end, we present a two-stage STMamba model for neural correlating-guided decoding via contrastive Mamba and spatiotemporal learning. First, a static-prototype contrastive Mamba module is designed to guide the samples to learn the neural correlating patterns from the task-relevant prototypes. Second, a multi-scale spatiotemporal framework is introduced to capture both local and global spatiotemporal neural representations for decoding. To validate the performance of STMamba, we conduct the experiments on two public datasets (BCI Competition IV 2a and BCI Competition IV 2b) and one laboratory dataset. The visualization analyses show the effectiveness of the contrastive learning of neural correlating patterns. Experimental results demonstrate that STMamba achieves average classification accuracies of 82.21%, 87.59% and 82.50% for 4-class, 2-class and 6-class classifications, respectively, outperforming a suite of state-of-the-art methods reported in recent years. These findings underscore the efficacy of STMamba in advancing EEG-based neural decoding for BCI applications.