Abstract
Hybrid Brain–Computer Interface (BCI) systems that combine electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) achieve higher classification accuracy but face computational bottlenecks due to the high dimensionality of the data. This study proposes a computationally efficient channel selection framework based on time-series distance metrics—specifically Euclidean and Manhattan distances—to address the “curse of dimensionality” while preserving signal integrity. The proposed method uses a systematic electrode-pairing strategy and compares two thresholding criteria: Mean and Median. The framework was evaluated on two public datasets covering Motor Imagery (MI), Mental Arithmetic (MA), and P300 Speller paradigms. Classification performance was assessed using Support Vector Machines (SVMs), Linear Discriminant Analysis (LDA), and Random Forests (RFs), validated using 10-fold cross-validation and Wilcoxon Signed-Rank tests. The results demonstrated that the proposed method reduced the channel count by over 50% across all tasks. Notably, the Median-based thresholding proved statistically superior (p < 0.05) on noise-sensitive tasks, achieving peak accuracies of 94.36% for the P300 Speller with the LDA classifier and 72.93% for hybrid MA tasks with the SVM classifier. Furthermore, topographic heatmap analysis confirmed that the selected channels aligned with task-relevant brain regions, such as the sensorimotor and prefrontal cortices. Crucially, the channel reduction reduced decision latency to 0.11–0.20 s, validating the system’s feasibility for real-time implementation. These findings suggest that the distance-based channel selection method, particularly with Median thresholding, offers a robust and high-speed alternative to complex heuristic algorithms for multimodal BCI systems.