Enhancing Subject-Independent Accuracy in fNIRS-based Brain-Computer Interfaces with Optimized Channel Selection

TL;DR

This work tackles the challenge of achieving high subject-independent accuracy in fNIRS-based BCIs while minimizing hardware channels. It introduces a multi-domain feature extraction pipeline combined with a Pearson-correlation-based channel-selection strategy, enabling high performance with as few as two channels. Empirical results on a 52-channel MA dataset show a 28.09% average accuracy improvement over a prior approach and a peak 95.98% accuracy with channels 26 and 43, highlighting strong potential for efficient, online subject-generalizable fNIRS-BCI systems. The approach offers a practical blueprint for deploying compact, interpretable, and accurate fNIRS-based binary BCIs in real-world settings.

Abstract

Achieving high subject-independent accuracy in functional near-infrared spectroscopy (fNIRS)-based brain-computer interfaces (BCIs) remains a challenge, particularly when minimizing the number of channels. This study proposes a novel feature extraction scheme and a Pearson correlation-based channel selection algorithm to enhance classification accuracy while reducing hardware complexity. Using an open-access fNIRS dataset, our method improved average accuracy by 28.09% compared to existing approaches, achieving a peak subject-independent accuracy of 95.98% with only two channels. These results demonstrate the potential of our optimized feature extraction and channel selection methods for developing efficient, subject-independent fNIRS-based BCI systems.

Figures (8)

• Figure 1: Flow chart of our methods.
• Figure 2: Temporal variations in the concentration changes of oxy-hemoglobin across our selected 9 channels during a MA task for a single subject. Each channel's data is represented by a distinct color, with colors cycling through a predefined set for clarity. Vertical red lines indicate the start or end of an experimental trial, serving as markers to differentiate between rest periods and active MA phases. The x-axis denotes the sample number, serving as a proxy for time, while the y-axis indicates the amplitude of the oxy-hemoglobin concentration changes. Considering the time delay of the hemodynamic response, there are data rounding operations. The observed fluctuations provide insights into the cerebral hemodynamic responses associated with the cognitive demands of the task.
• Figure 3: Histogram comparison of values derived from adjacency matrices based on mutual information (left) and Pearson correlation (right). While the average values from the mutual information are notably smaller than those from the Pearson correlation, both histograms exhibit a congruent trend in counts as the values increase, underscoring the parallelism in their distribution patterns.
• Figure 4: Comparative visualization of adjacency matrices constructed using mutual information (top) and Pearson correlation (bottom). While the mutual information-based matrix exhibits a generally lighter colormap, the alignment of its bright and dark regions with those of the Pearson-based matrix is evident. Specifically, regions of low Pearson correlation align well with areas of low mutual information, and vice versa for high values, highlighting the consistency between the two methods.
• Figure 5: This figure illustrates the 4x2 set of confusion matrices representing the performance of SVM and LDA models trained with various types of fNIRS data: oxy-Hb, deoxy-Hb, total-Hb, and a combination of all. The first row displays the confusion matrices of the SVM models: from left to right, the models are trained using oxy-Hb, deoxy-Hb, total-Hb, and a combination of all types of fNIRS data respectively. Similarly, the second row represents the LDA models trained with the same sequence of data types. Each matrix provides insights into the model's ability to correctly classify the instances, offering a visualization of true positive, true negative, false positive, and false negative counts.