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

Yuxin Li, Hao Fang, Wen Liu, Chuantong Cheng, Hongda Chen · 2025

Read Enhancing Subject-Independent Accuracy in fNIRS-based Brain-Computer Interfaces with Optimized Channel Selection by Yuxin Li, Hao Fang, Wen Liu, Chuantong Cheng, Hongda Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-18719.txt.

Introduction

Yuxin Li, Hao Fang, Wen Liu, Chuantong Cheng, and Hongda Chen raise the accuracy of a light-based brain-computer link, in a 2025 paper. The link reads blood-flow changes near the scalp with near-infrared light, letting users steer devices by thought. A system that works for a brand-new user, with no personal tuning, is the hard goal — and fewer sensors would make it cheaper.

Their method pairs new signal features with a channel-picking algorithm based on correlation between sensors. On a public data set, average accuracy rose by about twenty-eight percent over earlier methods. With only two channels, accuracy peaked near ninety-six percent for unseen users. Sparse, accurate setups bring thought-controlled aids closer to daily use.

Catalog record

CREATORYuxin Li, Hao Fang, Wen Liu, Chuantong Cheng, Hongda Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-18719

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enhancing Subject-Independent Accuracy in fNIRS-based Brain-Computer Interfaces with Optimized Channel Selection Yuxin Li, Hao Fang, Wen Liu, Chuantong Cheng, Hongda Chen 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. \FAILED \FAILED I Introduction The advancements in the field of brain-computer interfaces (BCI) have fueled scholarly enthusiasm for the study of cerebral activities. Studying cerebral activity enables interaction between the human cerebral and external devices, with significant contributions to fields such as rehabilitation medicine, psychology, and military. Commonly employed methodologies for evaluating cerebral activity across distinct task states encompass functional near-infrared spectroscopy (fNIRS), functional magnetic resonance imaging (fMRI), and electroencephalography(EEG) [ 1 ] . Notably, fNIRS boasts distinct advantages over alternative neuroimaging modalities, primarily manifesting in its inherent portability, safety profile, and capacity to withstand environmental electrical noise. Additionally, fNIRS exhibits heightened resilience to motion artifacts and electromyographic (EMG) interferences when juxtaposed with methodologies such as EEG [ 2 ] . Conversely, while fMRI has been prolific in elucidating the blood oxygenation level-dependent (BOLD) signal, particularly during visual stimuli, fNIRS has unequivocally established correlations between alterations in blood hemoglobin concentration and the presence or absence of external stimulation [ 2 , 3 , 4 ] . This positioning of fNIRS as a compelling choice for scrutinizing diverse task states within the intricacies of the human brain is thus underscored. fNIRS devices have a simple structure, and the measurement principle is not overly complex. An fNIRS device typically includes multiple light sources and detectors, placed on the scalp. The emitted near-infrared light passes through the scalp and brain tissue before being captured by the detectors. By examining how light is absorbed and scattered, it becomes possible to deduce alterations in oxygen levels in different brain regions, thus enabling the study of brain function. fNIRS employs near-infrared light, situated between visible and infrared wavelengths, to measure changes in the concentration of hemoglobin and oxyhemoglobin in brain tissue [ 5 ] . When specific brain regions are active, blood flow increases, leading to variations in oxygen levels [ 6 ] . fNIRS calculates these oxygenation changes by analyzing differences in the intensity of incoming and outgoing light. The pivotal pursuit of discerning neural activation patterns in the human brain during diverse task states constitutes a fundamental thrust in the practical advancement of BCIs. By measuring hemoglobin concentration changes, fNIRS provides insights into neural activities, positioning it as a promising tool for various BCI tasks, including cognitive, motor, visual, and visuo-motor functions. [ 7 , 8 , 9 , 10 , 11 , 12 , 13 ] . In 2009, Huppert, T. J., et al [ 14 ] , utilized Support Vector Machines to classify brain activity during the execution of working memory tasks, achieving an accuracy rate of approximately 85%. In 2017, Hong, K. S. [ 15 ] , employed Convolutional Neural Networks (CNN) for classifying experiments that involved controlling external devices through brain activity, achieving an accuracy rate exceeding 90%. In 2018, Santosa, H. [ 16 ] , used Linear Discriminant Analysis to classify subjects based on different emotional states, achieving an accuracy rate of over 80%. While numerous studies [ 17 ] have delved into enhancing the performance of fNIRS-based BCIs, the dual challenge of subject-independent accuracy and online classification remains largely unaddressed. Most existing methodologies either focus on one aspect, sacrificing the other, or require a large number of subjects and extensive individual calibration sessions, which are not feasible for real-world applications [ 18 , 19 , 20 ] . Depending on specific research questions and interests, different types of tasks and classifiers can be selected for fNIRS users. However, there are still some challenges in this field: One of the paramount challenges in the development and deployment of fNIRS-based BCI devices is achieving high subject-independent accuracy. This refers to the algorithm’s ability to maintain consistent classification accuracy across different individuals, ensuring that its recognition capabilities are not confined to subjects-dependent but can be generalized across a diverse population [ 21 ] . Such generalization is crucial, as individual differences in brain anatomy, physiology, and cognitive strategies can introduce variability in neural signals. Moreover, the demand for online classification, which involves real-time processing and interpretation of neural data, further complicates the BCI design. Our study pioneers a novel optimization scheme specifically tailored for binary classification tasks to improve the accuracy of classification. This scheme aims to discover the characteristics of fNIRS data and develop a feature extraction scheme that aligns with hemodynamics for training machine learning classifiers. These features include statistical features, time-domain features, frequency-domain features, and principal component features. Extracting statistical features for different brain regions can construct more accurate classification models. The strength of our scheme lies in its holistic approach, amalgamating diverse feature sets that collectively offer a rich, multi-faceted view of the data. This ensures that our models are informed by temporal, statistical, and spectral variations of the signals, positioning them for superior performance in classification tasks. When conducting channel selection, confronted with the complexity of a vast number of channels, we employ a Pearson correlation-based approach to construct an adjacency matrix. This enables us to quantify the information content within each channel, with the objective of minimizing information loss while reducing the number of channels. This reduction is aimed at enhancing classification accuracy across distinct task states. Central to our approach is a strategic reduction in hardware components, ensuring their optimal positioning, coupled with an innovative feature extraction methodology that adeptly captures the unique attributes of fNIRS signals. To validate the effectiveness of our scheme, we conducted experiments on a 52-channel binary classification fNIRS dataset related to mental arithmetic. It’s noteworthy that our experimental setup maintained consistency with G. Bauernfe et al. [ 22 ] in terms of utilizing the Linear Discriminant Analysis (LDA) algorithm and the same dataset. With this setup, the introduction of our feature extraction scheme led to a notable 28.09% improvement in average accuracy. This improvement is the mean value derived from three sets of experiments using oxy-Hb data from three ROIs. Specifically, for ROI 1 , ROI 2 , and ROI 3 , the accuracy enhancements were 25.44%, 29.56%, and 29.28%, respectively. Furthermore, our channel selection algorithm enabled us to achieve a peak subject-independent accuracy of 95.98% using only two channels: channel 26 in the anterior prefrontal cortex (aPFC) and channel 43 in the right dorsolateral prefrontal cortex (r.DLPFC) on the same dataset. The rest of this paper is organized as follows. The second section presents the methods used in our scheme. The third section describes the experimental setup of our research, including the introduction of the dataset used, experimental parameter settings, and evaluation metrics. The fourth section presents the experimental results. The fifth section

Read the complete text