MetaNIRS: A general decoding framework for fNIRS based motor execution/imagery

• Published in: Neural networks Vol. 192; p. 107873
• Main Authors: Li, Yu, Sun, Yu, Wan, Feng, Yuan, Zhen, Jung, Tzyy-Ping, Wang, Hongtao
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.12.2025
• Subjects: Brain-computer interface (BCI); fNIRS; Motor execution (ME); Motor imagery (MI); PoolFormer; Motor imagery (MI); Brain-computer interface (BCI); PoolFormer; fNIRS; Motor execution (ME)
• ISSN: 0893-6080; 1879-2782; 1879-2782
• DOI: 10.1016/j.neunet.2025.107873

Functional near-infrared spectroscopy (fNIRS) is a crucial brain activity monitoring tool with remarkable potential applications in brain-computer interfaces (BCI), particularly in rehabilitation therapy for disabilities. However, the performance of fNIRS-based BCI systems remains suboptimal, such as motor execution (ME) and motor imagery (MI) decoding. Inspired by the successful application of the PoolFormer framework in visual tasks, we first proposed a novel long-range dilation multilayer perceptron (LongDilMLP) to utilize the hemodynamic characteristics of fNIRS. Furthermore, the LongDilMLP was integrated with the PoolFormer framework, called as MetaNIRS in this study. The proposed framework MetaNIRS was employed for both ME and MI classification tasks, achieving rigorous validation of its effectiveness and practical applicability. To evaluate the performance of MetaNIRS, two publicly available ME datasets (A and C) and one self-collected MI dataset (B) were employed. The experimental results demonstrated that the average accuracy were 76.00 %, 57.45 %, and 84.14 %, with cross-subject accuracy of 77.24 %, 58.55 %, and 85.52 %, respectively. Moreover, sensitivity experiments of model parameters showed the robustness. Ablation experiments highlighted the significance of each MetaNIRS component and the efficacy of LongDilMLP over traditional MLP. Additionally, visualization techniques enhanced the interpretability of MetaNIRS, indicating the main contribution of the first half signals for classification. Using the first half of signals, the average accuracy only reduced 4.30 %, 1.69 %, and 1.11 %, respectively. These findings suggest that the superior performance of MetaNIRS, which provide an efficient general decoding framework for ME and MI.