Document resource
Objectives High-dimensional data collected by functional near-infrared spectroscopy (fNIRS) often faces the challenge of dimensional catastrophe. To this end, a cross-modal computational framework is constructed to optimize the classification model for systemic lupus erythematosus (SLE) brain function abnormalities by integrating co-spatial mode (CSP) algorithm and dynamic resampling strategy.Methods The study included 76 SLE patients and 54 healthy controls (HCs), and their 37-channel fNIRS data were collected. Firstly, oxygenated hemoglobin (HbO) signals from 14 brain regions were extracted during the 60-second task period. Then, the CSP space filter (two spatial modes are selected, m=2) is used to reduce the characteristic dimensionality of the signal. Finally, the K nearest neighbor (KNN) classifier is trained based on 20 randomly generated equilibrium subsets, and the performance is evaluated.Results The features extracted by CSP improve the classification accuracy of the model from 52.68% to 72.06% of the original data. The characteristic brain regions that contributed the most included the left frontal lobe (F3), left frontal pole (Fp1), right central frontal region (Fc4), right frontal pole (Fp2), and right frontal lobe (F2).Conclusions This study confirms that the CSP algorithm can break through the limitations of traditional frequency domain analysis, and the robustness of the classification model under small sample conditions can be significantly improved by combining the dynamic resampling strategy. This framework provides a new tool with translational potential for clinical rapid screening of SLE-related neurocognitive impairment.