Functional near-infrared spectroscopy (fNIRS) is a non-invasive multi-channel imaging tool for assessing brain activities, which has shown its high potential in brain-computer interface (BCI) technique. Most previous studies have focused on constructing high dimensional features from whole channels, adding to the complexity of their classifiers. Another multi-channel source for BCI is electroencephalograph (EEG), which possesses different spatial and temporal features from fNIRS. In EEG field, Common Spatial Pattern (CSP) algorithm is widely used aimed at dimensionality reduction. In our article, we modified it based on the characteristics of fNIRS and evaluated its effectiveness in discriminating Mental Arithmetic (MA) against resting status in an open-access dataset. The Modified Common Spatial Pattern algorithm significantly outperforms CSP algorithm in fNIRS-based BCI and shows its potential in further BCI related explorations.