Functional near-infrared spectroscopy (fNIRS) has been applied to brain-computer interfaces (BCIs) in many studies because of its simplicity of use and portability. However, scalp-hemodynamics creates artifacts that often contaminate fNIRS signals and substantially degrade the signal-to-noise ratio of functional signals. Although some studies have reported methods for reducing these artifacts, no study has investigated their effects on BCIs. Previously, we proposed to remove these artifacts using a method that estimates the global scalp-hemodynamic component from a minimal number of short source-detector distance channels (Short-channels), and removes its influence from standard source-detector distance channels using a general linear model (GLM) that incorporates the scalp-hemodynamics in the design matrix. Here, we investigated the effects of applying scalp-hemodynamic reduction to classify four actions: grasping a ball with the right, left, or both hands, or resting. We used a support vector machine (SVM) and binary-tree multi-classification, and compared five types of Delta Oxy-Hb features: time samples of raw data, of data after subtracting out scalp-hemodynamics, and of the estimated scalp-hemodynamics themselves, and GLM beta values for the cerebral-hemodynamic component obtained using a standard GLM without the scalp-hemodynamic model and those obtained using our proposed GLM. Results showed that the proposed method successfully improved the signal-to-noise ratio of Delta Oxy-Hb signals, and the beta values estimated by the proposed method showed the highest accuracy for classification. Thus, reduction of scalp-hemodynamic artifacts using our method may make fNIRS-BCIs more accurate.