In this paper, a novel technique for determination of the optimal feature combinations and, thereby, acquisition of the maximum classification performance for a functional near-infrared spectroscopy (fNIRS)-based brain-computer interface (BCI), is proposed. After obtaining motor-imagery and rest signals from the motor cortex, filtering is applied to remove the physiological noises. Six features (signal slope, signal mean, signal variance, signal peak, signal kurtosis and signal skewness) are then extracted from the oxygenated hemoglobin (HbO). Afterwards, the hybrid genetic algorithm (GA)-support vector machine (SVM) is applied in order to determine and classify 2- and 3-feature combinations across all subjects. The SVM classifier is applied to classify motor imagery versus rest. Moreover, four time windows (0–20�s, 0–10�s, 11–20�s and 6–15�s) are selected, and the hybrid GA-SVM is applied in order to extract the optimal 2- and 3-feature combinations. In the present study, the 11–20�s time window showed significantly higher classification accuracies – the minimum accuracy was 91% – than did the other time windows (p�<�0.05). The proposed hybrid GA-SVM technique, by selecting optimal feature combinations for an fNIRS-based BCI, shows positive classification-performance-enhancing results.