This paper determined the best combination that maximizes the classification accuracy of single-channel electroencephalogram (EEG)-based motor imagery brain–computer interfaces (BCIs). BCIs allow people including completely locked-in patients to communicate with others without actual movements of body. Whereas EEGs are usually observed by multiple electrodes, single-channel measurement has been recently studied for gaining the simplicity of use. However, existing single-channel BCI studies have evaluated the performance on their own, private datasets that are not accessible from other researchers. Therefore, it remains a practical challenge to determine the optimal combination of channel, feature, and classifier using a public dataset. For the assessment, we used an open-access database (BCI competition IV dataset 2a) and a 10-fold cross-validation procedure. We found that support vector machine or multilayer perceptron with power spectrum or single-channel common spectral patterns of C3 or C4 position showed high classification accuracies in all subjects (mean: 63.5±0.4%, maximum: 86.6±0.4%).