The performance of a brain computer interface (BCI) system is usually degraded due to the outliers in electroencephalography (EEG) samples. This paper presents a novel outlier detection method based on robust learning of Gaussian mixture models (GMMs). We apply the proposed method to the single-trial EEG classification task. After trial-pruning, feature extraction and classification are performed on the subset of training data, and experimental results demonstrate that the proposed method can successfully detect the outliers and therefore achieve more reliable result.