Mental workload estimation has been under extensive investigation over the years, because the capability of monitoring the cognitive workload enables the prevention of cognitive overloading and improvement of workplace safety. Electroencephalogram (EEG) signals has been found to be an objective and nonintrusive measure of mental workload. However, the evaluation of cognitive workload based on singletrial EEG data, which is an essential step towards realtime workload monitoring and braincomputer interface, has been a major challenge. Recently, a number of advanced feature extraction methods and machine learning algorithms have been employed in EEGbased mental workload assessment. In this study, we performed singletrial workload classification using the EEG data recorded during the performance of nback tasks with 2 levels of difficulty (corresponding to low and high levels of workload respectively), examined the effectiveness of 3 types of feature extraction (spectral power, discrete wavelet transform and common spatial filtering), and evaluated the performance of 4 classification algorithms (support vector machine, Knearest neighbors, random forest and gradient boosting classifiers). Our findings indicate that common spatial filtering was the bestperforming individual feature extraction method for singletrialbased workload classification, and the optimal performance was achieved by combining the features from either spectral power or discrete wavelet transform with those from common spatial filtering, and adopting the random forest classifier. This study might provide some helpful guidance on the selection of feature extraction methods as well as machine learning algorithms in mental workload evaluation based on singletrial EEG data.