Long-term psychological stress can highly influence brain structure and functions. However, there are only a few studies using electroencephalogram (EEG) that have examined this fact. The current study demonstrates a brain-computer interface (BCI) to classify EEG correlates of long-term mental stress in various mental states. The study was performed on 26 healthy right-handed university students and the examination period was considered as a long-term mental stressor. Two groups of subjects were selected based on their stress levels evaluated by the perceived stress scale (PSS-14). The subjects' EEG data were collected during their eyes-open resting state and while they were exposed to positive and negative emotional stimuli scored by self-assessment manikin questionnaire (SAM). Several types of features were extracted from EEG data including power spectrum density (PSD), laterality index (LI), correlation coefficient (CC), Canonical correlation analysis (CCA), magnitude square coherence estimation (MSCE), mutual information (MI), phase-slope index (PSI), Granger causality (GC) and directed transfer function (DTF). Subsequently, the extracted features were discriminated using several types of classifiers including k- nearest neighbor (KNN), support vector machine (SVM) and naive Bayesian (NB) classifiers. The proposed BCI was validated by one leave-out method and investigation was done in different time windows using low and high-frequency resolutions, 7 and 36 frequency bands respectively. The results showed that the proposed system can accurately recognize subjects' stress levels in various mental states. Moreover, the MI as a functional and DTF as effective connectivity methods yield the highest classification accuracy compared to other feature extraction methods.