Over the past several years, electroencephalography (EEG) has become widely used in many domains, e.g., epilepsy and seizure testing, sleep assessment, brain dysfunction assessment and brain-computer interface. In this thesis, EEG is used as a method of communication for patients, especially those with Total Locked-in Syndrome. In this condition, the patient cannot move or communicate verbally due to a complete paralysis of nearly all voluntary muscles in the body. Therefore, we propose the classification method of the patient's imagination for representing YES/NO answer. Although there were many classification techniques, none of them considered the non-stationary characteristic of brainwaves; thus, they cannot really be employed in real-world situations due to low classification accuracy. This research aims to tackle the non-stationary issue by modifying the traditional classifier to be stacked classifiers, the proposed new classifier constructed by a group of EEG signals that partitioned based on data distribution. We also employ an adaptive process for adjusting test data to reflect train distribution closely. The experiment was conducted on the BCI Competition IV (2b) data set. Our proposed method was compared to nine baseline techniques in three groups: (1) static algorithms, (2) variants of adaptive Pool Mean algorithms, (3) the state of the art Adaptive CSP method, Corrected Sequential EM and PMean LDA. The proposed method has shown to reduce the non-stationary characteristic of EEG signal in both training and testing data. The results showed that our proposed method significantly outperform all baselines and yield 3% accuracy better than the best baseline method.