Performance analysis of ensemble methods for multi-class classification of motor imagery EEG signal

Recent advances in the field of Brain-computer Interfacing (BCI) has opened wide potentials in neuro-rehabilitative applications. Electeroencephalography (EEG) is the most frequently used brain measure in BCI research. Mental states are distinguished from classifiers which uses features extracted from the raw EEG as inputs. Ensemble classifiers combine a number of classifiers or learners to improve the classification results. It is more suited for multi-class classification of time-varying EEG signal. In this paper, we have used AdaBoost, LPBoost, RUSBoost, Bagging and Random Subspaces for classification of 3-class motor imagery EEG data. For this purpose, we have employed adaptive autoregressive coefficients as features and feed forward neural network (FFNN) as the base learner of the ensemble methods. The results show that the classification accuracies of the ensemble classifiers except RUSBoost performs better than a single FFNN classifier.