At present, more number of electrodes are used to develop brain computer interface (BCI) devices based on motor imagery. However, the number of trials for a given subject is usually less. Under this situation, the performance of motor imagery task classification may degrade. In this research work, we propose a combination of graph theoretic spectral method and quantum genetic algorithm (QGA) to obtain a subset of relevant and non-redundant electrodes for effective motor imagery task classification. Stationary Common Spatial Pattern method, which can handle non-stationarity issue, is used for extraction of features from the reduced set of electrodes. Support Vector Machine (SVM) is used as a classifier. Improvement in classification performance on publicly available dataset signifies efficacy of the proposed method. Friedman statistical test demonstrates that the performance of the proposed method is significantly better in comparison to existing CSP and its variants. © 2019 Elsevier B.V.