Performance Comparison of Machine Learning Algorithms in P300 Detection Using Balanced Mind-Speller Dataset

Visual P300 mind-speller is a brain-computer interface that offers an easy and effective approach to track human brain responses. One major challenge in the design of this system is the unbalanced nature of its dataset, which can bias the classification process. In this work two distinct methods viz. nontarget undersampling and target oversampling were used to balance the mind-speller dataset. Since the choice of classification algorithm can impact the performance of mind-speller, the effect of dataset balancing was analyzed for a set of classifiers. The error rate, accuracy, true positive and false negative rates, true negative and false positive rates, positive predictive value, Matthews correlation coefficient, F-score, G-mean, and time consumption were the metrics used in this study. Among the various evaluated classifiers, k-nearest neighbor, support vector machine, and artificial neural network demonstrated significantly improved classification performance for the balanced (by target oversampling) mind-speller dataset while the Gaussian support vector machine yielded the highest metric scores.