A robust framework for optimum feature extraction and recognition of P300 from raw EEG

This chapter proposes a framework geared toward the conundrum of detection of P300 signals from raw time-series electroencephalogram (EEG) signals using a four-step technique. The raw data is preprocessed and molded into a dataset capable of running through a classification pipeline. Features are then extracted from the dataset using methods previously used in similar literature, popularly available signal-processing techniques and a few methods newly proposed by the authors. The extracted features are arranged into vectors, both stacked and individually. These feature vectors are trained on a multiple classifiers—linear, nonlinear, ensembled learners, and genetic algorithms. The large number of classification models generated by feature-feature and feature-classifier combinations is evaluated using data from the BCI Competition III Dataset V. The classification results of the best performing classification models are delineated and visualized. The results show how the Wavelet Transform, the proposed Converse Wavelet Transform, and the Time Series data, stacked with other representative features and among one another, in combination with ensembled linear classifiers, are the key elements to achieving high P300 classification accuracy.