Classification of P300 Component Using a Riemannian Ensemble Approach

We present a framework for P300 ERP classification on the 2019 IFMBE competition dataset using a combination of a Riemannian geometry and ensemble learning. Covariance matrices and ERP prototypes are extracted after the EEG is passed through a filter bank and an ensemble of LDA classifiers is trained on subsets of channels, trials, and frequencies. The model selects a final class based on maximum probability of evidence from all ensembles. Our pipeline achieves an average classification accuracy of 81.2% on the test set.