Information fusion and artifact detection to improve the performance of multi-channel brain waveform classifiers

The goal of this dissertation is to improve the classification of multi-channel evoked potential (EP) brain waveforms through multi-channel information fusion and artifact detection. Improving the estimation of EPs from single-trial EPs is also investigated.; The information fusion strategies developed include vector decision fusion, vector data fusion, dynamic decision fusion, and dynamic data fusion. In the vector fusion methods, a ranking strategy is developed to rank and select channels. If the decisions are fused, the resulting decision fusion vector is classified using a Bayes discrete vector classifier. If the data of the selected channels are concatenated into a data fusion vector, the resulting data fusion vector is classified using a single vector classifier. The dynamic information fusion strategy establishes a rule to dynamically select either the data or the decisions of different channels at different time instants. If the decisions of different channels at different time instants are fused, the resulting decision vector is classified using a discrete Bayes classifier. If the samples are fused, the resulting data fusion vector is classified with a vector classifier. The information fusion classification strategies are tested on a neonatal EP data base in a dyslexia prediction problem and it is shown that, through dynamic decision fusion, dyslexia can be predicted with an accuracy of 99.88% and with a zero false positive rate.; An offline/real-time temporal domain artifact detection strategy is introduced and is used to determine how the performance of classifiers is affected when the training and test set ensembles are contaminated with artifacts and when only the test set ensemble is contaminated with artifacts. The artifact detection strategy consists of two steps in which the easily identifiable artifacts are first detected using a sequence of within-channel standard deviation and clipping tests. The more difficult to identify artifacts are detected using a multi-channel channel median distance based approach. The effects of artifacts on classifier performance are demonstrated by incorporating artifact rejection in classification experiments. The results show that artifacts, as expected, have detrimental effects on the performance. The performance improves significantly when artifacts are removed from the training and test set ensembles.; Two variations of the non-linear alignment averaging algorithm aimed at improving the estimation of the underlying EP from single-trial EPs are introduced. The variations exploit the advantages of both the conventional averaging and the non-linear alignment averaging algorithms. In order to evaluate the performance, a model is introduced to simulate single-trial EPs. Experimental results show that the best estimate is obtained using the algorithm which averages the pre-aligned single-trial EPs with the conventional average.; Finally, it must be noted that the strategies developed in this dissertation are quite general. Therefore, they can be applied, in general, to other multi-sensor classification problems involving multi-variate - multi-category data.