Effective Seizure Detection through the Fusion of Single-Feature Enhanced-k-NN Classifiers of EEG Signals

Electroencephalogram (EEG) physiological signals are widely used for detecting epileptic seizure. To reduce complexity stemming from the dimensionality problem, EEG signals are often reduced into a smaller set of discriminant features. The effectiveness of the detection techniques relies on the discriminant power of these features. However, since these features summarize the actual EEG signals, some valuable information can be lost. If unaddressed, this inherited incompleteness property can have a negative impact on the performance of the detection techniques, depending on the importance of the lost information. In this work, the evidence theory is utilized at two levels; (1) to enhance k-nearest-neighbor (k-NN) classifier, and (2) to combine decision collected from these enhanced-k-NNs. To effectively handle the incompleteness problem, the enhanced-k-NN has each feature pool k neighbors and consider each of these neighbors as a piece of evidence regarding its discriminant quality. Within the framework of evidence theory, these neighbors are combined together using the Dempster's rule to form a feature evidence structure. The feature evidence structures are then fused together to produce an overall pattern evidence structure which is used for the classification decision. To demonstrate the effectiveness of the proposed approach, five simple time domain features are obtained from EEG signals, and higher classification accuracy of 90% is achieved compared to other approaches which use more complex features.