Performance analysis of fuzzy clustering approaches for the identification of a visual evoked potential called the lambda wave

Abstract

A performance analysis of the fuzzy clustering approaches of c-means (FCM), Gustafson-Kessel (FGK) and maximum likelihood estimates (FMLE) for the identification of a saccade related visual evoked potential (EP) called the lambda wave was carried out. The lambda waves were extracted from suitably pre-processed electroencephalogram (EEG) waveforms using independent component analysis (ICA). The ICA-extracted EEG components were represented by a spatial and three temporal features that suitably characterised the lambda wave. The feature sets were processed by the three clustering algorithms. The performances of the algorithms for separating the lambda waves from nonlambda waves were compared against each other as well as against the classification results achieved by visual inspection. FMLE managed to identify 97.6 % of the lambda waves and 99.6 % of nonlambda waves in accordance with visual classification. FCM identified 99.7 % of the lambda waves and 82.1 % of the nonlambda waves. FGK managed to identify up to 98.7 % of the lambda waves and 94.7 % of the nonlambda waves. Comparing the three approaches, FMLE as a whole was the most accurate method for differentiating between lambda and nonlambda waves however it required significantly larger number of iterations for the analysis. (8 pages)