Feature Selection and Classifier Parameters Estimation for EEG Signals Peak Detection Using Particle Swarm Optimization

Table 1

Summary of different peak models on different style of framework.

Peak model	Type of signal	Description of framework
Dumpala et al. (1982) [8]	Electrical control activity (ECA)	The theory of maxima and minima using three-point sliding window approach has been applied to detect a candidate peak. Two flowcharts of peak detection have been proposed. A predicted peak can be identified if the feature values satisfied the decision threshold values. The strength and weakness of the proposed approach are described as follows: (1) strength: the authors claimed that the proposed peak detection algorithm can be used for other biological signals, (2) weakness: the utilization of peak-to-peak amplitude on the peak model is hard to distinguish between noise and actual peak. In addition, large variation of peak width in the signal may drop the classification performance.
Dingle et al. (1993) [9]	Epileptic EEG	Based on the defined peak model, the features are grouped into two: (1) epileptiform transient parameters and (2) background activity parameters. Two-threshold systems have been employed to detect a candidate peak or candidate epileptiform transient. Expert system which considered both spatial and temporal contextual information has been used to reject the artifacts and classify the transient events. The strength and weakness of the proposed approach are described as follows: (1) strength: moving average amplitude is good in rejecting false peak points. The employed features are claimed to offer good performance in the proposed expert system, (2) weakness: inconsistency of feature slope information as the proposed work claimed that the proposed framework fails to provide slope information.
Liu et al. (2002) [10]	Epileptic EEG	Wavelet transform has been used to decompose the EEG signal. Based on the decomposed signals and the defined peak model, seven features are calculated. These features are used as the input of ANN classifier. Expert system which considered both spatial and temporal contextual information has been used to reject the artifact. Several heuristic rules have been employed to distinguish the type of artifact. After all artifacts are recognized and rejected, the decision will be made to classify the epileptic events. The strength and weakness of the proposed approach are described as follows: (1) strength: the employed features is claimed to offer good performance in the proposed expert system, (2) weakness: it considers that almost all the features may deteriorate the classification performance.
Acir et al. (2005) [11]	Epileptic EEG	A three-stage procedure based on ANN is proposed for the detection of epileptic spikes. The EEG signal is transformed into time-derivative signal. Several rules have been used to detect a peak candidate. The features of peak candidate are calculated based on the defined peak model. These features are fed into two discrete perceptron classifiers to classify into three groups: definite peak, definite non-peak, and possible/possible non-peak. The peak that belongs in the third group is going to be further processed by nonlinear classifier. The strength and weakness of the proposed approach are described as follows: (1) strength: the employed features are claimed to offer good performance in the proposed system, (2) weakness: inconsistency of feature slope information as the proposed work claimed that the proposed framework fails to provide slope information.
Acir (2005) [26]	Epileptic EEG	A two-stage procedure based on a modified radial basis function network (RBFN) is proposed for the detection of epileptic spikes. The EEG signal is transform into time-derivative signal. Several rules have been used to detect a peak candidate. The features of peak candidate are calculated based on the defined peak model. These features are fed into discrete perceptron classifiers to classify into two groups: definite non-peak and peak-like non-peak. The peak that belongs to the second group requires further process by modified RBFN classifier. The strength and weakness of the proposed approach are described as follows: (1) strength: the employed features are claimed to offer good performance in the proposed system, (2) weakness: inconsistency of feature slope information as the proposed work claimed that the proposed framework fails to provide slope information.
Liu et al. (2013) [21]	Epileptic EEG	A two-stage procedure is proposed for the detection of epileptic spike. k-NEO has been used to detect a candidate peak. The peak features are calculated based on the defined peak model. These features are then used as the input of the AdaBoost classifier. The strength and weakness of the proposed approach are described as follows: (1) strength: the peak model considers feature based on peak area, (2) weakness: the definition of area integration is not presented in the paper.