Finding Non Dominant Electrodes Placed in Electroencephalography ( EEG ) for Eye State Classification using Rule Mining

Electroencephalography is a measure of brain activity by wave analysis; it consist number of electrodes. Finding most non-dominant electrode positions in Eye state classification is important task for classification. The proposed work is identifying which electrodes are less responsible for classification. This is a feature selection step required for optimal EEG channel selection. Feature selection is a mechanism for subset selection of input features, in this work input features are EEG Electrodes. Most Non Dominant (MND), gives irrelevant input electrodes in eye state classification and thus it, reduces computation cost. MND set creation completed using different stages. Stages includes, first extreme value removal from electroencephalogram (EEG) corpus for data cleaning purpose. Then next step is attribute selection, this is a preprocessing step because it is completed before classification step. MND set gives electrodes they are less responsible for classification and if any EEG electrode corpus wants to remove feature present in this set, then time and space required to build the classification model is (20%) less than as compare to all electrodes for the same, and accuracy of classification not very much affected. The proposed article uses different attribute evaluation algorithm with Ranker Search Method. Keywords—Electroencephalography (EEG); Most Non Dominant (MND); Ranker algorithm; classification; EEG


INTRODUCTION
The MND feature subset selection is a part of corpus preprocessing, and it is useful for classification model building as a supervised learning .Classification is one of the task performed by data mining tools and applicable in different area of biomedical electrical devices such as EEG, ECG(Electrocardiograms), EMG(Electromyography), EOG(Electrooculography), Actigraph devices etc.These devices are popular devices for recognizing of different types of disease like Sleep Apnea diagnosis [1] using ECG, driving drowsing using EEG [2],EEG and electromyography (EMG) enable communication for people with severe disabilities [20],muscles activity using EOG [3],and military operation using EEG [21] etc.These are the motivational points for proposed work because the article finds those positions electrode they are less responsible for classification then the removal of those electrodes minimize the size of devices.The present work is performed with EEG electrode data having 16 electrodes and 14892 instances [4,5].This uses the instance based classifier (K*), because based on statistic of data and nature of data spread over the corpus found it is best among other classifier the result of this present in literature [6,7], [28], [33], [38].Method selects either one electrode, two electrode or three electrodes based on how much search space the corpus wants to reduce.Its outcome generated from different attribute selection search with attribute evaluation techniques [8], [37].Here it is 11 different combination of search with evaluation techniques.Then generating rules using Apriori algorithm [9], it gives frequent electrodes which are placed in ranked as a last four sequences, it also depends how many last feature ranked matrix the corpus wants to create.Here it is 11*4 , where 11 are a Row value and 4 is a column value.Ranker Search with different attribute evaluation algorithms shown in Figure [1].Rankers Algorithm is an algorithm useful for ranking of attributes by their individual evaluation [10].Here three attribute evaluation methods are defined.

1) Info Gain Attribute evaluation:
Evaluate the worth of an attribute by measuring the gain ratio with respect to the class.
2) Classifier Attribute Evaluation: Evaluate the work of n attribute by using a user specified classifier.
3) OneR Attribute Evaluation: Evaluate the work of an attribute by using the oneR classifier.

II. ASSOCIATION RULE MINING
Association Rule Mining is used here for obtaining frequent set they are correlated with each other using support and confidence parameters [11][12][13].
Support is define as how frequently a specific item set occur in the data base (the percentage of transactions that contain all of the items in the item set, here the set of items are electrodes present in corpus and the transaction is the different method used for evaluation).
Confidence is the probability that items in RHS (Right Hand Side) will occur given that the items in LHS (left hand side) occurs.It Computed as Confidence (LHS) =>Support (LHS U RHS)/ Support (LHS) Electrode1 => Electrode2 [0.588, 0.88] If Electrode1 is selected in MND set, then Electrode2 also selected in MND set if it will satisfies minimum support and minimum confidence value.Left hand side electrode as Antecedent and Right hand side electrode [RHS] as consequent www.ijacsa.thesai.orgfrequency. .

III. ELECTROENCEPHALOGRAPHY (EEG)
EEG is useful for measuring brain activity.During the test very little electricity is passed between the electrodes and skin.EEG usually takes 30-60 minutes.The technician will put a sticky gel adhesive on 16 to 25 electrodes at various spots on our scalp [14].There are various spatial resolution of EEG systems like 10/20, 10/10, 10/5 systems as relative had surface based positioning system.The international 10/20 system a standard system for electrode positioning with 21 electrodes extended to higher density electrode such as 10/10 and 10/5 systems allowing more than 300 electrode positions [15].

V. EXTREME VALUE REMOVAL
The extreme value removal is a part of data cleaning step for data mining.The procedure for applying the extreme value theorem is to first establish that the function is continuous on the closed interval [16].The next step is to determine the critical points in the given interval and evaluate the function at these critical points and at the end points of the interval.If the function f(x) is continuous on closed interval [a, b] then f(x) has both a maximum and a minimum on [a, b] [17].In proposed method inter-quartile range [IQR] is used for extreme value calculations.IQR is major of variability based on dividing the dataset into quartiles [18].

VI. FEATURE SUBSET SELECTION
Feature Subset Selection is a task of data mining tool [25,26] ,it selects optimal feature subset for classifying the dataset but the literature shows the subset of optimal feature may or may not be optimal [19], [22][23][24].The proposed work is searching Most Non Dominant features (MND) from the feature set.This performed by ranker algorithm and with different search methods.The outcome of this step is ranks of electrodes placed in scalp.Proposed work used different 11 algorithms for obtaining the ranks of electrodes (most to least dominant).

VII. CLASSIFICATION
Classification is the task of data mining and it is a supervised learning.To classify EEG signals, various classification techniques present in literature [34][35][36][37][38].The instances present in corpus for eye state recognition using EEG, these instances are classified in to two different classes, zero is for eye opened state and one is for eye closed state.The instance based classifier is a type of lazy classifier [27], and proposed method uses K* is a type of instance base classifier, after extreme value removal and attribute selection.The literature shows there are various statistical measures are used for analysis of classification outcomes generated from classification process [29][30][31][32].

VIII. PROPOSED METHODOLOGY FOR MND SET
The proposed methodology is use full for finding nondominant feature from feature set.If "n" number of features are used for classification of eye state recognition then the space and time requirement is very high but if using less no of features obtained from proposed method then this will save time and space requirement.MND set electrodes are always a most non-dominant electrodes they are less responsible for classification accuracy.The flowchart shows in figure [1], and described steps shows, how to get MND from feature subset results generated from previous step.RESULT AND ANALYSIS This study used Ranker Search with Attribute Evaluation technique for MND set creation shown in table [1], then for rule generated using association rule mining this task performed by using Apriori algorithm ,all the generated rules are shown in figure [3],and the lattice shown in figure [4],shows how many frequent set to be considered for rule generation , the rules which is having minimum support and confidence is highlighted in figure [3],this gives frequent items (Electrode) set ,here it is {FC5,O2,F7}.This set declared as MND set, removing of this electrodes from EEG corpus sufficiently decrease the space and time requirement to built the classification model.The accuracy towards the classification changed very less and this analysis outcome shown in table [3], figure [6].The Confusion matrix shown in figure [5] and ROC curve shown in figure [7], evaluate the classifier performance here the classifier is Instance based classifier (K*), the classification accuracy is computed and it is mapped in table [3].

Fig. 6 .
Fig. 6.Time duration with Removal of Different Attributes

TABLE I .
SEARCH METHOD USED WITH DIFFERENT ATTRIBUTE EVALUATORS

TABLE III .
RESULT ANALYSIS AFTER REMOVAL OF ATTRIBUTES FROM FEATURE SET FROM EEG DATA SET