AIEFS and HEC based emotion estimation using ......AIEFS and HEC based emotion estimation using physiological measurements for the children with autism spectrum disorder. K Kiruba1*,

AIEFS and HEC based emotion estimation using physiological measurementsfor the children with autism spectrum disorder.

K Kiruba1*, D Sharmila2

1Department of Electronics and Communication Engineering, K.S.R. College of Engineering, Tiruchengode, TamilNadu, India2Department of Electronics and Insturmentation Engineering Bannari Amman Institute of Technology,Sathyamangalam, Tamil Nadu, India

Abstract

The emotion of the children with Autism Spectrum Disorder (ASD) cannot be identified and recognizedeasily. The research in automated emotion recognition methods is steadily growing thrust in the last fewyears due to applicability in various domains which would benefit from a clear understanding of humanemotional states. The studies have shown that a human’s physiological parameters are directly related tohis/her psychological reaction from which the emotions can be estimated. There is a strong relationshipbetween human emotion and physiological signals. The major aim of this work is to identify preferableArtificial Intelligent Ensemble Feature Selection (AIEFS) framework and Heterogeneous EnsembleClassification (HEC) model for such a concept. The experiment was necessary to achieve the uniformityin the various aspects of emotion elicitation, data processing, feature selection using EFS, and estimationevaluation using HEC and in order to avoid inconsistency problems. Here, three base classifiers such asSupport Vector Regression with Genetic Algorithm (SVR-GA), Multinomial NaiveBayes (MNB) andEnsemble Online Sequential Extreme Learning Machine (EOS-ELM) that learn different aspects of theemotion dataset samples are used together to make collective decisions in order to enhance performanceof health-related message classification. The results indicate that the combination of AIEFS with HECexhibited the highest accuracy in discrete emotion classification based on physiological featurescalculated from the parameters like ECG, respiration, skin conductance and skin temperature. Specificdiscrete emotions were targeted with stimuli from the IAPS database. This work presents experimentbased comparative study of four feature selection methods and five machine learning methodscommonly used for emotion estimation based on physiological signals.

Keywords: Emotion estimation, Feature reduction, Machine learning, Artificial intelligent ensemble feature selection,Heterogeneous ensemble classification, Support vector regression with genetic algorithm, Multinomial naïve bayes,Ensemble online sequential extreme learning machine, Mean based weighted for quaternions firefly algorithm,Modified trapezoidal fuzzy membership genetic algorithm.

Accepted on August 23, 2016

IntroductionFacial expressions convey the feelings and state of mind ofothers and enable us to adjust the behavior and to reactappropriately. Therefore, the ability to interpret facialexpressions accurately and to derive socially relevantinformation from them is considered a fundamentalrequirement for typical reciprocal social interactions andcommunication. Autism spectrum disorder is a seriousneurodevelopmental disorder that weakens child's ability tocommunicate with others. It also includes repetitive behaviors,interests and activities. Generally it is difficult to analyze theemotions of the children with autism. At the same time,emotion recognition using facial expressions is difficult withthe autistic children because they donot prefer eye to eye

contact and the difficulties in recognizing, identifying, andunderstanding the meaning of emotions are often considered asone of the trademarks of their social problems. Differentprocedures have been used to examine emotion processingabilities in children and adults with ASD, with or withoutintellectual disability: sorting, (cross-modal) matching, andlabeling tasks (for a literature review and a meta-analysis onthis topic, see [1], resp.). Each of these procedures has revealedproblems with affect processing in individuals with ASD.Other studies, however, failed to find atypical emotionrecognition skills in individuals with ASD (e.g., [2]).Inconsistencies may be due to differences in sample andparticipants’ characteristics, task demands [3], and stimuli.Performances of individuals with ASD seem to be especially

Biomedical Research 2016; Special Issue: S237-S250 ISSN 0970-938Xwww.biomedres.info

Biomed Res- India 2016 Special Issue S237Special Section: Computational Life Science and Smarter Technological Advancement

impaired for negative, more subtle, or more complex emotionsor expressions embedded in a social context [4].

In the last few years the research in automated emotionrecognition methods is steadily growing momentum due toapplicability in various domains which would benefit from anaccurate understanding of human emotional states, likeentertainment, ASD ,safe driving, training and e-learning,telemedicine and home robotics [5,6]. Furthermore, variousmental health applications may benefit from automatedestimation of patient's emotions, like treatment of stress-relateddisorders [7]. Psychophysiology attempts to achieve humanemotions by study of the interrelationships between thephysiological and psychological aspects of behavior.Physiological affect in general refers to responses that comefrom the body, more especially those associated withautonomic nervous systems in the body. The various types ofphysiological signal that can be obtained from the human bodyby using the sensors those are wearable. The principal level offocus of on the physiology is at the level of organs and thesystems.

Physiological parameters like heart rate, skin conductance,temperature and respiration rate can be used to analyse theemotions of the human beings because these variables respondto signals from the nervous system, which is not underconscious control. For a variety of these applications,individually adjusted emotion estimators rather than genericemotion estimation may achieve higher accuracy [8],particularly if the estimator can learn emotional responseidiosyncrasies of a particular individual over the course ofmultiple sessions. Such personalized adaptive emotionestimator system should perform real-time estimation of user'semotion and con-currently adapt itself over time based on themeasured user's responses. The aim of the study was toexamine the differences of boredom, pain, and surprise. Inaddition to that, it was conducted to propose approaches foremotion recognition based on physiological signals. Threeemotions, boredom, pain, and surprise, are induced through thepresentation of emotional stimuli and Electrocardiography(ECG), Electrodermal Activity (EDA), Skin Temperature(SKT), and photoplethysmography (PPG) as physiologicalsignals are measured to collect a dataset from differentparticipants when experiencing the emotions. The DiscriminantFunction Analysis (DFA) as a statistical method [9] and fivemachine learning algorithms such as Linear DiscriminantAnalysis (LDA), Classification And Regression Trees (CART),Self-Organizing Map (SOM), Naïve Bayes algorithm, andSupport Vector Machine (SVM) are used for classifying theemotions. The highest recognition accuracy of 84.7 % isobtained by using DFA.

Most literature about emotions claims that the emotions have acomplex nature. Even though several feature reduction andmachine learning methods have been so far successfullyemployed in the previous research to build emotional stateestimators from physiological indices, a comparison of variousmethods used by different research groups, has been precludeddue to the following reasons:

• Emotion elicitation method diversity.• Emotional state representation method - discrete emotions

or dimensional (valence-arousal) space.• Properties of used physiological signals and features.• Referent emotional state selection-subjective ratings or

stimuli annotations.• Estimator evaluation method.

With these issues, finding a common ground for comparingmethods and analyzing their features is demanding. Therefore,this paper uses appropriate dataset to compare accuracy,execution and learning times of four feature selection and fivemachine learning methods commonly employed in emotionestimation based on physiological features. In the comparativeanalysis, each feature selection method is paired with everylisted machine learning method. Finally the results areevaluated using those machine learning classifiers.

Literature SurveyData mining methods have been proposed in the literature tosolve these problems which is discussed as follows. Wagner etal. [10] discussed the most important stages of a fullyimplemented emotion recognition system including dataanalysis and classification. For collecting physiological signalsin different affective states, a music induction method wereused which elicits natural emotional reactions from the subject.Four-channel biosensors were used to obtain electromyogram,electrocardiogram, and skin conductivity and respirationchanges from which the emotion were estimated. Then severalfeature selection/reduction methods were tested to extract anew feature set consisting of the most significant features forimproving classification performance. Three well-knownclassifiers, linear discriminant function, k-nearest neighbor andmultilayer perceptron, were used to perform supervisedclassification. An advantage of this method is that most peopleare used to listen to music during other activities and for thisreason tend to associate different moods with specific songs.

Kim [11] collected a physiological data set from multiplesubjects over many weeks and used a musical inductionmethod that spontaneously lead subjects to real emotionalstates, without any deliberate laboratory setting. Four-channelbiosensors were used to measure electromyogram,electrocardiogram, skin conductivity, and respiration changes.A wide range of physiological features from various analysisdomains, including time/frequency, entropy, geometricanalysis, subband spectra, multiscale entropy, etc., wasproposed in order to find the best emotion-relevant featuresand to correlate them with emotional states. The best featuresextracted were specified in detail and their effectiveness wasproved by classification results. Classification of four musicalemotions (positive/high arousal, negative/high arousal,negative/low arousal, and positive/low arousal) was performedby using an extended linear discriminant analysis (pLDA).Furthermore, by exploiting a dichotomic property of the 2Demotion model, we develop a novel scheme of emotion-specific multilevel dichotomous classification (EMDC) and

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S238

compare its performance with direct multiclass classificationusing the pLDA.

Biosignals can reveal the emotions and, as such, can serve asan advanced Man-Machine Interface (MMI) for empathicconsumer products. However, such a MMI requires the correctclassification of biosignals to emotion classes. A state-of-the-art review is presented on automatic emotion classification[12]. Moreover, guidelines werepresented for affective MMI.Subsequently, a research was presented that explores the use ofEDA and three facial EMG signals to determine neutral,positive, negative, and mixed emotions, using recordings of 21people. A range of techniques is tested, which resulted in ageneric framework for automated emotion classification withup to 61.31% correct classification of the four emotion classes,without the need of personal profiles.

Integrated system was proposed in [13] which provide aninnovative and intelligent solution for the monitoring ofpatients with anxiety disorders during therapeutic sessions. Itrecognizes an individual's affective state based on 5 pre-defined classes such as relaxed, neutral, startled, apprehensiveand very apprehensive, from physiological data collected vianon-invasive technologies such as blood volume pulse, heartrate, galvanic skin response and respiration. The system wasvalidated using data obtained through an emotion elicitationexperiment based on the International Affective PictureSystem. Four different classification algorithms such asArtificial Neural Networks (ANNs), Support Vector Machines(SVM), Random Forests (RF) and a Neuro-Fuzzy System wereimplemented. The overall classification accuracy achieved was84.3%.

Park et al. [14] suggested an optimal algorithm for emotionclassification which classifies seven different emotional statessuch as happiness, sadness, anger, fear, disgust, surprise, andstress using physiological features. Skin temperature,photoplethysmograph, electrodermalactivity andelectrocardiogram were recorded and analyzed as physiologicalsignals. For classification problems of the seven emotions, thedesign involves two main phases. At the first phase, ParticleSwarm Optimization selects P % of patterns to be treated asprototypes of seven emotional categories. At the second phase,the PSO is instrumental in the formation of a core set offeatures that constitute a collection of the most meaningful andhighly discriminative elements of the original feature space.The emotion stimuli used to induce a participant’s emotionwere evaluated for their suitability and effectiveness. Theysuggested that the use of the prototype is also justifiableconsidering that this classification scheme is the simplest thatcould be envisioned in pattern classification.

In the recent work [15] analyzed various types of physiologicalsignals of a person with respect to the stress developed withinhim/her. The analysis of stress was done using ECG, EEG andrespiratory signals acquired from the automobile drivers whowere made to drive on different road conditions to get differentlevels of stress. As a part of the analysis, two features wereextracted from the physiological signals and it shows thechanges in the feature with respect to the stress of the driver.

From the features that are extracted, stress is classified usingSVM classifier. The performance of the networks was testedand compared with other physiological signal and producebetter result with high accuracy.

Niu et al. [16] applied novel feature selection method torecognize human emotional state from four physiologicalsignals such as Electrocardiogram (ECG), electromyogram(EMG), Skin Conductance (SC) and Respiration (RSP). Theraw training data was collected from four sensors, ECG, EMG,SC, RSP, when a single subject intentionally expressed fourdifferent affective states, joy, anger, sadness, pleasure. Thetotal 193 features were extracted for the recognition. A musicinduction method was used to elicit natural emotional reactionsfrom the subject, after calculating a sufficient amount offeatures from the raw signals, the genetic algorithm and the K-neighbor methods were tested to extract a new feature setconsisting of the most significant features which representsexactly the relevant emotional state for improvingclassification performance. The numerical results demonstratethat there is significant information in physiological signals forrecognizing the affective state. It also turned out that it wasmuch easier to separate emotions along the arousal axis thanalong the valence axis.

Valenzietal [17] conducted offline computer aided emotionclassification experiments using strict experimental controlsand analyzed EEG data collected from nine participants usingvalidated film clips to induce four different emotional statessuch as amused, disgusted, sad and neutral. The classificationrate was evaluated using both unsupervised and supervisedlearning algorithms, in total seven algorithms were tested. Thelargest classification accuracy was computed by means ofSupport Vector Machine (SVM) which is a machine learningalgorithm. The experimental protocol effectiveness was furthersupported by very small variance. This small variance isobtained among individual participants classification accuracy.Classification accuracy and rate evaluated on reduced numberof electrodes suggested, consistently with psychologicalconstructionist approach that classified human emotions. Theexperimental protocol therefore appeared to be a key factor toimprove the classification outcome by means of data qualityimprovements.

Kukolja et al. [18] suggested the preferable methods foridentifying an experiment-based comparative study of sevenfeature reduction and seven machine learning methodscommonly used for emotion estimation based on physiologicalsignals. The results of the performed experiment indicate thatthe combination of a Multi-Layer Perceptron (MLP) withSequential Floating Forward Selection (SFFS) exhibited thehighest accuracy in discrete emotion classification based onphysiological features calculated from ECG, respiration, skinconductance and skin temperature. In order to identify whichmethods may be the most suitable for real-time estimate oradaptation, execution and learning times of emotion estimatorswere also comparatively analyzed. Based on this analysis,minimum Redundancy - Maximum Relevance (mRMR) wasidentified the fastest approach. In combination with mRMR,

AIEFS, heterogeneous ensemble classification

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S239

highest accuracies were achieved by k-nearest neighbor (kNN)and MLP with negligible difference however, they suggestedmRMR+kNN is preferable option for real-time estimatoradaptation due to considerably lower combined execution andlearning time of kNN versus MLP.

Studies have shown that a human being’s physiologicalchanges are directly related to his/her psychological reaction.A wearable wristband for acquiring physiological signalsandan algorithm, using a Support Vector Machine (SVM)classifier [19] that will predict emotional states such as neutral,happy &involvement of children with autism was proposed.The psychological reactions (or emotions) were recognizedbased on the changes in thebodily parameters (physiologicalbasis) such as the Galvanic Skin Response (GSR) and HeartRate Variability (HRV). For thispurpose, vital featuresextracted from the recorded physiologicalsignals wereclassified into different emotional states usingSVM, whichresulted in an overall accuracy of 90 %. This helps the parentsand the care takers to understand the emotionalpatterns of thechild better.

The portability of the system ensures ease of use and real-timeemotion recognition and aid for immediate feedback whilecommunicating with caretakers. Principal Component Analysis(PCA) had been identified as the least complex featureextraction algorithm to be implemented in hardware. To solvethis problem a detailed study of the implementation of serialand parallel implementation of PCA in order to identify themost feasible method for realization of a portable emotiondetector for autistic children was presented [20].

MethodologyIn the recent years, the research in automated emotionrecognition methods is steadily growing due to applicability invarious domains. For a variety of these applications,individually adjusted emotion estimators rather than genericemotion estimation may achieve higher accuracy [18]. Thestudies reveal that a human being’s physiological changes arerelated directly to his/her psychological reaction. There is astrong correlation between human emotion and physiologicalsignals. Using physiological signals, the emotions can berecognized and classified by feature reduction and machinelearning algorithms. Emotions such as happiness, sadness,disgust, fear, anger, surprise, and stress are classified. Theabove emotions are classified using the physiological signalssuch as skin conductance; skin temperature, ECG, Respiration,blood pressure, blood oxygen saturation etc., The featurereduction provided many insights into affective experience.Here Artificial Intelligent Ensemble Feature Selection (AIEFS)framework is proposed to feature reduction to increase theclassification accuracy. To analyze an efficient algorithm inorder to recognize emotions, the comparative analysis ofHeterogeneous Ensemble Classification (HEC) model is to beperformed which classifies the reduced features into severalclasses. In cooperation with the Department of Psychology atUniversity of Zagreb, Faculty of Humanities and SocialSciences, an emotion elicitation experiment was conducted

with the goal of evaluating accuracy, execution and learningtimes of emotion estimators based on data mining of acquiredphysiological signals [18].

Dataset collection: Emotion elicitation stimuli: During theexperiment design, it was decided that emotion would beelicited with a standardized database of emotionally annotatedmultimedia. Therefore, the International Affective PictureSystem (IAPS) database [21,22] was selected as the preferredsource of stimuli for the experiment since it is the most widelyused and referenced database in the field of emotion elicitation.The IAPS contains more than 1000 static pictures which areemotionally annotated regarding the dimensions of valence,arousal and dominance [18].

In the emotion elicitation experiment the target discreteemotions were sadness, disgust, fear and happiness in additionto neutral. IAPS pictures suitable for elicitation of discreteemotion states were selected based on research in categorizingthe dimensional model to normative emotional states [23,24].Due to the categorization of discrete emotions in these studies,we were aiming at Ekman's basic emotions: happiness,surprise, sadness, anger, disgust and fear. However, althoughall aforementioned studies categorize negative emotions asEkman's sadness, anger, disgust and fear, just a few imagescould be labelled with only anger when taking a closer look atpicture labels [25] and multidimensional normative ratings[24]. These findings are consistent with the definition of angerbeing a combination of appraisals of extreme unpleasantnessand high certainty which are difficult to achieve with passiveviewing of static pictures [25]. Therefore, five emotions wereconsidered: disgust, sadness, anger, surprise and fear.

To identify the emotions actually elicited, after exposure toeach IAPS picture the participant expressed her judgmentsabout the elicited emotions using a written questionnaire. Ineach session, every participant was exposed to two consecutivesequences separated by a pause of at least 150 s, which wasintended to bring the participant back to the neutral state. Eachsequence of pictures was designed to specifically elicit oneparticular emotional state. The targeted emotional states weresadness, disgust, fear and happiness, in addition to the neutralemotional state during the neutral-stimulus period. First eightparticipants were shown sequences of fear and happiness in thefirst session, while during the second session they wereexposed to sequences of disgust and sadness. Exposuresessions for the remaining six participants were reversed - theywere the first exposed to sequences of disgust and sadness, andin the second session they were exposed to fear and happinesssequences. However, one of these participants dropped outafter the disgust sequence, so only data from her disgustsequence were included in analysis. Order of stimulipresentation in each sequence was the same for all participants[18].

To counter the physiological signals drift [26] the elicitationprotocol always included a neutral stimulus before everyemotionally non-neutral stimulus. Therefore, a session beganwith a 30 s neutral stimulus, a simple blue-green neutralbackground, to establish participants' baseline response. This

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S240

particular appearance of the neutral screen was selected basedon a study that identified blue-green, i.e. cyan, as a color withthe best ratio between elicited positive and negative emotions.

Electrocardiogram (ECG) represents electrical activity ofhuman heart. ECG is composite from 5 waves - P, Q, R, S andT. This signal could be measured by electrodes from humanbody in typical engagement. Signals from these electrodes arebrought to simple electrical circuits with amplifiers andanalogue - digital converters. The main problem of digitalizedsignal is interference with other noisy signals like powersupply network 50 Hz frequency and breathing muscleartefacts. These noisy elements have to be removed before thesignal is used for next data processing like heart rate frequencydetection. Digital filters and signal processing should bedesigned very effective for next real-time applications inembedded devices. Heart rate frequency is very importanthealth status information. The frequency measurement is usedin many medical or sport applications like stress tests or lifetreating situation prediction. One of possible ways how to getheart rate frequency is compute it from the ECG signal. Heartrate frequency can be detected d from ECG signal by manymethods and algorithms. Many algorithms for heart ratedetection are based on QRS complex detection [27] and hearrate is computed like distance between QRS complexes. Fromoriginal and normalized heart rate, skin conductance,respiration rate and skin temperature signals for eachstimulation, 288 statistical features were calculated, based on14 statistical methods: mean, standard deviation, mean of thefirst derivative, minimum, maximum, difference betweenmaximum and mini-mum, mean of the offset, minimum of theoffset, maximum of the offset, difference of means betweentwo consecutive segments, difference of standard deviationsbetween two consecutive segments, difference of means of thefirst derivative between two consecutive segments, mean of theabsolute values of the first differences and mean of theabsolute values of the second difference[18].

Skin conductance response featuresThe skin conductance signal includes two types of electrodermal activity: the DC level component and the distinctiveshort waveforms. The latter is usually called the SkinConductance Response (SCR) and is considered to be useful asit signifies a response to internal/external stimuli. Thealgorithm measures the magnitude and the duration of the risetime. From this information, the following features werecalculated: the frequency of occurrence (FREQSCR), themaximum magnitude (SM_MAXSCR), the mean magnitudevalue (SM_MEANSCR), the first SCR magnitude(SM_FIRSTSCR), the mean duration value (SD_MEANSCR)and the area of the responses (SMnSDSCR). In order toincrease robustness, features SM_MAXSCR, SM_MEANSCR, SM_FIRSTSCR and SMnSDSCRwere alsocalculated from normalized skin conductance signals using all7 normalization methods [18].

Heart rate variability featuresHeart Rate variability (HRV) is one of the most often usedmeasures for ECG analysis. HRV is a measure of thecontinuous interplay between sympathetic and parasympatheticinfluences on heart rate that yields information aboutautonomic flexibility and thereby represents the capacity forregulated emotional responding [28]. Therefore, HRV analysisis emerging as an objective measure of regulated emotionalresponding. In the time domain, calculated the followingstatistical features: the standard deviation of all NN intervals(SDNN), the square root of the mean of the sum of the squaresof differences between adjacent NN intervals (RMSSD),standard deviation of differences between adjacent NNintervals (SDSD), the proportion derived by dividing NN50differing by more than 50 ms, the proportion derived bydividing NN20 differing by more than 20 ms and Fano factor(FF).

Features calculated from respiration signalFeatures are also calculated from the raw respiration signal.Here calculated the power mean values of four subbandswithin following ranges: 0-0.1 Hz, 0.1-0.2 Hz, 0.2-0.3 Hz and0.3- 0.4 Hz. The power spectral densities of detrendedrespiration signal were obtained using the Burg algorithm. Toincrease the robustness of emotion estimation, features werealso calculated from detrended respiration signal normalizedby dividing with mean peak-to-peak magnitude of respirationsignal in baseline. In this way, a total of 8 features fromrespiration signal were calculated. From the above extractedfeatures dimensionality reduction is performed by usingEnsemble Feature Selection (EFS) framework.

From this dataset the above mentioned feature selection isperformed using Ensemble Feature Selection (EFS), here EFScombines the methods of Artificial Intelligence methods so it isnamed as AIEFS framework. Total 368 different physiologicalfeatures were computed for each stimulus presented to theparticipant. From this features irrelevant or reduce thedimension of the features using AIEFS framework which isdiscussed as follows:

Artificial intelligent ensemble feature selection (aiefs)framework: Many of the existing mechanisms for featureselection follow the general principle of supervised learning,be they filter or wrapper based approaches. As such, they workby relying on identified correlations between class or decisionlabels and the underlying feature values [29]. However, inmany real-world applications, the thorough interpretation of alarge data may become infeasible and hence, the amount oflabelled training samples is often limited. This makesunsupervised feature selection algorithms [30], and semi-unsupervised learning [31] techniques potentially beneficialand desirable [30]. Although the performance of this newdevelopment is promising, it merely contributes to the familyof FS techniques as yet another single method that produces asingle feature subset of features when presented with a training

AIEFS, heterogeneous ensemble classification

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S241

dataset. The performance of such techniques may varysignificantly over different problem domains.

Ensemble Feature selection (EFS) is an ensemble-basedmethod that constructs a group of feature subsets, and thenproduce an aggregated result out of the group. In so doing, theperformance variance of obtaining a single result from a singleapproach can be reduced. It is also intuitively appealing thatthe combination of multiple subsets may remove less importantfeatures, resulting in a compact, robust, and efficient solution.Ensembles of feature ranking techniques have been studied inthe literature for the purpose of text classification and softwaredefect prediction; they work by combining the ranking scoresor exploring the rank ordering of the features. Additionally,feature redundancy elimination has been achieved by the usedof tree-based classifiers ensembles. In this, three steps arecarryout to implement the EFS concept these are: 1) buildingensembles using optimization search algorithms, 2) generatingdiversity by partitioning the International Affective PictureSystem (IAPS) database, and 3) constructing ensembles bymixing various different FS approaches.

In this section, the proposed implementations of the AIEFSframework concept are specified as follows with the aid ofillustrative diagram. In the context of AIEFS framework, anIAPS databaseis represented using wherebe the number ofsamples and are finite, non-empty sets of dataset samples andphysiological signalfeatures, respectively.Physiological signalsfeatures might be either discrete-valued or real-valuedattributes. Here, a Physiological signals feature subset isrepresented by a binary string FS of length M, if , otherwise.An AIEFS can therefore be represented by a set of such featurestrings, where K denotesthe size of the ensemble here K=2.The finally selected physiological signals features subset by theAIEFS framework is the outcome of aggregating the elementsof EPSF, which is denoted by hereafter.

Figure 1. Flow chart for AIEFS framework.

By employing multiple FS algorithms, the ensemble diversitycan be naturally obtained from the differences in opinionsreached by the evaluators themselves. The ensembleconstruction process may be further randomised by the use of apseudo random generator, as illustrated in Figure 1, so that theavailable FS algorithms are randomly selected when formingthe ensemble beneficial when the available feature selectors arefewer than the desired number of ensemble components, wherecertain selectors are expected to be used multiple times. Theflowchart for AIEFS framework is given in Figure 1.

Modified trapezoidal fuzzy membership genetic algorithm(MTFGA): Genetic Algorithms (GA) [32] are a class ofevolutionary algorithms that use evolution as a source ofinspiration to find the solution for many optimizationproblems. All the possible solutions of the given problem arecalled the chromosomes. The chromosomes can be consideredto be a physiological signal feature vector and each dimensionof this physiological signal feature vector can be considered tobe a gene. Each generation has a specific number ofInternational Affective Picture System (IAPS) databasechromosomes also called as the population. In the traditionalGA the size of the population of each generation kept same allthe features in the dataset is reduced or irrelevant.

The most important procedure of GA is the fitness function;here the fitness function is determined based on theclassification accuracy for selected physiological signalsfeature. This fitness function is also known as the objectivefunction. Each physiological signal feature (chromosome) fromthe generation is passed through the fitness function and thus,they get their fitness value. The classification accuracy (fitnessvalue), then determine the proximity of the physiologicalsignal feature from chromosome to the highest classificationaccuracy value. The physiological signal features fromchromosomes with high highest classification accuracy(fitness) values is selected for reproduction. The modes ofreproduction are mainly depends on crossover and mutation.Crossover is the interchange of two physiological signalfeatures between the IAPS database and mutation is therandom change in the physiological signal features. Mutation isusually done on a comparatively weak physiological signalfeatures from IAPS database, so that it adds diversity to thephysiological signal features (population) without actuallyimpeding the progress towards the optimal solution. Thechromosomes that have reproduced are replaced by the newphysiological signal features, irrespective of the fitness valuesof the new physiological signal features. This results in theformation of the new physiological signal features generation.The physiological signal features in this generation, whichwere the offspring of the previous generation, are now thephysiological signal features of the next generation. Thesephysiological signal features (chromosomes) are now passedthrough the highest classification accuracy(fitness function)again and the strongest physiological signal features areselected to reproduce, which results in a new physiologicalsignal features generation, with a new set of physiologicalsignal features (chromosomes) and ideally nearer to theoptimal solution with highest classification accuracy.

OperatorsThe following are the most important GA operators:

The Selection operator selects physiological signal features inthe IAPS database (population) for reproduction. The selectionfunction is usually stochastic and designed to selectphysiological signal features the highest classification accuracy(fitness) of the chromosomes from the IAPS database(population).

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S242

The Crossover operator chooses a physiological signal featuresand exchanges the physiological signal features in thechromosomes before and after the physiological signal featuresto create new physiological signal features offspring.

The Mutation operator randomly flips the physiological signalfeatures thereby creating a new physiological signal featuresoffspring. Mutation adds diversity to the IAPS database(population).

ElitismWhile replacing physiological signal features (chromosomes)from IAPS database in the iteration N to iteration N+1, aphysiological signal feature with a fairly good highestclassification accuracy value may be replaced by aphysiological signal feature with a mediocre or a poorclassification accuracy value. Thus, this may result in theselection of optimal physiological signal feature. In elitism, thetop physiological signal feature of each generation is classifiedas elite individuals. These physiological signal features willparticipate in the reproduction, but will not be replaced by anyphysiological signal feature from the next generation. This iscalled Simple Elitism. In Global Elitism, each physiologicalsignal feature from generation N+1 can replace its parent fromgeneration N, if and only if, its performance is found to besuperior. The drawback in this case is that, the comparison isstill being done only on a physiological signal feature tophysiological signal feature basis and not on a generation togeneration basis [32].

Dynamic representation of the featuresTrapezoidal Fuzzy membership function: Then TrapezoidalFuzzy membership function is introduced to automaticrepresentation of the physiological signal feature based on theattribute value into equal ranges [0-1]. The trapezoidal curve isa function of a gene expression matrix, Y, and depends on fourscalar parameters a, b, c, and d, as given by

� �,�, �, �,� =  0, � ≤ �� − �� − � ,� ≤ � ≤ �1, � ≤ � ≤ �� − �� − �0,� ≤ � ,   � ≤ � ≤ � (1)

Dynamic population size: The basic problem with thetraditional GA is the static population size. So thecomputational complexity drastically increases if out of kphysiological signal features (chromosomes), the fitness valuesof k/2 chromosomes are below par. GAs would consider thesephysiological signal features (chromosomes) for reproductionusing crossover and mutation, thus increasing the timecomplexity. In the modified GA, a cut-off on the classificationaccuracy has been considered and every physiological signalfeature that has a fitness value less than this cut-offclassification accuracy is discarded. If at any point after thecutoff, the number of physiological signal features is greater

than the initial population size, the size is reset to initialpopulation size with the less fit physiological signal featuresbeing discarded. Thus, in this way the number of physiologicalsignal features at any point will never be greater than the sizeof the initial population from IAPS database, thus ensuringcomputational efficiency [32].

Dynamic elitism: The global elitism is either done on aphysiological signal feature to physiological signal featurebasis or a fixed number of physiological signal features areconsidered as elite individuals. The current approach that isbeing used in this Modified Genetic Algorithm (MGA), thenumber of elite physiological signal features is dynamic, i.e. itis changing from generation to generation. The advantage ofthis method is that, the life of the physiological signal featuresis directly proportionality with the fitness [32] (classificationaccuracy).

Aging factor: A parameter called the age of the physiologicalsignal feature (chromosome) has been introduced. Theunderlying principle behind the inclusion of this parameter isthat, the physiological signal features (chromosomes) that arefit to live on for a large number of generations have alreadyreproduced in the previous generations [32]. Thus, allowingthese physiological signal features (chromosomes) toreproduce again will decrease the diversity of the populationfrom IAPS database and hence may cause a prematureconvergence. Thus, the fitness values of the physiologicalsignal features that are considered for the sake of reproductionare indirectly proportional to the age of the physiologicalsignal features (Figure 2).

Figure 2. Flow chart of a MTFGA.

Mean based weighted for quaternions firefly algorithm(MWQFA): Firefly Algorithm (FA) as being one of the morefamous representatives of this class of algorithm. Fireflies areinsects, the main characteristic of which is their flashing lightsthat can be admired in the summer sky at night. These lightshave two fundamental functions, i.e., to attract mating partnersand to warn off potential predators. The flashing

AIEFS, heterogeneous ensemble classification

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S243

lights’intensity I decreases as the distance r increases accordingto the term to formulate the FA [33]. To avoid prematureconvergence in FA algorithm introduce a quaternion’srepresentation of physiological signal features. In mathematics,quaternions extend complex numbers. Quaternion algebra isconnected with special physiological signal features ofgeometry of the appropriate Euclidian spaces. However, fitnessvalue is determined not only based on the classificationaccuracy, here the fitness value is determined based on thestandard deviation value of the physiological signal features.The light-intensity I is considered as the fitness value(classification accuracy) of flashing firefly increases as thedistance between two physiological signal features in thepopulation (IAPS database) r decreases in terms of IαI/r2.Additionally, the air absorption causes the light to becomeweaker classification accuracy and weaker as the distance fromthe physiological signal features in the population (IAPSdatabase) increases. Here, the light-intensity is proportional tothe fitness function of the problem being optimized (i.e.,� ��� ∝ ��� ��� where s=S(psf) represent a candidatesolution[34] . In order to formulate the FA, some flashingcharacteristics of fireflies were idealized, as follows:TheMWQFA is based on the original FA, where the representationof virtual fireflies (physiological signal features) is movedfrom a Euclidian space to a quaternion space. In the Euclidianspace, each virtual firefly (physiological signal features) isrepresented as D-dimensional real-values physiological signalfeatures vector psf=(psfi0,…psfin), where ����� ∈ ℝ�, while inquaternion space as a D-dimensional vector of quaternionsqi={qi0,…qin}, where ��� ∈ ℍ�. So the search-process couldbe directed towards the more promising areas of the search-space.

The MWQFA differs from the original FA by using thequaternion’s representation of physiological signal features. Onthis quaternion’s representation of physiological signalfeatures, however, the quaternion algebra is applied.Quaternions are formal expressions q=x0+x1i+x2j+x3k, wherex0, x1, x2, x3 real values of physiological signal are featuresand they constitute the algebra over the real numbers generatedby basic units i, j, k (also the imaginary part) that satisfyHamilton’s equations:

ij=k, jk=i, ki=j → (2)

ij=-k, kj=-i, ik=-j → (3)

i2=j2=k2=-1 → (4)

The quaternions � ∈ ℍ describes a 4-dimensional space overthe real numbers. Using this notation, a pair of quaternions isdenoted as q0=x0+x1i+x2j+x3k and q1=y0+y1i+y2j+y3k. Thequaternion algebra defines the following operations such asaddition and subtraction, scalar multiplication, multiplication,on quaternions. In addition to pure quaternion algebra, twounary functions are added as follows: qrand() is a quaterniondefined as

qrand()={xi=N(0,1)|for i=0,…3} → (5)

where N(0,1) denotes a random number drawn from aGaussian distribution with zero mean and standard deviationone. In other words, each component is initialized with therandom generated number. qzero: is a quaternion defined as

qzero()={xi=0|for i=0,…3} → (6)

Where each component of quaternion is initialized with zero.The QFA algorithm acts as follows. The population ofquaternions is initialized in InitQFA() using the qrand()function. The solution psfs=(psfs0,…psfsD) in the Euclidianspace is obtained from i-th quaternions’ vector qi using thenorm function as follows:

psfs=||qij||, for j=1 to D → (7)

Calculating the distance between the fireflies (physiologicalsignal features) in the search-space is expressed in themodified algorithm based on the weight value of the twofeatures���2 = ���(���) = (��−��)2 (8)Where qi is the i-th virtual firefly ((physiological signalfeature) position, and qj is the j-th virtual firefly (physiologicalsignal feature) position in the search-space. Moving the fireflyi to another more attractive firefly j is expressed as follows:���(���) = (��−��)2 (9)�� = ��+ �0�−����2 (��− ��) + � . � .�����() (10)Where r2

ij represents the distance between the i-th and j-thfireflies in the quaternion’s space, α is the randomizationparameter, ε the scale, and the Qrand() is a random generatedquaternion vector. After moving the virtual fireflies(physiological signal features) a verification function islaunched. It ensures that the new firefly (physiological signalfeature) position is under the prescribed limitations, i.e., lbi ≤ ||qi|| ≤ ubi. In this work the fitness value is updated based on theweight values of the physiological signal features, a fitnessvalue (fitmessi) for a physiological signal features selectionproblem can be assigned to the solution qij by (10).

�������� = 11 + ���� .�� ���� .�� ≥ 01��� ���� .�� ���� .�� < 0 (11)

Where fiti is the classification accuracy. According to theSDWFA declaration, assigned a weight W(psfi) to eachattribute psfi. The value of weight W(psfi) for each psfi, whichis set to zero initially, is calculated sequentially throughout thewhole matrix using the mean value of the attribute and updateusing the following formula when a new entry ai is met in thediscernibility matrix:

wi=w(psfi).μ( psfi) → (12)

When the optimization problem involves more than oneobjective function is described in equation (11-12), the task is

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S244

to find more optimal physiological signal features solutionsregarding each objective.

Decision support: One of the commonly used approaches fordealing with EFS is majority voting, where the most agreedfeature is selected as the final ensemble prediction. Similarly, amajority voting scheme with threshold may be adopted forFSE. Using the notations introduced earlier, for a givenensemble E, the decisions of the ensemble components can beorganized in a K × MBoolean decision matrix D, where K isthe size of the ensemble, and M is the total number ofphysiological signal features. In this representation, thehorizontal row Di denotes the feature subset fsi, and the binarycell value Dij indicates whether psfi Є fsi.Borrowing theterminology of ensemble EFS, the ensemble agreement j forthe physiological signal features psfj can therefore becalculated by:γj = ∑������ (13)A agreement threshold , 0<, can then be defined to control thenumber of physiological signal featuresbeing included in thefinal result psf*, such that: psf*=1; if j. From this, the commonmajority vote can be assimilated by setting =0.5. The value αmay be adjusted according to the problem at hand, if theamount of agreement is very high, a higher α value can be usedto control the size of the resultant feature subset. Alternatively,if a highly diverse FSE is obtained, there may exist no featurewith j to combat this, it may be necessary to employ a loweredα value. Finally less error values are selected using intersectionoperation in mathematical framework.

Heterogeneous ensemble classification (HEC)To propose and test the efficacy of HEC methods wherein threebase classifiers that learn different aspects of the emotiondataset samples with selected physiological signal features areused together to make collective decisions in order to enhanceperformance of health-related message classification.

Support vector machine (SVM) [35] is a function basedclassifier built upon the concept of decision planes that definedecision boundaries. In this experiment we use the linearkernel SVM with C=1.0. SVM has long been known forsuperior performance for all examples let us consider in textclassification with word features. In this work follows theprocedure of the hybrid prediction algorithm comprised ofSupport vector regression (SVR) and Genetic Algorithm (GA)is proposed for parameter optimization. The SVR modelsutilize the nonlinear mapping feature to deal with nonlinearregressions. However, general SVR suffers from parameterproblem algorithm so, in the proposed method the GA isemployed to obtain the SVR parameters accurately andeffectively. Which is similar to the previous work [36], hereinstead of Modified Firefly Algorithm (MFA), GA is used inthis research work. Multinomial NaiveBayes (MNB) [37]implements the Naïve Bayes algorithm for multinomiallydistributed data, and is one of the two classic Naive Bayesvariants used where the data is typically represented as

physiological signal features vector. McCallum and Nigamcite[31] found Multinomial NaïveBayes to perform better thansimple NaiveBayes, especially at larger physiological signalfeatures.

OS-ELM is developed on the basis of Extreme LearningMachine (ELM) [38] that is used for batch learning withemotion classification based on physiological features and hasbeen shown to be extremely high emotion classificationperformance. Compared to ELM, OS-ELM can learn at a one-by-one with fixed or varying chunk size for emotion datasetsamples with selected physiological signal features. Theparameters of OS-ELM in the hidden nodes, input weights andbiases for additive nodes for RBF nodes are randomly selectedand the output weights are analytically determined. OS-ELM isperformed based on physiological signal featuresand it issimilar to ELM with SLFNs and RBF hidden nodes, expectsequential manner. Consider N arbitrary emotion datasetsamples (xi,ti) Є Rn × Rm If a SLFN with L hidden nodes canapproximate these N emotion dataset samples with mphysiological signal features and equals to zero error, it thenimplies that there exist bi, ai and bi such that there exists βi, aiand bi such that

�� �� =∑� = 1� ��� ��, ��, �� = ��, � = 1, ..� (14)

Where bi,ai are the learning parameters of the hidden nodes, biis the output weight, and G(ai,bi,xj) denotes the output of theith hidden node with respect to the emotion dataset samples xj.When using additive hidden node,� ��, ��, �� = � �� . ��+ �� ,   �� ∈ � (15)Where ai is the input weight vector, bi is the bias of the ithhidden node, and ai.xj denotes the inner product of the two.When using RBF hidden node,� ��, ��, �� = �(�� |��+ ��), �� ∈ �+ (16)Where ai and bi are the center and impact width of the ith RBFnode, and R+, indicates the set of all positive real values.Assume the network has L hidden nodes and the data. Thereare two phases in OS-ELM algorithm, an initialization phaseand a sequential phase. In the initialization phase, rank H0= Lis required to ensure that OS-ELM can achieve the samelearning performance as ELM, where H0 denotes the hiddenoutput matrix for initialization phase. It means the number oftraining data required in the initialization phase N0 has to beequal to or greater than L, i.e. N0>L. And if N0=N, OS-ELM isthe same as batch ELM. Hence, ELM can be seen as a specialcase of OS-ELM when all the data present in one iteration.

(a) Randomly assign the input parameters: for additive hiddennodes, parameters are input weights ai and bias bi; for RBFhidden nodes, parameters are center ai and impact factor bi;i=(1,…L).

(b) Calculating the initial hidden layer output matrix H0

AIEFS, heterogeneous ensemble classification

Biomed Res- India 2016 Volume Special Issue IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S245

�0 = � �1, �1, �1 … � ��, ��, ��⋮� �1, �1, ��0 � ��, ��, ��0 �0 × �(17)

Estimating the initial output weight β0, Set k=0. (k: a parameterindicates the number of chunks of data that is presented to thenetwork).

Sequential learning phase: Present the (k+1) th chunk of newobservations

��+ 1 = ��, �� � = (∑� = 0� ��) + 1∑� = 0�+ 1�� (18)and Nk+1 denotes the number of observation for emotiondataset samples with m physiological signal features in the (k+1) the chunk .Compute the partial hidden layer output matrixHk+1� �1, �1, � ∑� = 0� �� + 1  ⋯� ��, ��, � ∑� = 0� �� + 1� �1, �1, � ∑� = 0� �� + 1 � ��, ��, � ∑� = 0� �� + 1 ��+ 1 × �

Calculate the output weight β(k+1). Have

��+ 1 = �(∑� = 0�+ 1��)  + 1,…,�(∑� = 0�+ 1��) ���+ 1 ×� (20)

Set k=k+1, go to (a) in this sequential learning phase.

Ensemble of EOS-ELM consists of many OS-ELM networkswith same number of hidden nodes and same activationfunction for each hidden node. All EOS-ELMs are trained withnew emotion dataset samples with m physiological signalfeatures in each incremental step. The input parameters foreach OS-ELM network are randomly generated and the outputweights are obtained analytically based on the sequentialarrived input emotion dataset samples with m physiologicalsignal features. Then compute the average of the outputs ofeach OS-ELM network, which is the final output of the EOS-ELM. Assume the output of each OS-ELM network is f(i)

(xi),j=1,.P. Hence,

� �� = 1�∑� = 1� � � �� (21)

Expect that EOS-ELM works better than individual OS-ELMnetwork because the randomly generated parameters makeeach OS-ELM network in the ensemble distinct. Therefore, theOS-ELM networks composing the ensemble may havedifferent adaptive capacity to the new emotion dataset sampleswith m physiological signal features. When the emotion datasetsamples with m physiological signal features come into theensemble network sequentially, some of OS-ELM networksmay adapt faster and better to the new data than others.

Simulation results in [38] have shown that EOS-ELM is fasterthan other OS-ELM and produces better generalizationperformances. When N0=N, EOS-ELM becomes an ensembleof batch ELM networks [39]. Therefore, the ensemble of ELMproposed in [39] can be seen as a special case of EOS-ELMwhen all the training emotion dataset samples with mphysiological signal features are available at one time. Finallyall the classification methods are combined into ensemblemethod by using majority voting. When compared to allclassifier ensemble method provides higher classificationaccuracy which is discussed in the experimentation results.

Experimentation ResultsIn order to classify discrete emotions based on participants'subjective ratings, every segment of participant's physiologyacquired during stimulation was associated with the referentemotional state. During the experiment, each participant gavesubjective ratings regarding the emotional state that a particularstimulation elicited in her. She was supposed to give perceivedintensity for all discrete emotions, and in many instances co-occurring emotions [40,41] appeared, in which the participantperceived more than one discrete emotion as very intense.Therefore, an algorithm was developed for finding referentemotions that were elicited even in such ambiguous cases. Thealgorithm resolves the referent emotion based on the intensityof sadness, disgust, fear, happiness and other reported discreteemotions, depending on the intended emotion that a particularstimuli sequence was expected to elicit. By conducting thealgorithm over all subjective ratings of the participants, thefollowing numbers of samples for each discrete emotion wereobtained: 91 samples of sadness, 95 samples of disgust, 38samples of fear, 78 samples of happiness, 65 samples of anger,210 samples that algorithm annotates with “surprise, stress andengagement” such as 85 samples for surprise, 55 samples forstress, and 70 samples for engagement.

To apply the data mining algorithms, MATLAB tool was used.MATLAB is an environment for machine learning, datamining, text mining and business analytics. It is used forresearch, education, training and industrial applications. In thisstudy, version 2013 of MATLAB is used. All algorithms wereused in the default state. In what follows the obtained resultsand discussions are presented. The experiments are designed sothat the different parts of the work could be evaluated. Theseinclude the evaluation of the features of the dataset and thefeature selection. To this aim, first the features which wereselected by the feature selection method and their importanceare discussed. Second, all the two possible combinations of thefeature selection and classification methods are tested over thedataset. Finally, results techniques are presented in this section.Accuracy, Precision, and Recall are the most importantperformance measures in the medical field, which arecommonly used in the literature. So for measuring theperformance of algorithms, these measures are used.

Confusion matrix: A confusion matrix is a table that allowsvisualization of the performance of an algorithm. In a two classproblem (with classes C1 and C2), the matrix has two rows and

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S246

two columns that specifies the number of False Positives (FP),False Negatives (FN), True Positives (TP), and True Negatives(TN). These measures are defined as follows: TP is the numberof samples of class C1 which has been correctly classified. TNis the number of samples of class C2 which has been correctlyclassified. FN is the number of samples of class C1 which hasbeen falsely classified as C2. FP is the number of samples ofclass C2 which has been falsely classified as C1. Table 1 showsconfusion matrix. In this eight classes are used Sadness,disgust, fear, happiness, anger, surprise, stress and engagement.

Table 1. Confusion matrix.

Class Actual class C1 Actual class C2

Predicted class C1 True positive (TP) False positive (FP)

Predicted class C2 False Negative (FN) True Negative (TN)

Accuracy: Accuracy shows ratio of correctly classifiedsamples to the total number of tested samples. It is defined as:�������� = ��+ ����+ ��+ ��+ �� (22)Precision and recall: According to confusion matrix,precision and recall are explained as following,��������� = ����+ �� (23)������ = ����+ �� (24)Table 2. Comparison of classification accuracy using differentfeatures reduction and machine learning methods.

Featureselection

Accuracy (%) Average

MLP SVR-GA MNB EOS-ELM HEC

SFFS 60.65 62.392 64.296

65.85 67.2464.0856

MTFGA 68.63 71.231 73.83 74.523 76.083 72.8594

MWQFA 77.296

79.203 82.322

83.247 83.36281.086

Proposed AIEFS 84.005

85.269 87.002

88.562 91.68187.3038

Believe that the choice of evaluation method is an importantreason why the best-case accuracies obtained in Table 2 are inthe range 55-60%, even though related work analysis hasshown the accuracies around 64-89% for classification of acomparable number of distinct emotions. From the table it isconcluded that the proposed AIEFS-HEC provides averageaccuracy results of 91.681 % which is 7.676%, 6.412%,4.679% and 3.119% high when compared to MLP, SVR-GA,MNB and EOS-ELM methods respectively. Believe that thechoice of evaluation method is an important reason why thebest-case recall obtained in Table 3 are in the range 55-60%,even though related work analysis has shown the recall around65-91% for classification of a comparable number of distinctemotions. From the table it is concluded that the proposed

AIEFS-HEC provides precision results of 90.99% which is8.2%, 6.27%, 4.92% and 3.27% high when compared to MLP,SVR-GA, MNB and EOS-ELM methods respectively.

Table 3. Comparison of classification recall using different featuresreduction and machine learning methods.

Featureselection

Recall (%) Average

MLP SVR-GA MNB EOS-ELM HEC

SFFS 58.43 60.234 62.22 63.81 65.49 62.036

MTFGA 66.75 69.33 72.25 72.674 74.57 71.11

MWQFA 75.63 77.50 80.73 81.81 81.83 79.5

AIEFS 82.79 84.72 86.07 87.72 90.99 86.45

Table 4. Comparison of classification recall using different featuresreduction and machine learning methods.

Featureselection

Precision (%) Average

MLP SVR-GA MNB EOS-ELM HEC

SFFS 60.6560.34 62.311 63.96

65.506 62.553

MTFGA 67.342 69.64 71.63 73.06 74.84 71.302

MWQFA75.82 77.521 80.7178 81.676

81.711 79.489

AIEFS82.96 84.39 86.09 87.455

90.763 63.894

Figure 3. Comparison of classification methods using SFFS featurereduction.

Believe that the choice of evaluation method is an importantreason why the best-case Precision obtained in Table 4 are inthe range 55-60%, even though related work analysis hasshown the Precision around 62-91% for classification of acomparable number of distinct emotions. From the table it isconcluded that the proposed AIEFS-HEC provides Precisionresults of 90.763% which is 7.803%, 6.373%, 4.67% and3.308% high when compared to MLP, SVR-GA, MNB andEOS-ELM methods respectively.

AIEFS, heterogeneous ensemble classification

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S247

Figure 4. Comparison of classification methods using MTFGAfeature reduction.

From the experimental results, it is concluded that theperformance comparison results of accuracy for the emotiondataset the proposed HEC with SFFS algorithm performs6.59% better than the MLP algorithm, 4.848% better than theSVE-GA algorithm, 2.944% better than the MNB algorithmand 1.39% better than the EOS-ELM algorithm is illustrated inFigure 3. Similarly recall for the emotion dataset the proposedHEC with SFFS algorithm performs 7.06% better than theMLP algorithm, 5.256% better than the SVR-GA algorithm,3.27% better than the MNB algorithm and 1.68% better thanthe EOS-ELM algorithm is illustrated in Figure 3. Similarly itworks better for precision parameter.

Figure 5. Comparison of classification methods using MWQFAfeature reduction.

From the experimental results, it is concluded that theperformance comparison results of accuracy for the emotiondataset the proposed HEC with MTFGA algorithm performs7.453% better than the MLP algorithm, 4.852% better than theSVE-GA algorithm, 2.253% better than the MNB algorithmand 1.56% better than the EOS-ELM algorithm is illustrated inFigure 4. Similarly recall for the emotion dataset the proposedHEC with MTFGA algorithm performs 7.82% better than theMLP algorithm, 5.24% better than the SVR-GA algorithm,2.32% better than the MNB algorithm and 1.896 % better than

the EOS-ELM algorithm is illustrated in Figure 4. Similarly itworks better for precision parameter.

From the experimental results, it is concluded that theperformance comparison results of accuracy for the emotiondataset the proposed HEC with MWQFA algorithm performs6.066% better than the MLP algorithm, 4.159% better than theSVE-GA algorithm, 1.04% better than the MNB algorithm and0.115% better than the EOS-ELM algorithm is illustrated inFigure 5. Similarly recall for the emotion dataset the proposedHEC with MWQFA algorithm performs 6.2% better than theMLP algorithm, 4.33% better than the SVR-GA algorithm,1.11% better than the MNB algorithm and 0.2% better than theEOS-ELM algorithm is illustrated in Figure 5.

Figure 6. Comparison of classification methods using AIEFS featurereduction.

From the experimental results it is concluded that theperformance comparison results of accuracy for the emotiondataset the proposed HEC with AIEFS algorithm produces91.681%, it performs 7.676% better than the MLP algorithm,6.412% better than the SVE-GA algorithm, 4.679% better thanthe MNB algorithm and 3.119% better than the EOS-ELMalgorithm is illustrated in Figure 6. Similarly recall for theemotion dataset the proposed HEC with AIEFS algorithmperforms 8.2% better than the MLP algorithm, 6.27% betterthan the SVR-GA algorithm, 4.92% better than the MNBalgorithm and 3.27% better than the EOS-ELM algorithm isillustrated in Figure 6. From the experimentation results inFigures 3-6 it concludes that the proposed the proposed HECwith AIEFS algorithm produces higher accuracy results whencompared to all the classifiers and feature selection, since theproposed work feature selection and classification is donebased on the ensemble construction.

Conclusion and Future WorkThis paper has addressed the gap in feature reduction andmachine-learning methods in physiological parameter basedreal time emotion estimation and therefore it provides an addedinsight in creating an adaptive emotion estimator. The majoraim of this work is to identify preferable Artificial IntelligentEnsemble Feature Selection (AIEFS) framework andHeterogeneous Ensemble Classification (HEC) model for such

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S248

a concept. To realize the comparative analysis of methods forphysiology-based emotion estimation, an emotion basedexperiment was conducted using a set of static data from theIAPS database to elicit Sadness, disgust, fear, happiness, anger,surprise, stress and engagement state. Emotional responseswere calculated by standard set of physiological signals whichincluded ECG, respiration, skin conductance and skintemperature, as well as by individual subjective ratings ofdisplayed images.In order to establish the most suitableemotion estimation methods for real time applications,comparative analysis based on estimation accuracy wasperformed.The highest classification accuracy was achievedwith the HEC model and AIEFS feature selection algorithm.From the experimental results it is concluded that theperformance comparison results of accuracy with the emotiondataset for the proposed HEC with AIEFS algorithm produces91.681%, it performs 7.676% better than the MLP algorithm,6.412% better than the SVE-GA algorithm, 4.679% better thanthe MNB algorithm and 3.119% better than the EOS-ELMalgorithm. Because of its high accuracy AIEFS isrecommended for feature analysis. Future work involves an in-depth analysis using real-time estimator. This analysis willinclude an experimental verification of improvements that canbe obtained under certain conditions such as the number ofsessions, the number of data types that can be obtained duringthe session, the homogeneity of the participant's group, etc.

References1. Harms MB, Martin A, Wallace GL. Facial emotion

recognition in autism spectrum disorders: a review ofbehavioral and neuroimaging studies. Neuropsychol Rev2010; 20: 290-322.

2. Kuusikko S, Haapsamo H, Jansson-Verkasalo E. Emotionrecognition in children and adolescents with autismspectrum disorders. J Autism Dev Disord 2009; 39:938-945.

3. Evers K, Noens I, Steyaert J, Wagemans J. Combiningstrengths and weaknesses in visual perception of childrenwith an autism spectrum disorder: perceptual matching offacial expressions. Res Autism Spectrum Disord 2011; 5:1327-1342.

4. Da Fonseca D, Santos A, Bastard-Rosset D, Rondan C,Poinso F, Deruelle C. Can children with autistic spectrumdisorders extract emotions out of contextual cues? ResAutism Spectrum Disord 2009; 3: 50-56.

5. Gepner B, F´eron F. Autism: a world changing too fast for amis-wired brain? Neurosci Biobehav Rev 2009; 33:1227-1242.

6. Rani P, Liu C, Sarkar N, Vanman E. An empirical study ofmachine learning techniques for affect recognition inhuman-robot interaction. Pattern Anal Appl 2006; 9: 58-69.

7. Cosic K, Popovic S, Kukolja D, Horvat M, Dropuljic B.Physiology-driven adaptive virtual reality stimulation forprevention and treatment of stress related disorders.Cyberpsychol Behav Soc Netw 2010; 13: 73-78.

8. Picard RW. Emotion research by the people, for the people.Emot Rev 2010; 2: 250-254.

9. Jang EH, Park BJ, Park MS, Kim SH, Sohn JH. Analysis ofphysiological signals for recognition of boredom, pain, andsurprise emotions. J Physiol Anthropol 2015; 34: 25.

10. Wagner J, Kim J, André E. From physiological signals toemotions: Implementing and comparing selected methodsfor feature extraction and classification. In 2005 IEEEInternational Conference on Multimedia and Expo, 2005,940-943.

11. Kim J, André E. Emotion recognition based onphysiological changes in music listening. IEEE TransPattern Anal Mach Intell 2008; 30: 2067-2083.

12. van den Broek EL, Lisý V, Janssen JH, Westerink JHDM,Schut MH, Tuinenbreijer K. Affective man-machineinterface: unveiling human emotions through biosignals.Biomed Eng Syst Technol 2010.

13. Katsis CD, Katertsidis NS, Fotiadis DI. An integratedsystem based on physiological signals for the assessment ofaffective states in patients with anxiety disorders. BiomedSignal Process Control 2011; 6: 261-268.

14. Park BJ, Jang EH, Kim SH, Huh C, Sohn JH. Sevenemotion recognition by means of particle swarmoptimization on physiological signals: Seven emotionrecognition. 2012 9th IEEE International Conference onNetworking, Sensing and Control (ICNSC), 2012, 277-282.

15. Subi D, Anuja HS. Detection of Automobile Drivers Stressfrom Physiological Signals. Int J Comput Sci MobileComput 2014; 3: 454 - 458.

16. Niu X, Chen L, Xie H, Chen Q, Li H. Emotion PatternRecognition Using Physiological Signals. Sens Transduc2014; 172: 147.

17. Valenzi S, Islam T, Jurica P, Cichocki A. Individualclassification of emotions using EEG. J Biomed Sci Eng2014; 7: 604.

18. Kukolja D, Popovic S, Horvat M, Kovac B, Cosic K.Comparative analysis of emotion estimation methods basedon physiological measurements for real-time applications.Int J Human-Comput Studies 2014; 72: 717-727.

19. Krupa N, Anantharam K, Sanker M, Datta S, Sagar JV.Recognition of emotions in autistic children usingphysiological signals. Health Technol 2016.

20. Smitha KG, Vinod AP. Facial emotion recognition systemfor autistic children: a feasible study based on FPGAimplementation. Med Biol Eng Comput 2015; 53:1221-1229.

21. Lang PJ, Bradley MM, Cuthbert BN. InternationalAffective Picture System (IAPS): Affective Ratings ofPictures and Instruction Manual. University of Florida,2008, Gainesville, FL (Technical Report A-8).

22. http://csea.phhp.ufl.edu/media.html23. Barke A, Stahl J, Kröner-Herwig B. Identifying a subset of

fear-evoking pictures from the IAPS on the basis ofdimensional and categorical ratings for a German sample. JBehav Ther Exp Psychiatry 2012; 43: 565-572.

AIEFS, heterogeneous ensemble classification

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S249

24. Libkuman TM, Otani H, Kern R, Viger SG, Novak N.Multidimensional normative ratings for the InternationalAffective Picture System. Behav Res Methods 2007; 39:326-334.

25. Mikels JA, Fredrickson BL, Larkin GR, Lindberg CM,Maglio SJ, Reuter-Lorenz PA. Emotional category data onimages from the International Affective Picture System.Behav Res Methods 2005; 37: 626-630.

26. Jurin T, Jokic-Begic N, Korajlija AL. Factor structure andpsychometric properties of the anxiety sensitivity index in asample of Croatian adults. Assessment 2012; 19: 31-41.

27. Rottenberg J, Ray RD, Gross JJ. Emotion elicitation usingfilms. In: Handbook of Emotion Elicitation andAssessment. Oxford University Press, 2007, New York,USA.

28. Christov I. Real time electrocardiogram QRS detectionusing combined adaptive threshold. BioMed Eng Online2004; 3: 28.

29. Liu H, Motoda H. Computational Methods of FeatureSelection (Chapman & Hall/CrcDataMining andKnowledge Discovery Series). Chapman & Hall/CRC,2007.

30. Boongoen T, Shen Q. Nearest-neighbor guided evaluationof data reliability and its applications. IEEE Trans SystMan Cybern B Cybern 2010; 40: 1622-1633.

31. Mac Parthal´ain N, Jensen R. Fuzzy-rough set based semi-supervised learning. IEEE International Conference onFuzzy Systems, 2011.

32. Yang XS. Firefly algorithm. In X.-S. Yang (Ed.), Nature-inspired metaheuristic algorithms. Wiley Online Library2008.

33. Fister I, Yang XS, Brest J, Fister I. Memetic self-adaptivefirefly algorithm. In Swarm intelligence and bio-inspiredcomputation: theory and applications. Elsevier, 2013,Amsterdam.

34. Bishop CM. Pattern recognition and machine learning(information science and statistics). Secaucus, NJ, USA:Springer-Verlag New York, Inc.; 2006

35. Kavousi-Fard A, Samet H, Marzbani F. A new hybridmodified firefly algorithm and support vector regressionmodel for accurate short term load forecasting. Expert SystAppl 2014; 41: 6047-6056.

36. Mccallum A, Nigam K. A comparison of event models forNaive Bayes text classification. In: AAAI-98 workshop on‘Learning for Text Categorization’; 1998.

37. McCallum A, Nigam K. A comparison of event models forNaive Bayes text classification. AAAI-98 workshop onlearning for text categorization 1998.

38. Huang GB, Chen L, Siew CK. Universal approximationusing incremental constructive feed forward networks withrandom hidden nodes. IEEE Transact Neural Networks2006; 17: 879-892.

39. Golub GH, Loan CFV. Matrix Computations, 3rd Ed., TheJohns Hopkins University Press, Baltimore, MD,1996.

40. Harley JM, Bouchet F, Azevedo R. Measuring learners' co-occurring emotional responses during their interaction witha pedagogical agent in MetaTutor. In: Intelligent TutoringSystems, ITS '12, Lecture Notes in Computer Science2012; 7315. Springer-Verlag, Berlin.

41. www.easychair.org/publications/download/Feature_Selection_Ensemble

42. Sridharan B. Modifications in Genetic Algorithm usingadditional parameters to make them computationallyefficient. 2010 IEEE 2nd International Advance ComputingConference.

*Correspondence toK Kiruba

Department of Electronics and Communication Engineering

K.S.R. College of Engineering

India

Kiruba/Sharmila

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S250