Elsevier

Expert Systems with Applications

Volume 42, Issue 3, 15 February 2015, Pages 1644-1651
Expert Systems with Applications

Efficient classification system based on Fuzzy–Rough Feature Selection and Multitree Genetic Programming for intension pattern recognition using brain signal

https://doi.org/10.1016/j.eswa.2014.09.048Get rights and content

Highlights

  • We design a classification system for brain signal based on evolutionary algorithm.

  • The proposed methods base on fuzzy rough theory and Multitree Genetic Programming.

  • We examine the intension pattern of brain signals with fNIRS.

  • The proposed FRFS reduced the data volume and extracted the informative features.

  • The proposed GP classified with higher accuracy than conventional methods.

Abstract

Recently, many researchers have studied in engineering approach to brain activity pattern of conceptual activities of the brain. In this paper we proposed a intension recognition framework (i.e. classification system) for high accuracy which is based on Fuzzy–Rough Feature Selection and Multitree Genetic Programming. The enormous brain signal data measured by fNIRS are reduced by proposed feature selection and extracted the informative features. Also, proposed Multitree Genetic Programming use the remain data to construct the intension recognition model effectively. The performance of proposed classification system is demonstrated and compared with existing classifiers and unreduced dataset. Experimental results show that classification accuracy increases while number of features decreases in proposed system.

Introduction

The pattern analysis and recognition system of brain signals has a ripple effect on the whole world. Especially, the conceptual activities of a brain such as thoughts, imaginations are one of the most important thing in the brain signals to recognize the patterns (Izzetoglu, Yurtsever, Bozkurt, & Bunce, 2003). However, the recognition task is very difficult because of the noise from changing human body status and stray thoughts. The brain signal of human is influenced by the varying metabolism, respiration, fatigue, and so on. Also, the conceptual activities by about 14 billion neurons are too much complex to find out specific processes of them, even if the all external factors are excepted (Muroga, Tsubone, & Wada, 2006). These difficult problems can be relieved by Functional Near-InfraRed Spectroscopy (fNIRS) that is a method to collect brain signal data. The fNIRS system is non-invasive brain function analysis equipment that measure the consistency of oxy-hemoglobin and deoxy-hemoglobin of blood at cerebral cortex (Izzetoglu et al., 2003). The fNIRS can continuously measure the brain signal more than 10 times per a second and near-infrared spectroscopy have the more reduced noise than Electroencephalography (EEG). It applied medical engineering and neurology field (Muroga et al., 2006).

Recently, a considerable number of researches of brain-computer interface (BCI) are focused on rehabilitation to recovery brain activity for movements, or aid the conquest of partial uncomfortable actions by ordering via thinking to the motors (Abibullaev et al., 2014, Mattia et al., 2013, Pfurtscheller et al., 2008). These approaches opened new possibilities to various medical applications. Recent literatures study neuroscience research for neurorehabilitation in stroke (Ang et al., 2010); BCI for communication in real-time (Eklund, Andersson, Ohlsson, Ynnerman, & Knutsson, 2010); controlling orthosis (Ortner, Allison, Korisek, Gaggl, & Pfurtscheller, 2011); approach to hemodynamic signals underlying lower limb motor preparation (Rea et al., 2014); reduction of electrodes and stress to patient (Tam, Tong, Meng, & Gao, 2011); and inferring hand movement kinematics for hand plegia (Jerbi et al., 2011). More recent study have researched brain-controlled mobile robots (Bi, Fan, & Liu, 2013); detection of intention of patient (Lim, Hwang, Han, Jung, & Im, 2013); the effects of upper-extremity robot-assisted rehabilitation (Varkuti et al., 2013); motor imagery and passive movement (Yu, Ang, Guan, & Wang, 2013); and test the feasibility of a single trial classifier to detect motor execution (Zimmermann et al., 2013).

Many researches as above have been studied, but it is still difficult to design an effective classification system for BCI because the kinds of data have some problems such as vast volume and noises from the human body. For intension recognition with considerable accuracy, the high-density observation that uses a lot of sensors and frequent measurement is required but it dramatically increases the size of data. To overcome these problems effectively, this study approaches to the system in a perspective on machine learning by means of evolutionary computation that show outstanding performance to find global optimum model. In this paper, we proposed a classification system based on fNIRS, Fuzzy–Rough based Feature Selection (FRFS) and proposed Multitree Genetic Programming (Multitree GP) for the intension pattern recognition with higher accuracy. The main advantage of FRFS is that it reduces the brain signal data. The remains are information rich features which leads to higher performance of future processing. Then these pure data are given to the proposed GP based classifier as an input. Finally high-quality models are constructed based on the results of collaboration between these two methods.

The rest of the paper is organized as follows: Section 2 describes related work of fuzzy–rough sets and genetic programming. Section 3 presents the proposed method and in Section 4 the experimental results are depicted. Conclusion is placed in Sections 5.

Section snippets

Fuzzy–rough set theory

The amount of data which is produced daily is much higher than generated information. Therefore, more cost and time is needed to process, save and maintain those data for later processing. Many problems in machine learning, data mining and pattern recognition involve vast datasets. High dimensional data in terms of features needs huge effort to be processed. Therefore feature selection (FS) method can effectively reduce the size of datasets in one direction. This method selects most-informative

Proposed classification system

The proposed system consists of a preprocessing of feature selection and a Multitree GP classifier. The process of feature selection needs two components 1 – evaluation metric and 2 – search method. The former is based on fuzzy–rough set of Jensen and Shen (2009) and for the later the Shuffled Frog Leaping Algorithm (SFLA) (Eusuff, Lansey, & Pasha, 2006) based on heuristic search is newly employed instead of the typical search method. After preprocessing the data, a multi-class classifier will

Experimental results

The neural signal acquisition was done by a multi channel optical brain imaging system (fNIR-300) and the levels of oxy-, deoxy and total-hemoglobin were specified using 45 signal channels at 14 Hz sampling rate. The signals were collected through optical fibers which were attached to the pre-frontal cortex. As Fig. 5 shows, two cognitive activities of rest  right imagery movement and rest  left imagery movement were sampled in a dataset with three classes rest, right and left (Abibullaev, An,

Concluding remarks

The classification of intension pattern, based on neural signal data, is a challenging task in bioinformatics. Many machine-learning techniques have been developed to obtain highly accurate classification performance. This paper also targets the improvement of the neural signal recognition and proposes a new classification system for intension pattern recognition which is based on FRFS and proposed Multitree GP with considerations of the signal characteristics. The FRFS selects features by

Acknowledgments

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MEST) (NRF-2012R1A2A2A01013735) and by the DGIST R&D Program of the Ministry of Science, ICT and Future Planning (14-RS-02).

References (27)

  • A. Eklund et al.

    A brain computer interface for communication using real-time fMRI

  • M. Eusuff et al.

    Shuffled frog-leaping algorithm: A memetic meta-heuristic for discrete optimization

    Engineering Optimization

    (2006)
  • M. Hall et al.

    The WEKA data mining software: An update

    ACM SIGKDD Explorations Newsletter

    (2009)
  • Cited by (28)

    • Differential evolution algorithm as a tool for optimal feature subset selection in motor imagery EEG

      2017, Expert Systems with Applications
      Citation Excerpt :

      Their method utilizes mutual information for feature selection and SVM for classification. In another experiment, Lee, Anaraki, Ahn, and An (2015) proposed a feature selection algorithm based on fuzzy rough theory and multi-tree genetic programming to improve the classification accuracy of brain signals data measured by fNIRS. Adam et al. (2016) implemented angle modulated simulated Kalman filter to select features along with neural network as a classifier.

    • Self-Organizing and Error Driven (SOED) artificial neural network for smarter classifications

      2017, Journal of Computational Design and Engineering
      Citation Excerpt :

      Classification challenges are ubiquitous in different disciplines and areas of science, business, and industry. To name just a few, we have looked at healthcare and pharmaceuticals (Belacel, 2000; Hunger & Mullighan, 2015; Xu, Xu, & Wunsch, 2009) artificial intelligence and pattern recognition (Lee, Anaraki, Ahn, & An, 2015; Nieddu & Patrizi, 2000; Zehtaban, Elazhary, & Roller, 2016), and business, marketing, finance, and management (Baier & Decker, 2012; Keramati et al., 2014; Lima & de Castro, 2014). Understandably, the efforts to improve the performance of the existing classification techniques such as semi-supervised classification method (Fouss, Francoisse, Yen, Pirotte, & Saerens, 2012), imbalanced data classification (Datta & Das, 2015), or even introduce new methods such as Clustering-based ensembles (Krawczyk, Woźniak, & Cyganek, 2014), and Generalized classifier neural network (Ozyildirim & Avci, 2013) are many.

    • Detection of attention in multi-talker scenarios: A fuzzy approach

      2016, Expert Systems with Applications
      Citation Excerpt :

      In this work, we propose an automatic adaptive m-PSK detector for attentional paradigms based on fuzzy logic. Fuzzy has been thoroughly utilized in many research and engineering fields such as control systems (Wang, Tanaka, & Griffin, 1996; Feng, 2006; Zhou, Li, & Shi, 2015; Cerman, 2013), pattern recognition (Lee, Rahimipour Anaraki, Ahn, & An, 2015; Melin & Castillo, 2013), and aerospace (Napolitano, Cnsanova, Windon, Seanor, & Martinelli, 1999; Barua & Khorasani, 2011). In particular for BCI applications, fuzzy logic has been used for motor imagery classification (Hsu, 2012; Nguyen, Khosravi, Creighton, & Nahavandi, 2015) and mental task recognition (Lledo, Cano, Ubeda, Ianez, & Azorin, 2012; Palaniappan, Paramesran, Nishida, & Saiwaki, 2002).

    • Rough Cognitive Networks

      2016, Knowledge-Based Systems
      Citation Excerpt :

      This classifier is combined with consistency-based subset evaluation and fuzzy–rough instance selection method. Multi-tree Genetic Programming classifier (MT-GPC) [58] is a classifier system based on Fuzzy–Rough Feature Selection and Multi-tree Genetic Programming. In this model, the attribute dimensionality is reduced by the feature selection method, so the informative features are extracted using Fuzzy–Rough Feature Selection.

    View all citing articles on Scopus
    View full text