A new approach for EEG signal classification of schizophrenic and control participants

Abstract

This paper is concerned with a two stage procedure for analysis and classification of electroencephalogram (EEG) signals for twenty schizophrenic patients and twenty age-matched control participants. For each case, 20 channels of EEG are recorded. First, the more informative channels are selected using the mutual information techniques. Then, genetic programming is employed to select the best features from the selected channels. Several features including autoregressive model parameters, band power and fractal dimension are used for the purpose of classification. Both linear discriminant analysis (LDA) and adaptive boosting (Adaboost) are trained using tenfold cross validation to classify the reduced feature set and a classification accuracy of 85.90% and 91.94% is obtained by LDA and Adaboost, respectively. Another interesting observation from the channel selection procedure is that most of the selected channels are located in the prefrontal and temporal lobes confirming neuropsychological and neuroanatomical findings. The results obtained by the proposed approach are compared with a one stage procedure, the principal component analysis (PCA)-based feature selection, utilizing only 100 features selected from all channels. It is illustrated that the two stage procedure consisting of channel selection followed by feature reduction gives a more enhanced results in an efficient computation time.

Introduction

Schizophrenia is a severe and persistent debilitating psychiatric disorder that affects approximately 0.4–0.6% of the world’s population. Affected individuals frequently come to clinical attention during late adolescence or early adulthood. According to the diagnostic criteria of the American Psychiatric Association (DSM-IV) (American Psychiatric Association, 2000), patients show disturbances in thoughts, affects, and perceptions and difficulties in relationships with others. Schizophrenia is often described in terms of positive and negative symptoms. Positive symptoms include delusions, auditory hallucinations and thought disorder and typically regarded as manifestations of psychosis. Negative symptoms are the loss or absence of normal traits or abilities, poverty of speech and lack of motivation. Patients with schizophrenia have lower rates of employment, marriage, and independent living than other people.

Because schizophrenia represents a disturbance in some, but not all, brain functions, it is reasonable to suppose that specific brain regions or neural circuits are involved. Neuropsychological studies (Sadock & Sadock, 2004, chap. 12) have found evidence in patients with schizophrenia for decreased cortical gray matter in the prefrontal and temporal cortex; cerebral white matter fiber tract alternations, decreased volume of limbic system structures, for example, the amygdale, hippocampus and entorhinal cortex and the thalamus; and increased volume of basal ganglia nuclei. Also, neuroanatomical studies (Sadock & Sadock, 2004, chap. 12) propose that the disturbance of prefrontal and limbic system function lead to the positive and negative symptoms and cognitive impairments observed in patients with schizophrenia. Although clinical, neuroanatomical, neuropsychological and neuroimaging approaches have contributed to a better understanding of the illness, a more precise knowledge of the underlying mechanisms is still necessary.

Recently, much attention has been paid to analysis of electroencephalogram (EEG) signals of schizophrenic patients. In some research (Jeong et al., 1998, Koukkou et al., 1993, Roschke et al., 1995), nonlinear methods have been applied to EEG signals of schizophrenic patients and control participants. The results showed differences in dynamic process between the two groups. Hornero et al. (2006) asked the participants to press space bar key randomly to generate time series. The results showed that the time series generated by schizophrenic patients had a lower complexity than the control group. In an interesting test, for random number generation (Rosenberg, Weber, Crocq, Duval, & Macher, 1990), participants were asked to choose a number from one to ten several times. Numbers had to lack a generative rule, that is, to be as random as possible. They found that schizophrenic patients tended to be more repetitive. Pressman, Peled, and Geva (2000) showed lack of synchronization alternation ability in the schizophrenic patients during working memory task. They indicated a difference in brain activity, especially in frontal and temporal channels. Cherif, Naìt-Ali, Motsch, and Krebs (2003) indicated that the schizophrenic patients, compared with healthy participants, presented abnormality in eye fixation tasks. Gaser, Volz, Kiebel, Riehemann, and Sauer (1999) depicted structural brain changes in schizophrenic patients. Moreover, Kiel, Elbert, Rockstroh, and Ray (1998) showed rhythmic finger oscillations in schizophrenic patients. Paulus, Geyer, and Braff (1999) carried out a simple choice task, consisting of predicting 500 random right or left appearances of a stimulus in order to obtain binary response in patients with schizophrenia and control group. After applying mutual and cross-mutual information, they showed that the response sequences generated by patients exhibited a higher degree of interdependency than those in control group.

The objective of this study is to design an easy way to use algorithm producing reliable results in distinguishing schizophrenic patients form the control using information contained in EEG signals. A two stage procedure consisting of channel selection followed by feature reduction is adapted for classification of the two groups. It is shown that the selection of EEG recording locations (channel selection) can be done robustly in the absence of prior knowledge of brain activity. Mutual information as a good indicator of relevance between variables is employed for channel selection. Genetic programming (GP) is used to reduce the feature dimension by considering the classification error as fitness function. Two efficient classifiers are used for the classification purposes using the reduced feature set. The proposed approach is compared and contrasted with the principal component analysis (PCA)-based feature selection method. It is demonstrated that excellent results are obtained in an efficient computational time.