Journal of Biomedical Engineering and Biosciences (JBEB)

ISSN: 2564-4998

1. Introduction

Schizophrenia and Bipolar disorder are related to a neurocognitive deficit that affects different domains. According to the diagnostic criteria of the American Psychiatric Association (DSM-IV) [1], schizophrenia is characterised by significant impairments related to changes in behaviour and reality such as hallucinations, delusions and disorganized thinking. On the other hand, bipolar patients show disturbances in executive, working memory, verbal memory, visual memory, sustained attention, perceptions, and difficulties in relationships with others. Generally, initial diagnosis of both mental disorders is based on subjective observing, depending on patient’s actions, behavioural changes, history of mental illnesses in the family, etc., following DSM-IV criteria and compared to other patient groups and controls.

However, these methods present, in many cases, difficult diagnosis in early disorder stages. Therefore, additional analysis and tools are employed. In this regard, the use of electrophysiological activity can provide important information for the clinical diagnosis of patients. The measured electroencephalogram (EEG) signals appear to be very useful for the identification of brain rhythms, diagnosis of brain disorders, detection of brain impairments, and consequently the possibility to provide precise treatment to correct or improve certain brain-health conditions [2]-[5].

In this study, the utility of EEG recordings together with feature analysis and deep leaning classification to distinguishing brain disorders is presented. In particular, a radial basis function (RBF) neural network (NN) [6]-[8] has been employed. This type of network has some characteristics that make it ideal for schizophrenia or bipolar disorder. Good performance with different patterns can be achieved, also RBF employs fast training procedures, and it includes simple network configurations. In addition, its network structure can grow to the desired degree of accuracy [9].

The paper is divided into the following sections: In Section 2 the methodology employed for signal acquisition and signal processing are presented. It also shows the machine learning techniques that can be applied for classification. Section 3 outlines the main results obtained. Finally, the conclusions of the work are summarized in Section 4.

2. Methodology

In this study, real EEG recordings have been used to investigate the operation of the NN system proposed. Three hundred and twelve schizophrenic patients and one hundred and five euthymic bipolar disorder patients were tested for brain-disorder diagnosis measured by EEG recordings. The structured clinical interview for DSM-IV (SCID) was administered to all subjects to obtain DSM-IV diagnoses. All participants provided written informed consent after being explained the study and the procedures involved. The study was approved by the Clinical Research Ethics Committee of the Cuenca Health Area. The EEG records were recorded at the Psychiatric Service of the Virgen de la Luz Hospital in Cuenca (Spain). The equipment available at the Hospital was used to perform the EEGs. The International System 10-20 was used to place the electrodes by the medical staff. The EEG records of the different patients presented various noise samples, such as muscle noise, artefacts, baseline etc. To get a more accurate result of the neural network these signals were filtered out. Information about the position of the electrodes is used to create the maps. In our case, according to the 10-20 system for data acquisition.

Figure 1 shows the steps followed from EEG measurements to patient classification. First, electrophysiological data were recorded using a 32-channel Brain Vision system with a sampling frequency of 500Hz and signal gain equal to 75 K (150x at the headbox). According to the international 10-20 system, EEG data were continuously recorded from the 32 electrodes (Z: Midline; F_Z: Midline Frontal; C_Z: Midline Central; P_Z: Midline Parietal; Even numbers, right hemisphere locations; odd numbers, left hemisphere locations; F_p: Frontopolar; F: Frontal; C: Central; T: Temporal; P: Parietal; O: Occipital).

Thereafter, the stored data was pre-processed in order to remove external interferences. The noise present in the signal was due to the electrical distribution network, the surrounding electronic equipment, or may be due to the body functioning i.e., body movements, eye-blinking, breathing, or sweating. To remove these noise components, present in the signal, a notch and a high pass filtering were applied. Once the brain signal is clear of interferences and artifacts, a feature extraction (mean and entropy) was performed. It determined the characteristics of the signal for classification during the execution of ANOVAS and machine learning methods. Specifically, a deep learning method by means of a radial basis function (RBF) neural network (NN) based on fuzzy means (FM) algorithm was used for signal classification.

The FM algorithm was applied using MALTAB to choose the NN structure and the centres of the hidden nodes [7]-[8], where the proposed algorithm uses an initial fuzzy partition (FP) of input space and a number of fuzzy sets that are defined for each input variable. The novel RBF method realizes a uniform division of the discourse universe for its input p_j (j = 1,2, ..., M) into c_j fuzzy sets F_j1, F_j2, ..., F_jc with functions of form as follows:

where v_sj represents the central element to which the unit’s membership value is set and l_sj is half of the respective width. The sum of the degrees of correspondence at any point in the discourse universe is close to 1, for each input variable. Defining a FP into the M dimensional input space results in the initial FP of every input. From this, the following algorithm is proposed to find from the input data vector the nearest fuzzy subspace [7]-[8].

3. Results and discussion

The results obtained during the study indicate that patients with bipolar disorder can be classified accurately from recorded EEG signals. Table 1 and Table 2 present the values of Balanced Accuracy, Recall, Precision and F₁ score of well-known classification methods i.e., support vector machine (SVM) and K-nearest neighbors (KNN), and the proposed algorithm for schizophrenia and bipolar disorder, respectively.

Table 1. Balanced accuracy, recall, precision and F1 score for different machine learning models and the proposed method for schizophrenia patients.

Method	Balanced Accuracy (%)	Recall (%)	Precision (%)	F1 score (%)
SVM	84,74	84,84	84,64	84,14
KNN	87,94	88,08	87,81	87,68
Ours	93,40	93,49	93,30	92,73

 

Table 2. Balanced accuracy, recall, precision and F1 score for different machine learning models and the proposed method for bipolar disorder patients.

Method	Balanced Accuracy (%)	Recall (%)	Precision (%)	F1 score (%)
SVM	88,17	88,28	87,54	87,91
KNN	89,63	89,74	88,99	89,36
Ours	96,78	96,89	96,09	96,49

Systems based on SVM and KNN obtained values of accuracy, recall, precision and F₁ scores below 90 % in all cases. The proposed system, based on NN, obtained the highest performance, achieving values above 93% and 96 % for schizophrenia and bipolar disorder, respectively. It is important to mention that all figures of merits were obtained in all cases for real EEG records.

In addition, different parameters extracted from the cleaned EEG signals were subsequently analyzed using an analysis of variance (ANOVA) study with Bonferroni tests. Specifically, the EEG signals acquired were divided in windows of 5 seconds and subsequently processed with MALTAB to obtain the entropy for all electrodes and volunteers. Thereafter, the significance between patients and healthy controls was obtained by means of the software IBM SPSS Statistics.

The results exhibited significant differences in the left part of the frontal and occipital lobes as it is shown in Figure 2 and Figure 3.

The ANOVAs analyses with Bonferroni tests show significant modifications in the left part of the frontal and occipital lobes (see p-values below 0.005 in deep green), therefore demonstrating the deterioration suffered by patients with schizophrenia in synaptic connections of the aforementioned brain areas.

The results obtained in this paper, improve similar research performed in the past. For instance, other machine learning and deep leaning algorithms applied to classify schizophrenia and bipolar disorder obtained values of accuracy up to around 91% as it can be observed in the summary presented in table 3.

Table 3. Classification performance of last research papers including up to date machine learning and deep learning methods.

Paper	Mental Disorder	Method	Accuracy (%)
[10]	Schizophrenia	RF	70,80
	Schizophrenia	KNN	83,30
	Schizophrenia	RF	79,20
	Schizophrenia	SVM	91,70
[11]	Bipolar disorder	SVM	80,19
[12]	Schizophrenia	RF	81,10
[13]	Schizophrenia	LSTM	72,50
	Bipolar disorder	LSTM	67,50
[14]	Bipolar disorder	ANN	89,80
Ours	Schizophrenia	RBF+FM	93,40
Ours	Bipolar disorder	RBF+FM	96,78

These results suggested that patients with schizophrenia and bipolar disorder can be successfully classified employing deep learning methods using radial basis functions. In addition, our results led to improved accuracy values, which set useful strategies for the development of novel and future classification algorithms.

4. Conclusion

The study performed proves that EEG records can be used for discrimination of schizophrenia and bipolar disorder patients with healthy controls. Before classification, data obtained from EEG signals are processed so that the main features such as mean, or entropy can be extracted. Furthermore, machine learning techniques have been applied for classification. Two well-known classifiers namely, SVM and KNN, and a novel proposed NN have been employed. The RBF neural network based on FM algorithm shows the best results obtained with Balanced Accuracy, Recall, Precision and F1 score better than 93% for schizophrenia and 96 % for bipolar disorder. Furthermore, different ANOVAs studies with Bonferroni tests shows significant modifications in the left part of the frontal and occipital lobes. Finally, the experimental results presented in this paper demonstrate the usefulness of the proposed classification approach that could be employed during clinic analysis as a complementary tool to help psychiatrists diagnosing patients with mental illness.

References