Investigate Schizophrenia Classification Based on EEG Electrode Reduction Using Machine Learning Techniques

The human nervous system is the main system that determines lifestyle and general behaviour, including decisions and emotional control. The brain may be exposed to various accidents or diseases that lead to abnormalities in the brain's neural structure. The severity of the abnormalities varies depending on many factors, such as the patient's age, the family’s medical history, or the brain injury's location, which leads to multiple differences occurring between one patient and another in the behaviour and diagnoses. Neurological diseases that affect the neural cells, for instance, epilepsy, schizophrenia (SZ), Parkinson’s, and Alzheimer's, effect different locations of the brain and have a specific effect on brain function that is discovered and diagnosed through symptoms, tests, and using different tools. One of the worst diseases that may completely destroy the neural system in the brain is schizophrenia.

Delusions, hallucinations, depression, and anxiety are just a few of the symptoms that come along with schizophrenia, which causes a severe behavioral change in character [1]. These symptoms can be clearly seen and distinguished between the ages of 16 and 30 and can be utilized to determine the patient's state, where timely detection is a crucial part of the patient's healing process [2, 3]. Besides, according to the official data of the World Health Organization, there are around 21 million patients, which is around 1% of people worldwide who are afflicted with this illness [4]. The traditional process of diagnosing schizophrenia is considered a complex process, according to some factors such as the psychiatrists' experience and the different case responses between one patient and another.

Indeed, the neuroimaging techniques field provided help in discovering various brain disorders like schizophrenia, which is considered a time-consuming process, encouraging scientific researchers to accomplish new medical aims.

Electroencephalography (EEG) devices have made tremendous strides in the diagnosis of nervous system disorders in recent years, including epilepsy [5], Alzheimer's disease [6-9], and schizophrenia [10, 11]. The non-invasive nature of EEG has led to its adoption as the ideal tool for recording and collecting the electrical activity of the brain, which in turn provides brain dimensions that contain huge amounts of data, which gives the capability to diagnose different brain diseases. Meanwhile, deep learning and machine learning (ML) are the most important methods in the medical domain that offer advanced processing and evaluation of diverse medical datasets, including brain signals [12].

ML used the feature extraction approach to retrieve the hidden information of signals from the data, where these features provide the possibility of obtaining clearer signals and identifying and measuring the most relevant data, which will be used with the frequency domain or the time domain [13, 14].

Many studies have addressed the subject of classifying schizophrenia based on EEG and analyzing the brain signals captured using the global 10-20 electrode system. The analysis of the relationship between these electrodes represents the nature of neural communication in the brain, which in turn leads to the detection of a person's mental health. According to the brain regions and the different functions and channels associated with them, the features in each channel of EEG are analyzed and extracted separately, and these features are subsequently linked and then classified.

Consequently, this study investigates the feasibility and importance of electrodes through the model's ability to classify schizophrenia without compromising classification accuracy by reducing electrodes and selecting electrodes randomly.

This approach supports the ability of electrodes for non-sophisticated devices as well as for devices with few electrodes, such as Emotiv Insights (5 channels), Muse EEG (4-6 channels), Neurosky Mindwave (1 channel), and InteraXon Muse S (4 channels). Clinically, random electrode reduction offers advantages by mitigating bias related to electrode type and placement. Furthermore, the proposed methodology simplifies traditionally employed techniques.

The manuscript is written in the following format: related work in Section 2, which included a comparison with the literature that has classified schizophrenia using different datasets and techniques. In Section 3, we illustrated the materials and methods: Firstly, we described the EEG dataset; secondly, we mentioned the preprocessing methods used; thirdly, we summarized the techniques of feature extraction; and finally, the normalization techniques applied to the dataset. In Section 4, we explained the classification through the description of machine learning classifiers; then in Section 5, we presented and discussed the results, and finally, the conclusion is clarified in Section 6.

3.1 Dataset

The EEG dataset used was accessible to the public [49]. Table 2 presents the details of the total signal used containing 28 subjects, 14 from each group: schizophrenia and healthy controls. The sampling frequency of the EEG dataset recording is 250Hz. The montage was executed with a conventional 10-20 system. Moreover, the dataset was assembled utilizing the subsequent 19 EEG channels: Fp1, Fp2, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2 [50].

Table 2. The used datasets details

3.2 Preprocessing

Preprocessing is a crucial step in ML due to EEG signals often containing a combination of noise resulting from several different factors, which in turn affects classification accuracy. There are several ways to reduce the noise and improve the EEG signals, for instance, by applying filters to enhance the signal quality.

Thus, the bandpass filter was utilized to decompose the EEG data into five frequency sub-bands: beta rhythm (31-31Hz), alpha rhythm (10-14Hz), theta rhythm (5-9Hz), delta rhythm (0.1-4Hz), and gamma rhythm (32-100Hz).

The bandpass filter works on filtering out very high or low frequencies, as presented in Figure 1 with the 19 channels of the EEG signal after the band-pass filter was applied.

1.png

Figure 1. The impact of band-pass filtering on EEG signals

The brain's electrical activity generates billions of signals, so a multitude of electrodes are used to record these signals; adding more electrodes has a positive impact on the classification outcomes. It is important to note that there is no defined minimum or maximum number of electrodes, rules, or specific standards that should be used, as this varies depending on the kind of disorder being dealt with.

Hence, we examined our proposed models by the two electrode scenarios (8 and 5 electrodes); the electrodes were selected randomly to assess the suggested approach and determine whether or not it might yield encouraging outcomes. Additionally, we evaluated the computational efficiency of the classifier's final output, as well as estimated the training and speed time.

3.3 Feature extraction techniques

Different feature extraction techniques have been put out in the literature, such as entropy, approximate entropy, Shannon entropy, fuzzy entropy, fast Fourier transform, etc. Each one of these methods has a special mechanism for obtaining the feature from the signal. We identified the most commonly used methods with brain signals that helped identify subtle abnormalities associated with schizophrenia and combined these methods with our proposed approach to evaluate their effectiveness with varying constraints.

In this study, four extraction methods were calculated to extract the hidden features as presented in the following equations: Eq. (1) represents the mathematical formulation of Fast Fourier Transform (FFT), approximate entropy (ApEn) as shown in Eq. (2), log energy entropy (LogEn) by Eq. (3), and Shannon entropy (ShEn) as shown in Eq. (4).

FFT, LogEn, and ShEn were implemented with the band-pass filter; ApEn was applied in both cases with and without the band-pass filter. We selected these four techniques due to their established applications in EEG research and their accurate dealing with brain signals, which are explained in the following paragraphs.

FFT is one of the most important algorithms developed of all time, and the real reason for depending on FFT is its quick and effective method of denoising data. The FFT features have been implemented on the SZ EEG signals to convert the time domain to the frequency domain. Thus, it allows one to see the most prominent frequencies in the EEG signal and identify abnormalities or patterns that may appear in the brain wave frequencies [51].

$\begin{aligned} X(K)=\sum_{n=0}^{N-1} X[n] W_N^{k n} & =\sum_{n \text { even }} x(n) n w_N^{k n}+\sum_{n \text { odd }} x(n) w_N^{k n} \\ K & =0,1 \ldots \ldots, N-1,\end{aligned}$           (1)

$\begin{aligned} \operatorname{ApEn}(E, r, N)= & \frac{1}{(N-e+1)} \sum_{i-1}^{N-e+1} \log C_i^e(r) -\frac{1}{N-e} \sum_{i=1}^{N-e} \log C_i^{e+1}(r)\end{aligned}$          (2)

$\mathrm{ShEn}=\frac{H S h}{o g k}$          (3)

$\mathrm{E}=\sum_n \log \left(w_{i, j}^{n 2}\right)$         (4)

Entropy is the most frequently used feature to measure time-domain features. It is also widely used in disease detection by capturing any slight or subtle change in the brain signal, which assists when dealing with small data sizes. ApEn is defined as a measurement of the regularity or randomness of data in a time series and is used for short-length data due to its lower sensitivity to noise.

ShnEn is a time-domain complexity metric that does not rely on the signal spectrum and is similar to ApEn functionality. Since entropy may be used to ascertain the degree of randomness in the information, it is employed as the feature approach for schizophrenia due to ShEn decreasing with decreased neural complexity [52].

In signal processing, a common measure called LogEn, is used to extract relevant information. Since the frequency bands specific to schizophrenia have been identified, these bands can provide insights into the underlying neural systems. All these feature extraction methods are computed using a MATLAB routine.

3.4 Normalization

In general, normalization is the most common technique that deals with data in linear transformations. Its role is to rescale numerical features to a standard range to avoid larger values that may affect and bias the machine learning results. Although applying any normalization method is easy to implement, each method has strengths and weaknesses. Selecting the appropriate normalization method depends on some factors, such as the data and what is required from machine learning to obtain optimal results. In our work, we used L2 normalization to increase and enhance our accuracy results.

By using the public database [49], we explore how bandwidth filtering and electrode reduction affect our ability to find schizophrenia-related EEG signals. We selected electrodes randomly, unlike previous studies that focused on identifying brain regions and selecting effective electrodes, which introduced monotony into the work and repeated the results. Most current studies now use all channels to obtain the highest possible classification accuracy to achieve the best computational efficiency in less time, which is often not the case when the electrodes are reduced. Therefore, our study investigates the effect of minimizing the data channel numbers in two different scenarios to achieve comparable results to studies that used all data channels.

Random channel selection is computationally inexpensive when compared to the traditional method (algorithms used that have the ability to rank channels or feature importance analysis), which is considered computationally expensive and helps to influence without prior knowledge of which are the most important channels.

Our experiments were based on 14 subjects for SZ and 14 subjects for healthy, so to maximize the data, we used one epoch window size for our two scenarios.

We examined the efficacy of four feature extraction methods: FFT, ApEn, LogEn, and ShnEn. Among the features, ApEn was applied twice: with and without the band-pass filter on the data. Then, the data was standardized using the L2-normalization approach and fed to three ML classifiers: SVM, KNN, and QDA. The original signals were recorded with 19 channels and a 250Hz frequency.

5.1 Experiment I: EEG classification based on five randomly selected electrodes

The brain's electrical activity generates billions of signals, so a multitude of electrodes are used to record these signals, and using more electrodes has a favourable impact on the classification outcomes. It is important to note that there is no defined minimum or maximum number of electrodes that should be used, as this varies depending on the kind of disorder being dealt with. As a result, there is no guideline, rule or specific standard for this. Therefore, we first tested the proposed models to evaluate our approach by reducing the number of electrodes to five, as the goal of choosing this number was to determine whether or not it could lead to encouraging results using a very limited electrode setup. And demonstrate the ability of the approach to extract and obtain information contained in a small set of EEG channels.

Classification was performed using the SVM, KNN, and QDA algorithms on the five electrodes for each class; the obtained confusion matrix values for the three classifiers are presented in Table 3, and the performance values are listed in Table 4.

Based on Table 4, the results show that when randomly selecting five electrodes, EEG signal classification with KNN with implementing LogEn outperformed the other two classifiers with results of 98%, 98%, and 99% for accuracy, sensitivity, and specificity, respectively.

Table 3. The confusion matrix obtained by the three classifiers with random 5-electrodes

Table 4. Classification performance results utilizing five random electrodes

In contrast, SVM scored 98%, 98%, and 98%, and QDA scored 92%, 87%, and 98%, respectively, for accuracy, sensitivity, and specificity.

5.2 Experiment II: EEG classification based on eight randomly selected electrodes

In this scenario, we chose eight electrodes as an intermediate step between the first scenario of five electrodes and the previous studies mentioned in Section 2, which include more comprehensive electrodes. Here we verify the proposed approach to see if a small increase in the number of electrodes leads to a significant improvement in classification performance. Table 5 displays the confusion matrix obtained by using 8 electrodes, and Table 6 displays the performance values.

For this scenario reduction, the classification accuracy rates were 99%, 99%, and 95% for SVM, KNN, and QDA, respectively. The corresponding sensitivity and specificity of the SVM and KNN classifiers were identical at 99%, while QDA got 90% and 99%, respectively. Consequently, SVM outperformed the other classifiers.

Table 5. The confusion matrix was obtained by the three classifiers with random 8-electrodes

Table 6. Classification performance results utilizing eight random electrodes

5.3 Experiment III: EEG classification based on PCA

Finally, we applied the principal component analysis (PCA); it preserves the most important features and reduces the dimensionality of the data by its formula illustrated in Eq. (5), but at the same time, these features may not be the most relevant or important for classifying the EEG data. This belongs to the PCA's lack of knowledge of the patterns that are important and required to function.

Tables 7 and 8 compare and demonstrate classifying EEG data for schizophrenia disorder by the confusion matrix for the two PCA scenarios, while Tables 9 and 10 show the performance results.

$\operatorname{cov}_{x, y}=\frac{\sum\left(x_i-\bar{x}\right)\left(y_i-\bar{y}\right)}{N-1}$           (5)

Given the earlier studies mentioned in section 2 and our study that investigated the limitations of random electrode selection and compared them to the PCA feature selection methods, this work stands out as a real demonstration of the validity of achieving impressive and unexpected results with fewer electrodes than expected, which demonstrates the existence of substantial discriminative information within this restricted grouping. This highlights the need to further explore the potential of EEG systems and the extent to which they can be extended.

Table 7. The confusion matrix was obtained using PCA with the 5-electrodes

Table 8. The confusion matrix was obtained using PCA with the 8-electrodes

Table 9. Classification performance results with PCA for 5-electrodes

Table 10. Classification performance results with PCA for 8-electrodes

By comparing all the results obtained using 8 electrodes with the results using 5 electrodes in the two scenarios, our proposed methods result confirmed an acceptable balance between classification accuracy and the number of effective electrodes, taking into account the limited amount of data. We can set this random selection of electrodes in small numbers as a benchmark for comparison with more advanced electrode selection techniques. Where, the model based on PCA achieved a lower level of accuracy in identifying individuals with schizophrenia based on their EEG data, demonstrating the potential of our proposed approach as a valuable tool in the automated diagnosis of this mental disorder.

On the other hand, the datasets used are open-source data for all, and the number of subjects from whom data were collected was equal, which is an important factor in not biasing one group over another in our work. In addition, numerous previous studies have utilized the same dataset, as shown in Table 11.

To add reliability and achieve appropriate robustness to our approach, we calculated the prediction speed and training time, as computational efficiency is one of the indicators relied upon to determine the effectiveness of systems. Tables 12 and 13 show the prediction speed and training time for our approach when using 5 electrodes and 8 electrodes, while Tables 14 and 15 include the PCA prediction speed and training time.

These values are affected by several factors, such as the size of the medical dataset to be classified and the complexity of the model used. The model’s prediction speed was significantly slower using 5 electrodes, indicating faster classification and lower computational complexity. In the 8-electrode scenario, increasing the number of electrodes collected more features, enabled the model to make faster decisions. Thus, the 5-electrode model is faster in prediction but takes longer to train. This is a trade-off between performance speed and the actual time required by the system.

Table 11. Comparing the sensitivity (Sen), specificity (Spe), and accuracy (Acc) values of the studies using the same dataset, where Chan. Refer to channels

Table 12. The computational efficiency utilizing 5 electrodes

Table 13. The computational efficiency utilizing 8 electrodes

Table 14. PCA computational efficiency utilizing 5 electrodes

It is important to note that successful classification of machine learning-based data in MATLAB, such as neuroimaging, also depends on memory management, as dealing with such data is a challenge due to the large memory resources required for loading, preprocessing, feature extraction, and model training. Tables 16 and 17 show the memory utilization in megabytes for the random selection for each of the three ML models (SVM, KNN, and QDA) during the execution. In contrast, Tables 18 and 19 present the memory utilization values by PCA.

Upon calculating both M.A for all arrays and M.U by MATLAB for the model, the SVM model with random selection showed the highest memory usage when using five electrodes at 7356MB, while the KNN model with eight random electrodes required the most memory at 320MB. These results strengthen our conclusions and provide a more comprehensive understanding of the approaches for EEG-based schizophrenia classification.

Table 15. PCA computational efficiency utilizing 8 electrodes

Table 16. Summarizing the memory usage for the five electrodes, where M.A refers to memory availability and M.U refers to memory usage

Table 17. Summarizing the memory usage for the eight electrodes, where M.A refers to memory availability, and M.U refers to memory usage

Lastly, the visual examination aids clinicians in identifying the disorder's signal and location inside the brain. Figure 3 displays the power spectral density, highlighting the disparity between the signals of a patient with schizophrenia and those of a healthy individual. The healthy signals are mostly uniform and stable, whereas the patient signals display irregularities and unusual rhythms. Furthermore, a topographical picture delineates the electrode's position, enabling physicians to ascertain the etiology of the neurological issue more precisely and identify the indicators of schizophrenia.

3.png

Figure 3. Visualisation method for power spectrum density and topography image, whereas A represents the Schizophrenia person and B represents the patient

Above all, the five random electrodes using KNN with LogEn features outperformed the performances of the two other classifiers (SVM and QDA). Whereas SVM with LogEn achieved the best performance compared with KNN and QDA in the eight-electrode scenario. To conclude, the two random scenarios, based on the findings, indicate that superior outcomes may be achieved with a reduced number of electrodes.

In this study, some limitations are listed below:

(1) The dataset used was publicly available and not privately collected for this study.

(2) The one-epoch window size technique was an advantage that allowed us to cover and provide a much bigger amount of the data, with an overlap was 50%, which was fed into the machine-learning models.

Table 18. PCA memory usage with the five electrodes, where M.A refers to memory availability, and M.U refers to memory usage

(3) The model processed the data and diagnosed schizophrenia, not its three severity stages (early, active, or residual).

(4) The study results may lack the ability to generalize to patients with schizophrenia for several reasons that may be due to some characteristics of the data, such as age and gender.

Table 19. PCA memory usage with the eight electrodes, where M.A refers to memory availability and M.U refers to memory usage