Schizophrenia (SZ) is one of the most important mental disorders, leading to disruption in brain growth (Lewis and Levitt, 2002; Schmitt et al., 2011). This disorder seriously damages thoughts, expression of emotions, and also individuals' perception of reality (Elvevag and Goldberg, 2000). The reason for SZ is not fully understood, though most research has demonstrated that the structural and functional abnormalities of the brain play a role in its creation (Qureshi et al., 2019). According to the World Health Organization reports, nearly 21 million individuals suffer from such a brain disorder worldwide. The average age starting to get affected by this disorder is in youth age; in men 18 years old, and women 25 years old, and it is more prevalent among males (Sadeghi et al., 2021).

Numerous methods have been provided for automated SZ diagnosis; among these techniques, neuroimaging-based methods have a special potential for specialist physicians (Li et al., 2021; Yan et al., 2021). Generally, neuroimaging methods include various structural or functional modalities (Steardo et al., 2020; Hu et al., 2021). Structural MRI and diffusion tensor imaging-MRI are among the most important modalities of structural neuroimaging, providing important information regarding brain structure to specialist physicians (Sui et al., 2013; Lee et al., 2018; Oh et al., 2020). Contrarily, electroencephalography (EEG) (Boutros et al., 2008), magnetoencephalography (Fernández et al., 2011), functional MRI (Sartipi et al., 2020), and functional near-infrared spectroscopy (Chen et al., 2020) are the most important functional modalities of the brain. These modalities provide vital information on brain function to specialist physicians.

EEG is one of the most practical and inexpensive functional neuroimaging modalities, specifically capturing the interests of specialist physicians. In this modality, the electrical activities of the brain are recorded from the head surface with a high temporal resolution and an appropriate spatial resolution, which is influential in SZ diagnosis (Murashko and Shmukler, 2019). In addition to the mentioned merits, EEG signals regularly have various channels recorded in the long term (Murashko and Shmukler, 2019). In some cases, these reasons make specialist physicians face serious challenges in SZ diagnosis via EEG signals.

In recent years, various investigations have provided automated SZ diagnosis via EEG signals using artificial intelligence (AI) methods (Prasad et al., 2013; Shim et al., 2016; Chu et al., 2017; Alimardani et al., 2018; Devia et al., 2019; Jahmunah et al., 2019; Li et al., 2019; Naira and Alamo, 2019; Oh et al., 2019; Phang et al., 2019a,b; Aristizabal et al., 2020; Luo et al., 2020; Prabhakar et al., 2020; Shalbaf et al., 2020; Siuly et al., 2020; Sharma et al., 2021; Singh et al., 2021; Sun et al., 2021). The AI investigations in this field include conventional machine learning (ML) and deep learning (DL) methods (Khodatars et al., 2020; Shoeibi et al., 2020, 2021,a,b). The AI-based SZ diagnosis algorithm includes preprocessing sections, features extraction and selection, and in the end, classification. Feature extraction is the most important part of SZ diagnosis via EEG signals. In conventional ML, the extracted features from EEG signals are mainly categorized into four groups: time (Diykh et al., 2016), frequency (Faust et al., 2010), time-frequency (Madhavan et al., 2019), and non-linear (Gajic et al., 2015; Shoeibi et al., 2021a) fields. Siuly et al. (2020) used empirical mode decomposition (EMD) in preprocessing step. In the following, various statistical features were extracted from EMD subbands, and the ensemble bagged tree method was used for classification. In another study, Jahmunah et al. (2019) used non-linear features and support vector machine (SVM) with radial basis function kernel in the feature extraction and classification steps, respectively. Devia et al. (2019) have provided an event-related field features-based SZ diagnosis method via EEG signals. Extremely randomized trees (ERT) features were extracted from EEG signals in this effort, and then linear discriminant analysis was used in the classification step. In Prabhakar et al. (2020), statistical features of steady-state visual evoked potentials were extracted, and in the end, classification has been executed by the k-nearest neighbors (KNN) method. Li et al. (2019) used solitary pulmonary nodule features and SVM classification for SZ diagnosis via EEG signals. In another study, Shim et al. provided a new method of SZ diagnosis via EEG signals (Shim et al., 2016). This investigation used sensor-level and source-level features in the feature extraction step and then employed the Fisher's score for feature selection. Ultimately, the SVM method was used in the classification step, and they achieved promising results.

In conventional ML, selecting proper feature extraction algorithms for SZ diagnosis is a relatively demanding task, requiring a great deal of knowledge in signal processing and the AI field. To overcome this problem, DL-based methods have been provided in recent years for SZ diagnosis via EEG signals, where feature extraction operations are carried out without surveillance by deep layers (Shoeibi et al., 2021a). Shalbaf et al. (2020) define a transfer learning model for SZ diagnosis via EEG signals. In this study, the ResNet-18 model has been used for feature extraction from EEG signals. Besides, SVM has been used in the classification step. Some researchers have studied other convolutional network (CNN) models utilization in SZ diagnosis via EEG signals. CNN models have been used in Naira and Alamo (2019) and Oh et al. (2019) for SZ diagnosis, resulting in satisfactory achievements. CNN-recurrent neural network (RNN) models are an important group of DL networks and are significantly popular for their capability of various brain diseases diagnoses via EEG signals. In Aristizabal et al. (2020), Sharma et al. (2021), Singh et al. (2021), Sun et al. (2021), CNN-long short-term memory (LSTM) models have been used for SZ diagnosis, and the researchers have been able to achieve promising results.

In this paper, SZ diagnosis via EEG signals will be investigated by using various proposed DL and conventional ML-based methods. A summary of proposed methods is depicted in Figure 1.

In this study, the dataset of the Institute of Psychiatry and Neurology in Warsaw, Poland, is used (Olejarczyk and Jernajczyk, 2017). In the preprocessing step, the z-score and L2 normalization techniques will be applied to EEG signals. Next, to classify EEG signals, various conventional ML methods and DL-based proposed models will be used. The conventional ML methods employed, include various classification, SVM (Cortes and Vapnik, 1995), KNN (Cover and Hart, 1967), decision tree (DT) (Rokach and Maimon, 2007), naïve Bayes (Zhang, 2004), random forest (RF) (Breiman, 2001), ERT (Geurts et al., 2006), and bagging (Friedman, 2001) methods. Besides, the proposed DL networks include various one-dimensional (1D)-CNN, LSTM, and ID-CNN-LSTM models for executing the steps from feature extraction to classification. Generally, nine LSTM-, 1D-CNN-, and ID-CNN-LSTM-based DL methods will be investigated in this step.

In section Materials and Methods, we described our method in detail. In addition, we outline several baseline methods for comparison purposes in the same section. The statistical metrics to analyze and validate the proposed model are described in section Experiment Results. Experiment results are provided in section Limitation of Study, and some limitations of the proposed method are provided in section Conclusion, Discussion, and Future Works. Finally, a discussion, the conclusion, and future works are represented.

This section will discuss the proposed methods for SZ diagnosis via EEG signals and various conventional ML and DL models. First, the proposed dataset will be examined. Then, the preprocessing method of EEG signals will be explained. In the end, conventional ML and DL models will be introduced for SZ diagnosis via EEG signals.

The results of the proposed methods are presented in this section. First, the simulation results obtained from conventional ML techniques for SZ diagnosis via EEG signals are presented and discussed. The original dataset was flattened to have only a vector per sample, and then we used the flattened dataset to train several classification algorithms using the scikit-learn library (Pedregosa et al., 2011). Namely, we studied the performance of KNNs, DTs, SVMs, and naive Bayes; and three ensemble algorithms (bagging, extremely randomized trees, and RF). The algorithms were trained using the by-default hyperparameters provided by the implementation of the scikit-learn library. Moreover, we studied the impact of z-score normalization (Cheadle et al., 2003) on the performance of the models. All the experiments were conducted in an Intel (R) Core (TM) i7-4810MQ CPU at 2.80 GHz. In Table 8, the results obtained from conventional classification algorithms for raw input EEG signals or normalized by z-score normalization are indicated.

According to Table 8, the bagging conventional classification algorithms for EEG signals normalized using z-score normalization resulted in the maximum accuracy. Figure 5 shows the ROC curves for ML classification algorithms with different normalizations of EEG signals. The figure on the left shows the results of ML classification methods with z-score normalization; additionally, the ROC curves for ML classification algorithms with z-score + L2 normalization is presented in the figure on the right.

We also employed several DL architectures based on CNNs and LSTMs (Goodfellow et al., 2016), and the combination of both convolutions and LSTM layers. Namely, three CNNs, two LSTMs, and two CNN-LSTM networks (see Tables 1–7 for the concrete architecture of these networks) were studied. We also analyzed the relevance of using three different activation functions (ReLU, Leaky ReLU, and seLU), and the impact of z-score normalization. To avoid overfitting, we applied two regularization techniques that are Dropout (Goodfellow et al., 2016) and weight regularization (Goodfellow et al., 2016). In particular, dropout was applied after each convolutional and LSTM layer using a dropout value of 0.5, and after dense layers using a dropout value of 0.25. Weight regularization was employed in all the convolutional, LSTM, and dense layers of our architectures using L2 regularization with a value 0.01. The final selected values for batch size and hyperparameters of our networks are all available in Table 9. All the experiments were conducted using the Keras library (Gulli and Pal, 2017) and using a GPU NVidia RTX2080 Ti.

In the following, the results obtained from the DL proposed methods for different activation functions are demonstrated in Tables 10–12. First, the results obtained from the proposed DL method with the Leaky ReLU activation function are demonstrated in Table 10.

As indicated in Table 10, the second proposed CNN-LSTM model with the Leaky ReLU activation function and combined normalization of z-score with L2 could obtain the maximum accuracy. Table 11 presents the results obtained from the proposed DL method with the seLU activation function.

Table 11 indicated that the second proposed LSTM method could result in maximum accuracy. The results of all proposed DL models with the ReLU activation function and z-score and L2 normalizations are presented in Table 12.

According to Table 12, it can be seen that compared to all classification methods with different activation functions, the second proposed CNN-LSTM model with ReLU activation function and combined normalization technique of z-score and L2 could lead to the maximum accuracy. In the following, the ROC diagrams for the DL models with ReLU activation functions and z-score and z-score + L2 normalization methods are drawn in Figure 6. Firstly, on the left of Figure 6, the results of the DL algorithms with z-score + L2 normalization are presented. Also, the ROC curves for DL algorithms with the z-score normalization for EEG signals is shown in the right of Figure 6. Furthermore, learning curves of the CNN-LSTM method with ReLU activation and z-score normalization and also with z-score + L2 normalization are shown in Figures 7, 8, respectively.

The simulation results of the proposed models for SZ diagnosis via EEG signals were investigated in this section. Compared to all DL and conventional ML methods, the CNN-LSTM models with 13 layers have higher accuracy and efficiency among the proposed methods. Selecting the number of layers in this model and the type of the activation functions are presented in this research for the first time, which is the novelty of the article. Besides, simultaneously using z-score and L2 normalizations along with the proposed CNN-LSTM model is another novelty of this article. Figure 9 shows the DL models with different activation functions and z-score normalization. Also, Figure 10 displayed the DL architectures with different activation functions and z-score and L2 normalization. According to Figures 9, 10, the second version of CNN-LSTM with z-score and L2 normalization has the best performance compared to other methods.

The limitations of the study are investigated in this section. The available EEG datasets for SZ diagnosis consist of a limited number of cases which has made access to the tools of SZ diagnosis via EEG signals and DL models challenging. The dataset in this research was not used to determine the severity of the disorder but to diagnose the disorder. This dataset is unsuitable for prognosis or early diagnosis, and other appropriate datasets must be gathered for these purposes. Another limitation of this study is that the classifiers are not separately designed and compared for different age and gender groups, and other suitable datasets must be gathered for this purpose. Classifiers are of the two-class type and can become multiclass by adding the classes of brain disorders with similar symptoms to SZ.

SZ is a mental disorder that negatively affects brain function, causing various problems for the patient. Different screening methods have been introduced for SZ mental disorder diagnosis, among which the EEG functional imaging modality has captured the interest of neurologists and specialist physicians. SZ diagnosis via EEG signals has always been challenging. In recent years, various investigations into using AI techniques for SZ diagnosis and interpretation via EEG signals have been conducted to tackle this challenge. These methods are proposed to help physicians and neurologists with a quick and accurate diagnosis of SZ disorder via EEG signals.

Various AI approaches are presented for the diagnosis of SZ mental disorder via EEG signals. These approaches include using different conventional ML (Alizadehsani et al., 2021) techniques and also DL models (Martinez-Murcia et al., 2019; Górriz et al., 2020; Gorriz et al., 2021; Jiménez-Mesa et al., 2021). The AI models for SZ diagnosis via EEG signals consist of the following steps: dataset selection, preprocessing, feature extraction and selection, and classification.

In this study, the dataset consisted of EEG data of 14 normal individuals and patients with SZ (Olejarczyk and Jernajczyk, 2017). The EEG signals of this dataset are of a 10-channel type and have a sampling frequency of 250 Hz (Olejarczyk and Jernajczyk, 2017). In the preprocessing step, first, the EEG signals were divided into 25 s frames. Afterward, z-score and z-score-L2 were used for the normalization of EEG signals. In this section, each frame of EEG signals had a dimension of 19 × 6,250. It should be noted that the preprocessing of EEG signals for the DL models included two z-score and z-score-L2 normalization techniques.

Different conventional ML-based classification algorithms were used for SZ diagnosis via EEG signals. In this section, the normalized EEG signals were considered as features to be applied in classification algorithms. The employed classification algorithms included the following methods: SVM (Cortes and Vapnik, 1995), KNN (Cover and Hart, 1967), DT (Rokach and Maimon, 2007), naïve Bayes (Zhang, 2004), RF (Breiman, 2001), ERT (Geurts et al., 2006), and bagging (Friedman, 2001). The bagging classification via EEG signals normalized using z-score could obtain an accuracy of %81.22 ± 1.74, which is the highest accuracy compared to other classification methods.

In the following, different DL methods of SZ diagnosis via EEG signals were employed. The proposed DL methods in this section included three 1D-CNN architectures, two LSTM models, and ultimately two 1D-CNN-LSTM networks. Different activation functions, namely, Leaky ReLU, seLU, and ReLU were used to implement the proposed DL models. Besides, in all models, the sigmoid activation function was used for classification. The results of DL models for different normalization methods and activation functions were indicated in Tables 10–12. Among the proposed DL models, the 1D-CNN-LSTM architecture consisting of 13 layers with the ReLU activation function and z-score + L2 normalization could obtain an accuracy of %99.25 ± 0.25. This model is presented for the first time in this research, as this article's novelty. The comparison between the proposed 1D-CNN-LSTM model with the proposed models of the previous studies conducted on SZ diagnosis via EEG signals is indicated in Table 13.

As shown in Table 13, the model proposed in this research could obtain higher accuracy compared to a vast majority of conducted studies. The proposed model can be implemented on special software and hardware platforms for quick SZ diagnosis via EEG signals and may be employed as an assistant diagnosis method in hospitals.

In the following, some future investigations into SZ diagnosis via EEG signals are presented. The CNN-AE models can be employed for SZ diagnosis via EEG signals as the first future work. Several researchers indicate that CNN-AE models are highly efficient in neural disorders via EEG signals (Shoeibi et al., 2021a). As mentioned in the section of limitation of the study, the dataset used in this study is for SZ disorder diagnosis. However, providing EEG datasets for SZ disorder diagnosis can be of paramount importance for future investigations. One of the future works is to provide classification models based on DL for different age and gender groups, which requires researchers to have access to relevant data.

Another future work is using a combination of conventional ML and DL models for SZ diagnosis such that different non-linear features are extracted from EEG signals first. Afterward, the features are extracted from raw EEG signals by DL models. Ultimately, manual and DL features are combined, and the classification is carried out. Graph models based on DL are one of the new fields in diagnosing brain disorders. Accordingly, in future works, using graph models based on DL can be suitable for SZ diagnosis via EEG signals (Cao et al., 2016).

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author/s.

AS, NG, and JH: methodology. AS, YK, and JH: software. AS and JH: validation. AS, RA, and JG: formal analysis. AS, DS, and PM: resources. AS and NG: writing—original draft preparation. AS, NG, and RA: writing—review and editing. AS, Y-DZ, SN, and YK: visualization. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.