Multiresolution feature fusion for smart diagnosis of schizophrenia in adolescents using EEG signals

Abstract

Numerous studies on early detection of schizophrenia (SZ) have utilized all available channels or employed set of a few time domain or frequency domain features, while a limited number of features may not be sufficient enough to perform diagnosis efficiently. To encounter these problems, an automated diagnosis model is proposed for the efficient diagnosis of schizophrenia symptomatic adolescent subjects from electroencephalogram (EEG) signals via machine intelligence. A publicly accessible EEG dataset featuring 16-channels EEG obtained from 84 adolescents (45 SZ symptomatic and 39 healthy control) is used to demonstrate the work. Initially, the signals are decomposed into sub-bands using two multi-resolution signal analysis methods: Empirical Wavelet Transform and Empirical mode decomposition. 75 unique features from each sub-bands are extracted and the few selective prominent features are applied to machine learning classifiers for optimal sub-band selection. Subsequently, a hybrid model is proposed, combining convolutional neural network (CNN) and ensemble bagged tree, incorporating both deep learning and handcrafted features to perform SZ diagnosis. This innovative model achieved superior classification performance compared to existing methods, offering a promising approach for SZ diagnosis. Furthermore, the study explores the impact of different brain regions and combined regional data in SZ diagnosis comprehensively. Hence, this computer-assisted decision-making model minimizes the limitations of prior studies by providing a more robust and efficient diagnostic system for schizophrenia.

Introduction

Schizophrenia is marked by persistent and severe impairments in emotional, social, and cognitive functionalities, making it a chronic and devastating neuropsychiatric disorder (Dvorak et al. 2018; Lillo et al. 2022). People with schizophrenia have a severe mental condition that causes them to view reality in a peculiar way. Hallucinations, delusions, disorganized thinking, and other cognitive and behavioural abnormalities often seen in people with schizophrenia may severely impair their ability to function in daily life (Thilakavathi et al. 2019; Sharma and Acharya 2021). It usually manifests in adulthood and continues to have a detrimental impact on fundamental functions, including social participation, exhibiting desirable behaviour, thinking, communication, educational performances and other functional activities (https://psychiatry.org/patients-families/schizophrenia/what-is-schizophrenia). World Health Organization (WHO) estimates more than 24 million people or 1 out of 300 (0.33%) are affected by SZ worldwide (https://www.who.int/news-room/fact-sheets/detail/schizophrenia?) (Chen et al. 2018). Compared to the average healthy individual, those with SZ have a mortality rate two to three times higher and a lifespan ten to twenty years shorter (Laursen 2011). Most cases of SZ are seen in people in their early 20s to late adolescent (Buckley and Miller 2015). SZ is more common in the early to midlife female population as compared to teenage male population (Gogtay et al. 2011). Schizophrenia can lead to mental deterioration, which may cause significant intellectual impairment. As a result, SZ causes a decline in both academic and professional achievement. The conventional diagnosis approach involves manual inspection of the patient by a qualified psychiatrist. However, this method heavily relies on the subject’s response and the experience of psychiatrics, leading to a significant degree of subjectivity, potential bias, and time-intensive evaluations. Early and precise detection of SZ is critically essential to enhance the quality of life for affected individuals (Najafzadeh et al. 2021).

According to research, Clinical results for people with neuropsychiatric disorders are vastly improved if they are diagnosed early (García-Gutiérrez et al. 2020). There is no specific sign associated with schizophrenia, posing a challenge for detection and diagnosis. As of now, there is no convincing biological sample analysis (such as a blood test) for evaluating the condition, and a minimum of 6 months is needed to arrive at an accurate diagnosis; it is highly unlikely that a positive diagnosis would be reached. The doctors usually have to rely on a patient’s symptom report, interview findings, or the existence of observable and recorded behavioural patterns to make a diagnosis. The periodic treatment of SZ places a significant financial and emotional strain on the families of patients and healthcare systems (Shalbaf et al. 2020). It is important to remember that schizophrenia is a chronic condition that cannot be cured but must be controlled (Messias and Garcia-Rill 2019). Hence, it is necessary to develop an alternative automated diagnostic method for early-stage detection of schizophrenia. This approach would empower psychiatrists to diagnose symptomatic patients efficiently for clinical practice. The electroencephalogram (EEG) is preferably employed to make diagnosis of mental disorders due to the fact that it is a non-invasive, cost-efficient, and realistically realizable less complex instrument for establishing an objective diagnosis of mental health than other neuroimaging technologies (such as MRI, CT, fMRI, and PET), which encourage us to employ it for the development of computer-assisted decision-making model for SZ classification (Aslan and Akin 2022).

Related work

In 1924, Hans Berger began recording electrical waves of the brain using an EEG, marking the beginning of the quest for a pattern or comparable genetic features in EEG signals (Dvey-Aharon et al. 2015). In the modern era, EEG plays a crucial role in the medical field by enabling the visualization and identification of patterns in human brain through EEG recordings (Ranjan et al. 2018). Current research efforts are focused on developing an automated diagnostic approach that combines various signal processing techniques and artificial intelligence (AI) algorithms with EEG recordings. This innovative method aims to overcome the limitations of conventional diagnosis and enhance the accuracy of preliminary diagnosis for neuropsychiatric disorders. (Raghavendra et al. 2020). Recently, there has been remarkable increase in interest in the diagnosis of schizophrenia using EEG as a result of technological advancements, robustness of machine learning or transfer learning approaches, low-cost equipment, and more accessibility to high-performance computing hardware. Several reported studies in the literature have used informative feature-based machine learning methods.

It follows the decomposition of pre-processed EEG through nonlinear methods such as EMD, EWT, wavelet transform(WT), iterative filtering, variational mode decomposition and other signal transformation methods, into sub-bands and then different sets of handcrafted features are extracted (Siuly et al. 2020a; Das and Pachori 2021; Khare and Bajaj 2021, 2022; Zülfikar and Mehmet 2022; de Miras et al. 2023; Gosala et al. 2023; Kumar et al. 2023). The researchers have observed that these various kinds of feature sets are effective in diagnosing SZ from EEG signals. In (Das and Pachori 2021), EEG signal decomposed into modes using multivariate iterative filtering (MIF), then modes are further separated into EEG rhythms and Hjorth parameters are extracted. The set of various domain features sets (temporal, spatial, statistical, time–frequency features, nonlinear, and derived features) are applied to support vector machine (SVM) with cubic kernel and 98.9% accuracy is reported from EEG recording of 28 subjects. In a recent study (Hassan et al. 2023) on the same dataset, Hassan et al. has used hybrid model as CNN for feature extraction and logistic regression for classification. 98% accuracy was reported for non-subject based testing and 90% accuracy for subject based testing. In another study, Calhas et al. (2020) applied discrete short time Fourier transform (DSTFT) on resting state 16-channel EEG recording of 84 subjects and 84% accuracy is reported using random forest classifier. Balasubramanian et al. proposed an improved adaptive neuro-fuzzy inference system model to detect schizophrenia from EEG using statistical, time-domain, frequency-domain, and spectral features are extracted from pre-processed EEG, and reported 97.8% accuracy. In a research study (Sairamya et al. 2022), the relaxed local neighbour difference pattern (RLNDiP) features from both time domain (TD) and time–frequency domain (TFD) are calculated and the features are fused together to achieve maximum accuracy of 100% with artificial neural network (ANN). Apart from these studies, there are many reported studies which is suggesting that nonlinear features such as fractal dimensions, entropy features, complexity features, exponent features, Hjorth parameters, correlation features, norm features, have significant potential in SZ detection using EEG signals. A few ML classifiers such as SVM, K-nearest neighbour (KNN), logistic regression, decision tree, random forest, and ensemble bagged tree, are found proficient under subject independent testing strategy with tenfold cv for SZ classification task (Lanillos et al. 2020; Cortes-Briones et al. 2022).

The present research trajectory is moving towards the adoption of deep learning (DL) methods, which have substantially revolutionized areas such as computer vision and the medical field, particularly in the assessment of brain signals and diagnosing mental health disorders. DL shows the capability of automatically extraction of features and perform the classification task using raw signals (Göker 2023; Kasim 2023). Various deep learning models have been proposed for diagnosis of schizophrenia in adolescents using EEG signals. In (Naira and Del Alamo 2019), Naira et al. applied Pearson correlation coefficient heat map derived from EEG signal epochs, to six layered CNN and achieved accuracy of 90%. Aslan et al. (2020, 2022),(Zülfikar and Mehmet 2022), Calhas et al. (2020), Sohabi et al. (2022), and Nsugbe et al. (2022) have converted EEG signal into time–frequency representation (TFR) images (such as spectrogram, scalogram, Hilbert spectrum etc.) applied it into DL or hybrid model (ML + DL) and achieved good accuracy (Ranjan et al. 2024). In (Singh et al. 2021), Singh et al. attempted to diagnose SZ using long short-term memory (LSTM) based model and achieved accuracy of 98.56%. In another work on the same dataset, Sharma and Joshi (2022) have developed schizophrenia hybrid neural network (SzHNN) made up of hybridization of CNN and LSTM models for detection of SZ using EEG data. 99.5% accuracy was reported for non-subject based testing and 89.60% accuracy for subject based testing. In the current literature, it is observed that there is scope to explore feature fusion, realisation of significantly affected brain region, instantaneous impact of SZ on two different portion or region of brain. The impact of handcrafted features, automatically derived features, or combinations of both features on SZ diagnosis using EEG still needs attention from researchers.

Objectives and key contributions

Researchers have explored numerous combinations of signal processing approaches along ML or DL algorithms, aiming to achieve optimal accuracy in the diagnosis of SZ through EEG data. Many of the approaches discussed in the literature exhibit limitations regarding both their efficiency and efficacy. Certain existing methods overlook the significance of selecting the appropriate EEG channels, which can offer crucial information for the detection of SZ. As a result, all the available EEG channels are processed to ensure comprehensive analysis. EEG signals exhibit non-stationary and aperiodic characteristics, offer valuable information in both temporal and spectral domains. Traditional machine learning approaches may overlook crucial features during manual extraction, potentially reducing diagnostic accuracy, whereas deep learning models excel in automatically extracting features, demonstrating impressive classification accuracies in EEG-based diagnoses. In recent years, the feature-fusion approach has been received much attention as a potential solution for enhancing the classification performance (Goshvarpour and Goshvarpour 2020; Jana et al. 2022; Yang et al. 2022). The existing methods provide promising performance in classification, but these algorithms are typically more computationally intensive and take longer time to execute. Furthermore, technological advancement has significantly changed the mental health management of patients with chronic disorders.

In this study, an innovative solution is proposed for the precise and effective diagnosis of SZ using EEG signals. Taking inspiration from recent studies on brain regions with high discriminant abilities, the proposed method significantly reduces data volume compared to current approaches, leading to minimal transmission and post-processing delays while maintaining higher SZ diagnosis accuracy. The methodology involves dataset preparation, signal decomposition using adaptive multi-resolution signal analysis methods (EWT and EMD), and extraction of seventy-five unique multi-domain EEG features from each sub-band. The feature set is reduced using Kruskal–Wallis (KW) test (Kulkarni et al. 2022), and significant features are selected for as input to ML classifiers for optimal sub-band selection. Three feature fusion schemes to merge multiresolution features of EEG data has been proposed at the final stage of SZ diagnosis. An extensive comparison is presented to observe the potential of traditional feature-based fusion, deep learning features based fusion, and hybrid features (deep learning as well as handcrafted features) based fusion on full channel EEG data, regional EEG data, and combined regional EEG data derived from a publicly accessible EEG dataset featuring 16-channels EEG obtained from 84 adolescents (45 SZ and 39 healthy). Figure 1 illustrates the temporal view of all 16 channels sample EEG data from both the group. The following is a brief summary of the key contributions of the proposed method:

• Introducing an effective multiresolution features fusion approach for multi domain discriminative features of EWT sub-bands and EMD IMFs to differentiate between healthy control and SZ subjects.

• Proposing hybrid feature combiner with three different feature fusion schemes to merge multiresolution features of EEG data and perform the best SZ diagnosis outcomes.

• Developing a light weight hybrid model (hybridization of CNN and ML classifier) for SZ diagnosis using subject independent (tenfold cross validation) and subject dependent (Leave-One-Subject-Out: LOSO) testing strategy to validate the effectiveness of proposed methodology.

• Comparing the performance extensively to explore the brain-region, and combined brain-regional biomarkers for the detection of SZ symptomatic adolescents through EEG signals.

Fig. 1.

Sample EEG signals of all 16 channels from the considered dataset, a Normal EEG signals from healthy control and b EEG signals from SZ symptomatic subject

The subsequent sections of this article are organized in the following manner. Materials and methods used in this study are discussed in Sect. "Materials and methods". The experimental result and discussion are elaborated in Sect. "Results and discussions". Ultimately, the last section offers the conclusive remarks of this article.

Materials and methods

The proposed methodology is briefly illustrated in Fig. 2. The automated classification model used in this study starts with signal pre-processing followed by EEG signal decomposition through EWT and EMD methods separately. Multi-domain Features from each sub-bands of EWT and EMD are extracted, and further feature selection and classification task is performed to obtain optimal sub-band from each multiresolution category based on classification accuracy. ML based band selection process is opted to make the diagnosis more robust. Later, three different feature fusion schemes are proposed to merge multiresolution features of EEG data. The fused feature set after feature selection is fed to various ML classifiers for precise diagnosis of SZ. EEG dataset, signal pre-processing, preliminary terminologies, feature sets, and other prominent details are comprehensively covered in the upcoming subsections.

Fig. 2.

Framework of the proposed methodlogy for computer-assisted diagnosis model for diagnosis of SZ symptomatic adolescents

EEG dataset

The dataset utilized in this study was created by Prof. N.N. Gorbachevskaya and Dr. S.V. Borisov in Neurophysiology and Neuro-Computer Interfaces laboratory at M.V. Lomonosov Moscow State University, Russia (Borisov et al. 2005; Phang et al. 2020). The dataset comprises resting-state EEG recordings from adolescents, with a total of 16 channels (10–14 years old, mean age = 12 years 3 months) who passed a psychiatrist’s screening were divided into two groups: Healthy control EEG (HC = 39) and those who exhibited symptoms of schizophrenia (SZ = 45) (http://brain.bio.msu.ru/eeg_schizophrenia.htm). 16 channels (viz., F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1 and O2 referenced to coupled ear electrodes) data per subject cover all five regions of brain.

The signals were recorded for a duration of 60 s at a sampling rate of 128 Hz, yielding a total of 7680 samples in one epoch (Sampling Frequency (128) x duration of recording (60 s)) (Ranjan and Sahana 2022). The study employs multi-channel EEG due to its high spatial and temporal resolution. The primary physiological artifacts are related to the actions of subjects such as muscular artifacts, ocular artifacts, and cardiac artifacts. Two experts’ opinions were used to manually eliminate artifacts, most of which were caused by head and eye movement. All patients with SZ (including childhood SZ, schizotypical, and schizoaffective disorders) were diagnosed at Mental Health Research Centre (MHRC) according to SZ diagnostic international standard criteria F20, F21, F25 of ICD-10). To validate this study, three different versions of dataset are derived from original dataset. These versions of dataset are created by rearrangement of data to analyse the impact of regions and combination regional brain areas on the diagnosis of SZ.

Full channels dataset

All subject’s EEG data from all 16 channels of two distinguished classes (HC and SZ) are concatenated individually to form full channels dataset. Keeping size of EEG signal epoch fixed (7680 samples) throughout the analysis, the number of EEG signal epochs for HC and SZ adolescents are 624 (No. of subjects (39) x No. of channel (16)) and 720 respectively. This full channels dataset is formulated to explore the capabilities of the SZ classification model on the entire EEG dataset, consistent with the approach considered in state-of-the-art methods.

Regional dataset

This dataset consists of channels that have been grouped together based on their association with a specific region in the brain. There are a total of five regions, each representing a distinct subset. These subsets consist of EEG signals obtained from different scalp areas within each region. Sixteen EEG channels considered in this study are segregated into five distinct regions of brain based on structural properties. This dataset version has been curated specifically to investigate the role of brain regions in the diagnosis of SZ through the analysis of EEG signals. Table 1 provides an overview of different brain regions, associated channels, and number of epochs.

Table 1.

Channel cluster associated with different brain regions

Brain region	Associated channels	No. of epochs
Central ( C)	C3, Cz and C4	117 HC, 135 SZ
Frontal (F)	F7, F3, F4 and F8	156 HC, 180 SZ
Occipital (O)	O1 and O2	78 HC, 90 SZ
Parietal (P)	P3, Pz and P4	117 HC, 135 SZ
Temporal (T)	T3, T4, T5 and T6	156 HC, 180 SZ

Combined regional dataset

This version of the dataset is created by combining two brain regions. The total number of possible combinations are 10. The combinations are: CF, CO, CP, CT, FO, FP, FT, OP, OT, and PT. The significance of combined brain-regional data in the diagnosis of SZ is observed through this version of dataset. Table 2 provides an overview of different combined regional brain regions, associated channels, and number of epochs.

Table 2.

Channel cluster associated with combined regional dataset

Brain region	Associated channels	No. of epochs
Centro frontal (CF)	C3, Cz, C4. F7, F3, F4 and F8	273 HC, 315 SZ
Centro occipital (CO)	C3, Cz, C4, O1 and O2	195 HC, 180 SZ
Centro parietal (CP)	C3, Cz, C4. P3, Pz and P4	234 HC, 270 SZ
Centro temporal (CT)	C3, Cz, C4. T3, T4, T5 and T6	273 HC, 315 SZ
Fronto occipital (FO)	F7, F3, F4, F8, O1 and O2	234 HC, 270 SZ
Fronto parietal (CP)	F7, F3, F4, F8, P3, Pz and P4	273 HC, 315 SZ
Fronto temporal (CT))	F7, F3, F4, F8, T3, T4, T5 and T6	312 HC, 360 SZ
Occipito parietal (CP)	O1, O2, P3, Pz and P4	195 HC, 225 SZ
Occipito temporal (CT)	O1, O2, T3, T4, T5 and T6	234 HC, 270 SZ
Parieto-temporal (PT)	P3, Pz, P4, T3, T4, T5 and T6	273 HC, 315 SZ

Data pre-processing

During the study, one-minute signal epochs of both healthy and SZ symptomatic adolescent EEG data underwent filtering using a 5th order band pass FIR filter with cut-off frequencies of 0.5 and 50 Hz. The purpose of this filtering was to remove unwanted high and low-frequency components, slow drifts, and DC components from the signals. Subsequently, a notch filter was employed to attenuate noise components caused by power line interference at 50 Hz. The application of these filters aimed to enhance the data quality, reducing artifacts and improving the reliability of the EEG recordings for further analysis and interpretation. Afterward, continuous wavelet transform (CWT) based wavelet denoising and average referencing are applied to eliminate the other artifacts and high-frequency white noise artifacts (Hamaneh et al. 2014; Ranjan et al. 2021).

Signal decomposition using EWT method

Empirical wavelet transform is one such adaptive multivariate signal processing mechanism to decompose the signal into a number of components with a solid mathematical basis (Gilles 2013; Akbari and Sadiq 2021).

The impressive signal-analysis capability has brought EWT a lot of interest in many other fields, including medical signal analysis, seismic data analysis, biomedical image processing, and numerous other domains (Ranjan et al. 2022a). In this approach, the time series data undergoes a fast Fourier transform (FFT), followed by the calculation of boundaries based on spectrum information segmentation (Kumar et al. 2022). Each component derived by EWT has a compact support Fourier spectrum. The separation of these modes is analogous to Fourier spectrum segmentation and filtering (Gilles 2013; Sadiq et al. 2021; Khan and Pachori 2021). Alike traditional wavelet transform, EWT contains both detailed and approximation coefficients. The inner product of a signal with its scaling function yields approximation coefficients (W_a(0,n)). Contrarily, the detailed coefficients W_d(m,n)) are the result of the inner product of the signal and its scaling function. These coefficients may be expressed mathematically as:

$$W_d(m,n)=⟨x(n),{\stackrel{⌢}{\mathrm{\Psi }}}_i(\omega )⟩={\sum }_{\tau =0}^{A-1}x(\tau ).\overline{{\stackrel{⌢}{\mathrm{\Psi }}}_i(\tau -n)}$$ 1

$$W_a(0,n)=⟨x(n),{\stackrel{⌢}{\mathrm{\Phi }}}_1(\omega )⟩={\sum }_{\tau =0}^{A-1}x(\tau ).\overline{{\stackrel{⌢}{\mathrm{\Phi }}}_1(\tau -n)}$$ 2

where, x is the input signal. On the basis of Fourier spectrum segment boundaries and spectral data, P wavelet filters are developed, with one being a low-pass filter (scaling function, ${\stackrel{⌢}{\mathrm{\Phi }}}_1(\omega )$) and remaining P-1 being band pass filters (wavelet function, ${\stackrel{⌢}{\mathrm{\Psi }}}_i(\omega )$). The mathematical expressions for scaling and wavelet function of EWT are as follows:

$${\stackrel{⌢}{\mathrm{\Phi }}}_1(\omega )=\left\{\begin{array}{c}1;if|\omega |\le ({\omega }_i-T_i),\hfill \\ cos\left(\frac{\mathrm{\Pi }}{2}\beta (\frac{1}{2T_i}(|\omega |-{\omega }_i+2T_i))\right);\hfill \\ if({\omega }_i-T_i)\le |\omega |\le ({\omega }_i+T_i)\hfill \\ 0;Otherwise\hfill \end{array}\right)$$ 3

and,

$${\psi }_i(\omega )=\left\{\begin{array}{c}1;if{({\omega }_i+T)}_i\le |\omega |\le ({\omega }_i-T_i),\hfill \\ cos\left(\frac{\mathrm{\Pi }}{2}\beta (\frac{1}{2T_{i+1}}(|\omega |-{\omega }_{i+1}+T_{i+1}))\right);\hfill \\ if({\omega }_{i+1}-T_{i+1})\le |\omega |\le ({\omega }_i+T_i).\hfill \\ sin\left(\frac{\mathrm{\Pi }}{2}\beta (\frac{1}{2T_i}(|\omega |-{\omega }_i+T_i))\right);\hfill \\ if({\omega }_i-T_i)\le |\omega |\le ({\omega }_i+T_i).\hfill \\ 0;Otherwise\hfill \end{array}\right)$$ 4

where, 2T_i be the transition width for center frequency ω_i, and T_i is equal to γω_i. The term γ and β be the arbitrary parameter and function respectively. The EWT coefficients divide the input signal based on energy into several sub-bands. Analysis coefficients, also known as sub-bands (SB) or multiresolution analysis (MRA) components, are obtained by inverting the transform after the wavelet filtering of the signal in the frequency domain. From the pre-processed EEG signal (x_f), the specific number (K) of sub-bands has been determined, which is mathematically expressed in Eq. (5). Figure 3 illustrates waveform plot of the sub-bands obtained from EWT for both HC and SZ cases.

$$x_f=\sum _{i=1}^{K}MR{A}_i=\sum _{i=1}^K{(SB)}_i$$ 5

Fig. 3.

Waveform plot of the sub-bands obtained from EWT for normal healthy subject (1st row), and SZ subject (2nd row)

Signal decomposition using EMD method

Empirical Mode Decomposition is one such adaptive multivariate signal processing method to decompose non-stationary and non-linear signals into finite sets of highly orthogonal pseudo-monochromatic waves, known as Intrinsic Mode Functions (IMFs) (Ranjan et al. 2022b). The fundamental advantage of the EMD approach is that it does not rely on a fixed set of mathematical functions. Given a set of n data points, the IMF may be written as X = {x = IMF₁, IMF₂….,IMF_N}. Then cubic spline interruption is used to get the envelopes after identifying the local maxima e_max and minima e_min of mean value (m_old). This is mathematically expressed as:

$$m_{\mathit{old}}={(e_{max}+e_{min})}^2$$ 6

Successively, new mean (m_new) is calculated by subtracting the original signal X from the older mean value (m_old), as shown in Eq. (7). If the spectral density (S_d) meets the stopping criteria as stated in Eq. (8), then the decomposition is stopped.

$$m_{\mathit{new}}=X-m_{\mathit{old}}$$ 7

$$S_d=\frac{\sum {\left|m_{\mathit{new}}-m_{\mathit{old}}\right|}^2}{\sum {\left(m_{\mathit{old}}\right)}^2}<\alpha$$ 8

The most frequent range for S_d is between 0.2 and 0.3 (0.2 ≤ α ≤ 0.3). If it is not achieving repeat the process to get IMFs. The EMD method has the problem of producing inter-slice discontinuities in the modes, especially when dealing with multi-dimensional or unstructured data. To resolve this problem, fast and adaptive extension of EMD (FAEMD) is adopted here in this study, which is a promising approach (Thirumalaisamy and Ansell 2018). The capability of FAEMD is to significantly reduce the computational cost of processing multidimensional signals by several orders of magnitude. This approach is made possible by the fact that order statistics filters are used to generate envelopes as opposed to the more computationally expensive spline interpolation methods (Cicone and Pellegrino 2022). Instead of using spline interpolation methods, which are slow and inflexible, FAEMD generates envelopes with the help of order statistics filters. The window size that accounts for node scales, which are derived from the EMD technique, is the most important parameter in this method (Thirumalaisamy and Ansell 2018). Extracting the size of the oscillatory scales contained in the data has the potential to provide comparable functionality to the spline-based extrema interpolation used in conventional EMD approaches. Finally, pre-processed EEG signal (x_f) is decomposed into specific number (K) of IMFs and residual (Re) using FAEMD, which is mathematically expressed in Eq. (9). Figure 4 illustrates waveform plot of the IMFs obtained from signal decomposition of EEG signals through fast and adaptive EMD for both normal and SZ cases.

$$x_f={\sum }_{i=1}^{K}IM{F}_i+\text{Re}$$ 9

Fig. 4.

Waveform plot of the IMFs obtained from FAEMD method for healthy control subject (1st row), and SZ subject (2nd row)

Band selection from multiresolution components

In the previous stage, EEG signals undergo decomposition using both EWT and EMD techniques, leading to the segmentation of these signals into four distinct sub-bands. Following this step, discriminative multi-domain EEG features are then extracted from each of these sub-bands. To identify the most relevant features, a statistical approach involving the KW test is employed, where approximately half of the prominent feature set are selected. These selected features are then utilized as input to ML classifiers, streamlining the process of selecting the optimal sub-band. For a comprehensive understanding of the methodology, the subsequent subsection provides detailed discussions on feature extraction, the feature selection, and the classification techniques applied.

Feature extraction

EEG signal classification through machine learning requires feature extraction as an essential intermediate step. A total of 75 features (statistical temporal feature, spectral feature, nonlinear features, and other EEG features) are calculated from each channel as well as sub-bands. Table 3 lists all extracted features from various domain. There are 36 features from temporal domain, 17 features from spectral domain, 13 nonlinear entropy features, 3 fractal dimension features, 2 complexity features, 2 exponent features, and 2 other EEG features (Correlation dimension, and Auto-Regressive coefficient). The features listed in Table 3, some features are widely recognised while others have not been extensively exploring in the area of EEG signal analysis. Consequently, those least popular features, their corresponding formulas, and remarks utilised in this study is represented in Table 4.

Table 3.

List of extracted features

Statistical temporal features		Spectral features	Nonlinear entropy Features	Other features
Mean (Siuly et al. 2020a)	Zero crossing (Budak et al. 2019)	Energy (Şen et al. 2014)	approximate entropy (Krishnan et al. 2020)	Katz fractal dimension
Variance (Siuly et al. 2020a)	Asymmetry (Hu et al. 2015)	Mean teager energy (Şen et al. 2014)	Spectral entropy	Higuchi fractal dimension
Standard deviation (Siuly et al. 2020a)	Hjorth activity (HA) (Das and Pachori 2021)	Total power	Tsallis entropy (Jahmunah et al. 2019)	Petrosian fractal dimension
Skewness (Siuly et al. 2020a)	Hjorth mobility (HM) (Das and Pachori 2021)	Avg. band power delta (Amin et al. 2017)	Shannon entropy (Sairamya et al. 2022)	Kolmogorov complexity (Goldblum et al. 2023)
Kurtosis (Siuly et al. 2020a)	Hjorth complexity (HC) (Das and Pachori 2021)	Avg. band power theta (Singh and Malhotra 2021)	Log energy entropy (Sairamya et al. 2022)	Lempel–Ziv complexity (Akar et al. 2016)
Minimum (Siuly et al. 2020a)	Normalized First Difference (Hamaneh et al. 2014)	Avg. band power alpha	Singular value decomposition En. (Krishnan et al. 2020)	Correlation dimension
Maximum (Siuly et al. 2020a)	Normalized second difference (Hamaneh et al. 2014)	Avg. band power beta	Permutation entropy (Krishnan et al. 2020)	Hurst exponent (Şen et al. 2014)
Normalized 6th moment (Riaz et al. 2016)	Log root sum of sequential variation	Avg. band power gamma	Sample entropy (Krishnan et al. 2020)	Lyapunov exponent (Kutepov et al. 2020)
Normalized 8th moment (Riaz et al. 2016)	crest factor (Yakoubi et al. 2019)	Relative band power delta	Fuzzy entropy (Sairamya et al. 2022)	Auto-Regressive coefficient (Wu et al. 2023)
Mode (Şen et al. 2014)	Shape factor (Isham et al. 2019)	Relative band power theta	Renyi entropy (Jahmunah et al. 2019)
Root mean square (Siuly et al. 2020a)	Impulse factor (Isham et al. 2019)	Relative band power alpha	Differential entropy (Pan et al. 2020)
Median (Şen et al. 2014)	Clearance factor (Yakoubi et al. 2019)	Relative band power beta	Dispersion entropy (Amezquita-Sanchez et al. 2021)
Mean cube	Mean curve length (Şen et al. 2014)	Relative band power gamma	Multiscale sample entropy
Coefficient of variation	L1-norm	Spectral flatness
Lower quartile (Q1) (Siuly et al. 2020a)	L2-norm	Signal to noise ratio
Upper quartile (Q3) (Siuly et al. 2020a)	L-infinity-norm	Total harmonic distortion
Intermediate quantile range (IQR) (Siuly et al. 2020a)	Logarithmic squared norm	Signal to noise and distortion ratio (Saini et al. 2020)
Pearson’s coefficient of skewness	Linear predictive coefficients (Khare et al. 2021)

Table 4.

Short explanations of features

Sr. No	Feature	Formula	Domain	Remark
1	Mean cube (MC)	$MC=\frac{1}{N}\sum _{i=1}^N[{(x_i)}^3]$	Statistical	It is the average value of a set of cubic values of data points
2	Coefficient of variation (Cv)	$Cv=\frac{\sigma }{\mu }$ where, $\sigma$ is Standard deviation and μ is the mean of the signals	Statistical	It is measure of how far apart the points in a set of data are from the mean
3	Pearson’s coefficient of skewness	${\mathit{Skew}}_{\mathit{pc}}=3\ast \frac{x_{\mathit{mean}}-x_{\mathit{mode}}}{x_{\sigma }}$	Statistical	It is used to study the direction and force of the asymmetry of the feature distribution of data points
4	Log root sum of sequential variation	$LRSSV={\text{log}}_{10}\sqrt{{\sum }_{i=1}^N{(x_i-x_{i-1})}^2}$	Statistical	It is the logarithmic value of root-sum-of-squares of two consecutive values
5	L1-norm or manhattan norm (Li et al. 2022)	$L1-Norm={\sum }_{n=1}^N|x(n)|$ where, x(n) represents instantaneous amplitude of a data point, and N is the total number of samples in an epoch	Statistical	It is the sum of absolute value of a signal. It is robust to outliers
6	L2-norm or euclidean norm (Zheng et al. 2022)	$L2-Norm=\sqrt{{\sum }_{n=1}^N{|x(n)|}^2}$	Statistical	It measures the magnitude of the signal
7	Infinity norm or maximum norm (Wilches-Bernal et al. 2022)	$Inf-Norm=\underset{N}{\text{max}}\left|x,\left(n\right)\right|$	Statistical	It calculates the maximum value of EEG signal epoch
8	logarithmic squared (LS) norm (Sofri et al. 2023)	$LS-Norm=\text{ln}[{\sum }_{n=1}^N{|x(n)|}^2]$ where, ln is denotation for natural logarithm	Statistical	It measures the energy of a signal and calculated by logarithm of the sum of the squares of the signal values
9	Total power (PTotal)	$P_{\mathit{Total}}={\sum }_{\forall freq.}\mathcal{F}\{x(t)\}.\mathcal{F}\{x^{\ast }(t)\}$ where, x(t) is input signal	Spectral	It is determined by summing the power of each frequency
10	Relative power (PR)	$P_g(f_1,f_2$) = $\frac{P(f_1,f_2)}{P_{\mathit{Total}}}\text{X}100$ where, f1, f2 indicate the low and high frequency	Spectral	It is the ratio of power of each given band to the total power
11	Spectral Flatness	$SF=\frac{\sqrt[N]{\prod _{k}x(k)}}{\frac{1}{N}\sum _{k}x(k)}$ where, N is the number of non-zero elements in spectrum	Spectral	It is the ratio of geometric mean of the power by the arithmetic mean of power spectrum
12	Signal-to-Noise Ratio (SNR)	$SNR=\frac{P_{\mathit{signal}}}{P_{\mathit{noise}}}$	Spectral	It is the ratio of signal power to noise power
13	Total harmonic distortion (THD)	$THD=\sqrt{{\sum }_{i=2}^N{\left(\frac{P_i}{P_1}\right)}^2}$ where, Pi represents instantaneous power and P1 is the fundamental power	Spectral	It is the ratio of total harmonic component power to fundamental power
14	Spectral entropy (SpE)	$SpE=\frac{1}{N}\sum _{k=1}^N|X(k).X^{\ast }(k)|$ where, X(k) is the signal in frequency domain	Spectral-domain nonlinear entropy feature	It quantifies the signal’s spectral power distribution
15	Multiscale sample entropy (MsEn)	$MsEn=\frac{1}{\tau }\sum _{i=(j-1)\tau +1}^{j\tau }{x}_i,1\le j\le \frac{N}{\tau }$ where, τ is the time scale	Nonlinear entropy feature	It calculates the entropy of a signal at various time-scales
16	Higuchi fractal dimension (HFD)	$HFD=\frac{ln\left(L\left(k\right)\right)}{ln\left(1/k\right)}$ where, L(k) represents mean value of the curve length, and k is the time interval	Fractal	It is a nonlinear measure of waveform complexity in time domain
17	Katz fractal dimension (KFD)	$KFD=\frac{ln\left(L/a\right)}{ln\left(d/a\right)}$ where, L, and a are sum and average of the Euclidean distances between the successive points of the sample. While d is any random point of the sample	Fractal	This FD is calculated by calculating the distance between two successive points
18	Petrosian fractal dimension (PFD)	$PFD=\frac{{log}_{10}k}{{log}_{10}k+{log}_{10}(k/(k+0.4N_{\delta }))}$ where, k is the number of signal’s samples and Nδ is the number of sign changes in the signal derivative	Fractal	It is a chaotic method that calculates the complexity of a signal
19	Correlation dimension [CR]	$C_R=\underset{\in ,{\in }^{\prime }\to 0^{+}}{lim}\frac{ln\left[\frac{C(\in )}{C({\in }^{\prime })}\right]}{ln\left(\in /{\in }^{\prime }\right)}$ where H is the Heaviside step function, and $C(\in )$ correlation integral $C(\in )=\underset{N\to \infty }{lim}\frac{1}{N^2}{\sum }_{i,j=1(i\ne j)}^{\infty }H\left(\in -\left|x_i-x_j\right|\right)$	Time-domain nonlinear	The complexity of a signal, an indicator of its self-similarity, is quantified by the correlation dimension

Feature selection and classification

Among all extracted handcrafted feature set, half of effective features are chosen to be tested with various classifier models in order to distinguish between the two defined classes. A non-parametric statistical test (i.e. Kruskal–Wallis test) is used to determine significance score of each and every features, then those features having higher significance scores are rearranged in descending order and first 38 features are considered for further stage of classification. After this, ML classifier is applied for optimal sub-band selection from EWT and EMD blocks respectively. The comparative studies of various classifiers are also carried out to determine the best classifier that produces outstanding performance in this specific task. In this study, we utilized seven potential classifiers selected for their widespread use, rapid learning capabilities, and proven effectiveness in biomedical applications. The classifiers used in this process are: KNN (Ranjan and Sahana 2019), SVM (Preity et al. 2023), Naive Bayes (NB) (Siuly et al. 2016), Decision Trees (DT) (Song and Lu 2015), Random forest (RF) (Ranjan and Sahana 2023), AdaBoost (ADB), and EBT (Bühlmann 2012). KNN algorithm is a simple yet powerful classifier which classifies a new data point by analysing the predominant class among its k-nearest neighbours within the feature space. On the basis of efficient SZ diagnosis literature, the parametric value of K for neighbours is consistently set at five throughout this study. SVM stands out as a robust machine learning algorithm designed to identify the optimal hyperplane that effectively separates data points belonging to different classes, maximizing the margin between them. SVM exhibits versatility by adeptly handling both linearly separable and non-linearly separable data. The cubic kernel of SVM classifier is used in this study. Further, NB follows probabilistic algorithm specially Bayes theorem with strong independence assumptions to perform classification task. It is frequently employed as a baseline classifier in machine learning due to its efficiency and remarkably performance in various biomedical classification problems. Another classifier, decision tree utilises a tree structure to divide data recursively based on the most effective features for separating the classes (Siuly et al. 2016, 2020b; Savas and Dovis 2019). Random Forest, an alternative classifier, is a widely used ensemble learning technique that uses multiple decision trees during training. Each tree functions as a classifier, and the output of all trees is given specific weights. ADB and EBT are two other ensemble tree-based algorithms that similarly utilize multiple trees to enhance accuracy and reduce overfitting. Ensemble classifiers basically uses weak models sequentially to minimize the errors of previous models to improve predictions. EBT classifier specially has shown the most promising results consistently when applied to the EEG data. In this study, the number of splits for each of tree-based classifiers is configured to use 100 splits.

Hybrid feature combiner

A hybrid feature combiner (HFC) is a feature engineering approach to generate novel features by merging multiple existing features and auto generated features in a hybrid manner. The primary objective of this approach is to augment predictive capabilities in machine learning models by capturing supplementary data. The study uses a range of multiresolution feature fusion techniques in conjunction with deep learning models as feature extractors, and employs KW tests as feature selector. Subsequently, various machine learning classifiers are employed to carry out the classification task. The primary concept underlying the hybrid feature combiner is to exploit the complimentary information inherent in several features, thereby generating more informative representations. Through the integration of various features, the model has the capability to effectively capture nonlinear interactions, higher-order relationships, and intricate patterns that would be missed if the features were considered separately. In order to avoid overfitting or creating redundant features, the complexity and dimensionality of the feature space should also be considered. In this study, there are hybrid feature combiner with three feature fusion schemes to merge multiresolution features of EEG data has been proposed. The effectiveness of proposed feature fusion approach is tested on three different versions of dataset which is derived from original dataset. The data forms are: (a) full channel EEG data, (b) selective brain regional data, and (c) selective combined brain regional data. The details about the proposed fusion schemes are as follows:

Feature fusion with handcrafted features

The extracted handcrafted feature sets of selected sub-band from EWT and EMD blocks are combined in a manner such that the resulting feature set is the mean of both sub-band features. The feature fusion scheme resulting new feature set (S_f) follows the Eq. 10 as:

$$S_f=\frac{1}{\mathit{Ns}}(S_{\mathit{EMD}}+S_{\mathit{EWT}})$$ 10

where Ns is the number of sidebands. $S_{\mathit{EMD}}$ and $S_{\mathit{EWT}}$ are selected handcrafted feature sets of optimal EWT and EMD Sideband respectively. For instance, if KFD is common for both the optimal multiresolution components, then the new hybrid feature would be the average of KFD features derived from both KFD_{optimal EMD sub-band} and KFD_{optimal EWT sub-band}. Since not all the selected features are common for both $S_{\mathit{EMD}}$ and $S_{\mathit{EWT}}$, the uncommon features are retained in their original form during the feature fusion stage. Therefore, following the fusion process, the new feature set have same dimension as of the EMD or EWT optimal sub band. The KW test is applied for feature reduction in the next step. One third of features having highest significant score in KW test is further fed to the classifiers for SZ classification from EEG data.

Feature fusion with deep features

The selected sub-bands from EWT and EMD block are directly supplied as input to deep learning based CNN Architecture in sequential manner for deep features extraction. The optimal sub-bands obtained from EWT and EMD blocks are processed separately. In each case, individual data samples of dimensions (7680, 1) from all subjects and defined channels are sequentially fed as input to CNN. As the data pass through the layers of the CNN, the network learns hierarchical representations of features. These features become more abstract and complex as the signal propagates through deeper layers. The entire dataset is utilized for the training of CNN. In the architecture, after passing through CNN architecture, the signal is processed up to the last layer, which consists of a fully connected (FC) layer containing 150 neurons. At this particular FC layer, the resultant feature set is aggregated into a single dimensional array of dimensions 150. As a result, distinct deep features are generated separately for the optimal EMD and EWT sidebands, each having the same dimensions. No classification task is executed in this process. The proposed light-weight CNN consists of 4 convolution layer in which each layer comprises of one 1D convolution block, one activation function (ReLU) block, one batch normalization and one max pooling block. Figure 5 includes the illustration of CNN Architecture for deep features extraction. A CNN takes full benefit of the inherent hierarchical patterns in data, assembling more complex patterns by using smaller and smaller patterns (Goodfellow et al. 2016; Shrestha and Mahmood 2019). The CNN architecture comprises a stack of convolutional layers, pooling, and fully connected layers in the sequence followed by each other (Roy et al. 2019). Features can be derived through the utilization of convolutional and pooling layers. Subsequently, these extracted features are linked to the output through a fully connected layer.

Fig. 5.

Block diagram of final stage classification after optimal multiresolution band selection. Hybrid model (CNN and ML classifier) and hybrid features set (Deep bottleneck features and selective handcrafted feature sets) work efficiently in the diagnosis of SZ

1D-CNNs have found widespread application in a variety of signal-processing tasks since their debut. It has shown excellent results in EEG signal classification study. The 1D-CNN comprises kernels (filters) that convolve with the input EEG signal. As the kernel matrix slides over the EEG signal, a convolution operation is performed, after this operation feature map has been generated. Let the feature map be f_m then it will be mathematically expressed for EEG signal (x) with N data points and l number of filters as:

$$f_m={\sum }_{n=0}^{N-m}{x}_n{l}_{m-n}$$ 11

The majority of research studies in EEG signal analysis utilizing convolutional layers within DL architectures have predominantly leaned towards adopting the rectified linear unit (ReLU) as the activation function for these layers. The fundamental concept underlying Batch Normalisation is to standardize the input of every layer within a mini-batch, which is a subset of the training data, in order to achieve a mean of zero and a variance of one. Max pooling typically selects the maximum input value while neglect the other values within the filter. The dropout layer is designed to randomly deactivate neurons during the training process in order to mitigate overfitting and increase the generalizability of the model.

The other layer of CNN as flatten layer takes a multi-dimensional tensor and flattens it into a vector, making it easier to connect to FC layers as an input. The FC layers, also known as dense layers, are an essential component of neural networks. Each neuron of FC is connected to every neuron gain an understanding of the intricate connections among input features, facilitating the development of hierarchical representations and the ability to make accurate predictions. In the FC layers, the softmax activation function is employed to generate a probability distribution across different classes. This distribution indicates the probability or likelihood of the input belonging to each specific class. The total number of parameters used in full channel based study is 4,93,449. In the later stage, the KW test is applied for feature reduction. One fifth of the deep features with the highest significance score in the KW test are then supplied to classifiers for SZ classification from EEG data. The ablation study of network structure required to configure the proposed CNN architecture is illustrated in Table 5.

Table 5.

Ablation study of network structure of proposed CNN architecture

Name of layer	Output shape
Input layer	(7680,1)
First Convolutional Layer Conv1D: 16 filters of size 7 ReLU activation Batch normalization MaxPooling1D: Pool size 3	(7674, 16) after Conv1D (2558, 16) after MaxPool
Intermediate Convolutional Layer Conv1D: 32 filters of size 8 ReLU activation Batch normalization MaxPooling1D: Pool size 3	(2549, 32) after Conv1D (849, 32) after MaxPool
Additional Convolutional Layer Conv1D: 12 filters of size 7 ReLU activation Batch normalization MaxPooling1D: Pool size 3	(843, 12) after Conv1D (281, 12) after MaxPool
Last Convolutional Layer Conv1D: 9 filters of size 7 ReLU activation Batch normalization MaxPooling1D: Pool size 3	(275, 9) after Conv1D (91, 9) after MaxPool
Dropout (0.5)	(91, 9)
Flatten	819
Fully-connected: 500 neurons	500
Fully-connected: 150 neurons (deep Features extraction Point)	150
Fully-connected: 2 neurons (softmax)	2

Feature fusion with hybrid features

One of the key contribution in this work is to observer the impact of multi-domain handcrafted features, and machine derived automated deep learning features on SZ diagnosis using EEG signals. The deep learning extracted features collected from fusion scheme-2 are combined together with handcrafted feature sets of selected sub-band from EWT and EMD blocks in sequential manner. 150 deep learning bottleneck features, and 38 significant handcrafted features from each optimal sub-band of corresponding signal decomposition blocks are combined together in sequence to form the new feature set for the classifiers with 226 features. Further, the KW test is applied for feature selection and only 10% features with the highest significance score in the test have been retained for final stage of classification.

Classification task at diagnosis end

A subset of features constituting just one-tenth of the total features is selected for testing after undergoing feature reduction using the KW test. These chosen features are then subjected to various classifier models to differentiate between the two specified classes as HC and SZ. The results clearly demonstrate that EBT classifier has consistently exhibited the most promising performance compared to other classifiers across various versions of EEG dataset utilized in this study which include the full channel EEG dataset, regional EEG dataset, and combined regional EEG dataset derived from a publicly accessible adolescent EEG dataset. In summary, EBT classifier consistently demonstrates exceptional performance throughout the study, making it a preferable choice for SZ classification. Figure 5 illustrates the block diagram of final stage classification after optimal multiresolution band selection. Hybrid model (CNN and ML classifier) is used for diagnosis of SZ symptomatic adolescents. CNN architecture is used for deep features extraction from optimal sidebands and feature fusion scheme-3 (deep learning bottleneck features and selective handcrafted features) has given the best classification outcome.

Performance evaluation indexes

The effectiveness of any classification system may be assessed through performance metrics. The most common approach to obtaining performance metrics is to acquire the confusion metric (CM). From CM in a binary classification task, true positives (TP), false positives (FP), true negatives (TN), and false negative (FN) can be computed. These values serve as the basis for various performance evaluation indexes. In this study, the effectiveness of the proposed methodology is measured using ten classification performance evaluation indexes, which is presented in Table 6.

Table 6.

Classification performance evaluation indexes

Sr. No	Performance indexes	Formula	Remarks	Desirable value
1	Accuracy (ACC)	$ACC=\frac{TP+TN}{TP+FP+TN+FN}$	It tells how the model works in general across all classes	High
2	Sensitivity (SEN)	$SEN=\frac{\mathit{TP}}{TP+FN}$	The ability of a model to identify positive instances	High
3	Specificity (SPE)	$SPE=\frac{\mathit{TN}}{FP+TN}$	The ability of a model to identify negative instances	High
4	Precision (PRE)	$PRE=\frac{\mathit{TP}}{TP+FP}$	It measures the proportion of samples that were correctly labelled as positive	High
5	False discovery rate (FDR)	$FDR=\frac{\mathit{FP}}{TP+FP}$	It quantifies the proportion of individuals classified as disease-positive who are actually disease-negative	Low
6	Negative predictive value (NPV)	$NPV=\frac{\mathit{TN}}{TN+FN}$	It is the likelihood that individuals with a negative screening test do not really have the illness	High
7	F1-Score	$F1-Score=2x\frac{(PRExSEN)}{PRE+SEN}$	It is a statistic that takes the harmonic mean of the accuracy and sensitivity	High
8	Matthews correlation coefficient (MCC)	$MCC=\frac{\left(T,P,,\ast ,T,N\right)-(FP\ast FN)}{\sqrt{\left(T,P,+,F,P\right)\left(T,P,+,F,N\right)\left(T,N,+,F,P\right)(TN+FN)}}$	The MCC provides a concise summary of the confusion matrix and is the best single-value classification statistic	High
9	Cohen’s Kappa (k)	$k=1-\frac{1-p_o}{1-p_e}$ Where, po is the observed agreement, and pe is the expected agreement	It is a statistic which measures interrater agreement for qualitative items	High
10	Geometric mean (G-mean)	$G-Mean=\sqrt{Sensitivity\ast Specificity}$	It is the root of the product of class-wise sensitivity and specificity	High

Testing strategies

Testing strategies pertain to the methodologies and approaches employed for assessing the effectiveness and ability of a trained model to make accurate predictions on new or unseen data. The primary objective of testing strategies is to evaluate the ability of classification model to perform accurate predictions on unobserved instances, while also identifying potential concerns such as overfitting or under-fitting. There are two testing strategies incorporated in this study which are subject independent testing and subject dependent testing. The objective of this study is to evaluate the classification accuracy of the proposed methodology by utilising train-test splits that are either dependent or independent on the subjects.

In subject independent testing, tenfold cross validation (cv) has been employed. tenfold cv follows the complete data split into 10 folds/parts in each fold, 8 folds are used for training, onefold for validation, and onefold for testing. Selection of input features, or folds is independent of the subjects in the dataset.

In subject dependent testing also known as leave –one subject-out, the train-test split is based on the subjects in the dataset (Dogan et al. 2022; Sharma and Joshi 2022; Hassan et al. 2023). The dataset under consideration comprises a total of 84 subjects, which have been divided into ten subsets. Six of these subsets contain eight subjects each, while the remaining four subsets consist of nine subjects each. The data is partitioned in a manner that ensures no subset shares a common subject. Later, 8-subsets are opted for training, 1 subset for validation, and one for testing. The leave one out approach ensures that there is no subject bias present. If the same subject appears in both the training and testing datasets, the model will learn too much about the subject, which is commonly observed in cross-validation. Since, the model in LOSO was never trained on the new subject, it is likely to produce subpar results. Therefore, LOSO accuracy outcome is comparatively lower than tenfold cv but it is practically realizable testing approach.

Algorithm 1.

Proposed computer-assisted model for automated diagnosis of schizophrenia from EEG signals.

The objective of employing these two testing strategies is to examine the performance efficacy of the proposed classification model for diagnosis of schizophrenia using EEG signals when utilizing train-test splits that are either dependent or independent to the subjects present in the dataset. Through a comparative analysis of the outcomes derived from both approaches, one can acquire valuable insights into the extent to which the model exhibits generalizability across different subjects, as well as assess its robustness to variations in subjects. At the end, the proposed methodology is summarized in the form of pseudocode as depicted in algorithm 1.

Results and discussions

Results achieved by classifying SZ and healthy control EEG signals using the proposed methodology are presented in this section. In this study, the effectiveness of the proposed methodology is measured across ten classification performance evaluation indexes, which is presented in Table 6. The first part of classification studies was carried out in MATLAB (version R2021a) environment on a 64-bit desktop machine outfitted with an Intel Core i7-8700 CPU (speed of 3.20 GHz), 8 GB RAM and Windows 10.1 home edition operating system. The implementation of second part of the simulation study performing feature fusion and deep learning bottleneck features extraction has been executed on a Google Colab GPU (Tesla K80) within a Python 3.9 environment. The model implementation in this study utilised TensorFlow 2.12.0 and Keras 2.12.0 libraries. The categorical cross entropy function is commonly employed as a loss function in the training and evaluation of models. The Adam optimizer is used with a learning rate of 0.0001, and keeping the batch size 32. The model has been trained for more than 100 epochs, utilising the default initialization parameters provided by the Keras libraries framework. Upon completing the training of the proposed hybrid model in the Google Colab platform, the trained model can be saved using the model.save() function. This step ensures that the model’s learned parameters and architecture are preserved for future use.

The study is conducted into 2 major parts: (a) optimal multiresolution signal components selection for SZ classification, and (b) features fusion-based SZ diagnosis. The first part of study is validated through subject independent testing strategy while the final classification tasks were completed using both subject dependent and subject independent testing. The proposed study has been explored to realise the impact of full channel data, region wise brain data, and combined regional data using different feature fusion schemes and testing strategies to diagnose SZ using EEG. The impacts of handcrafted features, automatically derived features, or combinations of both features on SZ diagnosis using EEG are also explored in this work. The superiority of proposed model is proven by comparing it among current state-of-the-art methods. The experimental findings in each case, and comparison are comprehensively elaborated in this section.

Optimal sub-band selection using subject independent testing

Initially, a total of 75 multi-domain handcrafted features are calculated from sub-bands and IMF from EWT and EMD group respectively. The process is done for all three EEG datasets defined in Sect. "EEG dataset", and the entire EEG analysis is carried out with fixed signal epoch size of 7680 samples. The extracted feature set is tested with different classifier models for optimal sub-band selection from EWT and EMD blocks respectively in order to distinguish between healthy control and schizophrenia classes. Seven potential classifiers have been employed in this study based on their popularity, quick learning capability, and effectiveness in biomedical applications. The comparative studies of various classifiers are also carried out to determine the best classifier that produces outstanding performance in this specific task. The classifiers follow tenfold cv testing strategy to discriminate between the defined classes. The results obtained in analysing each sub-bands of EWT and EMD blocks on full channel EEG data, brain regional data and combinational brain data are reported in Table 7, 8 and 9 respectively. The optimal values of each performance index are highlighted in bold. It is evident from Table 7, 8, and 9 that the last sub-band has shown exceptional performance in classifications among both the group (EWT and EMD group). The optimal sub-band is highlighted in bold in all three tables. Based on finding, EBT classifier has shown maximum classification efficacy in the selection of optimal multiresolution components from both EMD and EWT groups. The best reported accuracies in the first section of work using hand crafted features of EEG signals for optimal sub-bands selection from EMD and EWT blocks are 99.18% and 99.33% respectively in full channel EEG data.

Table 7.

Results obtained from various IMFs of EMD in terms of classification performance evaluation indexes

Sr. no	Signal part	Classifiers	Acc. (%)	SEN (%)	SPE (%)	Pre (%)	NPV (%)	F1-Score(%)	FDR	MCC	Kappa	G-mean
1	IMF-1	KNN	75.52	99.17	48.24	68.85	98.05	81.27	0.31	0.56	0.34	69.16
		SVM	78.79	97.92	56.73	72.31	95.93	83.19	0.28	0.61	0.38	74.53
		NB	97.69	98.61	96.63	97.13	98.37	97.86	0.03	0.95	0.61	97.62
		DT	96.06	95.69	96.47	96.91	95.10	96.30	0.03	0.92	0.60	96.08
		RF	97.47	97.92	96.96	97.38	97.58	97.65	0.03	0.95	0.61	97.43
		ADB	98.21	98.75	97.60	97.93	98.54	98.34	0.02	0.96	0.62	98.17
		EBT	98.66	98.47	98.88	99.02	98.25	98.75	0.01	0.97	0.62	98.68
2	IMF-2	KNN	69.49	84.03	52.72	67.22	74.10	74.69	0.33	0.39	0.31	66.56
		SVM	83.18	95.14	69.39	78.20	92.52	85.84	0.22	0.68	0.45	81.25
		NB	98.21	98.47	97.92	98.20	98.23	98.34	0.02	0.96	0.62	98.19
		DT	98.59	98.47	98.72	98.88	98.25	98.68	0.01	0.97	0.62	98.60
		RF	98.74	99.03	98.40	98.62	98.87	98.82	0.01	0.97	0.62	98.71
		ADB	98.36	97.64	99.20	99.29	97.33	98.46	0.01	0.97	0.62	98.42
		EBT	99.03	98.75	99.36	99.44	98.57	99.09	0.01	0.98	0.63	99.05
3	IMF-3	KNN	78.27	97.92	55.61	71.79	95.86	82.84	0.28	0.60	0.37	73.79
		SVM	80.06	96.53	61.06	74.09	93.84	83.84	0.26	0.63	0.40	76.77
		NB	98.07	98.33	97.76	98.06	98.07	98.20	0.02	0.96	0.62	98.04
		DT	98.36	98.19	98.56	98.74	97.93	98.47	0.01	0.97	0.62	98.38
		RF	98.88	99.44	98.24	98.49	99.35	98.96	0.02	0.98	0.62	98.84
		ADB	98.51	98.19	98.88	99.02	97.94	98.61	0.01	0.97	0.62	98.54
		EBT	98.96	99.03	98.88	99.03	98.88	99.03	0.01	0.98	0.63	98.95
4	IMF-4	KNN	79.17	94.44	61.54	73.91	90.57	82.93	0.26	0.60	0.40	76.24
		SVM	78.05	97.50	55.61	71.71	95.07	82.64	0.28	0.60	0.37	73.63
		NB	98.07	98.47	97.60	97.93	98.23	98.20	0.02	0.96	0.62	98.03
		DT	98.44	98.75	98.08	98.34	98.55	98.54	0.02	0.97	0.62	98.41
		RF	98.96	99.03	98.88	99.03	98.88	99.03	0.01	0.98	0.63	98.95
		ADB	99.03	99.03	99.04	99.17	98.88	99.10	0.01	0.98	0.63	99.03
		EBT	99.18	99.44	98.88	99.03	99.36	99.24	0.01	0.98	0.63	99.16

Table 8.

Results obtained from various sub-bands of EWT in terms of classification performance evaluation indexes

Sr. no	Signal part	Classifiers	Acc. (%)	SEN (%)	SPE (%)	Pre (%)	NPV (%)	F1-Score(%)	FDR	MCC	Kappa	G-mean
1	SB-1	KNN	73.44	75.56	70.99	75.03	71.57	75.29	0.25	0.47	0.39	73.24
		SVM	94.35	93.61	95.19	95.74	92.81	94.66	0.04	0.89	0.59	94.40
		NB	98.07	98.33	97.76	98.06	98.07	98.20	0.02	0.96	0.62	98.04
		DT	97.40	96.53	98.40	98.58	96.09	97.54	0.01	0.95	0.62	97.46
		RF	97.69	98.06	97.28	97.65	97.75	97.85	0.02	0.95	0.61	97.66
		ADB	98.59	99.17	97.92	98.21	99.03	98.69	0.02	0.97	0.62	98.54
		EBT	98.66	99.03	98.24	98.48	98.87	98.75	0.02	0.97	0.62	98.63
2	SB-2	KNN	72.25	78.89	64.58	71.99	72.61	75.28	0.28	0.44	0.36	71.38
		SVM	95.16	96.94	93.11	94.20	96.35	95.55	0.06	0.90	0.59	95.01
		NB	97.77	97.64	97.92	98.18	97.29	97.91	0.02	0.96	0.62	97.78
		DT	98.59	98.75	98.40	98.61	98.56	98.68	0.01	0.97	0.62	98.57
		RF	98.74	98.89	98.56	98.75	98.72	98.82	0.01	0.97	0.62	98.72
		ADB	97.92	98.33	97.44	97.79	98.06	98.06	0.02	0.96	0.62	97.88
		EBT	98.81	99.03	98.56	98.75	98.87	98.89	0.01	0.98	0.62	98.79
3	SB-3	KNN	67.41	72.92	61.06	68.36	66.15	70.56	0.32	0.34	0.32	66.72
		SVM	95.31	97.50	92.79	93.98	96.98	95.71	0.06	0.91	0.59	95.12
		NB	96.24	95.92	96.61	97.01	95.38	96.46	0.03	0.92	0.61	96.27
		DT	97.54	97.50	97.60	97.91	97.13	97.70	0.02	0.95	0.61	97.55
		RF	98.51	99.03	97.92	98.21	98.87	98.62	0.02	0.97	0.62	98.47
		ADB	98.59	98.89	98.24	98.48	98.71	98.68	0.02	0.97	0.62	98.56
		EBT	98.36	98.33	98.40	98.61	98.08	98.47	0.01	0.97	0.62	98.37
4	SB-4	KNN	79.39	84.44	73.56	78.65	80.39	81.45	0.21	0.59	0.44	78.81
		SVM	93.53	95.28	91.51	92.83	94.38	94.04	0.07	0.87	0.57	93.37
		NB	97.69	98.06	97.28	97.65	97.75	97.85	0.02	0.95	0.61	97.66
		DT	98.21	98.06	98.40	98.60	97.77	98.33	0.01	0.96	0.62	98.23
		RF	98.88	99.44	98.24	98.49	99.35	98.96	0.02	0.98	0.62	98.84
		ADB	98.96	99.31	98.56	98.76	99.19	99.03	0.01	0.98	0.62	98.93
		EBT	99.33	99.58	99.04	99.17	99.52	99.38	0.01	0.99	0.63	99.31

Table 9.

Comparative study of optimal EMD and EWT sub-band features in terms of performance indexes on different brain regions data

Sr. no	Brain region	Features	Best classifier & (sub band)	Acc. (%)	SEN (%)	SPE (%)	Pre (%)	NPV (%)	F1-Score(%)	FDR	MCC	Kappa	G-mean
1	Central	EMD optimal band Features	EBT (IMF-4)	98.81	99.26	98.29	98.53	99.14	98.89	0.01	0.98	0.62	98.77
		EWT optimal band Features	EBT (SB-4)	99.21	99.26	99.15	99.26	99.15	99.26	0.01	0.98	0.63	99.20
2	Frontal	EMD optimal band Features	EBT (IMF-2)	93.45	95.00	91.67	92.93	94.08	93.96	0.07	0.87	0.57	93.32
		EWT optimal band Features	RF/EBT (SB-4)	97.94	98.90	96.82	97.30	98.70	98.09	0.03	0.96	0.61	97.85
3	Occipital	EMD optimal band Features	EBT (IMF-4)	95.83	94.44	97.44	97.70	93.83	96.05	0.02	0.92	0.61	95.93
		EWT optimal band Features	RF/ADB (SB-3/4)	98.81	98.89	98.72	98.89	98.72	98.89	0.01	0.98	0.62	98.80
4	Parietal	EMD optimal band Features	RF (IMF-4)	94.05	94.81	93.16	94.12	93.97	94.46	0.06	0.88	0.58	93.98
		EWT optimal band Features	RF/EBT (SB-3)	98.41	98.52	98.29	98.52	98.29	98.52	0.01	0.97	0.62	98.40
5	Temporal	EMD optimal band Features	EBT (IMF-2)	93.11	94.41	91.61	92.86	93.42	93.63	0.07	0.86	0.57	93.00
		EWT optimal band Features	EBT (SB-4)	94.64	99.44	89.10	91.33	99.29	95.21	0.09	0.90	0.57	94.13

Based on extracted features, the studies have been carried out to see the significance of different brain regions data, and combinational brain regional data. It is observed from Table 9 that central brain region data has attained the maximum accuracy of 98.81 and 99.21% in EMD and EWT groups respectively. Occipital, and Parietal regions have also shown significant impact in the diagnosis of SZ using EEG signal. In the other side, the diagnosis could also rely on the combinational regional data as it shown good classification accuracy. It is evident from Table 10 that central-occipital combined brain region data has attained the maximum accuracy of 97.62 and 99.60% in EMD and EWT groups respectively. The considerable performances have also observed from Central- Parietal, and Occipital- Parietal combined regions data.

Table 10.

Comparative study of optimal EMD and EWT sub-band features in terms of performance indexes on different combined brain regions data

sr no	Combined brain region	Features	Best classifier & (sub band)	Acc. (%)	SEN (%)	SPE (%)	Pre (%)	NPV (%)	F1-Score(%)	FDR	MCC	Kappa	G-mean
1	CO	EMD optimal band Features	EBT (IMF-4)	97.62	96.89	98.46	98.64	96.48	97.76	0.01	0.95	0.62	97.67
		EWT optimal band Features	EBT (SB-4)	99.60	100.00	99.15	99.26	100.00	99.63	0.01	0.99	0.63	99.57
2	CP	EMD optimal band Features	EBT (IMF-4)	95.24	96.30	94.02	94.89	95.65	95.59	0.05	0.90	0.59	95.15
		EWT optimal band Features	EBT (SB-4)	99.29	99.56	98.97	99.12	99.48	99.33	0.01	0.99	0.63	99.26
3	CF	EMD optimal band Features	ADB (IMF-4)	90.31	92.06	88.28	90.06	90.60	91.05	0.10	0.81	0.55	90.15
		EWT optimal band Features	EBT (SB-4)	98.64	99.37	97.80	98.12	99.26	98.74	0.02	0.97	0.62	98.58
4	CT	EMD optimal band Features	EBT (IMF-4)	97.62	96.89	98.46	98.64	96.48	97.76	0.01	0.95	0.62	97.67
		EWT optimal band Features	EBT (SB-4)	99.60	100.00	99.15	99.26	100.00	99.63	0.01	0.99	0.63	99.57
5	OP	EMD optimal band Features	RF (IMF-4)	97.62	96.89	98.46	98.64	96.48	97.76	0.01	0.95	0.62	97.67
		EWT optimal band Features	EBT (SB-4)	99.60	100.00	99.15	99.26	100.00	99.63	0.01	0.99	0.63	99.57
6	OF	EMD optimal band Features	EBT (IMF-3)	97.62	96.89	98.46	98.64	96.48	97.76	0.01	0.95	0.62	97.67
		EWT optimal band Features	RF (SB-4)	99.60	100.00	99.15	99.26	100.00	99.63	0.01	0.99	0.63	99.57
7	OT	EMD optimal band Features	EBT (IMF-4)	97.62	96.89	98.46	98.64	96.48	97.76	0.01	0.95	0.62	97.67
		EWT optimal band Features	EBT (SB-4)	99.60	100.00	99.15	99.26	100.00	99.63	0.01	0.99	0.63	99.57
8	PF	EMD optimal band Features	ADB (IMF-4)	97.62	96.89	98.46	98.64	96.48	97.76	0.01	0.95	0.62	97.67
		EWT optimal band Features	EBT (SB-4)	99.60	100.00	99.15	99.26	100.00	99.63	0.01	0.99	0.63	99.57
9	PT	EMD optimal band Features	EBT (IMF-4)	92.86	96.19	89.01	90.99	95.29	93.52	0.09	0.86	0.56	92.53
		EWT optimal band Features	EBT (SB-4)	97.45	99.37	95.24	96.01	99.24	97.66	0.04	0.95	0.60	97.28
10	FT	EMD optimal band Features	ADB (IMF-4)	94.52	94.67	94.36	95.09	93.88	94.88	0.05	0.89	0.59	94.51
		EWT optimal band Features	EBT (SB-4)	99.05	99.56	98.46	98.68	99.48	99.12	0.01	0.98	0.62	99.01

Features fusion-based SZ diagnosis at final stage using subject independent and dependent testing

In the second part of the work, three feature fusion schemes to merge multiresolution features of EEG data have been introduced. An extensive comparison is presented to observe the potential of simple feature based fusion, deep learning features based fusion, and hybrid features (deep learning as well as handcrafted features) based fusion on full channel EEG data, regional EEG data, and combined regional EEG data derived from a publicly accessible EEG dataset featuring 16-channels EEG obtained from 84 adolescents (45 SZ and 39 healthy).

Both the subject independent (tenfold cv) and subject dependent (LOSO) testing strategy are opted to validate the study. Table 11 represents the obtained results using various feature fusion schemes in terms of performance indexes with subject independent testing. The fusion scheme is tested on full channel, selective brain regional, and selective combinational regional data. Among all the feature fusion schemes studied, the hybrid fusion of bottleneck features and handcrafted features extracted from EEG signals using optimal sub-bands of EMD and EWT blocks achieved remarkable results. It attained a maximum overall accuracy of 99.93%, indicating its excellent performance in predicting EEG signal patterns across all subjects and regions. Additionally, it achieved a perfect 100% region (central) specific accuracy, demonstrating its exceptional ability to accurately classify EEG signals using subject independent testing (tenfold cross validation). EBT classifier is performing exceptionally well in this portion of study.

Table 11.

Results obtained using various feature fusion schemes in terms of performance indexes with subject independent testing

Sr. No	Fusion scheme	Dataset	Best classifier	Acc. (%)	SEN (%)	SPE (%)	Pre (%)	NPV (%)	F1-Score(%)	FDR	MCC	Kappa	G-mean
1	Fusion scheme- 01 (feature fusion with handcrafted features)	Full Ch	RF / EBT	99.85	99.86	99.84	99.86	99.84	99.86	0.00	1.00	0.63	99.85
		Central region	EBT	99.21	99.26	99.15	99.26	99.15	99.26	0.01	0.98	0.63	99.20
		Occipital region	EBT	97.62	96.67	98.72	98.86	96.25	97.75	0.01	0.95	0.62	97.69
		Parietal region	EBT	97.02	98.89	94.87	95.70	98.67	97.27	0.04	0.94	0.60	96.86
		CO: Combo regional	RF / EBT	97.86	97.56	98.21	98.43	97.21	97.99	0.02	0.96	0.62	97.88
		CP: Combo regional	EBT	97.82	97.96	97.65	97.96	97.65	97.96	0.02	0.96	0.62	97.81
2	Fusion scheme-02 (feature fusion with deep features)	Full Ch	SVM	99.36	99.30	99.43	99.50	99.20	99.40	0.00	0.99	0.63	99.37
		Central region	EBT	99.47	99.35	99.62	99.67	99.25	99.51	0.00	0.99	0.63	99.48
		Occipital region	RF	99.12	99.02	99.25	99.34	98.87	99.18	0.01	0.98	0.63	99.13
		Parietal region	SVM	98.30	97.84	98.80	98.91	97.63	98.37	0.01	0.97	0.63	98.32
		CO: Combo regional	SVM	98.13	98.10	98.16	98.41	97.80	98.26	0.02	0.96	0.62	98.13
		CP: Combo regional	NB	96.45	97.93	94.67	95.70	97.42	96.80	0.04	0.93	0.59	96.29
3	Fusion scheme-03 (feature fusion with hybrid features)	Full Ch	EBT	99.93	99.92	99.95	99.96	99.91	99.94	0.00	1.00	0.63	99.94
		Central region	EBT	100.00	100.00	100.00	100.00	100.00	100.00	0.00	1.00	0.64	100.00
		Occipital region	EBT/SVM/NB	99.57	99.21	100.00	100.00	99.09	99.60	0.00	0.99	0.63	99.60
		Parietal region	RF	98.95	98.91	99.00	99.12	98.75	99.01	0.01	0.98	0.63	98.95
		CO: Combo regional	EBT	98.67	98.56	98.79	98.95	98.35	98.76	0.01	0.97	0.62	98.68
		CP: Combo regional	EBT	97.02	97.73	96.19	96.79	97.30	97.26	0.03	0.94	0.60	96.96

Similarly, Table 12 revealed the effectiveness of same hybrid feature fusion method with subject dependent testing scheme (LOSO). Although, the performance of classifiers in LOSO is not matching the performance of the aforementioned scheme, still demonstrated promising results. It achieved a maximum accuracy of 94.13% in overall accuracy with random forest classifier, indicating its proficiency in predicting SZ patterns in EEG signal across all subjects and regions. Furthermore, it achieved an accuracy of 94.63% in region (central) specific accuracy using EBT classifier, showcasing its competence in classifying SZ using EEG data.

Table 12.

Results obtained using various feature fusion schemes in terms of performance indexes with subject dependent testing

Sr. No	Fusion scheme	Dataset	Best classifier	Acc. (%)	SEN (%)	SPE (%)	Pre (%)	NPV (%)	F1-Score(%)	FDR	MCC	Kappa	G-mean
1	Fusion scheme- 01 (feature fusion with handcrafted features)	Full Ch	EBT	92.41	92.64	92.17	92.64	92.17	92.64	0.07	0.85	0.59	92.40
		Central region	EBT	92.86	92.96	92.74	93.66	91.95	93.31	0.06	0.86	0.57	92.85
		Occipital region	EBT	91.67	90.00	93.59	94.19	89.02	92.05	0.06	0.83	0.57	91.78
		Parietal region	SVM	90.08	90.74	89.32	90.74	89.32	90.74	0.09	0.80	0.55	90.03
		CO: Combo regional	EBT	90.83	90.67	91.03	92.10	89.42	91.38	0.08	0.82	0.56	90.85
		CP: Combo regional	SVM	89.48	91.30	87.39	89.31	89.69	90.29	0.11	0.79	0.54	89.32
2	Fusion scheme- 02 (feature fusion with deep features)	Full Ch	EBT	92.40	90.23	94.87	95.26	89.47	92.68	0.05	0.85	0.58	92.52
		Central region	RF	92.61	90.27	95.21	95.43	89.83	92.78	0.05	0.85	0.59	92.71
		Occipital region	EBT	92.77	92.86	92.66	93.60	91.82	93.23	0.06	0.85	0.57	92.76
		Parietal region	EBT	90.63	90.91	90.30	91.40	89.76	91.15	0.09	0.81	0.56	90.61
		CO: Combo regional	EBT	91.33	91.77	90.81	92.06	90.48	91.92	0.08	0.83	0.56	91.29
		CP: Combo regional	RF	89.65	90.41	88.71	90.65	88.44	90.53	0.09	0.79	0.53	89.56
3	Fusion scheme- 03 (feature fusion with hybrid features)	Full Ch	RF	94.13	95.01	93.13	94.08	94.20	94.54	0.06	0.88	0.58	94.06
		Central region	EBT	94.63	93.63	95.76	96.16	92.98	94.88	0.04	0.89	0.60	94.69
		Occipital region	EBT	93.52	92.48	94.72	95.29	91.61	93.86	0.05	0.87	0.58	93.59
		Parietal region	EBT	91.24	93.44	88.72	90.47	92.19	91.93	0.10	0.82	0.55	91.05
		CO: Combo regional	EBT	92.02	91.25	92.90	93.70	90.18	92.46	0.06	0.84	0.57	92.07
		CP: Combo regional	EBT	90.72	92.22	88.95	90.75	90.67	91.48	0.09	0.81	0.55	90.57

Figure 6 illustrates the accuracy plot for the IMFs obtained from EMD and the sub-bands derived from EWT. The plot shows the performance of various ML classifiers on these different signal decomposition methods in terms of accuracy for SZ classification task. The accuracy plot of optimal multiresolution component features set derived from EMD and EWT for brain regional data, and combined brain regional data is illustrated in Fig. 7. It is evident from Fig. 7 that central-occipital combined brain region data has attained the maximum accuracy in both EMD and EWT groups. These regional and combinational regional data have demonstrated significantly impactful results in differentiating between SZ and HC using the extracted EEG signal based handcrafted features. The practical implications of these findings are noteworthy, if the transmission of regional and combinational regional data from a mobile wearable device is required for real-time monitoring or diagnosis, the reduced data size resulting from the emphasis on specific brain regions can result in cost savings in terms of bandwidth and power usage. The process of transmitting a smaller amount of data from the wearable device to a central processing unit has the potential to optimise data management and analysis, making it more feasible for remote or continuous monitoring of SZ or other neurological conditions.

Fig. 6.

Accuracy plot for: a IMFs of EMD, and b sub-bands of EWT

Fig. 7.

Accuracy plot of optimal multiresolution component feature set derived from EMD and EWT for: a Brain regional data, and b Combined brain regional data

At last, Fig. 8 depicts the accuracy plot of various feature fusion scheme with full channel EEG data, selective regional channel data, and selective combinational regional data using both subject independent tenfold cv testing strategy, and subject dependent (LOSO) testing strategy. The subject dependent testing may be better at simulating real-world classification results of unseen data in models involving analysis of EEG signals. Although, the performance of classifiers in LOSO is not matching the performance of the aforementioned scheme, still demonstrated promising results. It achieved a maximum accuracy of 94.13% in overall accuracy with random forest classifier, indicating its proficiency in predicting SZ patterns in EEG signal across all subjects and regions. Furthermore, it achieved an accuracy of 94.63% in region (central) specific accuracy using EBT classifier, showcasing its competence in classifying SZ using EEG data. Subject-dependent testing can be more representative of real-world classification scenarios. Although the performance of classifiers in LOSO testing may not match that of tenfold cv, it still shows promising results in SZ detection using EEG signals. The subject-dependent approach is advantageous as it better simulates scenarios where the model encounters unseen data from new subjects, mimicking real-world applications and generalization capability.

Fig. 8.

Accuracy plot of various feature fusion scheme with full channel EEG data, selective regional channel data, and selective combinational regional data using: a subject independent (tenfold cross validation) testing strategy, and b subject dependent (LOSO) testing strategy

Performance comparison with state-of-the-art models

The effectiveness of an algorithm may be gauged via performance assessment. The approach used in this work has attained the highest classification performance compared to all state-of-the-art models implemented on the schizophrenia EEG dataset of adolescents. The proposed method, which utilizes feature fusion scheme and on selected multiresolution EEG signal components features with ensemble learning approach, has achieved a maximum overall accuracy of 99.93 and 94.13% using tenfold cv and LOSO respectively. The majority of studies in the literature have primarily focused on subject independent testing. However, a notable exception is the study (Sharma and Joshi 2022) which investigated both subject independent and subject dependent testing approach. In most of the reported studies used 2D CNN architecture. In this process, EEG signals are transformed into time frequency representation such as spectrogram, scalogram etc. Although the 2D CNN has demonstrated effectiveness in detecting various neural activity patterns present in the frequency domain, there are certain aspects where 1D CNN or hybrid methods excel. Table 13 summarises the differences between the proposed method and the current literature.

Table 13.

Summary of performance comparison with other methods for SZ diagnosis using EEG signals on adolescent’s schizophrenia EEG dataset

Method (study and year)	Methodology for feature extraction	Classifier	Testing strategy	Overall performance (%)
				Acc	Sens	Spec
Naira & Del Alamo, (2019) (Naira and Del Alamo 2019)	Convert multichannel EEG 10 s. epochs into Pearson Correlation Coefficient heat map	2D- CNN	Subject independent	90%	90%	90%
Phang et al., (2019) (Phang et al. 2019)	Features of 2D time–frequency connectivity matrices	Multi-domain Connectome CNN	Subject independent	91.69%	97.78%	92.50%
Aslan & Akin, (2020) (Aslan and Akin 2020)	Take 5 Sec. EEG segment and apply short time Fourier transform (STFT) to create Spectrograms, and apply it to deep learning models	2D- CNN [Pre-trained Model (VGG-16)]	Subject independent	95%	95.35%	94.72%
Calhas et al., (2020) (Calhas et al. 2020)	Pair-wise distance learning relies on the spectral properties of EEG	Siamese neural network (SNN)	Subject independent	95%	98%	92.7%
Kutepov et al., (2020) (Kutepov et al. 2020)	Nonlinear features (Largest Lyapunov exponent and spectrum of exponents) extraction from EEG using different algorithms	Statistical learning	Subject independent	Not mentioned	Not mentioned	Not mentioned
Singh et al., (2021) (Singh et al. 2021)	Features (spectral amplitude, power, and Hjorth parameters) are extracted from processed EEG of 5 s. of each epochs	1D- CNN	Subject independent (tenfold cv)	94.08%	92.70%	95.31%
Khodabakhsh et al., (2021) (Khodabakhsh et al. 2021)	Raw EEG time series data is applied to deep models	U-Net deep learning model	Subject independent	94.11%	91.66%	100%
Aslan & Akin, (2022) (Aslan and Akin 2022)	5 Sec. EEG epochs and, apply CWT with Morlet to create Scalogram, and apply it to deep learning models	2D- CNN	Subject independent	98%	98%	98%
Supakar et al., (2022) (Supakar et al. 2022)	Raw EEG time series data is applied to deep models	LSTM deep learning model	fivefold cv	98%	98%	98%
Balasubramanian et al., (2022) (Krishnan et al. 2020)	Statistical, time-domain, frequency-domain, and spectral features are extracted from preprocessed EEG	Improved adaptive neuro-fuzzy inference system model	Subject independent	99.51%	–	–
Sobahi et al., (2022) (Sobahi et al. 2022)	1D local binary pattern (LBP) images are constructed from EEG rhythms	Hybrid learning (ELM Auto encoder + 2D-CNN)	Subject independent	97.7%	97.8%	97.7%
Sairamya et al., (2022) (Sairamya et al. 2022)	Signal is decomposed using discrete wavelet transform (DWT) and relaxed local neighbor difference pattern features are extracted	Artificial neural network	Subject independent	97.14%	–	–
Sharma & Joshi, (2022) (Sharma and Joshi 2022)	One epoch 0f 25 Sec. of EEG, extract 5 rhythms from EEG data with normalization. Apply processed EEG time series data to deep models	CNN-LSTM	Subject independent and Subject dependent	99.5 and 89.6%	99.4 and 88.43%	99.59 and 91.27%
Khare et al., (2023) (Reinertsen and Clifford 2023)	Apply Margenau-Hill time frequency (TF) transform to create TF images from EEG	2D-CNN	Subject independent	99.74%	99.64%	99.87%
Proposed Method	Multiresolution optimal components and feature fusion scheme	Hybrid learning (bottleneck DL features + EBT classifier)	Subject independent and Subject dependent	99.93 and 94.13%	99.92 and 95.01%	99.95%and 93.13%

Discussion

It is important to remember that schizophrenia is a chronic condition that cannot be cured but must be controlled. The periodic treatment of SZ places a significant financial and emotional strain on patient’s families and healthcare systems. Moreover, starting early treatment may improve long-term prognosis by reducing symptoms and preventing more catastrophic consequences. Therefore, it is necessary to devise an alternative method for the computer-aided diagnosis of schizophrenia in the early stages, which will enable psychiatrists to efficiently diagnose symptomatic patients for clinical practice. The proposed methodology has a number of positive sides including its simplicity, ease of implementation, accuracy, automated process, and robustness. In spite of the impressive performance of proposed method in SZ classification, this approach has certain drawbacks. Acknowledging the limitations and challenges of applying algorithms to psychosis is essential. However, the feature fusion-based hybrid learning model demonstrated remarkable accuracy compared to other deep-learning models. The proposed algorithms surpassed existing methods, even with a reduced number of electrodes or specific region. But, the performance of the proposed model is constrained in terms of the dataset. In this study, the diagnosis model is tested over only available adolescent EEG dataset of 84 adolescents (45 SZ and 39 normal healthy). Evaluating the model on a larger dataset with greater heterogeneity and diverse brain activity among participants would be crucial to achieve optimal performance.

In comparison with 2D CNN, 1D CNNs and hybrid methods are well-suited for neural data analysis tasks that involve analyzing temporal and frequency features. These methods can offer improved performance, efficiency, and interpretability in comparison to 2D CNNs, which are better suited for tasks where spatial features are more prominent, such as image data analysis. Although the choice of architecture depends on the specific characteristics and objectives of the neural data still there are certain aspects where 1D CNN or hybrid methods excel. These advantages include:

• 1D-CNNs are purposefully crafted to capture temporal dependencies in time-series data. As data is transformed in TFR image to apply in 2D-CNN, there is a potential drawback in terms of temporal information loss. This is due to the fact that the process of transforming the data into TFR naturally entails a certain level of data aggregation.

• Since 1D CNNs are often lighter weight model and contains fewer parameters than 2D CNNs, their training and inference processes are less computational resource-intensive. If EEG dataset is massive, 1D-CNN can be effective in terms of computational load, and execution time.

• The filters learned by 1D-CNNs in the temporal domain are often more interpretable compared to 2D-CNNs. This interpretability is valuable in EEG signal analysis for understanding underlying patterns and identifying significant features associated with distinct neurological conditions.

• 1D CNNs are capable of handling variable-length sequences, whereas 2D CNNs applied to TFR usually demand fixed-size input data, which might necessitate additional pre-processing steps.

• Transfer learning with 1D-CNNs can be more effective in extracting relevant features from EEG signals than employing pre-trained 2D-CNNs on unrelated image datasets.

• Data augmentation for EEG signals is typically more straightforward and intuitive compared to augmenting TFR images. As a result, applying data augmentation to 1D-CNNs can be more effective, leading to enhanced generalization performance.

Besides these advantages, the process of selecting hyperparameters for the deep learning feature extractor and additional feature selection required manual intervention, necessitating automation. In the future work, this limitation is going to be overcome with some technical add-ons. The long-term goal is to translate this model into a web-based application in response to useful comments from psychiatrists. In the immediate future, we are planning to deploy the developed automated computer-based diagnosis model at the cloud to diagnose the SZ class correctly and quickly. The EEG signals that are recorded by the dedicated EEG acquisition system from the patient would first be stored on a local web server inside the hospital before being transferred to the cloud, where our trained SZ detection artificial intelligence hybrid DL-ML based model is positioned. Upon completion of the diagnosis, the report is promptly transmitted from the cloud directly to the hospital and the specialized doctor. Consequently, the specialized doctor carefully reviews the report and creates a personalized treatment plan. This process ensures that the patient receives appropriate medical attention and medication in a timely manner.

Further in practical scenario, the modern healthcare business has seen expenditures totalling billions of dollars in recent years (Chernew and Mintz 2021). Large corporations are conducting research projects with the goals of lowering the cost of current diagnostic tools and concentrating on accurate early diagnosis, both of which are important tenets of the customised medicine philosophy (Richens et al. 2020). In personalised medicine, patients get the right treatment for the disease at the correct time and at the lowest possible cost. In recent time, smart healthcare systems have gained widespread attention with the advancement of the Internet of Medical Things (IoMT) (Gatouillat et al. 2018). It seeks to interconnect medical devices and instruments, enabling real-time collection of patient data. The continuous advancements in communication technologies, the proliferation of sensing devices, and a strong emphasis on providing swift medical services have all played significant roles in shaping and fostering the development of smart healthcare systems. The IoMT-based approach for customised medicine has the goal of identifying SZ patients at an earlier stage, significantly improving their quality of life and reducing treatment costs. Figure 9 highlights the envisaged Internet of Medical Things (IoMT)-based smart schizophrenia diagnostic system using EEG signals to aid medical practitioners in an efficient way. Thus, incorporating IoMT advancements into the diagnostic process could be a future agenda for extending this work.

Fig. 9.

Conceptual diagram of IoMT-based smart schizophrenia diagnostic system using EEG signals for future work

Conclusion

Schizophrenia is one of the severe mental illnesses that afflict a significant portion of the people all over the globe. It causes a decline in both academic and professional aspects of a person’s life. The periodic treatment of SZ places a significant financial and emotional strain on patients' families and healthcare systems. Therefore, early computer-assisted diagnosis of SZ can only benefit patients by enhancing their quality of life. This study proposes an automated decision-making model for the precise and effective diagnosis of SZ symptomatic subjects from full-channel and limited channel EEG signals. The EEG signals from the adolescent’s schizophrenia EEG dataset are decomposed in two different time–frequency domain analyses: EWT and EMD methods. Multi domain unique features from each sub-bands are extracted and applied to ML classifiers for optimal sub-band selection followed by feature reduction. Later in this study, various feature fusion scheme is proposed. Among all the feature fusion schemes, the hybrid fusion of bottleneck features and handcrafted features extracted from EEG signals using optimal sub-bands of EMD and EWT blocks achieved remarkable results. It attained a maximum overall accuracy of 99.93% and 94.13% using subject independent (tenfold cross validation) and subject dependent (Leave-One-Subject-Out: LOSO) testing strategy, indicating its excellent performance in predicting EEG signal patterns across all subjects and regions. Additionally, it achieved a perfect 100, and 94.63% region (central) specific accuracy using tenfold cv and LOSO, demonstrating its exceptional ability to accurately classify EEG signals using limited channel EEG data. EBT classifier is performing exceptionally well throughout the study. To conduct trials of the proposed classification model on clinically larger dataset and incorporating advanced security features is part of the future agenda. Furthermore, the immediate future research direction ahead of this work is to plan IoMT-based smart schizophrenia diagnostic system.

Data availability

The research is based solely on the analysis of publicly available data which is accessible at http://brain.bio.msu.ru/eeg_schizophrenia.htm.

Declarations

Conflict of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this article.

Ethical approval

This study did not involve the use of human participants or animals. The research is based solely on the analysis of publicly available data, and no new data were collected from humans or animals for the purposes of this study.

Informed consent

Not applicable.