ECGPsychNet: an optimized hybrid ensemble model for automatic detection of psychiatric disorders using ECG signals

Background. Psychiatric disorders such as schizophrenia ( SCZ ) , bipolar disorder ( BD ) , and depression ( DPR ) are some of the leading causes of disability and suicide worldwide. The signs and symptoms of SCZ, BD, and DPR vary dynamically and do not have uniform detection strategies. The main causes of delays in the detection of psychiatric disorders are negligence by immediate caregivers, varying symptoms, stigma, and limited availability of physiological signals. Motivation. The brain functionality in the patients with SCZ, BD, and DPR changes compared to the normal cognition population. The brain – heart interaction plays a crucial role in tracking the changes in cardiac activities during such disorders. Therefore, this paper explores the application of electrocardiogram ( ECG ) signals for the detection of three psychiatric ( SCZ, BD, and DPR ) disorders. Method. This paper develops ECGPsychNet an ensemble decomposition and classi ﬁ cation technique for the automated detection of SCZ, BD, and DPR using ECG signals. Three well-known decomposition techniques, empirical mode decomposition, variational mode decomposition, and tunable Q wavelet transform ( TQWT ) , are used to decompose the ECG signals into various subbands ( SBs ) . Various features are extracted from the different SBs and classi ﬁ ed using optimizable ensemble techniques using two validation techniques. Results. The developed ECGPsychNet has obtained the highest classi ﬁ cation accuracy of 98.15% using the features from the sixth SB of TQWT. Our proposed model has the highest detection rates of 98.96%, 96.04%, and 95.12% for SCZ, DPR, and BD. Conclusions. Our developed prototype is able to detect SCZ, DPR, and BD using ECG signals. However, the automated ECGPsychNet is ready to be tested with more datasets including different races and age groups.


Introduction
The term 'psychiatric disorders' refers to a wide variety of mental conditions that have an impact on a person's mood, behavior, thinking, and overall functioning.Some common types of psychiatric disorders include schizophrenia (SCZ), depression (DPR), bipolar disorder (BD), etc. Numerous factors, such as genetics, environmental factors, and life events, might contribute to the development of these illnesses.Schizophrenia is a chronic and severe mental health disorder that affects how a person thinks, feels, and behaves.It is a disorder that can last lifelong and can interfere with the ability to function in daily life.Schizophrenia affects approximately 1% of the population, and typically develops in early adulthood (Khare et al 2023).Three types of symptoms, namely, positive, negative, and cognitive symptoms, characterize schizophrenia (Khare et al 2023).Hallucinations, delusions, distorted thinking, and strange motions are examples of positive symptoms.Negative symptoms include social retreat, a lack of enthusiasm, and apathy.Individuals with schizophrenia may also experience cognitive symptoms, such as difficulties with memory, attention, and decision-making.The life expectancy of patients with SCZ decreases compared to the general population.In the last decade, the death rate has grown 2.6 times more than that in the healthy population (Tasci et al 2022).BD (manic-depressive illness) is a mental disorder affecting an individual's mood, energy, and ability to function.It is characterized by episodes of manic or hypomanic symptoms (highs) and depressive symptoms (lows).BD is a chronic condition that can be managed with proper care.Its causes are unknown, but it is thought to be a combination of genetic, environmental, and psychological factors (Tasci et al 2022).A family history of the disorder, chronic stress, substance abuse, and traumatic life events are the common risk factors for BD.Symptoms of depression include sadness, hopelessness, fatigue, changes in appetite or sleep patterns, and difficulty concentrating.DPR is a mood, thinking, and behavioral disorder.It is marked by despair, hopelessness, and a loss of interest in previously loved activities.It is a prevalent and serious disorder that affects millions of people around the world.A familial history of depression, a history of stressful life experiences, chronic illness, substance addiction, and hormonal abnormalities are all common risk factors for depression (Tasci et al 2022).
Patients with psychiatric disorders like SCZ, BD, or DPR have a greater rate of suicide than the general population does.Studies have shown that up to 90% of individuals who commit suicide have a diagnosable psychiatric disorder at the time of their death (Brådvik 2018).The alarming increase in psychiatric disorders, comorbidity, and rising rate of suicide among psychiatric patients demands timely care.It is crucial for individuals with psychiatric disorders to receive proper treatment and support to manage their symptoms and reduce their risk of suicide.Therefore, this demands an urgent need for the accurate and effective detection of psychiatric disorders.Over the years, many automated models have been developed for the detection of psychiatric disorders.A statistical model using a mixture of factor analysis developed using electroencephalogram (EEG) signals to detect major depressive disorder (MDD), BD, and SCZ (Khodayari-Rostamabad et al 2010).A multi-modal deep denoising auto-encoder model comprising audio, visual, and textual features as been proposed for BD and DPR detection (Zhang et al 2020).A generic model combining filtering, principle component analysis, extraction of linear and nonlinear features, and traditional machine learning (support vector machine (SVM), decision tree (DT), logistic regression (LR), and k-nearest neighbor (knn)) models have been used for the detection of psychological disorders including SCZ, epilepsy, and Parkinson's disease using EEG signals (Anwar et al 2021).In addition to this, several automated models have been suggested by researchers to separately detect these psychiatric disorders.Many traditional machinelearning and deep-learning models utilizing signal analysis methods have been proposed for the detection of SCZ using EEG signals (Khare and Bajaj 2021c, Khare et  • Difficulty in the acquisition and high probability of introduction of artifacts.
• Weak measure of some neural activities in the brain.
• Low signal-to-noise ratio compared to other physiological signals like electrocardiogram (ECG) signals.
• Analysis of EEG requires additional signal analysis steps.
One of the most crucial organs in the human body is the heart, and it is tightly connected to the brain through a variety of physiological and neurological processes (Koh et al 2022).The brain regulates and synchronizes all body processes, while the heart regulates the circulation of blood throughout the body, supplying oxygen and nutrients to organs and tissues (Acharya et al 2006).The heart and brain communicate with each other in a variety of ways.The details of which are given below.
• Autonomic nervous system (ANS): Heart functions such as heart rate (HR), blood pressure (BP), and blood flow (BF) are controlled by the ANS.BP and HR go up during the 'fight or flight' response, which is controlled by the sympathetic nervous system, whereas the same parameters go down during the 'rest and digest' response, which is controlled by the parasympathetic nervous system.The brain controls the ANS and can influence the heart's function through this pathway (Acharya et al 2006, Silvani et al 2016).
• Stress and emotions: Stress and emotions also affect the rhythm of heartbeats.As an example, when an individual feels anxiety or fear the brain orders the heart to speed up and increases BP (Sgoifo et al 2009).
Similarly, sustained stress can alter the heart's physiological properties over time, increasing the risk of heart disease (Acharya et al 2006, Montano et al 2009).
• Heart-brain feedback loop: In a feedback loop known as the 'heart-brain axis', the heart also interacts with the brain.Serotonin and oxytocin are two neurotransmitters and the hormones produced have an impact on mood and brain activity (Montano et al 2009, Koh et al 2022).This feedback loop may also impact the ANS, resulting in modifications to BP and HR.
• Heart rate variability (HRV): HRV is the variation of heartbeat with respect to time.It is affected by the ANS and might provide details about mental disorders.Research has revealed that meditation and relaxation can enhance HRV, reflecting a favorable effect on the heart and the brain (Acharya et al 2006, Silvani et al 2016).
In general, the relationship between the brain and the heart is intricate and crucial to human physiology.

Literature review
The previous section shows that various brain states like DPR and stress can reveal important changes in the heart's activities.Therefore, ECG, which measures the electrical activities of the heart can be used to detect different psychiatric disorders.Table 1 presents the existing methods used to study psychiatric disorders, including SCZ, DPR, and BD.
A study of the literature indicates that patients with SCZ have a higher prevalence of increased heart rate, suppressed HRV, and higher risk of sudden cardiac death compared to the general population (Chang et al 2010, Blom et al 2014, Hou et al 2015, Polcwiartek et al 2020, Haigh et al 2021).In addition, ECG can also be used for tracking mood changes in BD patients and can be used as an effective tool for the detection of BD and DPR.There exists only one method that has combined the detection of psychiatric disorders (SCZ, DPR, and BD) using ECG signals (Tasci et al 2022).This opens up a new path for the detection of psychiatric disorders using ECG signals.Also, there is a wide scope to explore ECG signals for the detection of psychiatric disorders like SCZ, DPR, and BD.This motivates us to explore ECG signals for the detection of SCZ, BD, and DPR.The main contributions of the proposed model are as follows: (1) Explored the possibilities of using ECG signals for the detection of psychiatric disorders.
(2) Proposed system presents the development of an ensemble of powerful nonlinear decomposition for the analysis of ECG signals.
(3) Developed system investigates the performance of individual ECG leads to get the best-performing ECG lead.
(4) Introduced an optimizable ensemble classifier model to accurately classify nonlinear features.
(5) System generated is robust as it is developed using holdout and tenfold cross-validation techniques with an optimal feature set.
In our proposed system (figure 1), first, the ECG signals are acquired.Then we performed empirical mode decomposition (EMD), variational mode decomposition (VMD), and tunable Q wavelet transform (TQWT) decomposition techniques on various ECG leads.Feature extraction is performed on the decomposed coefficients of the subbands (SBs).Then we performed the features ranking.These ranked features are fed to the classifier for the automated classification of three classes.The remainder of the paper is organized as follows: the methodology is covered in section 3, and the results are presented in section 4. The results are discussed in section 5, and the conclusion is presented in section 6.

Methodology
The proposed ECGPsychNet model comprises four steps, which are explained in the following sections.The steps involved in the proposed model are shown in figure 1.

Dataset
The dataset used to develop our ECGPsychNet model is taken from the publicly available Kaggle repository (Burak Tasci 2022).The dataset is composed of a total of 233 subjects belonging to four classes.The ECG beat signals corresponding to SCZ, DPR, and BD are acquired from hospitalized patients at Elazig Fethi Sekin City Hospital.The signals from each patient were acquired using a Philips ECG TC20 device at a 5 kHz sampling frequency from 12 leads.The ECG data is stratified into a single ECG beat for further analysis.For the HC ECG beats, the publicly available dataset Physionet has been used (Goldberger et al 2000).Table 2 shows the details about the subjects and total signals corresponding to each class.As the dataset is taken from a public repository,

EMD
EMD is a data-driven approach that works based on the local characteristic timescale of the data, which does not assume any linearity and stationarity conditions (Acharya et al 2014).It decomposes any nonlinear signal into a group of intrinsic mode functions (IMFs).The IMF is a combination of amplitude and frequency-modulated components satisfying certain criteria (Huang et al 1998).At first, it identifies local extrema and connects it with a cubic spline to form two envelopes (upper and lower).This difference in the envelope is called the first IMFdenoting high-frequency component.This method is repeated until the residue is obtained or the stopping criteria are met.The resulting IMFs denote the bandpass frequency component of the original signal.The mathematical equation of EMD is written as where N denotes the number of IMFs, s(t) is the original signal, r N (t) is the residue, and f n (t) are the generated IMFs.
where   is a Lagrangian multiplier, α penalty factor, y l ¯is lth mode, and z t ¯( ) is the original signal.The relation between the original signals and generated modes is mathematically represented as (Dragomiretskiy and Zosso 2013, Khare and Bajaj 2021a) Decomposition using VMD provides better stability, accuracy, frequency resolution, and flexibility.

TQWT
TQWT is another powerful signal analysis tool that uses wavelet transform with variable quality factors.The fine-tuning of this quality factor helps control the frequency bandwidth of the wavelet for better frequency selection.It is a two-stage filter bank comprising the analysis stage or decomposition and synthesis or reconstruction stage.A signal can be decomposed into K 1 ¯+ SBs with a quality factor of (Q ¯), redundancy rate (R ¯), and decomposition levels (K ¯).The quality factor and redundancy rate are controlled by low and high pass scaling that controls the low-and high-frequency contents of the signals.The low and high pass frequency responses are generated to capture the respective frequency contents of the signal.The mathematical representation of the parameters of the TQWT is given in table 3 (Selesnick 2011, Khare et al 2020).

Features
We have extracted 42 statistical and nonlinear features from the decomposed coefficients of three methods.These features are the Higuchi fractal dimension, Hjorth parameters, Hurst exponent, average amplitude change, average energy, cardinality, the difference of absolute mean and standard deviation, kurtosis, Teager energy, mean absolute value of the square root, second difference, simple squared integral, skewness, Vth order, waveform length, Willison amplitude, variance, maxima, median, mean absolute value, mean absolute deviation, minima, peak, minimum deviation from mean, the maximum deviation from the mean, quartiles, temporal moments, modified mean absolute values, root mean square, mean, mean curve length, and spectral flatness (Khare and Acharya 2023).Once the feature set is obtained, the feature matrix of 42 features of each ECG lead is fed to the classifier.

Scaling factor
Tuning parameters Frequency response Where H p and L p are high and low pass scaling factors, and X 0 and X 1 are low and high pass frequency responses.

Optimizable ensemble classifiers
A single classifier algorithm may not always provide desired results in varying feature sets.Therefore, in our model, we have used an ensemble of classifiers to obtain the best set of classifiers for different feature sets (Dietterich 2000).This ensemble classifier model is composed of different variants of classification techniques, including a bagged tree, boosted tree, a discriminant tree, random under-sampling boosted trees, and subspace knn classifiers.These classifiers divide the feature set into subspace datasets on which these classifiers are trained.The final decision is obtained after getting the best-performing variant.The working of the ensemble classification techniques is shown in figure 1.However, a mere ensemble of these classifiers may not yield the best performance, as the performance is hyperparameter dependent (Thornton et al 2013).Tuning these hyperparameters of the ensemble classifier is timely and requires intense tuning.Even after several attempts of manual hyperparameter tuning, the desired results may not be attained (Khare and Bajaj 2021c).Therefore, in this paper, an optimizable classification technique has been employed to select the best-ensembled classifier using the Matlab classifier learner application.

ReliefF feature selection
ReliefF is a popular feature selection technique used for extracting relevant features from large (Urbanowicz et al 2018).The features are ranked by the algorithm based on their value or relevance to the target variable (Urbanowicz et al 2018).It works by determining the quality of features based on their ability to distinguish between instances with comparable target variable values (Urbanowicz et al 2018).It takes an instance at random from the dataset and looks for its nearest neighbors.These neighbors are made up of instances with the same and distinct target values.ReliefF computes the weight of each feature based on the variations between the feature values of current instances with its nearest neighbors.It provides more weight to features that exhibit significant variances between instances with the same and different targets.The weights for every feature are adjusted depending on the differences between the nearest neighbors that have the same target value versus those that have different target values.The method is repeated for each occurrence in the dataset, and the feature weights are gathered.Finally, the features are ordered according to their total weights, and a subset of the top-ranked features is chosen as the final feature set.

Performance evaluation
Performance evaluation of a model plays a crucial role in determining the outcome on various scales (Luque et al 2019).In our proposed model we have evaluated our model using holdout and tenfold cross-validation (TFCV) techniques.During holdout validation, 80% of the data is utilized for training and the remaining for testing.The holdout validation is repeated ten times and the final results are obtained by averaging the ten results.In TFCV, the entire dataset is divided into ten equal sets, nine sets are used for training and one for testing, which is repeated ten times to get the final output.In addition to this, we have evaluated the system performance by evaluating various performance matrices.These performance matrices are accuracy, specificity (SPF), recall, precision (PPV), F1 score, false positive rate (FPV), prevalence threshold (PT), and critical success index (CSI).

Results
This section presents a comprehensive performance analysis of the developed ECGPsychNet model.The ECG beats from a total of 233 subjects, including SCZ, HC, DPR, and BD, were fed to the ensemble of decomposition models.The ensembled decomposition model consists of the EMD, VMD, and TQWT-based techniques.Decomposition of ECG beats has been performed using the same experimental analysis.We have chosen the quality factor of 2, redundancy rate of 3, and level of decomposition to 5 for TQWT, while for VMD, the quadratic penalty factor of 2000, number of modes to 5, time step of the dual ascent to 0, DC to 0, and omegas are initialized from zero frequency, and tolerance of 1 × 10 −7 to obtain the optimum performance.For EMD, no parameters are required to be selected as it is a data-driven approach.It is important to note that modes for VMD are five, six SBs are obtained for TQWT, while, for EMD, we have selected two IMFs (a minimum of two IMFs obtained for each class) to maintain uniformity in our analysis.Forty-two features were extracted from the SBs of each decomposition technique.These lead-wise feature matrices of 42 features were fed to ensemble classifiers.The classification was performed using holdout and TFCV techniques.For each feature matrix, we have run an optimizable ensemble of classifiers to get the best classification accuracy of each multicomponent of decomposition techniques.A total of 30 iterations have been used to get the performance of each lead.Tables 4-6 present the lead-wise accuracy obtained using an optimizable ensemble classifier for the multicomponents of the EMD, VMD, and TQWT features.It may be noted from table 4 that ECG leads V2 and V4 have generated the most discriminable features in IMF-1, yielding the highest classification accuracy.The features of IMF-2 can be overlapping, generating the accuracy obtained for each lead that is less compared to IMF-1.The highest accuracies of 95.66% and 95.18% have been obtained for ECG leads V2 and V4 using holdout and TFCV techniques, respectively.Table 5 indicates that the third mode of VMD has generated the highest accuracy for the holdout and TFCV techniques using the ECG V2 lead.The highest accuracies of 94.68% and 93.95% have been obtained for the holdout and TFCV techniques, respectively.Similarly, a comparison of the accuracies of TQWT SBs shown in table 6 reveals that SB-6 has obtained the highest capability to generate the most representative features.The report shows that the highest accuracies of 98.6% and 98.15% have been obtained using the ECG V2 lead with the holdout and TFCV techniques.
The illustrative training and testing curves of misclassification errors obtained for the EMD, VMD, and TQWT features are shown in figure 3. Since an optimizable ensemble classifier has been used to evaluate the model, the hyperparameters obtained for each case are chosen automatically for the best-case scenario.For EMD-based feature classification, the adaptive boosting (AdaBoost) ensemble method having a learning rate of 0.96997, maximum splits of 102, and 448 learners are the optimal parameter settings.Similarly, for VMD-based feature classification, the random under-sampling boosting (RUSBoost) ensemble method with 591 splits, 473 learners, and a learning rate of 0.58815 have been obtained as optimal tuning parameters for classification.The classification using TQWT-based features employed a bagged ensemble method with 477 learners, maximum splits of 220, and 19 predictors to sample are the optimal parameters to get the best classification performance.Thus, by selecting the best parameters for classification based on the nature and type of feature extractor, the optimizable ensemble model provides the highest classification performance for psychiatric disorder detection.The considerable performance can be obtained with the minimum number of features (Khare and Acharya 2023).The statistical analysis of features not only reduces the input dimensionality of the feature matrix to the classifier but also reduces training time.Therefore, we have used a statistical feature analysis of our developed model.It can be noted from tables 4-6 that ECG lead V2 is the most discriminable.Also, the features extracted from TQWT have produced the highest classification accuracy using SB-6.Therefore, the feature matrix from ECG lead V2 in SB-6 has been used for the optimal feature selection.We have used the reliefF feature ranking method with the four nearest neighbors.The feature importance scores obtained using the reliefF technique are shown in figure 4. Based on the rank of the feature, we have arranged the features in their decreasing feature rank starting from two features.The variation in the accuracy with increasing feature combinations is shown in figure 5.The analysis shows that the highest accuracy of 95.85% is obtained with 13 features.It may be noted that the model's accuracy rises up to 13 features, after which it somewhat declines but stays constant up to 15 features.
To get insight into our developed ECGPsychNet model, we have used different performance measures to evaluate the best-performing lead and multicomponent of EMD, VMD, and TQWT using the holdout and TFCV techniques.Table 7 represents the performance in terms of recall, SPF, PPV, F1 score, FPR, PT, and CSI.It may be noted that 100% recall is obtained for HC ECG beats using the overall and optimal feature matrix of TQWT decomposition with the holdout and TFCV techniques.The HC class, followed by DPR, SCZ, and BD  have the highest performance matrices.The evaluation confirms that our developed model has obtained the highest performance parameters for different validation techniques for the HC, SCZ, DPR, and BD classes.The class-wise accuracy obtained for each class using EMD-, VMD-, and TQWT-based features classified with the TFCV technique is shown in figure 6.The class-wise accuracy obtained for EMD using holdout and TFCV is shown in figures 6(a) and (b), respectively.The accuracy obtained for the SCZ, BD, DPR, and HC classes using VMD and TQWT with holdout and TFCV for the ECG V2 lead is shown in figures 6(c)-(f), respectively.Finally, the sample distribution and accuracy for optimal features of the ECG V2 lead using the TQWT method are shown in figure 6(g).The distribution shows that our developed model is consistent in detecting HC accurately, followed by SCZ, with a mixed occurrence of DPR and BD.

Discussion
We have compared the performance of our developed ECGPsychNet model with available state-of-the-art methods developed using an ECG dataset as shown in table 8.Only one group has explored the detection of psychiatric disorders on the same dataset.Tasci et al (2022) used MDWT coupled with a ternary pattern for automated detection of psychiatric disorders.The features extracted from the SBs of MDWT using a ternary pattern have been classified using an ANN classifier.Their developed model has obtained an accuracy of 96.25% using an iterative majority voting scheme.The comparison of the performances shows that our developed model has yielded an accuracy of 2% more than that of Tasci et al (2022).   .Hence, we could classify the SCZ, BD, and DPR classes with high classification performance.Our study was carried out on a small dataset; however, this can pave the way for more studies on mental disorders using ECG signals.Also, ECG-based detection is advantageous as it is easy to acquire, has a portable system, and has a high signal-to-noise ratio.Our developed ECGPsychNet model has the following merits: • Explored a new ECG-based signaling model for the detection of neurological conditions.
• Model is robust as validated using holdout and tenfold cross-validation techniques.
• Obtained the highest performance using a single V2 ECG lead.
• Obtained the highest accuracy of more than 98% in classifying three classes using ECG signals.
The limitations of our developed ECGPsychNet are given below: • Model is developed using a small dataset.
• Used ECG signals from a single center.
• Leave one subject out cross validation was not done.
In the future, we will explore the possibility of using more ECG signals from different centers and age groups to validate our model.Also, with more datasets, we intend to explore the possibility of developing new deeplearning models to handle huge datasets while obtaining accurate results.We plan to employ explainable artificial intelligence to our model to increase trust in clinical experts and caretakers.Our developed model has used engineering features for the decision-making of psychiatric disorders.In the future, we will develop a more robust model using a hybrid combination of engineering and clinical features for real-time psychiatric disorder detection.

Conclusion
Psychiatric disorders are increasing at an alarming rate, especially SCZ, DPR, and BD.EEG-based automated detection systems are more frequently used for detecting such conditions.But EEG signals are prone to artifacts and suffer a low signal-to-noise ratio.Therefore, we have attempted to detect psychiatric disorders using ECG The ECGPsychNet model has also demonstrated that considerable performance can be obtained by 13 features arranged by their statistical rank.Performance analysis of ECGPsychNet reveals that our model has yielded the highest classification rates of 98.6% and 98.15% using holdout and tenfold cross-validation techniques, respectively.The analysis of ECGPsychNet indicates that electrical activities of the heart play a crucial role during psychiatric disorders like SCZ, DPR, and BD, strengthening the theory of the brain-heart interaction.Our developed ECGPsychNet is one of the first attempts to present an ensemble and optimized feature extraction and classification technique for the detection of multi-class psychiatric disorders.The limitation of this study is that we have developed the model using a small dataset.In the future, we plan to strengthen the evidence of brain-heart interactions reported in this study using a huge diverse dataset.Also, we intend to explore the possibility of classifying more classes of psychiatric disorders with our developed model.
al 2021, Lai et al 2021, Khare and Bajaj 2022, Baygin et al 2023, Khare et al 2023, Kumar et al 2023, Sharma et al 2023).Many models have also been presented for the detection of bipolar disorders and depression using EEG signals (Acharya et al 2018, Newson and Thiagarajan 2019, Teixeira et al 2019, Sharma et al 2021, Campos-Ugaz et al 2023).EEG signals may not always be viable for the detection of psychiatric disorders due to certain limitations.Some of the major challenges in EEG-based decision-making models are (Koh et al 2022), Richter et al 2022): • Acquisition of signals from a large number of electrodes.

Figure 1 .
Figure 1.Block representation of the proposed ECGPsychNet model.

Figure 2 .
Figure 2. Sample ECG signals used for various leads of SCZ, DPR, BD, and HC subjects.

Figure 4 .
Figure 4. Feature rank obtained for ECG lead V2 using the sixth SB of TQWT with the reliefF feature ranking method.

Figure 5 .
Figure 5. Variation of accuracy (%) vs. the number of features using the TQWT method.
The EEG signals in patients with psychiatric disorders vary from the healthy population (Khodayari-Rostamabad et al 2010, Zhang et al 2020).Changes in the brain during psychiatric disorders are also encoded in the ECG signals due to brain-heart interactions (Acharya et al 2014, Koh et al 2022, Tasci et al 2022).Hence, our study confirms the use of ECG signals for the automated detection of psychiatric disorders and the brain-heart interaction.Our study confirms that the changes that occur in the brain are faithfully reflected in the ECG signals.In addition, the nature of ECG signals is rhythmic due to which the changes in cardiac activities may be captured during SCZ, BD, ADHD, and DPR (Koh et al 2022, Tasci et al 2022).Also, the high signal-to-noise ratio

Figure 6 .
Figure 6.Confusion matrices obtained using an optimized ensemble model for various decomposition and validation techniques.
(Acharya et al 2006, Koh et al 2022ct with one another can help to detect DPR, anxiety, and other neurological conditions accurately and enhance mental health(Acharya et al 2006, Koh et al 2022, Khare et al 2023).

Table 1 .
Summary of state-of-the-art techniques for the detection of psychiatric disorders using ECG signals.
SCZ SCZ: 12, HC: 29, and ASD: 25 ECG Time, frequency, and nonlinear analysis using HRV features Statistical analysis Suppressed HRV with worse performance on neuropsychological tests of cognition in the SCZ group Valenza et al (2016) BD (hypomania, euthymania, and DPR) BD: 14 ECG Time, frequency, and nonlinear analysis using HRV features SVM ECG can be used as a mood forecasting tool in patients with BD with an accuracy of 83.3% Valenza et al (2014) BD (hypomania, euthymania, and DPR) BD: 15 ECG Mono-variate and multivariate feature analysis SVM ECG can be used as a mood forecasting tool in patients with BD with an accuracy of 95.81% Psychiatric disorders BP: 62, DPR: 17, SCZ: 119, and HC: 35 ECG Multilevel discrete wavelet transform (MDWT) and ternary pattern Artificial neural network (ANN) Performed single lead and iterative majority voting to obtain an accuracy of 96.25% HC: Health control.

Table 2 .
Khare and Bajaj (2021a)o (2013) used in this work.VMD decomposes the signal into a fixed number of modes that have varying frequency contents.It requires tuning parameters, including tolerance, number of modes, initialization, and quadratic penalty factor.The modes in VMD are obtained by optimizing the parameters for the defined frequency and amplitude of the modes.For this, a variational optimization problem is formulated to minimize the cost, which penalizes the deviation from the smooth frequency content.The constraint optimization helps in converting it to an unconstrained problem by using a penalty factor and a Lagrangian multiplier.The formulation of unconstrained optimization is given byDragomiretskiy and Zosso (2013)andKhare and Bajaj (2021a)as

Table 3 .
Tuning parameters and frequency response details of TQWT.

Table 5 .
Accuracy (%) obtained using different modes of VMD with an optimizable ensemble classifier.

Table 7 .
Performance parameters (%) obtained using our developed ensemble model using different feature extraction and validation techniques.

Table 8 .
Comparison of the performance (%) of the ECGPsychNet with the existing state of the art using the same dataset.by our developed ECGPsychNet model.The features extracted from TQWT-based decomposition have yielded the highest classification performance in classifying the SCZ, BP, and DPR classes using the V2 ECG lead. signals