1 Introduction

Sadness is one of the moods that any person can experience and is not a sign of disorder if it lasts for a short duration. However, prolonged sadness is one of the most obvious symptoms of depression. Sadness that lasts for at least two weeks is referred to as major depressive disorder (MDD) [1,2,3]. About 300 million people worldwide has suffer from MDD [4]. In addition, the number of female MDD patients is approximately 1.5–2 times higher than male MDD patients [5]. The widely seen symptoms of MDD are low self-esteem, sadness, tearfulness, burst of rage, suicidal tendency and hallucinations. Therefore, it will affect their social lives and hence, they cannot interact with other people easily [6, 7].

MDD is a mental disorder and it should be diagnosed by specialist physicians. It can be treated with therapy and medication (drug treatment). The commonly used therapies are cognitive behavioral therapy, electroconvulsive therapy, and interpersonal therapy. Drug treatment is used to prevent unwanted behaviors in advanced MDD cases [8, 9].

Usually it is very difficult to detect MDD manually in our daily life and also patients suffering from this disorder are reluctant to seek treatment due to social stigma. Therefore, early diagnosis and treatment of MDD is most important. Hamilton Depression Rating Scale and Beck Depression Inventory have been developed as a manual for the diagnosis of MDD and generally professionals use questionnaires to patients for diagnosis [6]. There is possibility of subjectivity in this and also patients may lie. Therefore, the most reliable way of MDD diagnosis is electroencephalogram (EEG) based diagnosis. The EEG signals consist of information about the neuronal activities of the brain. These EEG signals have the signatures of activities of the brain [10]. Using novel features extraction techniques the hidden information about the MDD can be obtained from the EEG signals [11]. However, it difficult to obtain salient information from the EEG signals as they are nonlinear and non-stationary in nature. Therefore, nonlinear features extraction methods coupled with machine learning techniques need to be employed to obtain high detection performances [12, 13]. Many machine learning techniques have widely used for automated detection of diseases and assist the clinicians [14,15,16,17].

In this study, we have developed MDD detection from EEG signals using machine learning techniques. We have used 34 MDD and 30 healthy subjects to develop the automated system. A new melamine pattern is proposed to generate the features from the EEG signals. These features are used to for the detection of MDD.

1.1 Literature review

In this section, depression studies conducted using speech, voice and facial biomarkers is provided Table 1. This table clearly shows that, authors did not obtain high detection accuracy using speech, voice and facial biomarkers as they are unable to extract unique features for normal and depression patients. Hence, many authors have used EEG signals in their studies and obtained higher detection performances [11].

Table 1 Summary of studies done detecting depression using speech, voice and facial biomarkers
Full size table

1.2 Motivation and our model

Table 1 shows the summary of studies conducted to develop the MDD detection systems using speech, voice and facial biomarkers. However, authors did not obtain higher detection accuracies. Therefore, in this work, we are proposing to develop an automated MDD detection model using EEG signals. In this work, we have used 20 channel EEG signals with 7339 EEG signals of 10 s duration in each channel. For this big dataset, many deep learning models have been used [36,37,38,39]. Generally, deep models have high computational complexities and take longer time to train. For instance, ten-millions parameters should be set in the convolutional neural networks such as residual network, dense network, and Google network [40,41,42]. Therefore, in this work, we have used hand-crafted features based classification model. This model includes feature generation, feature selection and classification phases. Our proposed melamine pattern and NCA- based depression detection model is shown in Fig. 1.

Fig. 1
figure 1

Proposed melamine pattern and NCA based depression detection model. The collected EEG signals are loaded to system. Multilevel DWT, melamine pattern and statistical features are used to extract both low-level and high-level features. Then, NCA chooses the top 256 features which are used to classify with kNN or SVM classifiers

Full size image

The novelty of this work is the use of melamine pattern (it is in the feature generation box). It uses molecular structure (molecular graph) of the melamine and used molecular graph called ChemSpider. Our aim is to show the superiority of molecular structured feature generation model. To gain high classification accuracies, features are extracted from both low and high frequency components of discrete wavelet transform (DWT) [43]. The neighborhood component analysis (NCA) is employed on the generated features and 256 features [44] are selected. These features are fed to support vector machine (SVM) and k nearest neighbor (kNN) [45, 46] for automated classification. Our classifiers are developed using hold-out validation (80:20) strategy.

1.3 Key contributions

Key contributions of the proposed melamine pattern and NCA- based model are given below:

  • A new molecular structure based feature generation model is presented.

  • Proposed model is accurate and robust as we have obtained classification accuracy of more than 95% for all channels.

2 Materials

Mumtaz [47] collected EEG dataset in 2016 to develop an automated MDD. The EEG signals were collected from 64 subjects between the ages of 12 and 77 years (average age = 20.54 years). Among the collected 64 subjects (30 healthy and 34 MDD patients), 40 of them were men and 24 were women. The EEG signals were collected between 23.06.2011 and 30.06.2013. In this dataset, the EEG signals were sampled at 256 Hz and segmented into 10-s length. Therefore, the length of each EEG signal is 2560 samples. This corpus includes 20 channels with 7339 (3893 normal and 3446 MDD) EEG signals in each channel. The twenty channels used are A2-A1, C3, C4, Cz, F3, F4, F7, F8, Fp1, Fp2, Fz, O1, O2, P3, P4, Pz, T3, T4, T5 and T6. The sample normal and MDD EEG signals are shown in Fig. 2.

Fig. 2
figure 2

Sample EEG signal: (a) normal, and (b) depression

Full size image

3 The presented melamine pattern

Molecule structure are accepted as DNA of the materials and can be used to identify the materials. Moreover, these structures are expressed using graphs. Therefore, machine learning model (especially deep learning models) are used in molecule shapes to reach high performance using a nature-inspired model [48]. However, there is no molecule shape for hand-designed feature generator. Hence, a molecule shape based feature generator called Melamine (the molecular structure of melamine is widely known and it is named ChemSpider) pattern is presented.

Hand-crafted feature generation methodology uses various techniques to generate the most relevant features from the signals. Histogram based feature generation and local binary pattern (LBP) are the two widely used feature generators [44, 49]. However, LBP has the following limitations [50,51,52].

  • It uses a linear pattern and by employing this pattern, few important features cannot be detected.

  • It employs signum function as kernel function which only compares the local values. This may result in losing few important features.

In order to overcome these drawbacks, we proposed two solutions. They are listed below:

  • We are proposing to use non-linear patterns for feature generation. In this work, molecular structure of melamine is used as pattern. The shape of molecular structure is called ChemSpider which has the ability to generate unique features.

  • Signum function is a good solution for feature generation but it is not able to solve some problems. Hence, we have used three nonlinear kernels.

Steps used to obtain the features from melamine pattern is given below.

  • Step 1: Divide EEG signal into 25 sized overlapping windows.

$$ window(j)= ES\left(i+j-1\right),i\in \left\{1,2,\dots, Ln-24\right\},j\in \left\{1,2,\dots, 25\right\} $$
(1)

Equation (1) defines overlapping block division. Where window is 25 sized overlapping block, Ln is length of EEG signal (ES).

  • Step 2: Apply vector to matrix transformation and obtain 5 × 5 sized matrix.

$$ m\left(k,l\right)= window(j),k\in \left\{1,2,\dots, 5\right\},l\in \left\{1,2,\dots, 5\right\} $$
(2)

where m is obtained 5 × 5 sized matrix.

  • Step 3: Create a pattern using ChemSpider. The typical sketch of ChemSpider is shown in Fig. 3.

    Fig. 3
    figure 3

    Molecular structure of melamine (ChemSpider)

    Full size image

Inspired by ChemSpider, a new pattern is presented and graphical sketch of our presented pattern is shown in Fig. 4.

Fig. 4
figure 4

Melamine pattern created using ChemSpider. The digits represent enumeration of relations. Kernels take two parameters: initial point of each arrow indicates the first parameter and final point of each arrow shows the second parameter of used kernel

Full size image

  • Step 4: Use three kernels for calculating binary features. The used kernels are defined in Eqs. 3–5.

$$ {k}^1\left( para{m}^1, para{m}^2\right)=\left\{\begin{array}{c}0, para{m}^1- para{m}^2<0\\ {}1, para{m}^1- para{m}^2\ge 0\end{array}\right. $$
(3)
$$ {k}^2\left( para{m}^1, para{m}^2\right)=\left\{\begin{array}{c}0, para{m}^1- para{m}^2>- para{m}^1\\ {}1, para{m}^1- para{m}^2\le - para{m}^1\end{array}\right. $$
(4)
$$ {k}^3\left( para{m}^1, para{m}^2\right)=\left\{\begin{array}{c}0, para{m}^1- para{m}^2< para{m}^2\\ {}1, para{m}^1- para{m}^2\ge para{m}^2\end{array}\right. $$
(5)

The used kernels are defines as k1(., .), k2(., .) and k3(., .). As seen from Eq. (3)–(5), k1(., .) is signum function. k2(., .) and k3(., .) kernels extract lower and upper signals. Each kernel generates nine bits deploying the presented melamine pattern.

  • Step 5: Extract three binary feature vectors using melamine pattern and the used three kernels. The mathematical explanation of the bit generation process using the defined three kernel and the presented melamine pattern is given in eqs. (6).

$$ \left[\begin{array}{c} bi{t}^h(1)\\ {} bi{t}^h(2)\\ {} bi{t}^h(3)\\ {} bi{t}^h(4)\\ {} bi{t}^h(5)\\ {} bi{t}^h(6)\\ {} bi{t}^h(7)\\ {} bi{t}^h(8)\\ {} bi{t}^h(9)\end{array}\right]={k}^h\left(\left[\begin{array}{c}m\left(2,3\right),m\left(3,4\right)\\ {}m\left(3,4\right),m\left(4,4\right)\\ {}m\left(4,4\right),m\left(5,3\right)\\ {}m\left(5,3\right),m\left(4,2\right)\\ {}m\left(4,2\right),m\left(3,2\right)\\ {}m\left(3,2\right),m\left(2,3\right)\\ {}m\left(2,3\right),m\left(1,3\right)\\ {}m\left(4,4\right),m\left(5,5\right)\\ {}m\left(4,2\right),\mathrm{m}\left(5,1\right)\end{array}\right]\right),h\in \left\{1,2,3\right\} $$
(6)

where bit1, bit2and bit3 are generated bits using k1(., .), k2(., .) and k3(., .) respectively. Equation (6) denotes the bit generation processes. Each kernel generates nine bits using melamine pattern. For instance, the first bit of the upper signal (bitupper(1)) is generated using k3(m(2, 3), m(3, 4)).

  • Step 6: Binary to decimal conversion is applied on the generated bits.

$$ SS(i)=\sum \limits_{j=1}^9 bi{t}^1(j)\ast {2}^{j-1},\kern0.75em i\in \left\{1,2,\dots, Ln-24\right\} $$
(7)
$$ SL(i)=\sum \limits_{j=1}^9 bi{t}^1(j)\ast {2}^{j-1} $$
(8)
$$ SU(i)=\sum \limits_{j=1}^9 bi{t}^1(j)\ast {2}^{j-1} $$
(9)

where SS, SL and SU represent signum signal, lower signal and upper signal consecutively.

  • Step 7: Extract histograms of these signals. Each signal coded with 9-bits. Therefore, the size of generated each histogram is 29 = 512.

  • Step 8: Merge the generated histograms and calculate feature vectors of length = 1536.

The given eight steps define our melamine pattern feature generator.

4 Proposed method

This work proposes a hand-crafted feature for automated detection of MDD using EEG signals. Therefore, a signal processing model is presented and this model has three main steps. In the first step, the recommended melamine pattern and DWT are used to extract the features. 25 statistical features are extracted from the presented model. The proposed feature generator extracts two types features. Melamine pattern generates textural features and the 25 statistical moments are extracted from statistical features. The six-leveled DWT is performed using Daubechies four (db4) mother wavelet function [53, 54]. The generated features are fed to NCA [55] and the most discriminative 256 features are selected. The selected features are classified using SVM and kNN classifiers developed using 80:20 hold-out cross validation strategy. The snapshot of the proposed automated major depression detection model is shown in Fig. 5.

Fig. 5
figure 5

Snapshot of the proposed automated major depression detection model

Full size image

Moreover, pseudocode of the introduced Melamine pattern and NCA based MDD classification model is shown in Algorithm 1.

Algorithm 1.

Pseudocode of the recommended melamine pattern and NCA based MDD model.

figure a

The steps of the presented melamine pattern and NCA-based MDD model is given below.

4.1 Feature generation

In each level of DWT, statistical feature generator and the presented melamine pattern are utilized together. The primary objective of this step is to extract features of both low and high levels. The DWT generates both levels (low and high frequencies), statistical feature generator extracts statistical features and proposed melamine pattern extracts textural features from the DWT coefficients. The steps involved to perform the features extraction are given below:

  • Step 0: Load EEG signal.

  • Step 1: Apply six-level DWT on the EEG signal.

$$ NL=\left\lfloor {\log}_2\left(\frac{Ln}{25}\right)\right\rfloor $$
(10)

Equation (10) defines the number of DWT levels. Melamine pattern uses 25 sized overlapping signal and length of the EEG signal used is 2560. Therefore, \( NL=\left\lfloor {\log}_2\left(\frac{2560}{25}\right)\right\rfloor =6 \) is calculated as number of levels.

$$ \left[ lo{w}^1, hig{h}^1\right]= DWT\left( ES, db4\right) $$
(11)
$$ \left[ lo{w}^k, hig{h}^k\right]= DWT\left( lo{w}^{k-1}, db4\right),k\in \left\{2,3,\dots, 6\right\} $$
(12)

where lowk and highk are kth low-pass and high-pass filters coefficients, DWT(., .) function represent one dimensional DWT.

  • Step 2: Generates textural features from EEG signal and features are generated using melamine patterns using low-pass coefficients of DWT.

$$ {tf}^1= MP(ES) $$
(13)
$$ t{f}^{t+1}= MP\left( lo{w}^t\right),t\in \left\{1,2,\dots, 6\right\} $$
(14)

where tft is tth level textural feature vector and MP(.) defines the presented melamine pattern and it generates 1536 features.

  • Step 3: Extract 25 statistical features from DWT coefficients.

$$ {sf}^1= St(ES) $$
(15)
$$ s{f}^{t+1}= St\left( lo{w}^t\right),t\in \left\{1,2,\dots, 6\right\} $$
(16)

where sft is tth level statistical features, St(.) is used statistical feature generation function and uses 25 statistical moments. Therefore, the length of each sf is 25. The used statistical moments are given in Table 2 [56].

Table 2 Details of 25 statistical moments extracted
Full size table

These moments are consisted of the St(.) Feature generation function.

  • Step 4: Merge the generated statistical and textural features.

$$ X\left(1561\ast k+j\right)= conc\left(t{f}^{k+1},s{f}^{k+1}\right),k\in \left\{0,1,\dots, 6\right\},j\in \left\{1,2,\dots, 1561\right\}\kern0.5em $$
(17)

where X defines the generated and merged statistical and textural features with a length of 1561 ∗ 7 = 10927.

4.2 Feature selection

Feature selection is one of the most critical steps of classification. The presented melamine pattern and NCA based model used a multileveled feature generation model. Therefore, 10,927 features (it is huge feature set) are generated using the presented multileveled (DWT- based) hybrid feature generation model. The size of this feature vector must be decreased. The general objectives of the feature selection methods are [57, 58]: (i) selection the most valuable features to increase the performance, (ii) decreasing number of features decrease the execution times of the classifiers. To meet these requirements, NCA is chosen as feature selector. The NCA is one of the simplest feature selector and it is a feature selection variance of the nearest neighborhood model [59, 60]. It generates non-negative (positive) feature weights which helps to select the most discriminative features. The steps involved are given below:

  • Step 5: Apply NCA to generate and merge the features (X).

  • Step 6: Select the most discriminative 256 features by using the generated weights.

4.3 Classification

The selected 256 features are fed to Quadratic SVM and Weighted kNN classifiers for automated classification. To select the most appropriate classifier, MATLAB classification learner toolbox (MCLT) which has 25 classifiers is used in this work. We have obtained highest classification accuracy using quadratic SVM and weighted kNN classifiers. The attributes of the classifiers used are given as below.

  • Weighted kNN: In this work, k value is selected as 10, distance metric is spearman and distance weight is selected as squared inverse [61].

  • Quadratic SVM: In this work, we have chosen second-degree polynomial kernel, box constraint level = 1 and one-vs-one as multiple classification model [45, 46] to obtain the maximum performance. We have developed the model using hold-out validation strategy with 80% is used for training and 20% for testing.

  • Step 7: Classify the selected features using Quadratic SVM or weighted kNN classifiers.

5 Experimental results

The presented melamine pattern and NCA- based model is implemented by using MATLAB (2020a) programming environment and it is programmed functionally. The used functions are named as melamine pattern, statistical feature generator, NCA and main functions. The main function and feature generation and selection functions generate features. The selected features are fed to MATLAB classification learner application containing 25 classifiers. We have used quadratic SVM and Weighted kNN classifiers in this work. We have developed the classification model for every channel using 80:20 hold-out validation strategy. To evaluate performance of this model for each channel, sensitivity, specificity, balanced accuracy, overall accuracy, geometric mean, F1-score and precision metrics were used. True positives (TP), false positives (FP), true negative (TN) and false negatives (FN) parameters are computed. Mathematical notations of the used measurements were given below (Eqs. (18)–(24)) [62, 63].

$$ sen=\frac{TP}{TP+ FN} $$
(18)
$$ spe=\frac{TN}{TN+ FP} $$
(19)
$$ gm=\sqrt{sen\ast spe} $$
(20)
$$ bacc=\frac{sen+ spe}{2} $$
(21)
$$ acc=\frac{TP+ TN}{TP+ TN+ FP+ FN} $$
(22)
$$ pre=\frac{TP}{TP+ FP} $$
(23)
$$ F1=2\frac{sen\ast pre}{sen+ pre} $$
(24)

where sen, spe, gm, bacc, acc, pre and F1 defines sensitivity, specificity, geometric mean, balanced accuracy, accuracy, precision and F1 score respectively. We have developed the model using 80:20 hold-out validation strategy and classifiers are executed 100 times. The minimum, average, maximum and standard deviation values of the results are listed in Tables 3-6 using two classifiers with various channels. There are 20 channels in the used EEG dataset and results of ten channels are given in each table. Furthermore, best results are highlighted in bold font in these tables.

Table 3 Summary of performance measures (%) obtained using our proposed system with quadratic SVM classifier for 1st -10th EEG channels
Full size table
Table 4 Summary of performance measures (%) obtained using our proposed system with quadratic SVM classifier for 11th -20th EEG channels
Full size table
Table 5 Summary of performance measures (%) obtained using our proposed system with Weighted kNN classifier for 1st -10th EEG channels
Full size table
Table 6 Summary of performance measures (%) obtained using our proposed system with Weighted kNN classifier for 11th -20th EEG channels
Full size table

The obtained results for different channels using quadratic SVM is given in Tables 3-4.

In this Table, Min, Av., Max and Std define the minimum, maximum, average and standard deviation values respectively. In Tables 3, 99.05% classification accuracy is achieved using A2A1 channel. The general classification accuracy of this channel is 98.23% ± 0.32%. The worst channel is F7, as it reached 95.57% maximum accuracy using SVM classifier.

It can be noted from Table 4 that, A2A1 channel reached the highest accuracy among all channels. The best accurate channel is T4 and it reached 97.61% maximum accuracy and general accuracy of 96.64% ± 0.42%. There is no channel with accuracy lower than F7 channel in Table 4. The channel-wise results obtained using weighted kNN are listed in Tables 5-6.

It can be noted from Table 5 that, 99.11% classification accuracy is achieved using A2A1 channel. The general classification accuracy of this channel is 98.29% ± 0.34%. The channel F7 reached the worst results with maximum accuracy of 95.53% using Weighted kNN classifier with general results = 93.39% ± 0.34%.

It can be seen from these results that, highest classification accuracy of 98.57% and 99.11% are obtained for channels A2A1 and T4 channels, respectively using weighted kNN classifier. Our same model obtained the accuracy of 99.05% and 97.61% using A2A1 and T4 channels, respectively with quadratic SVM classifier. In both classifiers, the best accurate channel is found to be A2A1 using both classifiers and F7 channel yielded the worst results for both classifiers. Tables 3-6 denotes the results of our proposed melamine pattern and NCA based model reached accuracies of >95% for all channels. Also, our MDD model reached specificity of 99.87% and recall rate of 99.85% for A2A1 channel using weighted kNN classifier.

Performance measures (sensitivity, specificity, geometric mean, precision and F1-score) obtained for various channels using SVM, and (b) weighted kNN classifiers are shown in Fig. 6.

Fig. 6
figure 6

Performance measures (sensitivity, specificity, geometric mean, precision and F1-score) obtained for various channels using classifier: (a) SVM, and (b) weighted kNN

Full size image

6 Discussions

The primary motivation of this work is to develop a novel feature generation model using molecular structure graph called melamine pattern. Our proposed model consists of melamine pattern, statistical feature generator, DWT, NCA selector and two conventional classifiers. In this work, we have used a huge dataset consisting of 64 subjects with 20 channels of EEG signals for each subject. The used testbed (corpus) has 20 channels with 7339 EEG signals in each channel. We have used seven performance measures to evaluate the performance of our model. The graph of accuracy (%) obtained versus various EEG channels for SVM and kNN classifiers is shown in Fig. 7.

Fig. 7
figure 7

Graph of accuracy (%) obtained versus various EEG channels for SVM and kNN classifiers

Full size image

The Fig. 7 demonstrates that the weighted kNN attained higher results than quadratic SVM for more number of channels. However, quadratic SVM reached higher results than Weighted kNN for 5th (F3), 7th (F7), 8th (F8) and 16th (Pz) channels. The highest accuracies are obtained using A2A1 and T4 channels. Confusion matrices of these channels are also shown in Fig. 8.

Fig. 8
figure 8

Confusion matrices of the best performing channels using kNN and SVM classifiers. These matrices denote number of true predicted and false predicted values. a Confusion matrix obtained using A2A1 channel with kNN classifier. b Confusion matrix obtained using A2A1 channel with SVM classifier. c Confusion matrix obtained using T4 channel with kNN classifier. d Confusion matrix obtained using T4 channel with SVM classifier

Full size image

The summary of state-of-the-art techniques developed for automated detection of depression using EEG signals is shown in Table 7.

Table 7 Summary of state-of-the-art techniques developed for automated detection of depression using EEG signals
Full size table

It can be seen from Table 7 that the presented model has obtained the high classification performance for depression detection. The presented model is a hand-modeled feature extraction based machine learning model and it selects the most informative features using NCA. Table 7 denotes that the deep learning based models have reached highest performance using big MDD datasets and hand-crafted models achieved high performance on the smaller MDD corpora. The best performing model among state-of-the-art classification methods is Sandheep et al.’s [75] method and they used one-dimensional CNN to classify 6000 EEG records using ten-fold cross-validation and their model achieved 99.31% classification accuracy. However, they reached this result using all channels of the dataset. Mumtaz and Qayyum [74] presented a EEG segmentation, one-dimensional CNN and two-layered LSTM based MDD classification model. They used 19 channels (all channels) of the used dataset and their model (1DCNN-LSTM) attained 98.32% classification rate, but the time burden of this model is very high. Sharma et al. [73] presented a hand-designed features based MDD classification model and reached 99.58% classification accuracy using small dataset. Moreover, they did not give channel-wise results. Our presented model achieved 99.11% and 99.14% accuracies using 80:20 split ratio and ten-fold cross-validation strategy, respectively. Our presented results are single channel results and the performance of the model is given for every channel. Furthermore, the presented model is developed using big MDD dataset and majority voting or channel concatenation is not employed to reach higher performance.

The advantages of this work are as follows:

  • A new public dataset is used to develop the machine learning model and obtain the highest classification performance.

  • A new feature generation model is presented for textural features using melamine pattern.

  • The effective and highly accurate model is developed.

  • Model attained >95% accuracies for all channels of EEG signals. These results imply that the developed model is robust and do not need to use all channels.

  • This model is lightweight and need not set millions of parameters like deep networks to attain highest classification accuracy.

  • This model used an effective feature generation method with a low computation complexity. This feature generation model used multiple kernel based melamine pattern. The time complexity of the melamine pattern is calculated as O(n). Moreover, the one dimensional DWT is used to decompose signals and it halves length of the signal in each level. Therefore, the time complexity of the presented multileveled feature generation model is calculated as O(nlogn).

  • Developed a simple model and can be implemented easily as it does not involve lot of calculations.

In this work, melamine pattern and NCA-based method is employed to detect MDD automatically using EEG signals. There are many molecules in nature and they can be defined as a graph. More molecular structure based feature generation models can be presented to improve the classification performance. The developed automated MDD detection model can be uploaded to the cloud. The EEG signals to be tested is sent to our model placed in the cloud to find the class of the EEG signal. The result of the model will be sent to the clinicians and after confirming from the psychiatrists will be sent to the patients. This will expedite the diagnosis process and also help to provide immediate treatment to the patients.

7 Conclusions

Automated depression detection using EEG signals is a complex and challenging problem in machine learning. Many deep learning models have been used to attain high performance. In this work a hand-crafted features-based depression detection model is proposed using novel melamine pattern with EEG signals. Our presented model attained >95% accuracies for all 20 channels of EEG signals. This clearly indicates the robustness of the developed system. We have obtained highest classification accuracy of 99.11% and 99.05% using Weighted kNN and Quadratic SVM respectively using A2A1 channel. In future, the developed model can be used to detect early (mild) stage of depression using EEG signals using bigger database.