We used two classifiers in this study: SVM and NB. Support Vector Machine (SVM) is one of the supervised machine-learning algorithms with kernel based model. Recently, SVM is highly used for pattern recognition, classification, and regression. The core fundamental of the SVM is to develop a hyper plane in infinite-dimensional space for data prediction. The hyper plane’s highest accuracy is achieved, which has the most considerable distance to the classes’ nearest training point [59].

Naive Bayes (NB) classifier is based on Bayes hypothesis with the independence assumptions between predictors. It is easy to be constructed without any problematic iterative parameter valuation that makes it individually useful for big datasets. Bayes hypothesis provides the calculation of posterior probability (P${}_{\text{tp}}$) from the prior probability of target (P${}_{\text{t}}$), the prior probability of predictor (P${}_{\text{p}}$), and the probability of predictor given target (P${}_{\text{pt}}$). NB classifiers admit the effect of the value of a prior probability of a predictor on a given target. The NB classifier is described in Eqn. 1:

(1)$$N_b=P_{tp}\frac{P_{pt}{P}_t}{P_p}$$

After normalized feature extraction, we used two types of classifications such as subject wise and mix (mental stress vs. control) classifications using SVM and NB machine learning classifiers in the parameters of Sensitivity (S${}_{\text{en}}$), Specificity (S${}_{\text{pe}}$), Accuracy (A${}_{\text{cc}}$), F${}_1$, Precision (P${}_{\text{re}}$), and Area Under the Curve (AUC) on the 14 recordings of the frontal lobe EEG signals. Additionally, we used seven models such as random sampling, leave one out, and five cross-validation techniques, including 20, 10, 5, 3, and 2 folds for the classifiers. These parameters’ performance, including S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, and P${}_{\text{re}}$, are described in Eqns. 2,3,4,5,6 and Table 2 [60, 61].

(2)$$S_{en}=\left(\frac{TP}{FN+TP}\right)$$

(3)$$S_{pe}=\left(\frac{TN}{FP+TN}\right)$$

(4)$$A_{cc}=\left(\frac{TP+TN}{TP+TN+FP+FN}\right)$$

(5)$$P_{re}=\left(\frac{TP}{TP+FP}\right)$$

(6)$$F_1=2\left(\frac{S_{en}\times P_{re}}{S_{en}+P_{re}}\right)$$

where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative.

As a preprocessing stage, alternating current (AC) power noise is removed and data is filtered using a notch filter, i.e., band stop filter, which is used to attenuate the specific band frequency range of the signal as given in Eqn. 7. Following that, to obtain frequencies within a specific range, EEG data is passed through a band pass filter with a cut-off frequency of 0.1–30 Hz. Finally, to decompose a multivariate EEG signal into independent signals, Independent Component Analysis (ICA) method is applied to remove several sources artifacts, such as Electrooculography (EOG) and Electromyography (EMG) signals. ICA is given in Eqn. 8. At data analysis and feature extraction stages, FFT, which is described in Eqn. 9, is applied in the frequency domain to extract the mean power spectrum for EEG main bands, delta (0.1–3.9 Hz), theta (4–7.9 Hz), alpha (8–13.9 Hz), beta (14–29.9 Hz), and gamma (30–80 Hz) for 16 channels covering the frontal lobe [62, 63, 64].

(7)$$H(s)=\frac{s^2+w^2}{s^2+w{c}^2+w^2}$$

(8)$$x=\sum _{i=1}^n{a}_i\cdot S_i$$

(9)$$X[k]=\sum _{n=0}^{N-1}x[n]e^{\frac{-j2\pi kn}{N}}$$

where, s is the vector form signal, wc is the range of attenuated band frequency, x is the signal with vector s, and N is the domain’s size.

To process our EEG data for frontal lobe channels, we used the EEG LAB, an interactive Matlab toolbox for processing continuous and event-related EEG. Table 3 shows the minimum and maximum values of the mean power spectrum of all EEG bands in the frontal lobe for stress and control groups, respectively. The differences between stress and control power spectrum value, expressed in $\mu$V${}^2$/Hz, for the EEG waves such as delta, theta, alpha, beta, and gamma are shown in Fig. 3. In the delta activity, the stress and control group’s differences in maximum and minimum normalized values were 0.065 and 0.388, respectively. However, in the theta activity, these differences were –1.094 and –0.426, respectively. In the alpha activity, the stress and control group’s differences in maximum and minimum normalized value were found to be –0.030 and –0.014, respectively. In the beta activity, the stress and control group’s differences in maximum and minimum normalized values were 0.101 and 0.235, respectively. Regarding gamma activity, the stress and control group’s differences in maximum and minimum normalized values were –0.375 and 0.518, respectively. At the final stage, the obtained feature vector is classified using SVM and NB classifiers.

Fig. 3.

Differences between mental stress and control group for the mean power spectrum.

We used two type of classifications: (1) subject wise and (2) mix (mental stress vs. control) using SVM and NB classifiers. In subject wise classification our system average performance in terms of S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, P${}_{\text{re}}$, and AUC were 0.9821, 0.9777, 0.9821, 0.9817, 0.9849, and 0.9855, respectively (Table 4 and Fig. 4).

Fig. 4.

Average of the subject wise classification using SVM and NB machne learning classifiers.

In the mix (mental stress vs. control) classification, Table 5 represents the performances of the SVM Radial Basis Function (RBF) Kernel classifier. The random sampling model gave the highest performance as compared to the others. The values of S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, P${}_{\text{re}}$, and AUC were found to be 0.814, 0.815, 0.814, 0.814, 0.815, and 0.861, respectively. Table 6 represents the SVM linear classifier’s performances, where Leave One Out model gave the highest performance compared to the other models. The values of S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, P${}_{\text{re}}$, and AUC were found to be 0.902, 0.902, 0.902, 0.902, 0.915, and 0.937, respectively. Similarly, Table 7 represents the SVM polynomial classifier’s performances, where the 3-fold model gave the highest performance compared to the others. The values of S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, P${}_{\text{re}}$, and AUC were found to be 0.799, 0.799, 0.799, 0.799, 0.799, and 0.846, respectively. Table 8 represents the SVM sigmoid classifier’s performances, where the random sampling model has the highest performance. The values of S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, P${}_{\text{re}}$, and AUC were found to be 0.814, 0.815, 0.814, 0.814, 0.815, and 0.774, respectively. The NB classifier’s performances are shown in Table 9, where the random sampling model gave the highest performance compared to other models. The values of S${}_{\text{en}}$, S${}_{\text{pe}}$, A${}_{\text{cc}}$, F${}_1$, P${}_{\text{re}}$, and AUC were found to be 0.817, 0.817, 0.817, 0.817, 0.817, and 0.897, respectively.

The highest and average results obtained by the proposed classifiers. As shown in Fig. 5, the SVM linear classifier achieved the highest accuracy in both individual and average cases. Additionally, the differences between the mean of the highest and average performance of the classification shown in Fig. 6. The proposed SVM linear classification results for the leave one out model regarding sensitivity, specificity, and accuracy were found to be 0.902 for all of them. In addition, SVM linear classifier achieved the highest accuracy in the models’ average in terms of sensitivity, specificity, and accuracy were found to be 0.8967, 0.8971, and 0.8967, respectively.

Fig. 5.

Comparison of the (a) highest performance and (b) average classifiers for the mix (mental stress vs. control) classification.

Fig. 6.

Comparative analysis between the mean of the highest and average performance classifications.

Asif et al. [49] designed a stress classification technique using EEG signals regarding previous related work. As a feature extraction stage, five main features are extracted from the recorded EEG signals: absolute power, relative power, coherence, phase lag, and amplitude asymmetry. Minimal sequential optimization, decent stochastic gradient, logistic regression, and multilayer perceptron are applied in the classification stage. Importantly, logistic regression shows the highest classification accuracy. Ranjith et al. [50] introduced a stress detection technique based on Improved Elman Neural Network (IENN). Power Spectral Entropy (PSE) and Gray Level Different Statistics (GLDS) methods were applied in the feature extraction stage, where IENN was applied in the classification stage. Bairagi et al. [51], a hardware system to detect mental stress based on a preamplifier and filter using EEG signals was presented. The authors have found that the Theta band is associated with the frontal lobe. The accuracy of the proposed approach was approximately 88%. In a similar study by Hasan et al. [65], a mental stress detection technique based on a multi-domain hybrid feature pool has been proposed. Moreover, statistical analysis and wavelet-based analysis are both applied to extract features from recorded EEG signals. Following that, the KNN algorithm is used in the classification stage. The accuracy of the proposed method was 73.38%. Many wearable smart devices such as CGX-Quick 30, dry EEG head set, EMOTIV-EPOC X, MUSE-2, mBrainTrain-SMARTING mobi, OpenBCI-Cyton, and sensing-DSI 24 based on flexible electronics are used in the recording of the EEG signal in modern world.

In Shon et al. [53], mental stress detection based on the Genetic Algorithm (GA) technique is proposed, where GA is applied as a feature extraction and selection technique. KNN classifier is applied as a classification technique. The accuracy of the proposed method was 72%. In Priya et al. [54], a stress detection technique based on power ratio for the frontal lobe using EEG is proposed. PSD for delta, alpha, beta, theta, and gamma bands are analyzed and computed. Kernel-based SVM and KNN classifiers are applied. As a result, KNN shows better accuracy results than SVM. Wijsman et al. [66] used the Piezoelectric Film Sensor (PFS) to record the participants’ stress. They applied the KNN classifier, and they were able to achieve 80% accuracy. Sun et al.[67] used thirty subjects using an ECG signal to record the cardiac signal. Additionally, they classified the signal using a Decision Tree (DT) classifier, which achieved 80% accuracy, where DT is easy to understand efficiently [68]. Previously, many smart devices, including accelerometer, Galvanic Skin Response (GSR), smart watch, and electrodes for the recording of the physiological signal such as ECG, EEG, and EMG, are used to detect stress [69, 70] . They used DT, KNN, and SVM classifiers to classify the subjects mentioned in Table 10 (Ref. [53, 65, 66, 67, 69, 70, 71]). However, our proposed method achieved better results (subject wise classification: 98% and mix (mental stress vs. control) classifications: 90%) than previously selected methods, as shown in Fig. 7.

Fig. 7.

Comparison between accuracy of the previously selected and proposed system.

The main advantages of the proposed method are free over fitting, fast, low complexity, high accuracy, simplicity and ease to use. Finally, it could be used as a real-time and continuous monitoring technique [72, 73]. However, one gender research, one modality, and one lobe are the limitation of this study. Moreover, this study is applied only for 14 subjects that might increase the classification accuracy of the study.