EEG Based Cigarette Addiction Detection with Deep Learning

2.png

Figure 2. Visual stimulation presentation

3.png

Figure 3. Signal acquisition setup

The electrode gel used is designed for the EEG cap and provides conductivity for impedance adjustment. These features provided by the EEG measurement system can be modified using the interface (System Plus Evaluation) produced by Micromed.

As shown in Figure 4, there are separate slots for each electrode belonging to the international 10/20 measurement system and a separate socket slot for cap connection. The 10/20 system is the accepted standard for electrode placement in EEG applications and has been used for half a century. This system defines the locations on the scalp using the relative distances between the cranial landmarks on the scalp [20].

Figure 5 shows the EEG device and sample EEG data acquisition process. The device used for data acquisition has 32 channels and a preamplifier gain of 1600 µV/cm. The cut-off frequency (fc) can be adjusted between 0.008-2000 Hz, and a notch filter (50 Hz) setting is available to suppress power line noise. The system has a maximum sampling frequency of 4096 Hz and 16-bit resolution ADC structure. In addition, the system offers bipolar and unipolar measurement modes for EEG channels.

4.png

Figure 4. 10/20 Measurement system electrode arrangement and brain regions [21, 22]

5.png

Figure 5. EEG recording system

The following precautions were taken before data acquisition:

1-Individuals who are smokers were asked not to smoke at least 2 hours before data acquisition.

2-Attention was paid to the cleanliness of the individuals' hair.

3-All devices that could cause artefacts (e.g. phones) were turned off in the room where data acquisition was performed.

4-Data acquisition was performed with a laptop computer to reduce city grid artefacts.

5-Individuals were asked not to eat or take medication/caffeine at least 2 hours before data acquisition.

The following precautions were taken during data acquisition:

1-Care was taken to ensure that individuals did not move.

2-Attention was paid to the use of conductive gel for the electrodes. (Gel from one electrode did not touch another electrode.)

3-The EEG cap and electrode cleaning were performed for each individual after the EEG procedure.

4-The electrode fixation process was carefully performed and EEG acquisition was performed after the electrode impedance values fell below 10 ohms.

5-If unusual changes were observed in the signals during EEG acquisition, the procedure was stopped.

6-The data acquisition hours were the same for all individuals (14:30-16:30).

The EEG data recorded in this study has a sampling frequency of 4096 Hz. EEG data was collected from 13 addicted and 17 non-addicted individuals, resulting in 118 seconds of EEG data per individual by showing each of the 59 slides in Figure 2 for 2 seconds each. Two-second segments of the individuals' EEG signals at the time when smoking images were presented were extracted from the signal and labelled as addicted or non-addicted according to the individual's FTND test result.

In the process of training and evaluating the model, the dataset has been evenly split into a training set comprising 70% of the data and a test set comprising the remaining 30%. This balanced distribution has facilitated the adoption of a comprehensive approach during the model's training and testing phases, thereby enhancing the representational capacity of the dataset used in the learning process. The results obtained on the test set reflect the final performance of the model and unveil its potential success on real-world data. These results are considered a critical indicator of the model's generalizability and its potential performance on external datasets. Additionally, the study has taken into account the participants' age, height-weight, and physiological characteristics, which are presented in Table 1.

Table 1. Individual characteristics

2.3 Discrete Wavelet Transform (DWT)

WT (Wavelet Transform) is a technique used to analyse a signal in the time-frequency domain. In this method, the signal is decomposed into different frequency components using a scaling function and a specific wavelet function [23]. DWT (Discrete Wavelet Transform) works on the same principle, but breaks down the signal into smaller parts and processes each part separately. DWT is a mathematical operation used to determine the frequency components of a signal. This process separates the high and low frequency components of a signal by breaking it down into smaller scales, and represents these components at different scales and times. DWT is widely used in fields such as image processing, signal processing, and data compression. This operation plays an important role in data compression because a large portion of the high-frequency components can be removed, reducing the signal size significantly. Additionally, DWT can be used in situations where data needs to be analysed at different scales [24]. DWT has been proven effective in the analysis of bio signals [25].

The mathematical formulation of the discrete wavelet transform is as follows:

${{C}_{j,k}}=\sum\limits_{n=0}^{N-1}{f(n).{{\phi }_{j.k}}(n)}$                   (1)

${{D}_{j,k}}=\sum\limits_{n=0}^{N-1}{f(n).{{\psi }_{j.k}}(n)}$                  (2)

where, f(n) is the n-th sample of the signal to be analysed, N is the number of sampling points of the signal. $\phi_{j, k}(n)$ and $\psi_{j, k}(n)$ are the j-th scale and k-th time shift of the scaling and wavelet functions, respectively. $C_{j, k}$ and $D_{j, k}$ are the outputs of the discrete wavelet transform for the scaling and wavelet components of the signal, respectively [26].

2.4 Feature extraction

Signal processing is a set of techniques used for the analysis and manipulation of time-varying or spatial data. Biomedical signals, such as electroencephalography (EEG), contain complex data on the electrical activity of the human body. In signal processing, feature extraction is the process of obtaining meaningful information from raw data and is usually a preliminary step for data analysis, recognition, or classification processes. Feature extraction involves determining and extracting metrics that represent the signals and are more useful for analysis, rather than directly processing the raw signals.

The 15 features mentioned in this study are widely accepted in signal processing, especially in the analysis of EEG data. These features include basic statistical properties of signals (mean, standard deviation, skewness, kurtosis, etc.), energy features (autocorrelation coefficients, Wilson amplitude), and time-domain features (number of zero crossings, sign changes in slope). Extracting such features simplifies the complex structure of EEG data, making the data more understandable and can highlight differences between brain states. Specifically, the ability to distinguish between specific brain states, such as addiction states, is a crucial factor in the selection of these features.

Feature extraction directly impacts the performance of classification and recognition systems. Well-chosen features can enhance the accuracy and reliability of the model, while less informative or misleading features can decrease the model's performance. Therefore, significant attention has been given to the feature selection stage in this study, including a testing phase to identify and remove features that decrease the overall performance of the model. This process ensures that only the most informative and meaningful features are included in the analysis, enabling more effective and accurate classification and recognition operations on EEG data.

The feature set used in the study is provided below along with their descriptions:

Mean: This feature is obtained by summing the amplitudes of the signal samples and dividing them by the number of samples. In statistics, this feature is called the first-order moment. The mean provides a general overview of the signal and allows us to obtain information about the behaviour of the signal. Therefore, the mean feature has an important place in signal processing and analysis.

$mean=\frac{1}{N}\sum\limits_{n=1}^{N}{{{x}_{n}}}$                     (3)

Mean absolute deviation (MAD): Mean Absolute Deviation is the sum of the absolute values of the distances of sample data points from the mean [27].

$mad=\frac{1}{N}\sum\limits_{n=1}^{N}{\left| {{x}_{n}}-mean \right|}$                  (4)

Variance: This value gives the distribution of the data around the mean value. It is calculated by taking the arithmetic mean of the squares of the distance of each value in the data set from the mean value, and expressed as follows:

$variance=\frac{1}{N-1}\sum\limits_{n=1}^{N}{{{({{x}_{n}}-mean)}^{2}}}$                      (5)

Standard deviation: Standard deviation is used to determine how close the samples in a dataset are to the mean and to measure how spread out the data is. Standard deviation is the square root of variance.

$sd=\sqrt{\frac{1}{N-1}\sum\limits_{n=1}^{N}{{{({{x}_{n}}-mean)}^{2}}}}$                        (6)

Skewness: It shows how much the distribution deviates from symmetry around the mean and is expressed as follows:

$skewness=\frac{\frac{1}{N}\sum\nolimits_{n=1}^{N}{{{({{x}_{n}}-mean)}^{3}}}}{s{{d}^{3}}}$                 (7)

Kurtosis: This provides information about whether the peak values of the data are flat or sharp.

$kurtosis=\frac{\frac{1}{N}\sum\nolimits_{n=1}^{N}{{{({{x}_{n}}-mean)}^{4}}}}{s{{d}^{4}}}$                      (8)

Autocorrelation coefficients: Autocorrelation coefficients are one of the commonly used methods in signal processing. Mathematically, it can be expressed as follows:

$x\left[ n \right]=-\sum\limits_{k=1}^{p}{{{a}_{k}}x\left[ n-k \right]+e\left[ n \right]}$                (9)

where, p is the degree of the model, X_n is the data signal consisting of n samples, a_k are the real-valued autocorrelation coefficients, and e[n] represents the error term of independent white noise from the previous samples. The aim here is to investigate the relationship of a data point with the previous few points and to create a model based on this.

Zero crossing rate: In this method, the points where the signal is 0 are counted and recorded.

$z c c=\sum_{N=1}^{N-1}\left[\operatorname{sgn}\left(x_n \cdot x_{n+1}\right] \cap\left|x_n-x_{n+1}\right| \geq\right.$ threshold                 (10)

Wilson amplitude: It is a value that indicates how many times the consecutive samples in the signal exceed a predetermined threshold value and is expressed as follows:

$\begin{aligned} & w a=\sum_{n=1}^{N-1}\left[\operatorname{sgn}\left(\left|X_n-X_{n+1}\right|\right)\right] \\ & f(x)=1 \Rightarrow f(x) \geq \text { threshold } \\ & f(x)=0 \Rightarrow \text { other_situations }\end{aligned}$                (11)

Slope sign change: The points where the slope of the signal changes from positive to negative and negative to positive are counted and recorded.

$\begin{matrix}  ssc=\sum\limits_{n=2}^{N-1}{\left[ f\left[ ({{x}_{n}}-{{x}_{n-1}})x({{x}_{n}}-{{x}_{n+1}}) \right] \right]}; \\  f(x)=1\Rightarrow f(x)>threshold \\  f(x)=0\Rightarrow other\_situations \\\end{matrix}$                      (12)

Singular value decomposition (SVD): SVD is a process of decomposing a matrix. It is achieved by calculating the eigenvalues and eigenvectors of the given matrix. SVD can compress the information contained in the matrix and emphasize its important features without changing the dimension of the matrix. Therefore, SVD is widely used in fields such as data analysis, dimensionality reduction, matrix approximation, and recommendation systems [28].

$M=U.d..{{V}^{T}}$                  (13)

where, M is an m x n sized matrix, U is an m x m sized orthogonal matrix, d is an m×n sized diagonal matrix, and the elements of d are the nonzero singular values. V^T is an n×n sized transposed orthogonal matrix.

Median: The value that falls in the middle when the samples in a dataset are arranged in order.

If there is an odd number of samples in the dataset;

$median=x\left[ \frac{(n+1)}{2} \right]$                          (14)

If there is an even number of samples in the dataset;

$median=\frac{\left( x\left[ \frac{n}{2} \right]+x\left[ \left( \frac{n}{2}+1 \right) \right] \right)}{2}$                     (15)

Minimum and maximum values: The smallest and largest values of the samples in a dataset.

$min=\left[ {{x}_{1}},{{x}_{2}},...,{{x}_{n}} \right]$                  (16)

$min=\left[ {{x}_{1}},{{x}_{2}},...,{{x}_{n}} \right]$                   (17)

Sum: The sum of the samples in a dataset.

$sum=\sum\limits_{i=1}^{n}{{{x}_{i}}}$                     (18)

2.5 Power Spectral Density (PSD)

PSD represents the power distribution of the frequency components of a signal. In other words, it measures the power levels of different frequency components of a signal. It can be used to analyse the characteristics of signals in the frequency domain. This method can be applied to EEG signals over a wide frequency range, as well as signals in defined sub-bands [29, 30].

The periodogram is a classical spectral estimation method and is calculated using the non-parametric FFT (Fast Fourier Transform) method. The Welch method, on the other hand, is a more advanced method and processes the signal by dividing it into overlapping windows. One of the PSD methods that increases signal power at different frequencies is the Welch method [31]. The Welch method has been shown to present very strong features of EEG signals and provide good discrimination between classes [32]. Welch PSD divides the data into overlapping segments, calculates their windows and FFTs, squares the resulting values, and finally takes the time average of the resulting periodograms. The advantages of the Welch method include reducing noise and power variances [33].

$P(f)=\frac{{{T}_{s}}}{K.M}{{\left| \sum\limits_{n=0}^{M-1}{x(n)w(n)}.{{e}^{-j2\pi fn}} \right|}^{2}}$                    (19)

The power spectrum of the signal sampled with Ts is equalized to the power spectrum of the continuous-time signal using the windowing function w(n). During this process, a normalization constant called K is used.

In Eq. (20), $S_{x x}^{\wedge(i)}(f)$ represents the i-th windowed periodogram. In Eq. (21), L is the number of windows. Taking periodograms in small parts instead of the entire signal and averaging them provides more accurate results.

$S_{xx}^{\hat{\ }(i)}(f)=\frac{{{T}_{s}}}{K.M}{{\left| \sum\limits_{n=0}^{M-1}{x(n)w(n)}.{{e}^{-j2\pi fn}} \right|}^{2}}$                  (20)

$P_{welch}^{\hat{\ }}(f)=\frac{1}{L}\sum\limits_{i=0}^{L-1}{S_{xx}^{\hat{\ }(i)}(f)}$                 (21)

2.6 Evaluation metrics

In scientific studies, evaluation metrics are necessary to monitor the performance of proposed systems. In classification problems, evaluation metrics can be easily calculated using a confusion matrix, as shown in Table 2. Cases where the outputs generated by the algorithm are correct are written as true positive (TP) cells and true negative (TN) cells in the confusion matrix. If an output is generated as negative but is actually positive, it is written in the false negative (FN) cell, and if an output is generated as positive but is actually negative, it is written in the false positive (FP) cell.

Table 2. The confusion matrix

Classification accuracy (ACC) is a commonly used quality measure to evaluate the performance of classification algorithms, and is calculated by dividing the number of correct predictions by the total number of predictions. In addition to ACC, which is the most basic measure, other commonly used measures include specificity (SPEC), sensitivity (SENS), precision (PREC), and F-score.

SPEC measures the rate of correctly predicting true negatives by the classifier, while SENS measures the rate of correctly predicting true positives. PREC is the ratio of the total number of true positive examples that are correctly classified to the total number of positive examples. The F-score is the harmonic mean of precision and recall values. These measures are calculated according to the following formulas:

$accuracy=\frac{tp+tn}{tp+tn+fp+fn}$                   (22)

$specificity=\frac{tn}{tn+fp}$                (23)

$sensitivity=\frac{tp}{tp+fn}$           (24)

$precision=\frac{tp}{tp+fp}$                (25)

$f-score=2\times \frac{\left( precision \right)\times \left( sensitivity \right)}{\left( precision \right)+\left( sensitivity \right)}$                    (26)

2.7 Artificial Neural Network (ANN)

Deep ANN is a subfield of artificial intelligence that enables algorithms to match learned models with data, similar to traditional machine learning algorithms. However, deep ANN works with more complex structures and larger datasets. Deep ANN uses a multi-layer architecture where data is passed through different layers to learn features and produces classification, prediction, or another output in the final layer. These features help the model understand the data and distinguish different features from each other.

In the 1950s, Frank Rosenblatt developed the simplest ANN model called Perceptron, inspired by the work of Warren McCulloch and Walter Pitts. Perceptron performed well on simple tasks, but failed to solve complex problems. In the early 1990s, advances in machine learning led to further development of ANN. During this time, new learning techniques like backpropagation algorithm were developed, allowing ANN to perform more complex tasks. In the early 2000s, the use of deep ANN with multiple hidden layers increased significantly. Today, deep ANN is considered one of the most important tools in machine learning and is used in many fields in the healthcare sector, ranging from disease diagnosis to medical image interpretation [34].

Table 3. Deep ANN layers

In this study, the deep ANN model used consists of 19 layers as shown in Table 3. ReLU function is used as the activation function in the model. Model training was performed using the ADAM optimization algorithm in 1000 epochs. During model training, the dataset was divided into two separate sets, 70% for training and 30% for testing.