EEG Signals Based Internet Addiction Diagnosis Using Convolutional Neural Networks

Abstract

Internet addiction (IA), as a new and often unrecognized psychosocial disorder, endangers people’s health and their lives. However, the common biometric analysis based on the combination of EEG signals and results of questionnaires is not quantitative, and thus difficult to ensure a specific biomarker. This work aims to develop a deep learning algorithm (no need to identify biomarkers) used for diagnosing IA and evaluating therapy efficacy. Herein, a five-layer CNN model combined with a fast Fourier transform is proposed to diagnose IA quantitatively. This algorithm is validated in the Lemon dataset by using it to process raw data, full spectral power, and alpha-beta-gamma spectral power (related to IA). In contrast to alpha-beta-gamma spectral power, the results based on full spectral power show better performance (87.59% accuracy, 88.80% sensitivity, and 86.41% specificity), which confirms that the proposed algorithm can diagnose IA without biomarkers. In addition, this proposed CNN model presents obvious advantages in processing raw data, achieving 81.1% accuracy. Such results verify that this method can contribute to the reduction of diagnosis time and be potentially used in real-time health monitoring systems. This work provides a quantitative approach to diagnose IA and evaluate therapy efficacy, as a general strategy, and can be widely used in other disorder diagnoses that affect EEG signals, such as psychiatric disorders, substance dependence, and depression.

1. Introduction

Internet addiction (IA), a biopsychosocial disorder, causes functional impairments and marked distress to oneself [1]. It has become a serious mental health problem in many countries, including China (16.4%), Vietnam (21.2%), South Korea (20%), and other parts of Asia such as the Philippines (21%) [2]. Considering the negative impact and prevalence of IA, it is critical to diagnose IA accurately for timely follow-up treatment. IA could cause neurobiological changes in the brain by transferring chemicals in the bio-environment and producing electric signals [3,4]. These electric signals are recorded by electroencephalography (EEG), which has been evidenced as valid in many scenarios such as epilepsy detection [5,6], driver drowsiness detection [7,8], and sleep analysis [9,10]. Due to the superiority of EEG in monitoring behaviors affecting brain activities, lots of studies consider focusing on finding relative EEG biomarkers to measure IA (Figure 1). Compared with the diagnostic methods that are only based on the score of the internet addiction test (IAT), EEG biomarkers, summarized by biometric analysis, could provide more objective measurements, which bring new hope for IA diagnosis.

However, the obstacle to obtain specific EEG biomarkers for IA is that the frequency bands screened by numerous research teams are different. For example, Both Jung-Seok et al. [3] and Yan et al. [11] found that IA patients have a lower beta band, but they have divergence in the gamma band. Jan et al. [12] provided another correlation verification result, i.e., the power of mid-alpha and high-beta bands are reduced by IA. In addition, traditional biostatistical analysis can only explain which frequency bands of EEG signals are related to IA, but cannot measure the effectiveness of these biomarkers quantitatively in the addiction diagnosis. Therefore, an effective method to obtain objective features is urgently needed.

With the development of advanced algorithms, a deep neural network (DNN) has been developed for effective EEG analysis, especially for feature extraction and classification [13,14,15]. Compared with other DNN methods, convolutional neural network (CNN) has been proven effective for biomarker classification and widely used to process EEG signals in many kinds of fields [16,17,18,19,20,21,22,23], but there are fewer investigations in the field of IA diagnosis. Although some algorithms for IA, such as support vector machine (SVM) [12] and random forest (RF) [24], have demonstrated good performance in preprocessed data, they cannot achieve high accuracy in processing raw data. Thus, considering the time-consuming process of obtaining preprocessed data, these algorithms are obviously not meet the requirement of the efficient and time-saving diagnostics. To solve these limitations, this work is aimed to develop an algorithm that has higher accuracy and a stronger ability to automatically discover hidden features from a huge number of samples and avoid the complicated process of manually extracting features.

In this work, a CNN combined with the fast Fourier transform (FFT) algorithm is proposed to analyze three kinds of EEG data (raw data, full spectral power, and alpha-beta-gamma spectral power) and diagnose IA. Among them, alpha-beta-gamma spectral power is based on the frequency bands associated with IA summarized from previous studies. Compared with the traditional questionnaire analysis [25] and biostatistics method [26], we have obtained more convincing quantitative results, and the accuracy is higher than the current SVM algorithm used for diagnosing IA [27]. In particular, the CNN model achieves 87.59% accuracy, 88.80% sensitivity, and 86.41% specificity based on full spectral power data. By comparing the results of full spectral power and alpha-beta-gamma spectral power, the proposed algorithm can determine whether a person has IA through EEG signals in the absence of biomarkers. In addition, the results show that the proposed algorithm has obvious advantages in processing raw data, which can reduce the diagnosis time and have the potential to be used in some real-time health monitoring systems [28,29,30].

The structure of this paper is: in Section 2, we briefly describe the used EEG signals from the LEMON dataset and propose our CNN model. In Section 3, LEMON dataset together with a baseline neural network is used to perform IA classification quantitatively. The results are compared and discussed based on existing studies in the literature. In Section 4, we conclude this paper.

2. Materials and Methods

2.1. Dataset Preparation

The EEG data from the LEMON dataset are used for extracting features, and a functional connectome phenotyping dataset is the label in our algorithm. Both datasets constitute parts of the Leipzig Mind-Brain-Body interaction research database (MPILMBB), and some of the participants in these two datasets are overlapped [31,32]. The LEMON dataset includes 203 valid resting-state EEG recordings, and each EEG session includes 16 separate parts (lasting 60 s: 8 closed-eyes (EC) parts and 8 opened-eyes (EO) parts) from each participant. The data are recorded by using a 62-channel (61 scalp electrodes plus 1 vertical electrooculogram (VEOG) electrode) EEG with a sampling rate of 2500 Hz. The EEG datasets were preprocessed by Babayan et al. [31]. They excluded data from 13 participants, considering the lack of event information, different sampling rates, mismatched header files, and insufficient data quality. The raw EEG data from 203 participants were down-sampled from 2500 Hz to 250 Hz, and band-pass filtered within 1–45 Hz by using a Butterworth filter with 8th order which can be further improved [33,34]. Then, it is divided into EO and EC conditions and to eliminate the influence of visual activity on EEG signals, EC EEG signals were used for this work. The functional connectome phenotyping dataset can be used to explore advanced cognitive abilities related to the internal functional structure of the brain, self-generated psychological experiences, and personality characteristics in the form of different questionnaires. In our algorithm, we use internet addiction test (IAT) results for classification based on different scores from individuals with or without IA. According to Cronbach’s Alpha coefficient (α) of 0.91, this IAT shows high reliability, and it reflects the severity of IA. Finally, 214 valid IAT results can be obtained in the MPILMBB database.

After comparing 203 EEG recordings with 214 IAT results, 107 participants not only provided EEG signals but also submitted valid IAT results. The results of the IAT are defined as valid if they show a total score of at least 19 points. Considering the possible misclassification due to the inaccuracy of the self-assessment, 24 results with the highest scores (37–59, including 37 and 59) and the lowest 25 scores (19–23, including 19 and 23) are used. In this way, the difference between the two groups of labeled people is as large as possible based on the results of the IAT. In the algorithm, individuals without IA are labeled as 0 and individuals with IA are labeled as 1. The total datasets used in this paper consist of 49 individuals, including 25 healthy control individuals and 24 online users. In the subsequent classification training, around 80% (39 subjects) and around 20% (10 subjects) are respectively chosen as the training set and test set.

2.2. Proposed CNN Architecture

The total CNN framework consists of input and output layers as well as multiple hidden layers. The hidden layers are convolutional layers, pooling layers, and fully connected layers. The input signal is convoluted by a trainable filter and an output map is formed through the output function f. The convolution operation is shown in Figure 2 and Equation (1).

$$y_k={\displaystyle {\sum }_{n=0}^{N-1}{x}_n{h}_{k-n}}$$

(1)

where y is the output vector in which n means nth element, x is the input signal from the previous layer, h is the filter, and N means the number of elements in x. The kth feature map at a given layer is obtained as Equation (2):

$$h_n^k=f(a)=f({(W^k\times x)}_n+b_k)$$

(2)

where W^k is the weight matrix for filter k, and b_k is the bias value.

During the training of the CNN model, the loss function is determined by the output and propagated back to change the interconnection of the network, which is updated based on the reduction of the loss function H(p, q), which is defined in Equation (3).

$$H(p,q)=-{\displaystyle {\sum }_n(p(n)\log q(n))}$$

(3)

where p(n) is the label value marked in advance, and q(n) is the predicted value obtained by the model operation. The pooling operations translate the output of the neuron clusters into one single neuron of the next layer, which can reduce the feature map resolution. For each feature map $a_{i,j}^n$, the output of the pooling layer is determined by Equation (4).

$$y_{i,j}^{n+1}=pool(a_{i,j}^{n+1}),\forall (m,n)R_{i,j}$$

(4)

where R_i,j is the local neighbor around the location (i, j). In this model, the max-pooling layer is used and $y_{i,j}^{n+1}$ is the max number in R_i,j.

The features obtained from the convolutional layer are applied as input to the fully connected layer, which is converted into a one-dimensional feature vector and used for classification. This step of converting the features into a one-dimensional (1D) vector is called flattening which is described in Equations (5) and (6).

$$y_i^{(n)}=f(z_i^{(n)})$$

(5)

$$z_i^{(n)}={\displaystyle {\sum }_{i=1}^{m_i^{(n-1)}}{w}_{i.j}^{(n)}{y}_i^{(n-1)}}$$

(6)

where f is the nonlinear transfer function, and $w_{i.j}^{(n)}$ represents the weight coefficients between the neurons of the fully connected layer. In the fully connected layer, all neurons in the previous layer are connected to the neurons in the current layer, the output of the fully connected network is the clustering result, and then the SoftMax function (Equation (7)) is used for classification in the fully connected network.

$$\sigma {(\mathit{z})}_i=\frac{e^{z_i}}{{\displaystyle {\sum }_{j=1}^n{e}^{z_j}}}$$

(7)

where z is a vector composed of all elements in the fully connected layer.

The proposed CNN model consists of 5 convolutional layers, 5 max-pooling layers, a fully connected layer, and a SoftMax classifier, the model structure is shown in Figure 3.

Conv and MP stand for the convolutional layer and max-pooling layer, respectively. The sizes of the first three convolution kernels are (1, 5), and the last two are (3, 5). The stride of all these five convolution kernels is (1, 3). After each layer of the convolution kernel, there are the Rectified Liner Units (ReLU) activation function and the max-pooling layer, and the ReLU function is defined by Equation (8).

$$f(a)=\mathrm{ReLU}(a)=\mathit{max}(0,a)$$

(8)

where a is defined in Equation (2). Compared with other excitation functions of saturated nonlinear functions (e.g., Sigmoid and Tanh), unsaturated nonlinear functions like ReLU can solve the problems of gradient explosion and gradient disappearance, and accelerate the convergence speed, the function curves are illustrated in Figure 4. The sizes of the five max-pooling layers are all 2 × 2, and the stride of these max-pooling layers is 2. The final classifier of this model includes a fully connected layer and SoftMax function.

To evaluate the performance of this CNN model, the raw EEG data from the LEMON dataset and the spectral power of this EEG dataset are used as input matrices. The raw data is a time-domain signal, which is converted to a frequency matrix after FFT processing. To verify the ability of the CNN model to extract features, we filter a specific range of frequency bands as the input matrix, and further compare the results produced by these three kinds of inputs.

3. Results and Discussion

Three experiments (based on raw data, full spectral power, and alpha-beta-gamma spectral power) were conducted to evaluate the performance of the proposed CNN architecture on the IA diagnosis. In the proposed CNN structure, the independent validation method was applied to ensure the objectivity of the model verification. A detailed representation of the data used for the proposed CNN model is shown in Figure 5.

The EEG session from LEMON datasets is recorded after the Multi 8/2 Drogen-Tauchtest drug strip test. Thus, a five-layer CNN structure was conducted to dig more features, reduce the impact of other addictions, and improve the accuracy of the proposed classification algorithm. Forty-nine subjects were divided into 5 groups, and each group contains 10 subjects except one group that contains 9 subjects. We run the CNN model and randomly divide these five groups as one testing set, one validation set, and three training sets. For the input signals used in the model, the EC-EEG dataset was chosen as the raw data to avoid the influence of insight activity on the EEG data. The total length of EEG time-domain signal from everyone is 8 min. To increase the number of training samples, an 8-s sliding window with a 2-s overlap was used to augment the data.

Two-dimension input samples with a dimension of 38 × 2000 were obtained through the data augmentation process, where 38 is the channel number and 2000 is sampled data points over 8 s. After applying FFT to raw data and corresponding amplitude processing, the signal amplitude for each frequency was obtained, which was squared to express the power spectral density. The processing of FFT is described in Figure 6. The dimension of this kind of input is 38 × 1000 two-dimension, one is 38 channels and the other is a 1000 frequency matrix. The alpha, beta, and gamma bands that may be affected by IA were then filtered according to the conclusions drawn from previous literature reviews [3,11,12], and this two-dimension input signal has 38 channels and 336 frequency matrices. The time-domain curve corresponding to the EEG raw data, the FFT-processing power spectrum, and the filtered alpha-beta-gamma power spectrum of an individual are shown in Figure 7. As in the figure, although the FFT-processed waveform is more clearly displayed in the frequency band than the original time-domain waveform, it is still difficult to identify features related to IA diagnosis through the human eyes. This also verifies the need for tools that are more objective and capable of extracting features.

To assess the performance of the proposed model, sensitivity, precision, specificity, and accuracy are used as the evaluation index for these three types of results (obtained from raw data, full spectral power after FFT, and filtered alpha-beta-gamma spectral power). These indexes are calculated by true positive (TP), true negative (TN), false positive (FP), and false negative (FN) according to Equations (9)–(12), and the results are listed in

Table 1. Results of three types of input signals.

Type of Inputs	Sensitivity	Precision	Specificity	Accuracy
Raw data	0.932	0.749	0.692	0.811
Full spectral power	0.888	0.865	0.864	0.876
Alpha-beta-gamma spectral power	0.619	0.746	0.725	0.677

.

$$\mathrm{sensitivity}=\frac{TP}{TP+FN}$$

(9)

$$\mathrm{precision}=\frac{TP}{TP+FP}$$

(10)

$$\mathrm{specificity}=\frac{TN}{TN+FP}$$

(11)

$$\mathrm{accuracy}=\frac{TP+TN}{TP+TN+FP+FN}$$

(12)

Alpha-beta-gamma spectral power is a previously summarized biomarker related to IA, which contains higher-density useful information for IA diagnosis. As illustrated in Table 1, the results of full spectral power are better than alpha-beta-gamma spectral power in sensitivity, precision, specificity, and accuracy. It verifies the superiority of this CNN model in the feature extraction and proves that the IA diagnostic method based on this model does not require EEG biomarkers, and the extra filtering process based on previous biometrics is useless to our algorithm. The lower accuracy of alpha-beta-gamma spectral power is because the selected spectrum data lose transient wave features that may be useful for classification. Meanwhile, although alpha-beta-gamma spectral power (as input matrices) shortens the model training time, the increased workload of data preprocessing makes it still time-consuming for diagnosis.

After the time domain signal is processed into a frequency domain signal, there will be more obvious peak features, which make the result of full spectral power show better performance. Currently, classification problems related to IA commonly use the methods: SVM [12,35], decision tree model [36], RF [24], etc. Here, we summarize the used dataset and final accuracy of some typical studies in

Table 2. Comparison with other state-of-the-art works.

Reference	Classification Model	Data Source	Accuracy Result
Internet addiction [this work]	CNN	EEG	87.59%
Internet addiction [12]	SVM	EEG	82.50%
Internet addiction [24]	RF	Survey Data	83.00%
Internet addiction [35]	SVM	Browsing Histories	66.67%
Internet gaming disorder [36]	Decision Tree	Survey Data	70.41%

. Compared with them, the algorithm proposed in this work has better performance in accuracy which shows its robustness in diagnosing IA. By contrast to those researches [37,38] that qualitatively diagnose IA through the biometric method, our work also holds the advantage to provide quantitative diagnosis results. Moreover, the accuracy in processing raw data also presents a good performance (81.1%), which evidences the advantages of proposed CNN model in processing raw data (due to its ability to learn high-level features).

4. Conclusions

This study proposes a new method for IA diagnosis based on EEG signals and the CNN model. The full spectral power results show that the proposed algorithm has better performance in terms of IA diagnostic accuracy and can diagnose IA in the absence of biomarkers. Furthermore, the results obtained from raw data also show that the proposed CNN model has an advantage in processing raw data (accuracy > 80%) due to the powerful ability of CNN to discover hidden features. This specific superiority of this CNN model is expected to reduce the diagnosis time and used in real-time health monitoring systems, meanwhile, its generality helps it to be used in a wide spectrum of addiction or neurological disorders diagnosis.