Stress management is related to public healthcare and quality of life; an accurate stress classification method is necessary for the design of stress monitoring systems. Therefore, the goal of this study was to design a novel stress classification model using a deep learning method.
In this paper, we present a stress classification model using the dataset from the sixth Korea National Health and Nutrition Examination Survey conducted from 2013 to 2015 (KNHANES VI) to analyze stress-related health data. Statistical analysis was performed to identify the nine features of stress detection, and we evaluated the performance of the proposed stress classification by comparison with several stress detection models. The proposed model was also evaluated using Deep Belief Networks (DBN).
We designed profiles depending on the number of hidden layers, nodes, and hyper-parameters according to the loss function results. The experimental results showed that the proposed model achieved an accuracy and a specificity of 66.23% and 75.32%, respectively. The proposed DBN model performed better than other classification models, such as support vector machine, naive Bayesian classifier, and random forest.
The proposed model in this study was demonstrated to be effective in classifying stress detection, and in particular, it is expected to be applicable for stress prediction in stress monitoring systems.
Stress is a prevalent risk factor for multiple diseases; therefore, an accurate and efficient prediction of stress levels could provide a means for targeted prevention and intervention in the personal healthcare domain [
Stress classification and prediction techniques using machine learning methods, such as support vector machine (SVM) and
Recently, prediction models have been based on artificial intelligence, and many methods using machine learning and statistics have been proposed for data mining in the healthcare sector [
In this paper, we propose a DBN-based stress classification model that uses stress-related physical activity and lifestyle data obtained from the 2013–2015 Korea National Health and Nutrition Examination Survey (KNHANES VI) database [
The remainder of this paper is organized as follows. Section II presents the proposed system and data processing method. Section III describes the design and implementation of the proposed stress classification by a deep learning model using the KNHANES VI dataset. Section IV discusses the experimental results and provides conclusions.
The research structure of this work is presented in
Clustering was conducted to investigate the possibility of stress classification through unsupervised learning with the statistical analysis data. In the DBN modeling, the statistical analysis data was used as a feature, and it was observed that stress classification was possible based on DBN. Finally, to evaluate the performance of the DBN model, we compared the stress classification results obtained using the statistical analysis data by existing models and the DBN model.
This study analyzed the records of adults aged from 19 to 80 years of age from the dataset obtained by the Health Questionnaire and Nutrition Survey conducted during the 2013–2015 KNHANES VI. In KNHANES VI (2013–2015), the questionnaire responses were classified into four categories of stress cognition. Of these, only the stress cognitive group (feeling a lot of stress and hardly feeling) was selected. The variables for classifying stress were extracted based on domestic studies on the relationship between stress, physical activity, and lifestyle. Input variables for learning included age, gender, sleeping time, pulse rate, systolic blood pressure (SBP), diastolic blood pressure (DBP), height, weight, and body mass index (BMI) as well as smoking and drinking behavior. Output variables included whether the subjects were stressed or not.
There was a total of 22,948 experimental records from KNHANES VI (2013–2015). Except for the uncertain (nonrespondent, null value) respondents, there were 14,622 records. Of the 14,622 records, the number of people who felt strongly stressed was 647, and the number of people who did not feel stressed was 2,533. Therefore, the number of stressed people was absolutely insufficient. Thus, the group that did not feel stressed was referred to as low-stress; and the group that felt very stressed was referred to as high-stress. A sample of 7 out of the 647 people who were severely stressed was excluded. The reason for this is that only 80% of the final dataset is learned with data, and we anticipated that the factor of the final dataset should be a whole number. Therefore, a total of 640 samples were extracted from the 2013–2015 data for each group, and the final dataset comprised 1,280 records.
A
IBM SPSS Statistics 24.0 (IBM, Armonk, NY, USA) was used for the statistical analysis. A significance level of
A confusion matrix was used to compare classification ability. The confusion matrix was mainly used as a performance evaluation indicator of the model. Accuracy, sensitivity, and specificity were measured as shown in
A DBN is a deep layer neural network with multiple layers of RBM [
A DBN is divided into two stages. The first stage is unsupervised pre-training that learns features only with input values without labels. The process is carried out as follows. The input value learns the first hidden layer x, which in turn, learns the second hidden layer. The second stage is tuning with an error back propagation algorithm using a label with supervised fine-tuning [
An RBM is a generative stochastic neural network that can learn a probability distribution over its set of inputs. A joint configuration (
In this study, DeepLearning4J (DL4J), a Java-based toolkit for building, training, and distributing neural networks, was used in the DBN. DL4J is a domain-specific language used to configure deep neural networks, which are made of multiple layers. In the DL4J platform, hyperparameters are variables that determine how a neural network learns with Java syntax [
The distribution of physical activity/lifestyle records, according to the 1,280 stresses recorded in the study, is shown in
The performance of the model depends on the number of hidden layers, nodes, and hyperparameters. To design a proper DBN-based stress classification model, we varied the number of layers, nodes and hyperparameters.
The DBN has one hidden layer and seven input nodes (gender, age, sleep time, pulse rate, SBP, BMI, drinking, and smoking), six hidden nodes, and two output nodes (low-stress, high-stress). The hyperparameters, namely, batch size, epoch, L2 regularization, learning rate, and momentum were set to 768, 150, 0.007, 0.0002, and 0.1, respectively. The statistically analyzed dataset was divided into a training set (which comprised 80%) and a testing set (which comprised 20%). As shown in
We compared the performance of the proposed DBN with various models, namely, naive Bayesian (NB), decision tree (DT) and SVM. The proposed statistical DBN (using seven input variables) sensitivity, specificity, and accuracy results are shown in
In this study, we classified stress based on DBN using physical activity and lifestyle data. Data were obtained from the KNHNES from the 2013–2015 period.
In this study, the goal was to design a novel stress classification model using a deep learning method. Therefore, we presented a stress classification model, which was evaluated by using a total of 14,622 experimental records of the KNHANES VI (2013–2015) dataset to analyze stress-related health data. The statistical analysis data was used as a feature, and it was observed that stress classification was possible based on the proposed DBN model. We designed profiles based on the number of hidden layers, nodes, and hyperparameters according to the loss function results. The experimental results showed that the proposed model achieved an accuracy and a specificity of 66.23% and 75.32%, respectively. The proposed DBN model performed better than other classification models, namely, support vector machine, naive Bayesian classifier, and random forest. The model proposed in this paper was demonstrated to be effective in classifying stress detection, and it is expected to be applicable for stress prediction in stress monitoring systems.
This work was supported by the Industrial Strategic Technology Development Program (No. 10073159, Developing mirroring expression based interactive robot technique by non-contact sensing and recognizing human intrinsic parameter for emotion healing through heart-body feedback) funded by the Ministry of Trade, Industry & Energy, Korea. This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. 2017R1D-1A1B03035606).
SBP: systolic blood pressure, DBP: diastolic blood pressure, BMI: body mass index.