Arousal And Valence Recognition Of Mental Stress In Academic Tasks By Using Electrocardiogram

In today’s society,social evaluation and performance assessment can be seen everywhere,and such assessments have become common sources of stress.The key impact factor on task performance is whether the individual’s stress response matches the task difficulty.For college students,examination is a typical and common evaluation method,and there are many factors that determine the quality of exam scores,among which students’ achievement motivation is a very important factor.Appropriate achievement motivation triggers positive stress and promotes learning performance.However,excessive or insufficient achievement motivation leads to negative stress and reduces learning performance.In specific populations,detecting high/low stress level,positive/negative attributes of stress responses,as well as the fluctuation of stress response intensity during the task process,provides individuals with objective and quantitative self-evaluation references.This is conducive to understanding the characteristics of their own stress responses in task execution,and further maintains their own good state and promotes task performance through training and adjusting achievement motivation.Aiming to collect and analyze stress neurophysiological response data under real performance evaluation task scenarios,this work collected electrocardiogram signals from college students during the examination period,and obtained heartbeat interval series that reflected the patterns of autonomic nervous activity through signal preprocessing.Statistical tests were used to analyze the inter-group differences in autonomic neurophysiological patterns of the students’ high/low stress levels,the changes of stress level during the examination,and the positive/negative stress attributes.Machine learning models were established to solve the classification problems of high/low stress,changes in stress level,and positive/negative stress attributes.The specific research contents and results of this paper are as follows.(1)Studied the physiological pattern recognition of high/low stress.First,this work marked the high/low stress labels of physiological data samples through the stress selfassessment of subjects,and established training-test set and validation set of high/low stress physiological data.Second,this work extracted 39 heartbeat interval features from the samples,and then used Mann Whitney U-test to analyze the inter-group differences of physiological features of high/low stress samples.Then,this work trained convolutional neural networks and classic classifiers to establish a high/low stress binary classification model.The results showed that there were significant inter-group differences in multiple physiological features of high/low stress.The support vector machine classifier and three-dimensional heartbeat interval features obtained a 92.06%best F1 score for distinguishing high/low stress on a validation set independent of model training.(2)Studied the changes of stress over time during a task.The examination was divided into the early,the middle,and the late phases.This work selected 500 heartbeat intervals as a sample from each examination phase.Binary classification models were established between each pair of the three examination phases through supervised machine learning.Independent sample t-tests were performed on the heartbeat interval features selected by the binary classification models.The results showed that the level of vagal nervous activity was lower at the early phase of the examination,and the level of sympathetic nervous activity was higher at the late phase than that at the middle phase of the exam.(3)Studied the physiological pattern recognition of positive/negative stress.According to the grade ranking order of students’ examination performance,the dataset was divided into high performance and low performance groups,corresponding to positive stress and negative stress groups.Thirty-nine heartbeat interval features reflecting autonomic nervous activity were extracted again,and inter-group differences of physiological features of positive/negative stress samples were analyzed using the Mann Whitney U-test.Then,a supervised machine learning was used to establish a binary classification model of positive/negative stress.The results showed that there were significant inter-group differences in multiple physiological features of positive/negative stress.The support vector machine classifier and two-dimensional heartbeat interval features obtained a best F1 score of 83.02% for distinguishing positive/negative stress on a validation set independent of model training.The research work in this dissertation had the following three findings.(1)High/low stress states had distinguishable neurophysiological patterns.Using machine learning methods to determine high/low stress states was feasible,and the binary classification model had good generalization performance in real examination scenarios.(2)In the analysis of the changes of stress in the task process,the trend of stress level changes of the subjects was that,the stress was the highest in the early phase of the task;the stress was the lowest in the middle phase of the task;the stress in the late phase of the task increased compared to that of the middle phase,but was lower than that of the early phase.(3)Positive/negative attributes of stress had distinguishable neurophysiological patterns.It was feasible that using machine learning methods to determine the positive/negative attributes of stress.But the positive/negative attributes of stress were more difficult to learn than the high/low attributes of stress for artificial intelligence machines.This paper has three innovations as follows.(1)This paper collected and analyzed stress induced by real performance evaluation events.Compared with the stress data collection and analysis under controlled task conditions in the laboratory,the stress information in the data was closer to common daily performance evaluation-type stress responses,and the stress classification model established had the value of generalized application to real scenarios.(2)This paper not only identified high/low stress,but also further identified the pattern of stress changes over time in performance evaluation events.(3)This paper’s analysis of the differences in the physiological characteristics of positive/negative attributes of stress,and the identification of positive/negative stress categories,were among the few similar studies in the literature.