Human Stress Inference Framework in a Non-Invasive Manner

Abstract

In our daily life, work related stress is recognized as one of the most serious health problem. It can a ect to the emotional and cognitive well-being of individuals as well as the quality of human life. Detection of stress at its early stage, may help to prevent most physiological as well as behavioral disorders. Since stress is a subjective measurement, detection is very challengeable. Use of standard psychological questionnaire is one of commonly used stress detection approach based on analysis of answers given by individuals. But it is not suitable for real time stress detection. Some of studies in this area are focused on physiological signals based stress detection such as analysis of signals like GSR and heart rate variability. But most of them are invasive and expensive techniques and also lead to disturbance of daily activities and changes of the normal behavior of users. Behavioral characteristics based stress detection is another possible way that researches are focused on. It has been demonstrated as a non-invasive stress detection approach. Signs of stress can be seen in the changes of the human behavior. It is visible to the outside environment if there are deviation from the normal behavior patterns when a person becomes stressed. Most of studies were used single models of behavioral characteristics to capture the behavior pattern changes of individuals in stress detection. This study is focused on detecting human stress while working with a computer by considering multiple behavioral characteristics, namely dynamic keystroke patterns, mouse movements and dynamic facial changes. As we all experienced there can be many reasons to become stress when working with a computer such as work overload, strict deadlines and etc. In generally, combination of several classi ers in order to produce an accurate output is most preferred rather than choosing a best individual classi er. This may help to produce an improved performance in more accurate manner. To address the drawbacks of stress detection using single model this study proposed a framework which consists of multiple feature based classi ers. Research study is demonstrated multiple behavioral feature extraction in a personalized manner from individuals to identify patterns and train those, under three di erent neural network classi er systems. It shows that, it can be obtained higher accuracy in stress detection by combining these three classi er systems. It come up with a solution, based on fuzzy integral which is one of most preferred fusion techniques and has high accuracy than measuring stress using single classi er. Finally research study provides cost e ective stress detection approach with increased accuracy.