Stress is a common condition that has major health impacts. Even though short-term stress helps people to react to danger immediately, chronic stress has larger health impacts such as early aging, post-traumatic stress disorder, and depression. In this thesis, I propose a continuous stress monitoring approach that can potentially lead to an improved understanding of stress. The approach applies machine learning algorithms to data collected from embedded sensors on a commercially-available wearable device in order to automatically recognize physiological symptoms of stress. Existing approaches for stress detection have typically required specialized sensor equipment or extensive manual labeling of data collected by a researcher in a laboratory setting. A distinguishing feature of our approach is the application of semi-supervised learning algorithms to a data set collected from study participants as they pursue their everyday activities in natural settings. Such an approach provides a foundation for a wearable system that can serve as a platform for use by researchers who wish to collect data about stress from a large population of study participants over extended periods of time, as well as a health and well-being application that alerts users to periods when they are experiencing physiological symptoms associated with stress. Preliminary results for a small study with 8 participants show that my approach can classify physiological symptoms of stress with an average accuracy of 86.3% in per-subject analysis and 75.9% in between-subjects analysis.
