Evaluating contextually-filtered and temporal-trend features for stress detection using mobile and wearable sensor data

Abstract

Stress detection in everyday life has always been a serious issue as more young adults suffer from different levels of stress. Mobile devices have made it possible to predict in-situ stress in everyday life using smartphone sensor data and smart wearables. In previous studies, researchers usually only leverage the sensor data right before the stress label and only extract features from low-level raw sensor data. In this thesis, we tried to “look back at the long past”, in other words, add long-term trend features. We also first mined routine behavior rules from the raw sensor data and predicted stress using high-level routine behavior features(contextually-filtered features). Our results show that adding long-term trend features and contextually-filtered features didn’t significantly improve the performance. Instead, compared with using current sensor data features right before the stress label, adding immediate past features, such as time-windowed features in a 160-minute time window right before the stress label, did improve the performance. We leverage Shap(SHapley Additive exPlanations) for feature importance interpretation. The results reveal that aside from other types of emotion labels such as valence and arousal which were also used for stress detection, UV index level and general app usage are the best predictors of stress level.