Stacked Ensemble Machine Learning Model for Psychological Stress Prediction Using Behavioral Smartphone Sensing Data

Abstract

This paper uses an Ensemble Machine Learning Methodology in order to create a predictive model of Subjective Stress Levels based on Passive Smartphone Sensing Data. In addition to behavioral characteristics (screen usage, mobility, etc.) from the StudentLife Dataset that reflect how people behave when they experience stress, the authors develop a modeling framework that utilizes multiple base learners with an XGBoost Meta-Learner to better generalize to User Variability. The results indicate that this methodology provides higher accuracy than each model individually and provide evidence that Scalable, Artificially Intelligent Systems can be utilized as Stress Monitoring Devices.