A Reproducible Stress Prediction Pipeline with Mobile Sensor Data

Abstract

Recent efforts to predict stress in the wild using mobile technology have increased; however, the field lacks a common pipeline for assessing the impact of factors such as label encoding and feature selection on prediction performance. This gap hinders replication, especially because of a lack of common guidelines for reporting results or privacy concerns that limit access to open codes and datasets. Our study introduces a common pipeline based on a comprehensive literature review and offers comprehensive evaluations of key pipeline factors, promoting independent reproducibility. Our systematic evaluation aimed to validate the findings of previous studies. We identified overfitting and distribution shifts across users as the major reasons for performance limitations. We used K-EmoPhone, a public dataset, for experimentation and a new public dataset---DeepStress---to validate the findings. Furthermore, our results suggest that researchers should carefully consider temporal order in cross-validation settings. Additionally, self-report labels for target users are key to enhancing performance in user-independent scenarios.