Integrative Analysis of Open Datasets for Stress Prediction

Abstract

Purpose: Stress is the physiological or psychological response to internal or external stressors. By causing physiological and behavioral changes, persistent stress affects psychological and physiological functionality, potentially reducing the quality of life. Therefore, it is of great interest to systematically monitor individuals daily to investigate their physiological reactions to potential stressors. The aim of this study is to investigate the potential of training reliable and accurate ML and DL models for the prediction of acute stress based on available open datasets. Methods: The open datasets include physiological signals collected using wearable sensors during stress characterization experiments and made available to the public. Various machine and deep learning (ML/DL) models were tested and compared based on their accuracy (ACC). Harmonizing techniques were also included to increase the generalizability of the models and reduce their dependence on sensors and experimental protocols variations. Results: The models achieved test-set accuracies over 80% for different classification problems, while the combination of the available data led to more accurate models for the binary classification problem (over 90%). Harmonizing differences between participants and datasets led to more generalizable models with encouraging initial results (around 70% accuracy) when evaluated with unseen subjects. Conclusion: The extracted models can all be used depending on the classification problem and the experimental settings/protocols. All models could play a significant role towards forming and developing a stress prediction ‘service’, taking into account the impact of using different sensors, and, thus, having to proceed to required adjustments.