High-Resolution Physiological Stress Prediction Models based on Ensemble Learning and Recurrent Neural Networks

High-resolution stress detection is an essential requirement for designing time- and event-based stress monitoring systems as a building block for mobile and e-health systems aimed at supporting personalised treatments, both in clinical and remote settings. However, most of the existing solutions focus on binary or few-class stress detection, thus providing a limited feedback and reducing their utility and applicability in real- world scenarios. In this paper we present an alternative approach that overcomes the standard formulation of stress detection as supervised classification problem, by using ensemble learners and recurrent neural networks (RNNs) as the most relevant models for solving time series regression tasks. We trained and tested models using WESAD, a public multimodal wearable dataset for stress and affect detection, and we defined and computed stress scores based on various validated questionnaires stored in the dataset to be used as ground truth. Leave-One- Subject-Out (LOSO) cross-validation scheme has been applied to test the generalisation capabilities of each model in predicting individual stress scores. Results show that Nonlinear AutoRegressive network with eXogenous inputs (NARX), Random Forest (RF), and Least-Squares Gradient Boosting (LSBoost) provide high-resolution personalised stress predictions for the majority of analysed subjects. The proposed predictive models may be integrated as support to decision making into a Decision Support System (DSS) for online stress monitoring, with the main goal to design personalised stress management and alleviation strategies related to the inferred stress severity.

I. Introduction

In the last decades, understanding stress has become a major concern worldwide, since research has extensively demonstrated its negative impact on health and well-being, especially on workers as well as older people [1]. The essential step to tackle the problem of stress is to accurately measure it. Currently used methods include psychological questionnaires and cortisol sampling, but the most objective, precise, unobtrusive, and continuous stress monitoring can be achieved through physiological signals [2]. A huge variety of physiological signals have been employed as stress markers, ranging from (but not limited to) heart, brain and eye activity, up to skin response, respiration, facial expressions, and body gestures [3]. Currently, m-health and e-health solutions exploit the sensing capabilities of both wearable and smartphone devices to provide solutions for stress monitoring and detection in several contexts. In a previous work, we developed a m-health solution to monitor and evaluate the relationship between cognitive performances and physiological stress response during specific cognitive and motor training protocols in institutionalised frail older adults [4]. The m-health solution is composed by commercial wearable sensors and a mobile application for smartphone or tablet. We conducted a pilot study on frail older adults suffering from Mild Cognitive Impairment (MCI) and we used the collected multimodal physiological dataset to perform binary stress detection using several machine learning (ML) techniques. We exploited the study protocol as implicit ground truth for labeling stressful and non-stressful events, since no self-report questionnaires or clinical evaluations were used to track the user perceived stress during the study. However, one of the major open challenges in this research field is the improvement of the stress detection resolution (i.e. increasing the number of detected stress levels), in order to provide more realistic and useful results [5]. For this reason, we proposed to enhance the m-health solution through the definition of a novel Decision Support System (DSS) for online stress monitoring with a higher resolution and accuracy. In the proposed DSS architecture, physiological data are processed in a time- or event-based manner on the mobile node and then fed to proper stress predictive models. The DSS also integrates several self-reports questionnaires and/or clinical evaluations as ground truth for both model training and evaluation.