Data and methods Accepted project proposals Prediction of Valence and Arousal from Subjective Measurements in Naturalistic Settings

Author: Selin Acan, Erno Hermans
Affiliation(s): Donders Center for Cognitive Neuroimaging
Keywords: Valence, Arousal, Machine learning, Stress Level, Artifact Detection, Feature Extraction
Research question(s): Is it possible to predict valence and arousal scores, activity levels and sleep scoring from a data collected in the daily life settings with the wearable devices?

Abstract:

In the contemporary world, stress and overall well-being have emerged as significant concerns due to social and economic challenges. Modern lifestyles and increasing pressures contribute to stress and sleep-related disorders, impacting productivity, healthcare costs, and quality of life. This public health issue affects economic stability and workforce efficiency globally, with organizations, particularly educational institutions, experiencing employee burnout, reduced productivity, and increased absenteeism.

Our research addresses this multifaceted issue by developing a valence and arousal prediction system to mitigate stress's adverse effects. Leveraging advances in wearable technology and machine learning, we aim to monitor and predict stress patterns, facilitating early intervention and personalized health management. Our study is distinct in differentiating between valence and arousal in stress prediction, mapping Ecological Momentary Assessment (EMA) data to a circumplex model of affect, and utilizing a large dataset of around 1000 participants, enhancing confidence in our findings. Our multimodal system integrates data from wearable devices and EMA, providing a comprehensive approach to stress detection and management. The developed models are developed to be further integrated into a single case experimental study. Activity detection and sleep staging algorithms will follow later.

We employ sophisticated data preprocessing, feature extraction, and machine learning techniques, including a modified AlexNet model and Random Forest classifier, to analyze raw and aggregated physiological data. Our methodology includes rigorous artifact detection, normalization, and statistical analysis, ensuring robust and accurate predictions. The study's insights can inform the development of adaptive, real-time stress management
solutions, contributing to healthier individuals and more resilient societies.