Unveiling Stress: A Comparative Analysis of Multimodal Sensor Fusion Techniques for Predictive Modeling

Abstract

Stress in professional environments is a significant concern. Medical professionals are particularly vulnerable to stress, especially during emergencies. Nurses hold a vital position in delivering care within hospital settings. It's crucial to anticipate stress levels among nurses to help them perform their duties effectively and avoid the long-term impacts of stress. This study seeks to explore how body sensors and machine learning techniques can be employed to monitor physiological signs and identify stress levels among nurses. It utilizes a benchmark dataset collected from 15 different nurses, including signals such as heart rate (HR), Electrodermal Activity (EDA), and Skin Temperature alongside location data extracted by an accelerometer. This study explores several fusion strategies, such as data, model, and prediction fusion levels to improve the accuracy and reliability of stress prediction models. Through a comparative analysis, this paper highlights the strengths and limitations of diverse fusion techniques. shedding light on their efficacy in capturing the nuanced features of stress. The findings offer valuable insights into the optimization of multimodal sensor fusion for enhanced stress prediction, creating pathways for reliable frameworks in the healthcare domain. This research conducted a comparative study between 3 different levels: data-fusion, model-fusion, and prediction-fusion. Prediction-level fusion outperformed both model-level fusion by 1.97% and data-level fusion by 1.26%.