Multimodal data fusion for accurate psychological stress assessment using HRV and facial features

In recent years, multimodal data analysis has gained widespread attention in psychological stress research. This study designs an experimental framework based on the Montreal Stress Model Theory, using mental arithmetic tasks as a stress induction method, to develop a stress assessment model based on physiological and behavioral features. The experiment is divided into three phases: no stress, moderate stress, and high stress. A comprehensive stress induction and analysis system is constructed using multimodal data collection techniques, including electrocardiogram (ECG) signals and facial expressions. In terms of data processing, the VGG16 deep learning model is used to extract facial video features, and a set of physiological features is constructed by calculating the time-domain and frequency-domain features of heart rate variability (HRV). A multimodal feature set containing both physiological and behavioral data is created. The TabNet model is employed for classification and identification of different stress states. Experimental results show that this method significantly improves the accuracy of stress state assessment, with the multi-parameter fusion model achieving an identification accuracy of 94.56%, providing crucial technical support for intelligent psychological stress assessment and personalized interventions.