StO2 Stress Recognition Based on Deformable Convolution and Multimodal Feature Blender

Stress is an emotional state that affects our lives, so accurately distinguishing different types of stress are important for maintaining individual health and social stability. A novel approach using Tissue Oxygen Saturation (StO2), detected via hyperspectral technology, offers a more accurate, non-invasive alternative. Unlike culturally influenced indicators such as facial expressions, StO2 is inherently culture-agnostic, as human physiological structures and blood oxygen responses remain consistent across ethnicities. We propose a new StO2-based stress recognition method named as MDNet. It combines Graph Convolutional Networks and Convolutional Neural Networks with Multimodal Feature Blender (MFB), to adaptively move candidate values to regions of higher importance, solving the image processing challenges in pressure recognition. ReLoss loss function framework is introduced, simplifying hyperparameter tuning and enhancing performance. Our study demonstrates an 80.48% stress classification accuracy, a 1.91% improvement over the previous best method, highlighting the robustness and cross-cultural applicability of StO2-based stress detection.