Pairwise Pain Intensity Classification from Multimodal Biosignals: Evaluating Sensor Fusion Across Physiological Modalities

Abstract

Objective pain assessment remains an open clinical challenge, as self-report measures are inherently subjective and unsuitable for patients with impaired communication. This paper investigates the feasibility of discriminating pain intensity levels from wearable physiological signals using multimodal feature fusion and tree-based ensemble classifiers. Four biosignal modalities-Blood Volume Pulse (BVP), Electrodermal Activity (EDA), Respiration (RESP), and peripheral oxygen saturation $\left(\text{SpO}_{2}\right)$-are evaluated across all 15 modality subsets and three pairwise binary classification t asks of i ncreasing d ifficulty: No vs. HIGH pain (NP vs. HP), No vs. LOW pain (NP vs. LP), and HIGH vs. LOW pain (HP vs. LP). Unimodal results reveal a consistent signal hierarchy, with EDA and BVP emerging as the most informative individual channels. Trimodal fusion of BVP+EDA+RESP achieves the best performance for NP vs. $\mathbf{H P}(\mathbf{A U C}=\mathbf{0. 9 4 4})$ and NP vs. $\mathbf{L P}(\mathbf{A U C}=\mathbf{0. 8 4 1})$, while the quadrimodal combination attains the best result for the interlevel HP vs.LP task $(\mathbf{A U C}=\mathbf{0. 7 9 2})$. Among the four classifiers evaluated-Random Forest, Gradient Boosting, XGBoost, and CatBoost-RF and CatBoost delivered the most consistent results, while modality selection proved to be the dominant factor governing classification p erformance a cross a ll configurations. These findings s upport m ultimodal p hysiological f usion a sa viable approach for objective, wearable pain intensity monitoring.