From Pain Detection to Pain Localization: Exploring the Potential of Multimodal Physiological Signals

Abstract

Pain assessment in non-verbal individuals remains a significant clinical challenge. While physiological sensing has shown promise for automated pain detection, identifying the anatomical location of pain from physiological responses remains largely unexplored. In this paper, we investigate pain localization using multimodal physiological signals collected in the AI4PAIN Grand Challenge. We propose a subject-independent physiological pain localization framework that distinguishes among No Pain, Pain Arm, and Pain Hand conditions using handcrafted features extracted from electrodermal activity (EDA), blood volume pulse (BVP), respiration, and peripheral oxygen saturation (SpO2) signals. Feature selection is performed using mutual information ranking and incremental feature evaluation within a GroupKFold cross-validation framework. Experimental results demonstrate that EDA is the most informative modality for pain localization, while the combination of EDA and BVP achieves the best overall performance, reaching 53.24% accuracy on the hidden test set. Although the proposed approach effectively differentiates pain from no-pain conditions, substantial confusion remains between arm and hand pain. These findings suggest that physiological signals reliably encode the presence of pain but provide limited information for discriminating between nearby pain locations. Our work establishes a benchmark for physiological pain localization and highlights important challenges and opportunities for developing more effective objective pain localization systems.