Knowing Where It Hurts? Multimodal Spatio-Temporal Feature Engineering for Pain Location Classification

Abstract

Pain assessment traditionally relies on self-report measures, limiting its applicability to non-verbal patients and individuals with communication impairments. While prior studies have focused on pain detection and intensity estimation, pain location classification from physiological signals remains largely unexplored. This paper presents a Multimodal Spatio-Temporal Feature Engineering Framework for pain location classification using physiological data from the AI4PAIN 2026 Grand Challenge. The proposed framework integrates electrodermal activity (EDA), blood volume pulse (BVP), respiration (RESP), and peripheral oxygen saturation (SpO2) signals sampled at 100 Hz. We extract 225 features, including modality-specific descriptors, 77 spatio-temporal features, and cross-modal interaction measures. To reduce inter-subject variability, we introduce subject-referenced transformations based on baseline-relative normalization and rank-based representations. The resulting features are refined through a variance-correlation-mutual information selection pipeline and classified using an Optuna-optimized soft-voting ensemble of XGBoost, LightGBM, and CatBoost. On the AI4PAIN test set, the best submission achieved 54.4% accuracy, outperforming the official multimodal baseline (39.8%) by 14.6 percentage points. Ablation studies show that the proposed spatio-temporal and multimodal features provide complementary information beyond conventional statistical descriptors. Further analysis reveals that the performance gain is driven primarily by robust pain detection (No Pain vs. Pain, 85% accuracy), whereas Hand Pain versus Arm Pain localization remains near chance level (∼52%), suggesting that fine-grained anatomical pain localization is substantially more challenging than pain detection under the current peripheral sensing setup.