Spatial Pain Recognition through Physiological Signal Fusion: A Deep Learning Perspective on Anatomical Localization

Abstract

Objective pain assessment remains a critical challenge in clinical practice, with most automated systems focusing solely on pain intensity while overlooking its anatomical location. This study presents a multimodal deep learning framework designed to simultaneously detect pain and distinguish its spatial origin, i.e. whether the stimulus occurred on the forearm or hand, using physiological signals from the AI4PAIN 2026 Challenge dataset. Electrodermal activity (EDA), blood volume pulse (BVP), respiration, and peripheral oxygen saturation (SpO2) were acquired from 65 healthy participants subjected to transcutaneous electrical nerve stimulation at two locations and two intensity levels. Several deep learning architectures were explored, including LSTM variants, temporal convolutional networks (TCN), hybrid TCN-LSTM models, and a novel Hierarchical Heterogeneous Pain Graph (HHPG) integrating graph neural networks with gradient boosting. Spectral EDA features, particularly TVSymp and phTFS, consistently outperformed other signal modalities, exceeding 53% accuracy. The parallel TCN-SBULSTM model achieved the highest test accuracy of 52.55%, outperforming the official Gaussian SVM baseline by approximately 13%. Pain localization on the arm remained the most challenging classification task across all models, highlighting the difficulty of distinguishing anatomically specific pain signatures from peripheral physiological signals. This solution was submitted to the AI4PAIN 2026 Challenge.