Assessing Pain Through Multimodal Physiological Signals Using Traditional Machine Learning in AI4Pain 2026

Abstract

We present our participation in AI4Pain 2026, an initiative focused on developing artificial intelligence (AI) models for automatic pain assessment through multimodal physiological signals, including Electrodermal Activity (EDA), Blood Volume Pulse (BVP), Respiration (RESP), and Peripheral Oxygen Saturation (SpO2). In contrast to recent deep learning approaches, this work relies on traditional machine learning methods combined with handcrafted signal representations for pain detection and localization, comparing two complementary pipelines: a multi-class stacking approach and a two-step hierarchical approach. Our participation pursued two main objectives. The first was to validate the usefulness of handcrafted representations for processing physiological signals and to evaluate their capability to identify the origin of pain. The second was to analyze the complementarity among these representations by assessing how the integration of features extracted from different physiological modalities affects the performance of traditional classification models. Overall, we aim to provide evidence regarding the potential of conventional machine learning and feature engineering strategies for multimodal pain assessment. Our results show that EDA and BVP are the two most relevant signals for this task, that their fusion yields a modest improvement over unimodal classifiers, and that most of the discriminative information is captured by temporal descriptors related to signal magnitude, variability, and statistical distribution.