A Feature-Based Multimodal Approach to Pain Detection and Localisation Using Wearable Biosignals

Abstract

Reliable pain assessment remains a major challenge in clinical care, especially for patients who cannot communicate their symptoms effectively. This paper presents a subject-independent, feature-based framework for classifying multimodal biosignals into three states: No Pain, Pain Arm, and Pain Hand using the AI4Pain 2026 benchmark. The proposed system integrates electrodermal activity (EDA), blood volume pulse (BVP), respiration, and peripheral oxygen saturation (SpO2) through segment-level preprocessing, handcrafted feature extraction, and classical machine learning classification. Six models were compared under strict subject-disjoint training, validation, and testing. A support vector machine with one-vs-rest (OvR) decoding achieved the highest validation accuracy of 60.19%, and the same model obtained 53.94% test accuracy (233/432) on unseen subjects. Modality ablation showed that no single channel matched the full four-modality fusion (best unimodal result: 53.94% with BVP), while Arm and Hand pain remained the main source of error. The proposed pipeline offers a transparent and reproducible baseline for multimodal pain detection and localisation.