A Two-Stage Hierarchical Classifier for Pain Detection and Localisation from Physiological Signals

Abstract

Assessing pain in patients who cannot self-report remains an open clinical problem. The AI4Pain 2026 Grand Challenge extends this task from detection to localization, requiring three-class discrimination of No-Pain, Pain-Arm and Pain-Hand from four peripheral physiological signals (blood volume pulse, electrodermal activity, respiration and oxygen saturation) recorded from 65 subjects. We propose a two-stage hierarchical classifier: the first stage detects the presence of pain and the second distinguishes arm from hand. Both stages are calibrated L2-regularized logistic-regression models, and every design choice was selected under repeated subject-fold cross-validation rather than on the leaderboard split. On the official test set the model attains 0.572 accuracy, 17.4 points above the multimodal SVM baseline, with a validation-to-test gap far smaller than those reported by recent challenge winners. A stage-wise analysis reveals a marked asymmetry: pain detection is accurate (AUC= 0.895), whereas arm-versus-hand localization remains near chance (AUC≈0.55), and no intervention (richer features, non-linear models, late fusion, or a variant given the ground-truth No-Pain labels) raises it. We interpret this as a ceiling on the arm–hand information carried by peripheral autonomic signals under the present sensing configuration: these signals reflect pain-related autonomic arousal rather than its somatotopic location. Whether longer temporal contexts or sensors closer to the stimulation site can exceed this ceiling remains an open question.