Subject-Independent Multimodal Pain Classification via a Calibrated XGBoost and Domain-Adversarial Neural Network Ensemble

Abstract

Subject-independent pain assessment from short multimodal physiological signals is a challenging generalisation problem because participants exhibit highly idiosyncratic autonomic baselines. We address the three-class task of the AI4Pain 2026 Grand Challenge — classifying 10-second segments of blood volume pulse (BVP), electrodermal activity (EDA), respiration, and peripheral oxygen saturation (SpO2) as No Pain, Pain Arm, or Pain Hand — using a calibrated soft-voting ensemble of (i) an XGBoost classifier on hand-crafted domain-knowledge features and (ii) a 1D Domain-Adversarial Neural Network (DANN) trained directly on the raw signals. Per-subject z-score normalisation is applied before feature extraction, and a gradient-reversal layer suppresses subject identity in the deep branch. On the 12-subject held-out validation set (fully disjoint from the 41 training subjects) the ensemble reaches 57.87 % accuracy / 58.28 % macro-F1. On the organiser-evaluated blind test set our primary submission scores 57.41 %, which is numerically close to the runner-up accuracy reported in the previous edition of the challenge; because the 2025 and 2026 subject cohorts differ, this is provided only as a qualitative reference point rather than a direct comparison. The DANN alone lifts a vanilla 1D-CNN by +4.4 pp, attributable solely to suppressing subject identity in the learned representation. Code, trained checkpoints, and four reproducibility scripts are released publicly.