Two-Stage Cascade Classification for Pain State Recognition from Physiological Signals

Abstract

We present our solution to the AI4PAIN 2026 challenge, which requires classifying 10-second long physiological signal into three pain states: No Pain (NP), Hand Pain (HP), and Arm Pain (AP), from four modalities (blood volume pulse, electrodermal activity, respiration, and blood oxygen saturation) recorded at 100Hz. The core challenge is subject-independence: the model must generalize to unseen individuals with substantially different physiological baselines. Our approach combines per-subject adaptive baseline normalization, multi-scale feature engineering across five feature groups (446 sub-window features with reference-delta, 48 temporal trend features, 96 signal complexity features, 55 base statistical features, and 8 SpO2specific features), and a two-stage cost-sensitive cascade classifier. Stage 1 distinguishes NP from Pain using an importance-masked LightGBM with a tuned cost weight; Stage 2 discriminates HP from AP using a greedy forward-selected subset of signal complexity features (DWT sub-band statistics and DFA exponents on RESP and EDA). Soft probability propagation between stages avoids hard routing errors. In a 5-fold subject-independent cross-validation, the system achieves 64.3% (±1.9%) accuracy. Performance drops to 56.0% on the official held-out validation split and to 50.7% on the blind challenge test set, remaining above the official test baseline of 39.8% but highlighting the difficulty of subject-independent physiological pain recognition.