Pain Presence, Not Location

Automated pain assessment from physiological signals has advanced rapidly, yet has focused almost entirely on detecting the presence or intensity of pain. Whether peripheral physiological signals also encode the anatomical location of a painful stimulus remains largely unexamined. We study this on the AI4Pain 2026 dataset, which provides electrodermal activity, blood volume pulse, respiration, and oxygen saturation for a balanced three-class problem: No Pain, Pain Arm, and Pain Hand. We build a simple, reproducible subject-independent pipeline that combines label-free per-subject normalization, interpretable features, and standard classifiers. On the official validation set our best model reaches 56.5% accuracy (95% CI 49.8 to 62.5) against a 41.7% challenge baseline, and 56.2% on the held-out test set, confirming that the accuracy generalizes to unseen subjects. A controlled analysis, however, traces this accuracy almost entirely to pain detection: No-Pain versus Pain separates cleanly, whereas Arm-vs-Hand localization stays at chance under standard features, a null that holds across four classifier families, individual subjects, a label-permutation control, and a model-independent information bound. The usable signal is carried by blood volume pulse and electrodermal activity, whereas respiration and oxygen saturation are uninformative and dilute naive fusion. Systemic autonomic responses thus index pain arousal but not stimulus site under standard peripheral features, with direct implications for physiological pain-localization systems.