Personalized Artificial Intelligence Detector for Individualized Pain Classification Using Wearable Heart Rate Data

Objective assessment of acute somatic pain remains a major challenge when patients cannot communicate, especially in neurosurgical and neurocritical care. Although acute pain produces reproducible autonomic activation, wearable heart-rate (HR) signals are highly individualized and nonstationary, making population-based thresholds unreliable. We developed a personalized representation-learning framework that maps short-term HR windows into a subject-specific physiologic manifold using a convolutional neural network trained with triplet metric learning. This approach was evaluated in two postoperative neurosurgical patients who together contributed several thousand continuously acquired HR samples, providing dense within-subject physiological coverage suitable for individualized modeling. The learned manifolds yielded clear geometric separation between pain and no-pain states and substantially improved classification performance compared with conventional HR-based models. These findings demonstrate that individualized autonomic embeddings derived from a single wearable sensor achieve strong within-subject separability (leakage-aware AUC: 0.974- 0.980), substantially outperforming raw HR and temporal baselines. This physiology-inspired framework offers a feasible path toward patient-specific pain monitoring, with prospective validation in larger cohorts required before clinical translation.