Pain is a prominent non-motor symptom of Parkinson's disease (PD); it may appear in various levels (elevated or diminished) during waking hours and substantially reduces quality of life. Although subthalamic nucleus (STN) signal analysis has dramatically advanced our comprehension of PD, the roles of bilateral STN, the relevant biomarkers, and objective recognition of the pain levels in PD patients remain less understood. We recorded bilateral STN signals from PD patients implanted with adaptive deep brain stimulation (DBS) systems and collected pain rating series during in-hospital recovery. Patients provided pain annotations prior to surgery that inform the location (specific or non-specific) and PD-association with pain (PD-related or non-PD-related). A machine learning model was trained to classify higher versus lower pain states, from the eight pain annotation series of the six patients, using features derived from STN signals. STN activity significantly classified the pain intensity in the PD-related pain group. Feature analysis indicated that STN activity from both sides can impact pain classification, with gamma and beta bands in the contralateral STN and delta and theta bands in the ipsilateral STN exhibiting a prominent role. Our observational study demonstrates a novel approach to decoding pain states and identifying STN biomarkers linked to PD-related pain.