Reliable shoulder function and pain assessment tools are critical for managing patients with adhesive capsulitis (AC). Particularly, objective pain assessment plays an important role, which could support just-in-time treatment or intervention, monitor short-term and temporal dynamic within-person changes, and provide real-time feedback. Currently, pain assessment for AC still relies on a self-report approach that often suffers issues in substantial recall biases, social desirability, and measurement error. To augment typical self-report for clinical decision-making and treatment in AC, the present pilot study proposed a novel pain assessment tool using wearable inertial measurement units (IMUs) and machine learning (ML) approaches. Twenty-three patients with AC performed 5 shoulder tasks and reported pain scores based on the shoulder pain and disability index. Two wearable IMUs were placed on the wrist and arm to collect upper limb movement signals while performing shoulder tasks. We analyzed correlations between pain scores and IMU feature categories (e.g., smoothness, power, and speed). The results revealed that smoothness-related features exhibited higher Spearman correlations with patient-reported pain scores than power and speed features. Meanwhile, we built pain prediction models with the extracted IMU features and different ML approaches. The ML-based pain prediction model using Gaussian process regression showed strong and significant Spearman correlations (0.795, p < 0.01), with a mean absolute error of 5.680 and root mean square error of 6.663.