Machine learning prediction of self-reported real-world knee osteoarthritis pain from wearable accelerometer-derived gait features

Osteoarthritis (OA) is the leading cause of chronic pain and mobility limitations, affecting over 32 million adults in the United States. Traditional assessments of OA pain and physical function rely on subjective self-reports and periodic clinical evaluations that are often limited by recall bias and insufficient sampling of real-world behavior. Recent advances in wearable technology offer new opportunities for objective, passive, and ecologically valid monitoring of daily movement patterns. This study investigates the feasibility of using gait quality features derived from seven-day continuous AX3 accelerometer recordings that are then analyzed using a structured analytical and ML framework to predict OA pain self-reported OA pain using the WOMAC questionnaire’s pain subscale. The data utilized in this study was from secondary analyses of data collected in the Multicenter OA Study (MOST), a large longitudinal cohort of adults aged 50 and older with or at risk of knee OA. We analyzed data of 1525 individuals including 1124 with OA pain and 401 without pain. Gait metrics were derived along the vertical (V), medio-lateral (M), and anterior–posterior (AP) axes and organized into two categories: activity-related measures (e.g., walking duration and step count) and indicators of gait quality, including rhythm, symmetry, variability, and amplitude. Using an Extreme Gradient Boosting (XGBoost) classifier, the framework achieved strong predictive performance (AUC = 0.91, accuracy = 0.93, precision = 0.94, specificity = 0.99, sensitivity = 0.91, F1 = 0.92). Gait quality metrics, particularly variability features such as stride regularity and dominant frequency characteristics, had superior predictive power compared to general activity measures. These findings indicate that subtle gait dynamics represent reproducible behavioral signatures and are candidates for digital biomarkers associated with perceived pain states. They reflect functional and behavioral adaptations to pain rather than direct measurements of underlying nociceptive processes. Accordingly, the proposed framework enables objective estimation of self-reported pain from real-world gait patterns. The primary contribution of this work is the application of a rigorous machine-learning framework for wearable-based gait analysis of a large real-world osteoarthritis cohort, and demonstrating that gait-quality features derived from extended free-living monitoring contain robust behavioral signatures associated with self-reported OA pain. Our findings suggest that real world gait signatures could be leveraged as scalable, non-invasive digital biomarkers for continuously monitoring perceived pain and functional impairment in real-world healthcare contexts.