Machine learning-based classification of nonspecific neck pain using movement trajectories of scapular upward rotation.

Objective: Nonspecific neck pain (NSNP) has been associated with altered scapular kinematics, but the relationship remains controversial due to inconsistent findings. This study aimed to develop and validate a machine learning (ML) model for predicting NSNP based on scapular upward rotation (SUR) trajectory patterns using a simplified 2D approach. Methods: A total of 332 public service office workers (240 with NSNP, 92 asymptomatic) participated. SUR was recorded using smartphone video and analyzed. Three kinematic variables were measured: horizontal displacement (HD), vertical displacement (VD), and horizontal-to-vertical ratio (HVR). Six ML algorithms were implemented and evaluated using leave-one-subject-out cross-validation (LOOCV) and multiple performance metrics. Partial dependence plots (PDPs) were generated to visualize relationships between movement parameters and NSNP probability. Results: K-Nearest Neighbors and Support Vector Machine algorithms demonstrated good classification performance (area under the receiver operating characteristic curve (AUC): 0.879 and 0.827 in LOOCV; 0.873 and 0.866 in independent testing, respectively) in identifying NSNP based on scapular movement patterns. The remaining algorithms showed substantially lower performance (AUC range: 0.487-0.560). PDPs revealed nonlinear relationships between kinematic variables and NSNP probability. 2D plots identified specific combinations of horizontal and vertical displacement associated with lower NSNP risk (HD: 4.00-6.00 cm, VD: 6.00-10.00 cm). Conclusions: ML analysis of SUR trajectories demonstrates powerful predictive capabilities for identifying NSNP. The nonlinear relationships and interaction effects revealed through this analysis help explain previously conflicting findings in the literature and provide quantitative targets for clinical assessment and intervention.