Machine learning models for classifying non-specific neck pain using craniocervical posture and movement

Objective Physical therapists and clinicians commonly confirm craniocervical posture (CCP), cervical retraction, and craniocervical flexion as screening tests because they contribute to non-specific neck pain (NSNP). We compared the predictive performance of statistical machine learning (ML) models for classifying individuals with and without NSNP using datasets containing CCP and cervical kinematics during pro- and retraction (CKdPR). Design Exploratory, cross-sectional design. Setting and participants In total, 773 public service office workers (PSOWs) were screened for eligibility (NSNP, 441; without NSNP, 332). Methods We set up five datasets (CCP, cervical kinematics during the protraction, cervical kinematics during the retraction, CKdPR and combination of the CCP and CKdPR). Four ML algorithms–random forest, logistic regression, Extreme Gradient boosting, and support vector machine–were trained. Main outcome measures Model performance were assessed using area under the curve (AUC), accuracy, precision, recall and F1-score. To interpret the predictions, we used Feature permutation importance and SHapley Additive explanation values. Results The random forest model in the CKdPR dataset classified PSOWs with and without NSNP and achieved the best AUC among the five datasets using the test data (AUC, 0.892 [good]; F1, 0.832). The random forest model in the CCP dataset had the worst AUC among the five datasets using the test data [AUC, 0.738 (fair); F1, 0.715]. Conclusion ML performance was higher for the CKdPR dataset than for the CCP dataset, suggesting that ML algorithms are more suitable than classical statistical methods for developing robust models for classifying PSOWs with and without NSNP.