Machine learning identifies distinct movement control impairment clusters in patients with chronic neck pain.

Patients with chronic neck pain experience various impairments, with reduced movement control suggested as a significant contributing factor. The heterogeneity of this patient population and suboptimal rehabilitation outcomes suggests the existence of latent subgroup characteristics. The aim of this study was to identify distinct groups among patients with neck pain based on the movement control test and pain intensity and to provide information on cluster-specific impairments. 135 patients with idiopathic neck pain performed a movement control test (the Butterfly test) at three difficulty levels and were assessed for pain intensity, providing 13 different parameters (classifiers). Louvain, hierarchical and k-means clustering methods were applied and the number of clusters determined by observing the symmetry and size of silhouette scores. Further, different machine learning algorithms were applied to develop and evaluate a classification framework (based on AUC, classification accuracy, sensitivity, and specificity) and to provide information on individual cluster characteristics using the Shapley Additive Explanations. The k-means and deep learning neural network methods provided the most efficient clustering and classification performance extracting 4 meaningful clusters. Patients between groups differed in the amount of impairment, with cluster 2 and 1 representing the most severe impairments and with clusters 3 and 4 the least severe impairments. Additionally, specific motor control impairments were observed in individual clusters suggesting distinct neck movement control adaptations. Identifying subgroups of patients with neck pain and their specific characteristics based on the results of the Butterfly test may inform future development of targeted rehabilitation strategies.