Feasibility of a Support Vector Machine Classifier for Myofascial Pain Syndrome: Diagnostic Case‐Control Study

OBJECTIVES: Myofascial pain syndrome (MPS) is the most common cause of chronic pain worldwide. The diagnosis of MPS is subjective, which has created a need for a robust quantitative method of diagnosing MPS. We propose that using a support vector machine (SVM) along with ultrasound (US) texture features can differentiate between healthy and MPS-affected skeletal muscle. METHODS: B-mode US video data were collected in the upper trapezius muscle of healthy (29) participants and patients with active (21) and latent (19) MPS, using an acquisition method outlined in previous works. Regions of interest were extracted and filtered to obtain a unique set of 917 images where texture features were extracted from each region of interest to characterize each image. These texture features were then used to train 4 separate binary SVM classifiers using nested cross-validation to implement feature selection and hyperparameter tuning. The performance of each kernel was estimated on the data and validated through testing on a final holdout set. RESULTS: The radial basis function kernel classifier had the greatest Matthews correlation coefficient performance estimate of 0.627 ± 0.073 (mean ± SD) along with the largest area under the curve of 91.0% ± 3.0%. The final holdout test for the radial basis function classifier resulted in 86.96 accuracy, a Matthews correlation coefficient of 0.724, 88% sensitivity, and 86% specificity, validating our earlier performance estimates. CONCLUSIONS: We have demonstrated that specific US texture features that have been used in other computer-aided diagnostic literature are feasible to use for the classification of healthy and MPS muscle using a binary SVM classifier.