This study aimed to predict temporomandibular disorder (TMD) using machine learning (ML) approaches based on measurement parameters that are practically acquired in clinical settings. 125 patients with TMD and 103 individuals without TMD were included in the study. Pain intensity (with visual analog scale), maximum mouth opening (MMO) and lateral excursion movements (with millimeter ruler), cervical range of motion (with goniometer), pressure pain threshold (PPT; with algometer), oral parafunctional behaviors (with Oral Behaviors Checklist), psychological status (with Hospital Anxiety and Depression Scale), and quality of life (with Oral Health Impact Profile) were evaluated. The measurements were analyzed via over 20 ML algorithms, taking into account an extensive parameter tuning and cross-validation process. Results of variable importance were also provided. Bagging algorithm using Multivariate Adaptive Regression Spline (MARS) algorithm (accuracy = 0.8966, area under receiver operating characteristic curve = 0.9387, F1-score = 0.9032) was the best performing model regarding the performance criteria. According to this model, the 5 most important variables for predicting TMD were pain intensity, MMO, lateral excursion and PPT values of masseter and temporalis anterior muscles, respectively. The Bagging algorithm using the MARS algorithm is a robust model that, in combination with clinical parameters, assists in the detection of patients with TMD in settings with limited capabilities. The clinical parameters and ML algorithm proposed in this study may assist clinicians inexperienced in TMD to make a preliminary detection of TMD in clinics where diagnostic imaging tools are limited.

Background: The data extracted from various fields inherently consists of extremely correlated measurements in parallel with the exponential increase in the size of the data that need to be interpreted owing to the technological advances. This problem, called the multicollinearity, influences the performance of both statistical and machine learning algorithms. Statistical models proposed as a potential remedy to this problem have not been sufficiently evaluated in the literature. Therefore, a comprehensive comparison of statistical and machine learning models is required for addressing the multicollinearity problem. Methods: Statistical models (including Ridge, Liu, Lasso and Elastic Net regression) and the eight most important machine learning algorithms (including Cart, Knn, Mlp, MARS, Cubist, Svm, Bagging and XGBoost) are comprehensively compared by using two different healthcare datasets (including Body Fat and Cancer) having multicollinearity problem. The performance of the models is assessed through cross validation methods via root mean square error, mean absolute error and r-squared criteria. Results: The results of the study revealed that statistical models outperformed machine learning models in terms of root mean square error, mean absolute error and r-squared criteria in both training and testing performance. Particularly the Liu regression often achieved better relative performance (up to 7.60% to 46.08% for Body Fat data set and up to 1.55% to 21.53% for Cancer data set on training performance and up to 1.56% to 38.08% for Body Fat data set and up to 3.50% to 23.29% for Cancer data set on testing performance) among regression methods as well as compared to machine algorithms. Conclusions: Liu regression is mostly disregarded in the machine learning literature, but since it outperforms the most powerful and widely used machine learning algorithms, it appears to be a promising tool in almost all fields, especially for regression-based studies including data with multicollinearity problem.