Predictive performance of supervised machine learning models for temporomandibular joint disc displacement in adolescents diagnosed by magnetic resonance imaging

Background: Temporomandibular joint (TMJ) disc displacement is a common condition among adolescents and often requires accurate imaging-based diagnosis. This study aimed to assess the predictive performance of supervised machine learning (ML) algorithms in classifying TMJ disc displacement in adolescents based on morphological and signal intensity features derived from Magnetic Resonance Imaging (MRI). Methods: MRI data from 260 TMJs of 130 adolescent patients (aged 12– 18 years) were retrospectively analyzed. Morphological variables included condylar shape, anteroposterior and mediolateral diameters, and a condylar ratio derived from these measurements. Signal intensity ratios of the superior and inferior heads of the lateral pterygoid muscle were also recorded. Disc status was initially categorized as normal, anterior displacement with reduction, or without reduction; for modeling purposes, dislocated groups were combined. ML analyses were performed using a patient-level data split to prevent information leakage, ensuring that both joints from the same patient were assigned to the same dataset. Fifteen supervised classification models were trained using group-aware cross-validation within the training set, followed by evaluation on a held-out test set. Hyperparameters were optimized exclusively within the training data. Model performance was assessed using standard classification metrics, with 95% confidence intervals estimated by patient-level bootstrap resampling. Results: Anteroposterior and mediolateral diameters differed significantly between disc status groups (p < 0.001). Among the evaluated models, MLPClassifier achieved the highest area under the receiver operating characteristic curve (ROC-AUC) (0.797), followed by Random Forest (0.777) and HistGradientBoosting (0.773). Although CatBoost ranked fourth in overall discrimination (ROC-AUC = 0.767), it demonstrated the highest recall (0.839), indicating superior sensitivity for detecting disc displacement. Ensemble and neural network-based models generally outperformed linear classifiers. Conclusions: Supervised ML algorithms showed balanced potential in diagnosing TMJ disc displacement in adolescents. Further validation in prospective clinical and multicenter studies is warranted to support clinical application.