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

1. Introduction

Temporomandibular joint disorders (TMDs), among the most frequent causes of orofacial pain, are increasingly prevalent across all age groups. TMDs encompass a spectrum of conditions affecting the temporomandibular joint (TMJ), the masticatory muscles, or both [1]. The TMJ is a complex structure composed of a fibrocartilaginous disc, joint capsule, and supporting ligaments connecting the mandibular condyle to the temporal bone. Clinically, TMDs are characterized by joint and muscle pain, joint sounds (e.g., clicking or popping), and limitations in mandibular movement, including joint locking [1, 2]. Prevalence estimates range from 10% to 70% in the general population and from 16% to 68% among pediatric and adolescent populations [3, 4]. Recent consensus statements emphasize that TMD diagnosis is primarily based on clinical assessment, rather than imaging, according to standardized criteria such as the Diagnostic Criteria for Temporomandibular Disorders (DC/TMD) framework [5, 6]. Imaging is considered complementary and is indicated when structural characterization of the joint may influence treatment planning or when clinical findings are inconclusive [7].

Accurate diagnosis of TMDs is essential, as they significantly impact quality of life. Clinical examination complemented by radiological imaging is critical for precise assessment. Computed tomography (CT) is mainly employed for evaluating fractures, postoperative changes, and osseous pathologies. Bone scintigraphy aids in detecting inflammatory and osteoarthritic changes, whereas ultrasonography has shown some potential for joint evaluation, though its diagnostic capacity remains limited [8]. The gold standard for assessing the soft tissue structures of the TMJ, especially for identifying disc displacements, is still Magnetic Resonance Imaging (MRI) [9]. MRI’s non-invasive nature and absence of ionizing radiation make it especially suitable for use in younger populations. Some studies have suggested that combining MRI with CT may improve diagnostic accuracy and inter-examiner reliability in TMJ evaluations [9]. Moreover, MRI is capable of identifying early indicators of joint dysfunction, such as osseous abnormalities, joint effusion, disc morphology changes, and thickening of the anterior or posterior disc bands [10]. MRI does not establish the clinical diagnosis of TMD, but provides valuable morphological information regarding internal derangement and structural changes when indicated [11]. Importantly, the presence of disc displacement alone is not sufficient to diagnose or manage TMD, as disc position and pain symptoms show only weak correlation in many cases [12]. Nevertheless, MRI has particular value in adolescents, where disc displacement may influence mandibular growth and remodeling due to incomplete skeletal maturation [13].

Recently, machine learning (ML) techniques have been increasingly explored to enhance diagnostic accuracy and automate the detection of TMJ disc displacement. Although MRI supports clinical evaluation, the diagnosis of TMD remains clinical, and artificial intelligence (AI) models should be interpreted as tools for standardized image assessment, rather than independent diagnostic systems [6]. Previous AI studies have primarily focused on adult populations, and limited evidence exists regarding the performance of supervised ML models in adolescents, whose ongoing skeletal growth may affect both morphology and imaging characteristics [14, 15]. AI-assisted diagnostic tools have shown promise, especially in cases where clinical findings are inconclusive or subjective [16]. Several studies have reported that AI models can detect TMJ-related abnormalities with accuracies as high as 0.93 [14]. Recent research has focused on segmenting the joint space and localizing disc position through deep learning-based neural networks applied to MRI data [16]. Additionally, ultrasonography-based approaches incorporating AI have demonstrated high diagnostic accuracy in assessing TMJ morphology by segmenting anatomical structures such as the mandibular condyle and joint space [17].

This study aimed to evaluate the predictive performance of supervised ML models in classifying TMJ disc displacement based on MRI findings in adolescents aged 12 to 18 years. Morphological and signal intensity-based features—including mandibular condyle shape, condylar anteroposterior and mediolateral dimensions, and signal intensities from the superior and inferior heads of the lateral pterygoid muscle—were extracted as input variables. Fifteen ML algorithms were developed by performing class balancing and hyperparameter optimization. Their classification performances were compared to assess the potential of ML models to support clinical decision-making in the diagnosis of TMJ disc displacement in the adolescent population. Rather than diagnosing TMD, our aim was to classify disc displacement patterns based on quantitative MRI features and explore whether supervised ML models could provide standardized image-based characterization to support clinical decision-making. This approach aligns with recent recommendations that emphasize imaging-based classification of internal derangement, rather than MRI-based “diagnosis” of TMD [18].

2. Materials and methods

Study approval was obtained from the local clinical research ethics committee (file number: 2024/316). We have read the Helsinki Declaration and have followed the guidelines in this investigation.

2.3 MRI evaluation

All MRI images were jointly evaluated by two radiologists during the same session, and consensus-based decisions were recorded. Morphometric measurements were performed using dedicated imaging software (Syngo.via, version VB30, Siemens Healthineers, Erlangen, BY, Germany). TMJs were categorized into three groups based on disc position, following standardized nonosseous component assessment criteria [15].

- Normal: Normal articular disc position.

- ADwR: Anterior disc displacement with reduction.

- ADwoR: Anterior disc displacement without reduction (Fig. 1).

Fig. 1.Oblique sagittal proton density-weighted magnetic resonance images of the temporomandibular joint. (A) Closed-mouth image showing an anteriorly displaced disc. (B) Open-mouth image showing a non-reduced disc position, consistent with anterior disc displacement without reduction.

On sagittal PD-W images, the shape of the mandibular condyle was classified as round (convex), flat (blunt), or beak (spur-like) (Fig. 2). The anteroposterior and mediolateral diameters of the condyle were measured from cortex to cortex on sagittal and coronal sections, respectively (Fig. 3). Measurements were performed in a consensus setting by two radiologists to minimize variability. A derived variable, CondyleRatio, was calculated by dividing the mediolateral diameter by the anteroposterior diameter.

Fig. 2.Oblique sagittal proton density-weighted magnetic resonance images demonstrating examples of mandibular condyle morphologies. (A) Closed-mouth image showing a round condyle. (B) Open-mouth image showing a flat condyle. (C) Open-mouth image showing a beak-shaped condyle.

Fig. 3.Condylar dimension measurements on magnetic resonance images. (A) Anteroposterior diameter measured on oblique sagittal proton density-weighted image. (B) Mediolateral diameter measured on coronal T2-weighted image.

Additionally, signal intensities of the superior and inferior heads of the LPM were measured on coronal T2-W images. Slice selection was standardized by identifying the coronal section demonstrating the maximum cross-sectional area of the LPM belly, avoiding partial volume effects at the muscle origin or insertion. Circular regions of interest (ROIs) with a fixed area of 4.97 mm² were manually placed within the central portion of each LPM head, taking care to avoid adjacent fat planes, visible blood vessels, and susceptibility artifacts. For signal normalization, reference ROIs of identical size were placed in visually homogeneous regions of cerebral gray matter (GM) and white matter (WM), avoiding cortical boundaries and cerebrospinal fluid contamination (Fig. 4). Signal intensity ratios (SIRs) were calculated separately for the superior and inferior heads of the LPM using the following normalisation formula [19]: SIR: (SI(measured) − SI(GM))/(SI(WM) − SI(GM)). All measurements were performed by two radiologists in a consensus reading setting.

Fig. 4.Coronal T2-weighted axial magnetic resonance image showing signal intensity measurements obtained using regions of interest. Measurements were taken from the superior and inferior heads of the lateral pterygoid muscle, as well as from the cerebral white and grey matter.

2.4 Machine learning for TMJ dislocation description, model assessment, and performance measures

This study aimed to assess the predictive performance of supervised ML algorithms in classifying TMJ disc displacement in adolescents using MRI-derived morphological and signal-based features. Predictor variables included anteroposterior and mediolateral condylar diameters, the derived CondyleRatio (mediolateral-to-anteroposterior ratio), condylar and disc morphology, and normalized SIRs of the superior and inferior heads of the lateral pterygoid muscle (LPMsupSIR and LPMinfSIR).

2.4.1 Outcome definition

Initially, TMJ disc position was categorized into three classes: Normal, anterior disc displacement with reduction (ADwR), and anterior disc displacement without reduction (ADwoR). Due to class imbalance and overlapping clinical characteristics between the ADwR and ADwoR groups, these two categories were merged into a single Dislocated disc group. This binary classification (Normal vs. Dislocated) provided a more stable predictive setting and improved model generalizability in the supervised learning context. A three-class classification (Normal, ADwR, ADwoR) was additionally explored as a secondary analysis.

2.4.2 Patient-level data splitting

To prevent data leakage arising from correlated bilateral TMJs, all data splitting was performed at the patient level. The dataset was divided into training (80%) and testing (20%) sets using GroupShuffleSplit, with the patient identifier used as the grouping variable. This ensured that TMJs from the same patient were never distributed across both the training and test sets.

2.4.3 Cross-validation and preprocessing

Model development and hyperparameter optimization were conducted exclusively within the training set using group-aware cross-validation. All preprocessing steps, including feature standardization with StandardScaler, were implemented within the cross-validation pipeline and fitted only on the corresponding training folds to avoid information leakage.

2.4.4 Model training and hyperparameter optimization

A total of 15 supervised ML classifiers were implemented: Random Forest, Logistic Regression, Ridge Classifier, Linear SVC, K-Nearest Neighbor, Gaussian Naive Bayes, Multilayer Perceptron, Stochastic Gradient Descent, AdaBoost, Gradient Boosting, HistGradientBoosting, Extra Trees, CatBoost, LightGBM, and XGBoost.

Hyperparameter optimization was performed using the Optuna framework with five-fold cross-validation, adopting a nested search strategy to minimize model selection bias. The test set was not accessed during model selection or optimization. After optimization, each model was refit once on the full training set using the selected hyperparameters. Final model performance was then evaluated a single time on the held-out test set, without any further tuning. Model performance was evaluated using Accuracy, Precision, Recall, F1-score, and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC), all calculated on the withheld test set. To quantify model uncertainty, 95% confidence intervals (CIs) for Accuracy and ROC-AUC were estimated using 1000 bootstrap resamples. The six best-performing classifiers based on ROC-AUC were selected for extended evaluation.

To further interpret error patterns, confusion matrices for these four classifiers were generated from test-set predictions (Fig. 5), providing a detailed visualization of true positive, true negative, false positive, and false negative distributions. Model explainability was assessed using SHAP (SHapley Additive exPlanations). Global feature importance was computed using mean absolute SHAP values derived from the positive (Dislocated) class, and feature ranking profiles were visualized for each model (Fig. 6).

Fig. 5.Confusion matrices of the four best-performing classifiers (MLPClassifier, Random Forest, HistGradientBoosting, and CatBoost) evaluated on the held-out patient-level test set. The split was performed at the patient level to prevent information leakage between contralateral temporomandibular joints. CatBoost demonstrated the lowest number of false-negative predictions for disc displacement, indicating high sensitivity for pathology detection, while MLPClassifier showed the highest overall discriminative balance. Random Forest and HistGradientBoosting exhibited comparable error patterns.

Fig. 6.SHAP-based global feature importance plots for the top four classifiers. Across all models, mediolateral (CondilML) and anteroposterior (CondilAP) condylar diameters consistently emerged as the most influential predictors of disc displacement, followed by condylar morphology and lateral pterygoid muscle signal intensity ratios. Demographic variables (age and sex) showed minimal contribution. SHAP values are presented on model-specific scales and are not intended for direct quantitative comparison across models. SHAP: SHapley Additive exPlanations; ML: Machine learning; AP: anteroposterior; SIR: Signal intensity ratio; LPM: lateral pterygoid muscle.

All preprocessing, model development, and statistical analyses were performed in Python (3.11.5) using the following libraries: pandas (1.5.3), numpy (1.24.0), scikit-learn (1.1.3), imbalanced-learn, Optuna (3.1.1), matplotlib (3.6.3), seaborn (0.11.2), and statsmodels (0.13.5).

3. Results

3.2 Machine learning model performance

The classification performance of fifteen supervised ML models in detecting TMJ disc displacement was assessed using Accuracy, Precision, Recall, F1-score, and ROC-AUC metrics. All performance estimates were derived from a patient-level held-out test set, and 95% CIs were computed using patient-level bootstrap resampling. A comprehensive summary of the leakage-free model performances is presented in Table 2.

Table 2.Performance metrics of fifteen supervised machine learning models in the binary classification of TMJ disc displacement diagnosed on MRI.

Model	Accuracy (95% CI)	Recall (95% CI)	Precision (95% CI)	F1-Score (95% CI)	ROC-AUC (95% CI)
MLPClassifier	0.675 (0.538–0.808)	0.619 (0.412–0.813)	0.788 (0.600–0.952)	0.689 (0.511–0.842)	0.797 (0.647–0.917)
RandomForest	0.671 (0.538–0.788)	0.649 (0.458–0.839)	0.763 (0.600–0.917)	0.697 (0.540–0.828)	0.773 (0.626–0.901)
HistGradientBoosting	0.631 (0.500–0.769)	0.648 (0.469–0.818)	0.710 (0.524–0.889)	0.673 (0.519–0.806)	0.770 (0.621–0.900)
CatBoost	0.694 (0.558–0.827)	0.842 (0.667–1.000)	0.705 (0.553–0.842)	0.764 (0.640–0.870)	0.765 (0.598–0.907)
LightGBM	0.653 (0.519–0.788)	0.682 (0.500–0.848)	0.721 (0.548–0.885)	0.697 (0.545–0.827)	0.748 (0.589–0.889)
XGBoost	0.652 (0.500–0.788)	0.714 (0.500–0.889)	0.705 (0.529–0.875)	0.706 (0.537–0.841)	0.738 (0.588–0.885)
DecisionTree	0.655 (0.519–0.788)	0.620 (0.429–0.808)	0.757 (0.583–0.920)	0.677 (0.520–0.815)	0.729 (0.569–0.864)
GradientBoosting	0.618 (0.481–0.769)	0.717 (0.516–0.906)	0.669 (0.500–0.839)	0.687 (0.542–0.815)	0.735 (0.572–0.892)
KNN	0.694 (0.577–0.827)	0.907 (0.750–1.000)	0.684 (0.537–0.826)	0.777 (0.656–0.884)	0.722 (0.552–0.868)
AdaBoost	0.689 (0.538–0.827)	0.776 (0.577–0.938)	0.722 (0.545–0.882)	0.744 (0.581–0.867)	0.723 (0.562–0.867)
LinearSVC	0.654 (0.500–0.808)	0.650 (0.423–0.871)	0.738 (0.545–0.905)	0.686 (0.490–0.842)	0.702 (0.524–0.861)
LogisticRegression	0.635 (0.481–0.788)	0.682 (0.471–0.882)	0.699 (0.517–0.880)	0.686 (0.518–0.828)	0.696 (0.517–0.860)
Ridge	0.634 (0.481–0.788)	0.713 (0.518–0.886)	0.685 (0.500–0.865)	0.695 (0.528–0.833)	0.690 (0.504–0.858)
SGD	0.637 (0.481–0.788)	0.713 (0.518–0.886)	0.688 (0.500–0.861)	0.697 (0.533–0.839)	0.681 (0.493–0.853)
ExtraTrees	0.638 (0.481–0.788)	0.619 (0.400–0.833)	0.731 (0.550–0.889)	0.666 (0.480–0.818)	0.675 (0.492–0.837)
Performance metrics are presented as means with 95% confidence intervals (CI). ROC-AUC: Receiver Operating Characteristic-Area Under the Curve; MLP: Multilayer Perceptron; SGD: Stochastic Gradient Descent; SVC: Support Vector Classifier; KNN: k-nearest neighbors.

3.2.1 Binary classification (normal vs. dislocated)

Among all classifiers, MLPClassifier achieved the highest ROC-AUC (0.797; 95% CI: 0.647–0.917), followed by Random Forest (0.773; 95% CI: 0.626–0.901), HistGradientBoosting (0.770; 95% CI: 0.621–0.900), and CatBoost (0.765; 95% CI: 0.598–0.907).

CatBoost demonstrated the highest recall (0.842; 95% CI: 0.667–1.000), indicating strong sensitivity for detecting dislocated discs. k-nearest neighbors (KNN) showed the highest recall overall (0.907), although with comparatively lower ROC-AUC (0.722).

Accuracy values ranged from 0.631 to 0.694 across models, with CatBoost (0.694), KNN (0.694), and AdaBoost (0.689) achieving the highest test accuracies.

Linear models (Logistic Regression, Ridge, LinearSVC, Stochastic Gradient Descent (SGD)) showed relatively lower discriminative performance (ROC-AUC range: 0.681–0.702), whereas ensemble-based tree models demonstrated more stable performance across metrics. ROC curves of the four best-performing models are shown in Fig. 7, and comparative metric visualizations are provided in Figs. 8,9.

Fig. 7.Receiver operating characteristic (ROC) curves of the four best-performing machine learning models for binary classification of temporomandibular joint disc status (normal vs. dislocated). Curves were generated using predictions from the held-out test set, which was not involved in model training, hyperparameter optimization, or threshold selection. ROC: Receiver Operating Characteristic; AUC: Area Under the Curve; MLP: Multilayer Perceptron.

Fig. 8.Heatmap summarizing the classification performance of the top six models across five evaluation metrics: accuracy, precision, recall, F1-score, and area under the ROC curve (ROC-AUC). Metric values are derived from the patient-level test set. The heatmap provides a comparative overview of model-specific performance trade-offs, rather than a ranking based on a single metric. ROC: Receiver Operating Characteristic; AUC: Area Under the Curve; MLP: Multilayer Perceptron.

Fig. 9.Bar plots showing ROC-AUC values of the four best-performing classifiers, along with 95% confidence intervals estimated using patient-level bootstrap resampling of the held-out test set (1000 iterations). Overall, performance differences between the top five models were moderate, and confidence intervals partially overlapped, suggesting comparable discriminative ability under the leakage-free evaluation framework. ROC: Receiver Operating Characteristic; AUC: Area Under the Curve; MLP: Multilayer Perceptron; CI: confidence intervals.

3.2.2 Multiclass classification (normal vs. ADwR vs. ADwoR)

For the three-class classification task, macro-averaged one-vs-rest ROC-AUC was used as the primary discrimination metric.

CatBoost achieved the highest ROC-AUC (0.677; 95% CI: 0.539–0.801), followed closely by Gradient Boosting (0.671; 95% CI: 0.544–0.789), AdaBoost (0.668; 95% CI: 0.540–0.794), and Random Forest (0.662; 95% CI: 0.533–0.796).

Overall accuracy values ranged from 0.369 to 0.540, with KNN (0.540) and AdaBoost (0.535) demonstrating the highest accuracies. However, these models did not correspondingly achieve the highest ROC-AUC values, reflecting differences in class-wise discrimination balance.

Compared with binary classification, multiclass performance was consistently lower across all algorithms, reflecting the increased complexity of distinguishing between ADwR and ADwoR subtypes.

4. Discussion

Temporomandibular disorders (TMDs) constitute a heterogeneous group of musculoskeletal conditions affecting the masticatory system, frequently emerging during adolescence and early adulthood. TMDs are common, multifactorial conditions that primarily affect individuals during adolescence and early adulthood, often presenting with joint clicking, pain, and limited mandibular mobility [2]. Although TMJ disc dislocation can occur in various directions, anterior displacement is the most frequently observed subtype [20]. Epidemiological studies consistently demonstrate a higher prevalence of TMD among females. A systematic review and meta-analysis reported that women have approximately twice the risk of TMDs compared with men [21]. In clinical series, TMD signs and symptoms have also been found to be significantly more frequent in females [22]. Pediatric and adolescent cohorts similarly show increased TMD prevalence in females, suggesting that sex-related factors may influence both symptom burden and imaging-based presentations. Early recognition of structural alterations may be relevant in adolescents, given the ongoing craniofacial growth. Pathological alterations in the mandibular condyle prior to skeletal maturity can lead to long-term asymmetries and growth disturbances in the mandible [23]. MRI is the reference imaging modality for visualizing TMJ disc morphology and joint structures; however, contemporary guidelines indicate that imaging findings should complement, rather than define, the clinical diagnosis of TMD. The International Association for Dental Research (IADR) recently articulated updated standards of care, highlighting that imaging findings must be interpreted within a clinical context, as many structural variations detected on MRI occur in asymptomatic individuals and do not, by themselves, mandate treatment [6]. In the present study, we investigated the predictive performance of fifteen supervised ML algorithms in classifying TMJ disc displacement in adolescents using MRI-derived morphometric and signal-based features.

While MRI is valuable for visualizing TMJ structures, contemporary orofacial pain guidelines emphasize that disc displacement is not a diagnostic or therapeutic target, and its association with pain is weak [11, 12]. Recent recommendations from the IADR also highlight that TMD management should not be based solely on structural imaging findings [6]. Therefore, MRI should be prescribed only when clinical findings warrant its use, in accordance with established imaging appropriateness frameworks [7]. Because disc displacement and other MRI findings may also be observed in asymptomatic individuals, disc position should be interpreted cautiously in the clinical context [24].

Several studies have reported that the mandibular condyle is smaller in joints exhibiting anterior disc displacement compared with those with a normal disc position, both in mediolateral and anteroposterior dimensions [25]. Experimental evidence further suggests that TMJ disc displacement can negatively affect mandibular development, potentially resulting in reduced mandibular height and length due to impaired condylar growth [13]. Our findings align with this body of evidence, as both anteroposterior and mediolateral diameters of the mandibular condyle were significantly smaller in joints diagnosed with ADwR and ADwoR compared with those with normal disc position. These results support the hypothesis that disc displacement, particularly in adolescents, may interfere with normal growth and remodeling processes of the TMJ.

It has been suggested that, due to the morphological and functional characteristics of the TMJ, adjacent anatomical structures in joints with disc displacement may be subject to changes in response to altered functional loading patterns [26]. Prior research has identified a relationship between osteoarthropathic changes and specific condylar morphologies, particularly flattened or beak-like condyle [27]. This association is believed to be of particular importance during adolescence, a critical period for craniofacial growth and joint development. In our study, approximately 50% of the TMJs exhibited a rounded condylar morphology. However, the prevalence of beak-shaped and flat condyles was markedly higher in joints diagnosed with ADwR and ADwoR. No statistically significant association was observed between condyle morphology and patient sex. These findings are consistent with the literature and further support the hypothesis that disc displacement may influence not only the size, but also the shape of the mandibular condyle, potentially contributing to early degenerative or adaptive joint changes.

In recent years, ML models have attracted increasing attention for their potential to enhance sensitivity and overall diagnostic performance in detecting TMJ disc displacement. As demonstrated by multiple studies, the integration of ML techniques into TMJ imaging analysis may contribute to more accurate diagnoses and better treatment planning. In a study by Orhan et al. [28], the diagnostic validity of MRI-based ML models was evaluated using support vector machines, random forest, and k-nearest neighbor algorithms to classify TMJ disorders associated with condylar changes and disc displacement. The study reported promising results, with an AUC of 0.89 in the training set and 0.77 in the test set, indicating strong generalizability of the model [28]. Similarly, Yüksel et al. [15] employed a radiomics-based ML approach to extract quantitative imaging features from MRI scans and classify TMJ disc displacement. Among the algorithms tested, the k-nearest neighbor classifier demonstrated the highest predictive performance and was identified as the most suitable model for this task [15]. In our study, non-linear and ensemble-based models demonstrated slightly superior overall discrimination compared with traditional linear classifiers. The MLPClassifier achieved the highest ROC-AUC (0.797), followed by Random Forest and HistGradientBoosting, while CatBoost showed the highest recall, indicating enhanced sensitivity for detecting disc displacement. These findings suggest that advanced non-linear approaches may provide a more adaptable and clinically balanced solution for classifying adolescent TMJ disc status based on MRI-derived morphometric and signal intensity features.

ML applications in TMDs are not limited to imaging-based evaluations such as MRI, but have also been applied to the analysis of clinical and functional parameters. In the present study, the proposed model is intended to function as a decision-support tool in cases where TMJ MRI is already clinically indicated, assisting radiologists in identifying morphometric patterns associated with disc displacement, rather than replacing expert interpretation. Taborri et al. [29] investigated the diagnostic performance of nine ML algorithms using mobility-related and baropodometric measurements and found that the k-nearest neighbors algorithm yielded the best performance, with an accuracy of 0.94 and an F1-score of 0.91 [29]. Complementing this, Yildiz et al. [30] evaluated over 20 ML models to predict TMDs using easily obtainable clinical parameters such as pain intensity and psychological factors. Their results indicated that a Bagging classifier using the Multivariate Adaptive Regression Spline (MARS) algorithm achieved the highest performance, with an accuracy of 0.89 and ROC-AUC of 0.90 [30]. In our study, non-linear ML models demonstrated modest, but consistent, advantages over linear classifiers in classifying TMJ disc displacement. The MLPClassifier achieved the highest overall discriminative performance based on ROC-AUC, while Random Forest and HistGradientBoosting showed comparable results with balanced accuracy and F1-scores. Notably, CatBoost exhibited the highest recall among the evaluated models, indicating enhanced sensitivity for detecting displaced discs, albeit with slightly lower specificity. HistGradientBoosting and LightGBM also performed competitively across multiple metrics, whereas linear classifiers, such as Logistic Regression and Ridge Classifier, yielded comparatively lower performance. Confusion matrix analysis provided complementary insight into model behavior, revealing that CatBoost minimized false-negative predictions, while MLPClassifier and Random Forest demonstrated more balanced error distributions. Collectively, these findings suggest that advanced non-linear and ensemble-based approaches may offer a more flexible and clinically relevant framework for TMJ disc displacement classification using MRI-derived morphometric and signal intensity features, particularly when sensitivity to pathology is prioritized.

Zhang et al. [31] conducted a systematic review and meta-analysis evaluating the diagnostic accuracy of ML models in distinguishing TMDs using imaging data. Although the review encompassed both traditional ML and deep learning algorithms, the meta-analysis did not provide a definitive conclusion as to which type of model demonstrated superior diagnostic performance for TMDs [31]. Similarly, Farook and Dudley reviewed twenty studies focused on automation and deep learning applications in TMJ radiomics. Their findings indicated that deep learning models performed similarly to clinicians in diagnosing TMJ osteoarthritis from both 2D and 3D radiographs. However, these models exhibited reduced performance when applied to MRI data for the detection of disc displacement or internal derangement [32].

This study has several limitations that should be acknowledged. First, the retrospective nature of the study limited the inclusion of important clinical variables, such as pain severity, functional limitations, and psychological status, which may influence the diagnosis and progression of TMDs. Second, although all TMJ MRI examinations were evaluated by two radiologists in consensus, inter-observer agreement was not formally assessed, and the evaluations were not performed independently at separate time-point factors that may introduce interpretative bias. Accordingly, model performance reflects agreement with expert consensus, rather than robustness to inter-reader variability, and clinical applicability should be interpreted cautiously. Additionally, all MRI data were obtained from a single institution, which may affect the generalizability of the findings to wider populations and imaging protocols. Future research should aim to incorporate multicenter datasets, integrate relevant clinical variables, and include inter-observer variability assessment to validate and strengthen the utility of ML models in TMJ diagnostics.

5. Conclusions

This study demonstrates the potential of supervised ML algorithms in classifying TMJ disc displacement in adolescents using MRI-derived morphometric and muscle signal features. Among the fifteen evaluated models, non-linear approaches showed modest, but consistent, advantages over traditional linear classifiers. The MLPClassifier achieved the highest overall discriminative performance, while tree-based ensemble models demonstrated comparable and stable results across evaluation metrics. Notably, CatBoost exhibited the highest sensitivity for detecting disc displacement, suggesting potential utility in scenarios where minimizing false-negative findings is clinically important.

Morphometric parameters—particularly condylar diameters—were strongly associated with disc displacement, underscoring their diagnostic relevance. To our knowledge, this is one of the largest ML-based TMJ studies conducted in an adolescent MRI cohort incorporating both structural and muscle signal characteristics. Future research should focus on external validation in multi-institutional datasets and on integrating automated segmentation and radiomics-based features to further enhance diagnostic robustness and generalizability.

Abbreviations

TMJ, Temporomandibular joint; TMDs, Temporomandibular joint disorders; ML, Machine learning; MRI, Magnetic Resonance Imaging; CT, Computed Tomography; AI, artificial intelligence; LPM, lateral pterygoid muscle; ADwR, Anterior disc displacement with reduction; ADwoR, Anterior disc displacement without reduction; GM, gray matter; WM, white matter; ROIs, regions of interest; SIRs, Signal intensity ratios; ROC-AUC, Receiver Operating Characteristic-Area Under the Curve; CIs, confidence intervals; FOV, field of view; T2-W, T2-weighted; ST/IP, slice thickness/inter-slice gap; TR, repetition time; TE, echo time; ETL, echo train length; PD-W, proton density-weighted; SHAP, SHapley Additive exPlanations; ANOVA, one-way analysis of variance; HSD, Honestly Significant Difference; SGD, Stochastic Gradient Descent; IADR, International Association for Dental Research; MARS, Multivariate Adaptive Regression Spline; DC/TMD, Diagnostic Criteria for Temporomandibular Disorders; NEX, number of excitations; KNN, k-nearest neighbors.

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Author contributions

SE—Conceptualization, Data Curation, Methodology, Writing–Original Draft, Software, Formal analysis, Visualization, Resources, Supervision. HÖ—Methodology, Formal Analysis, Visualization, Supervision. İD—Investigation, Supervision, Project administration. AG—Investigation, Validation, Data curation. AB—Investigation, Supervision, Validation, Resources. MÖ—Conceptualization, Supervision, Resources. All authors contributed to writing–review & editing and approved the final manuscript.

Ethics approval and consent to participate

Approval was obtained from the clinical research ethics committee of Selçuk University Faculty of Medicine (file number: 2024/316). Due to the retrospective design of the study and the use of anonymized MRI data, individual written informed consent was not obtained.

Acknowledgment

The authors would like to acknowledge the institutional support provided by Selcuk University Faculty of Medicine during this study.

Funding

This research received no external funding.

Conflict of interest

The authors declare no conflict of interest.