Artificial Intelligence in Diagnosing and Treating Temporomandibular Joint Disorders: A Review

TMD fundamentals

TMD diagnosis

Diagnosing TMD is challenging because its symptoms overlap with other conditions. Accurate diagnosis requires a combination of patient history, physical examination, and imaging. History should address the chief complaint, prior trauma, infections, degenerative or arthritic diseases, systemic factors, and lifestyle aspects such as stress, sleep, and recreational activities, which can influence symptom severity.⁴ During the COVID-19 pandemic, 51.4% of patients reported worsened facial pain and significantly higher bruxism rates (p < 0.001), highlighting how psychological stress can exacerbate TMD beyond purely physical causes.⁵

TMJ examination involves dental evaluation, occlusal analysis, and palpation of the mandibular condyle with approximately 1 kg of pressure to assess tenderness. Mandibular function should be recorded, including mouth opening and closing, range of motion (protrusion and lateral excursions), and joint sounds, with brief sounds indicating clicking and prolonged sounds indicating crepitus. Bilateral palpation of the masticatory muscles is performed to identify trigger points and muscle tenderness.⁶ Imaging is indicated when clinical findings are inconclusive, using panoramic or cone-beam computed tomography for hard tissues and magnetic resonance imaging for soft tissue assessment.⁴

TMD classes and classification

TMD is classified using several systems. The Helkimo clinical index assesses severity through three sub-indices: anamnesis for symptoms, dysfunction for mandibular mobility and muscle tenderness, and occlusion for dental assessment.⁷ The Wilkes classification, refined by Dimitroulis in 2012 into five stages, guides treatment based on clinical, radiological, and surgical findings.^8,9

The RDC/TMD, introduced in 1992, established a dual-axis system: Axis I categorized patients into muscular disorders, disc displacement, and joint disorders based on clinical and radiological findings, while Axis II assessed psychosocial factors and pain-related disability.¹⁰ The criteria were updated in 2014 as DC/TMD, retaining the dual-axis framework but refining assessment algorithms and expanding tools to evaluate pain behavior, psychological status, and psychosocial functioning.^2,9

In 2020, an international collaboration including OFHP SIG, INfORM, AAOP, and IHS released the first edition of the the International Classification of Orofacial Pain (ICOP), classifying all orofacial pain, including TMD-related pain, while excluding non-painful conditions. Painful TMDs are categorized as myofascial orofacial pain and TMJ pain, divided into primary and secondary disorders, with primary pain further classified as acute or chronic using a three-month threshold. Both types may be subclassified based on pain referral during palpation, following IASP definitions for primary and secondary pain.¹¹

Treatment modalities

TMD can progress to advanced stages requiring treatment that depends on the diagnosis.¹² Early-stage TMD is typically managed conservatively with a soft diet, drug therapy, physiotherapy, and occlusal guards tailored to each patient. Patients with anterior disc displacement who do not respond to conservative therapy may require minimally invasive procedures such as arthroscopy or arthrocentesis.^13,14 Both techniques reduce joint friction by releasing adhesions, removing inflammatory mediators, and increasing joint space.¹⁵ Arthrocentesis involves single- or double-needle lavage with physiological saline or Ringer’s solution, while arthroscopy additionally allows direct visualization of intra-articular structures, fibrocartilage degeneration, and pathologies not detectable by MRI.¹⁶

Advanced TMD may require synovectomy, discectomy, disc plication, eminoplasty, eminectomy, or open joint surgery, with total joint replacement as a last resort.¹² In 2025 systematic review of 684 videos and 550 social media discussions found that non-professional content, while more engaging, was generally low in quality (44.1% poor, 11.9% good), whereas professional content was more reliable but less engaging. Most content focused on treatment, with fewer than 6% addressing complications or prevention, highlighting the risks of misinformation and the need for evidence-based content in collaboration with healthcare professionals and influencers.¹⁷

Machine learning definition and subtypes

ML is a type of artificial intelligence that uses data to improve performance without deliberate programming and Samuel defined it in 1959 as a field of study that enables computers to learn from experience without being explicitly programmed.¹⁸ The aim is to develop automated systems that recognize patterns, predict outcomes, and make data-driven decisions.ML can be categorized into three categories: supervised, unsupervised, and reinforcement learning. The model is trained on annotated or labeled data, with every input linked to the associated output. Unsupervised learning examines unlabeled data to uncover patterns and relationships among them. Reinforcement learning comprises directing an agent to interact with the environment and gathering information via feedback, such as punishments or rewards.

Traditional ML algorithms, which predate deep learning, use statistical approaches rather than deep neural networks. Although traditional ML methods have been successful but have not performed well with image and audio data. The success of Krizhevsky and colleagues in the 2012 ImageNet Large Scale Visual Recognition Challenge (ILSVRC) sparked interest in the emerging field of deep learning algorithms. Consequently, researchers have commonly used DL methods in their studies.¹

Deep learning

DL is a subset of ML that uses artificial neural networks with multiple hidden layers, earning the name ‘deep’ due to their layered structure. This framework, a branch of artificial intelligence trains algorithms to recognize patterns and representations in data. Its capacity to learn complicated features from raw data has rendered it popularity and made it a useful tool. DL uses Artificial Neural Networks (ANNs) to replicate neural connections similar to those found in the brain. These networks consist of interconnected neurons or nodes that process and transfer data to the next layer. The term “deep” refers to networks that have multiple layers and can learn abstract and complex representations of input data. Deep learning has shown remarkable results in a variety of applications, including image processing, voice recognition, and natural language processing. Its ability to handle complex patterns and massive amounts of data has resulted in widespread usage across a variety of industries.¹

Articular disc displacements and deformities

Individuals with TMD may experience a range of symptoms such as intense headaches, jaw pain, limited mouth opening, and tinnitus. The most common joint abnormalities in TMD are articular disc displacements and deformations. In the following section, we will present a meticulous review of studies that used artificial intelligence in diagnosing TMD. A study done in 2021 identified anterior disc displacement (ADD) utilizing 9009 sagittal TMJ MRI images. The accuracy and Area Under the Curve (AUC) were the comparative parameters to evaluate the optionality of severe complications after treatments. Deep learning models were created utilizing a Residual Networks (ResNet) based convolutional neural network (CNN) and a variety of approaches, including 5-fold cross-validation, data augmentation and oversampling. The maximum open mouth position model demonstrated high accuracy of 0.970 (±0.007) and AUC values of 0.990 (±0.005). In closed mouth position models, the diagnostic Criteria one accuracy was 0.863 (±0.008) and AUC was 0.922 (±0.009). The heat map showed the correct ADD and non-ADD. The ADD group’s heat map focused on the joint disc, condylar head, and posterior disc tissue, whereas the N-ADD group’s heat map focused on the condyle area in the open mouth position. The CNN model detects ADD with high accuracy and AUC, making it beneficial to clinicians to identify ADD before orthodontic therapy and improve treatment outcomes.¹⁹

Orhan et al. evaluated the efficacy of a ML algorithm for detecting TMJ abnormalities in MRI images. The study included 214 TMJs from 107 patients with signs and symptoms of TMD. The researchers used a radiomic system to detect imaging features linked with TMJ abnormalities, modifications in the condylar bone and disc displacements. The radiomic features were then determined using a variety of ML algorithms, including Logistic Regression (LR). DT, Random Forest, k-NN, SVM, and XGBoost classifiers were used to predict TMD pathology. Specificity, sensitivity, and Receiver Operating Characteristic (ROC) curves were used to evaluate the efficacy of various techniques. Experimental results revealed that k-nearest neighbors and random forest classifiers are the most effective machine-learning methods to assess ADD and condylar bone.²⁰

A study done in Brazil by Diniz et al.,²¹ aimed to evaluate the performance of three ML techniques (radiomic, semantic, and radiomic-semantic) for detecting TMD in Infrared Thermography (IT) images. Radiomics in medical imaging entails extracting an enormous number of quantitative elements from medical images via data-characterization algorithms. Semantic features refer to high-level, human-readable features that are frequently manually annotated by specialists, such as size, location, and other clinically significant characteristics. Radiomic-semantic is basically the combination of both approaches to leverage the strengths of both high-dimensional quantitative data and high-level interpretative data. The study included 78 patients who met the Fonseca questionnaire and RDC/TMD criteria, 37 individuals were control individuals and 41 TMD patients. The approach involved obtaining each patient’s IT lateral projections and selecting the masseter and temporal muscles for the attributed extraction. The data was analysed using the classification algorithms support vector machine (SVM), Kernel nearest neighbor (KNN), and multilayer perceptron (MLP)Hopkins’ statistic, as well as the Tukey tests, Shapiro-Wilk and ANOVA. Semantic and radiomic-semantic association accuracy, precision, and sensitivity were significantly different from radiomic features (p = 0.008, p = 0.016, and p = 0.013, respectively). The radiomic-semantic association showed a statistically significant difference in testing accuracy value (p = 0.003). MLP stands out from the other classifiers in terms of radiomic-semantic association (p = 0.004). Accordingly, this study suggested detecting TMD via employing semantic and radiomic-semantic-associated ML feature extraction algorithms, as well as an MLP classifier, along with IT images and pain scale data.²¹

Kim et al.,²² published a study that aimed to establish a deep learning-based approach for predicting TMJ disc perforation using MRI data, and to confirm its performance by comparing it to previously published results. Data was collected by analyzing medical records at the Department of Oral and Maxillofacial Surgery, Gangnam Severance Hospital, Korea between January 2005 to June 2018. It included 299 joints from 289 patients and was divided into a perforated group and non-perforated group based on disc perforation detected during surgery. Experts evaluated the TMJ MRI scans to detect distinguishing features. Data containing these attributes were used to create and verify prediction models using random forest and MLP techniques, were MLP technique being based on the Keras framework, a new deep learning technology. The models’ performances were compared using the AUC curve. MLP scored the highest AUC (0.940), being followed by random forest (0.918) and disc shape alone (0.791). The MLP and random forest also previously published results utilizing MRI (AUC 0.808) and MRI-based nomogram (AUC 0.889) Figure 1. When predicting TMJ disc perforation, deep learning outperformed conventional techniques and previous findings.²²

Figure 1. Machine learning algorithms Area Under the Curve.

Multilayer perceptron had the highest AUC (0.940), being followed by random forest (RF) (0.918) and disc shape (ds) alone (AUC 0.630). Image adapted from Kim, J. Y. et al.²²

In Korea, Lee et al.,²³ utilized AI to investigate whether psychosocial and biological factors influence the onset of TMD. The study investigated factors like stress, socioeconomic status, and working conditions as potential causes of TMDs. This study included data from 4744 participants, who filed information on their TMD status, and 37 independent factors, including socioeconomic position, demographic status, working conditions, stress, and health determinants. Researchers employed six AI methods to identify variables associated with TMJ disorders, including random forest, decision trees, naïve Bayes, logistic regression, SVMs, and ANN. Following that, predictive performance and the models’ accuracy were assessed. The highest average accuracy identified using methods such as logistic regression, artificial neural networks, random forest and support vector machines, for both reported TMD (r-TMD) and diagnosed TMD by a doctor (d-TMD), the highest AUC was observed using an artificial neural network and logistic regression for r-TMD and d-TMD, respectively.²³

Juvenile idiopathic arthritis

JIA is the most prevalent rheumatic condition in children. The International League of Associations for Rheumatology (ILAR) defines JIA as “arthritis of unknown etiology persisting for more than 6 weeks with an onset at 16 years of age or younger, after excluding other causes of joint inflammation”.²⁴ Juvenile idiopathic arthritis can produce abnormal facial growth, pain, and impaired jaw function.²⁵ To diagnose JIA, a medical history, physical examination, and laboratory investigations are conducted to rule out other possible causes and confirm the inflammation in the joint. Early diagnosis and treatment are critical for reducing joint injury and improving long-term results.²⁵

Erika Van Nieuwenhove and colleagues investigated the use of machine learning (ML) to identify immunological signatures in juvenile idiopathic arthritis (JIA) patients. They conducted comprehensive flow cytometry profiling of the adaptive immune system in 85 JIA patients and 43 age-matched healthy controls. The results revealed distinct immunological differences between JIA patients and healthy individuals, with consistent immune alterations observed across all JIA subtypes. These alterations were especially pronounced in patients with systemic JIA and those with active disease, indicating a shared immunological mechanism underlying the various clinical forms of JIA.

To analyze and classify the immunological data, the researchers employed three ML models: random forests, artificial neural networks (ANNs), and potential support vector machines (pSVMs). The models were evaluated using tenfold cross-validation to ensure robust generalization. Random forests were implemented in R using the ‘randomForest’ package and included 10,001 trees with various hyperparameter and sampling scheme configurations to address class imbalance. ANNs were developed in Python using the Keras library with a TensorFlow backend, testing 672 hyperparameter combinations that varied in architecture, learning rate, dropout, momentum, class weighting, and regularization. The networks used ReLU activations in hidden layers and a sigmoid activation in the output layer, optimized using stochastic gradient descent and binary cross-entropy loss. The pSVMs, selected for their performance on imbalanced data, were tested in dyadic mode with different cost and shrinkage parameters, both with and without balancing. All models were assessed primarily based on the area under the ROC curve (AUC), and they achieved over 90% accuracy in distinguishing JIA patients from healthy controls. These findings highlight the potential of machine learning to enhance the diagnosis and understanding of immune mechanisms in JIA.²⁶

Another study investigated the design and implementation of an AI-driven system to improve patient education for families struggling with JIA. This was accomplished by utilizing an AI-based argumentation theory to generate interactive educational conversations via using the Toulmin model of argument, a framework for constructing arguments, to transform Patient Education Materials (PEM) into a dialogue system. This paradigm assists in managing the content and structure of educational conversations, making them more dynamic and user-friendly. Educational information is organized as arguments with evidence-based suggestions. The dialogue system enables users to interact with information in a contextually appropriate and timely manner, resulting in improved comprehension and management of the condition. The dialogue system’s effectiveness was assessed using intellectual walkthroughs and semi-structured conversations with JIA topic experts. These studies yielded encouraging results, demonstrating that the system can effectively provide quality information to families when needed.²⁷

TMJ osteoarthritis

The pathogenesis of TMJOA is defined by gradual cartilage deterioration, subchondral bone remodelling, and chronic synovial tissue inflammation. An increasing number of research has lately concentrated on subchondral bone remodelling and inflammation in the initial stages of TMJOA, which could offer light on the likely mechanism of initiation and development of TMJOA. The treatment approach for TMJOA focuses on decreasing discomfort and pain, halt subchondral bone and cartilage degradation, and improve the function of the joints. The most common treatments for TMJOA are nonsteroidal anti-inflammatory medications, physical therapy, splints and minor surgical intervention that includes low-energy laser and arthrocentesis. While these interventions are generally effective at treating symptoms and signs, their long-term therapeutic effects on degenerative articular structures are insufficient. As of now, there is no treatment for TMJOA damage that can reverse it. Therapies that stop the progression of subchondral bone damage and cartilage degeneration must be investigated, and TMJ regeneration may be the best long-term solution.²⁸

In the University of Michigan, Bianchi et al.²⁹ discussed the potential of early diagnosis of TMJOA using biomarkers and machine learning, via understanding the interactions between disease and health-related biomarkers. Biomarkers such as C-reactive protein (CRP), interleukins, and matrix metalloproteinases (MMPs) show potential in indicating early stages of OA and they are available in saliva, blood and synovial fluid. Advanced data science was employed to evaluate fifty-two clinical, biochemical, and high-resolution images CBCT (radiomics) parameters comprising TMJOA patients and control groups. The diagnostic effectiveness was assessed using four ML models XGBoost, logistic regression, lightGBM, and random forest. The XGBoost and the LightGBM model achieved (0.823) accuracy, F1-score (0.823) and AUC (0.870) in diagnosing TMJ OA Figure 2.²⁹

Figure 2. Disease states, specificity and sensitivity (A) Top characteristics for determining disease state, boxplots of normalized features; (B) diagnostic specificity and sensitivity for each characteristic with top mean relevance and the mean prediction of LightGBM, XGBoost and their collective method in the 10-times 5-fold CV are demonstrated on ROC curves.

Image adapted from Bianchi, J et al.²⁹

In 2020, Lee et al employed artificial intelligence to detect TMJOA in CBCT images from patients with TMJ dysfunction. Single-shot detection, an object detection approach, was optimized on 3514 sagittal CBCTs. TMJ radiographs revealed osseous variations in the condylar area. The relevant area (Condylar head) was characterized and divided into two categories, indeterminate for TMJOA and TMJOA based on the radiographic image analysis criteria for TMD diagnosis. The model was evaluated with two sets of 300 images in total. The average precision, accuracy, F1 scores, and recall for the two test groups were 0.85,0.86, 0.84, and 0.84, respectively. A deep neural network model can automatically identify TMJOA from sagittal CBCT images, which can greatly support the clinical diagnosis.³⁰ Employing the same concept, in 2023, another study was done to use AI to classify TMJOA using (CBCT) sagittal images. The efficacy of the YOLOv5 architecture, a model of artificial intelligence, in segmentation TMJ and classifying osteoarthritis was assessed using 2000 sagittal sections (500 healthy, 500 erosion, 500 osteophytes, and 500 flattening images) from 290 patients’ CBCT DICOM images. The classification model predicts healthy joints with an 88% accuracy, 70% flattened joints, 95% erosion, and 86% osteophytes. The YOLOv5 model for TMJ segmentation has an F1 score of 0.99976, sensitivity of one, and precision of 0.9953. The model for TMJ segmentation has an AUC of 0.9723. Additionally, 0.9953 model’s accuracy in TMJ segmentation was found. The AI model utilized in this research can assist clinicians in saving time and providing convenience in disease diagnosis, with excellent results in TMJ segmentation and osteoarthritis classification.³¹

In the United Arab Emirates in 2023, Talaat et al.,³² published a study that aimed to create a DL-based model to accurately and efficiently identify TMJOA from 2737 CBCT images of 943 patients. They utilized a single convolutional network as a model base and a regression model for object detection. Two experienced evaluators conducted a DC/TMD assessment to generate a model-testing collection of 350 images, with the final diagnosis considered as the gold standard. The model’s diagnostic capabilities were compared to those of an experienced oral radiologist. The AI diagnosis revealed a substantially higher agreement with the golden reference. Cohen’s kappa revealed statistically significant differences in the AI and radiologist agreement with the gold standard for diagnosing subcortical cysts (P=0.0214) as well as all the remaining signs (P=0.0079), Table 1. In comparison to the human expert, the AI model utilized in this study performed equally well or better in the diagnostic process for TMJ osteoarthritis. They concluded that the use of AI in radiographic diagnosis of osteoarthritis is anticipated to speed up the diagnostic process and minimize the subjectivity involved in human interpretation.³²

Audio-based diagnostic

A research paper published in 2021 described a unique decision support system for detecting TMD that used deep learning and neural network techniques. A non-invasive apparatus was developed to record TMJ sounds. An interface was created that allowed the dentist to work on the captured audio data. The collected data comprised patient’s left and right TMJ sounds, ambient noise sounds, clinical data, and comments from the dentist about the patient’s diagnosis and therapy. The readings were subsequently categorized using artificial neural networks, signal processing, and DL algorithms, which resulted in the development of a system that determined the patient’s condition. Classification success was compared across frequency, statistical, and deep learning methods. The results revealed that the deep learning-based classification approach consistently outperforms the previous two methods, with a classification success rate of over 94.5 %. The proposed system can provide physicians with information regarding the efficacy of treatment approaches used on patients to address joint sounds, which are one of the most common symptoms of TMD.³³

Decision support systems

A cross-sectional study done by Waked et al. aimed to design a model with predictive skills that uses a statistical analysis classification tree to forecast the presence of TMD by categorizing participants as high or low risk of having the condition. A total of 776 persons requested dental or medical care in Brazil, Recife’s Family Health Units. The sample has been submitted to anamnesis using RDC/TMD. The data was imported to Statistical Package for the Social Sciences 20.0 software and processed for bivariate analysis using the Pearson Chi-square test and the multivariate analysis for the categorical tree method. Orofacial pain, aging, and depression may all be indicators of TMD. Individuals with orofacial pain, those aged 25 to 59, and those suffering from depression made up the high-risk group. Individuals who did not have orofacial pain made up the low-risk group. The authors concluded that orofacial discomfort was the best predictor of TMD and that the classification tree’s prediction model might be used as a tool to facilitate decision-making for TMD patients.³⁴

Additional research concerning dentistry covering the clinical decision support systems includes Mago et al.³⁵ Polášková et al.,³⁶ Mendonça,³⁷ and Machoy et al.³⁸ Here is a summary of the articles mentioned in Table 2.