Introduction

Temporomandibular disorders (TMD) encompass diverse conditions affecting the temporomandibular joint (TMJ), masticatory muscles, and associated structures, often resulting in pain, joint noise, and limited mandibular function1,2. TMD is a common musculoskeletal pain disorder involving the orofacial region, affecting approximately 31% of adults and older individuals, with women being 1.5–2.24 times more likely to be affected than men3,4. Pain ranges from acute cases, where symptoms persist for <6 months, to chronic cases that extend beyond 6 months5. Chronic TMD poses greater management challenges owing to its persistent nature, which often leads to worse clinical outcomes6. Therefore, identifying the factors associated with the progression from acute to chronic TMD is crucial for effective intervention and prevention.

Chronic TMD, like other chronic pain conditions, has significant clinical implications. Persistent pain in patients with TMD is often associated with reduced quality of life, impairing essential functions such as chewing, speaking, and sleep7. In addition, chronic TMD can lead to psychological distress, including anxiety and depression, contributing to the perpetuation of pain through central sensitization mechanisms8,9. This cycle of pain, aggravated symptoms, and psychological distress necessitates early recognition and management to prevent chronicity and improve long-term outcomes. However, research exploring signs and symptoms that distinguish acute from chronic TMD, along with the key predictors driving the transition to chronic TMD, remains limited.

Structural abnormalities observed on magnetic resonance imaging (MRI) and behavioral factors such as sleep disturbances and bruxism may contribute to chronic TMD10. However, the complex interaction between these factors makes the accurate prediction of chronicity challenging. Recent advancements in artificial intelligence, including machine learning (ML) techniques, offer new opportunities to systematically analyze large datasets and uncover hidden patterns that may not be evident through traditional clinical assessments alone11,12. Identifying predictive markers for chronic TMD is essential for timely intervention, especially when behavioral and structural factors can be addressed to prevent disease progression.

In this study, we aimed to identify clinical, sleep-related, and MRI-derived predictors of chronic TMD by comparing classical statistical methods with AI-based models. We employed logistic regression as a representative statistical and ML approach, and a deep neural networks (DNNs) with multiple hidden layers and regularization techniques as a more expressive non-linear model. While our DNN is relatively simple compared to advanced deep neural architectures such as convolutional neural networks, recurrent neural networks, or transformer-based models, we designed it to be suitable for analyzing our structured tabular data13. This comparative analysis between logistic regression and DNN highlights the trade-off between interpretability and predictive power. The DNN demonstrates an ability to capture more complex feature interactions, whereas logistic regression offers greater transparency and ease of clinical interpretation, which may facilitate broader clinical adoption. This model selection reflects a balance between feasibility, data structure, and clinical applicability. Through this approach, we aim to improve the early detection of chronicity-related risk factors and provide clinical insights to support personalized treatment planning.

In this study, we identify key clinical and MRI-based predictors of chronic TMD. The logistic regression model demonstrates that bruxism, TMJ noise, pain intensity, sleep disturbances, and structural changes such as anterior disc displacement and joint space narrowing are significant predictors of chronicity. The deep neural network model improves prediction performance by approximately 4% compared to logistic regression. These findings highlight the utility of integrating behavioral and imaging features and support the potential of AI-based models for early risk stratification in TMD.

Methods

The research protocol for this study was reviewed to ensure compliance with the principles of the Declaration of Helsinki and was approved by the Institutional Review Board of Kyung Hee University Dental Hospital in Seoul, South Korea (KHD IRB, IRB No-KH-DT24025). Informed consent was obtained from all participants prior to their inclusion in the study.

Study population

The study population comprised 239 consecutive patients with TMD (161 women and 78 men; mean age 35.60 ± 17.93 years) who visited Kyung Hee University Dental Hospital between January 2020 and September 2024. All diagnoses were made by a specialist with more than 10 years of clinical experience following the diagnostic criteria for TMD Axis I14. Patients with TMD were identified, and all clinical reports and MRI images of the TMJs were retrospectively reviewed. The duration of TMD symptoms, as reported by the patients, was recorded in months, with 6 months used as the threshold to classify the patients into two groups: acute TMD (symptom duration <6 months) and chronic TMD (symptom duration ≥6 months)15. This study compared the clinical and MRI characteristics associated with chronic TMD to those observed in acute TMD and examined factors contributing to the prolonged duration of symptoms. The exclusion criteria were history of severe injuries, such as unstable multiple trauma to the orofacial area and maxillary and mandibular fractures; systemic diseases potentially affecting the TMJ, such as rheumatic diseases, systemic osteoarthritis, pregnancy, psychological problems, psychiatric or neurological disorders; and cases in which the structure of the TMJ complex was not clearly distinguishable on MRI16. Participants with missing values in key clinical, behavioral, and imaging variables were excluded from the analysis. Minor missing data in auxiliary variables, if encountered, were handled using listwise deletion during preprocessing.

Sample size

The required sample size was calculated using G*Power version 3.1.9.7 (Heinrich-Heine-Universität Düsseldorf, Düsseldorf, Germany). A minimum of 134 participants was needed to achieve a power of 0.95 with an alpha level of 0.05. In total, 239 patients with TMD were recruited, exceeding the minimum requirement. Although the sample size was relatively small for deep learning applications, several strategies were implemented to mitigate the risk of overfitting and enhance model generalizability. These strategies included applying dropout regularization in the DNN, employing L2 regularization (weight decay) in logistic regression, and adopting a revised data splitting strategy (60:20:20 for training, validation, and test sets). In addition, 5-fold cross-validation was performed, and the final model parameters were selected based on the epoch with the lowest validation loss. Furthermore, training and validation loss curves were carefully monitored to confirm that no severe overfitting occurred during model training.

Pain intensity

Pain intensity was assessed using the visual analog scale (VAS). The VAS score ranged from 0 to 10, with 0 indicating no pain and 10 indicating the worst imaginable pain.

Clinical symptoms and contributing factors for TMD

Six clinical symptoms were investigated in the patients with TMD: TMJ noise, TMD pain, locking, muscle stiffness, bruxism, and tinnitus. The presence of each parameter was recorded based on patient reports and assessed dichotomously as either “yes” or “no.”. The criteria for each significant complaint were as follows: (1) TMJ noise: sounds such as clicking and crepitus originating from the TMJ during both functional and nonfunctional movement of the mandible, (2) TMD pain: pain associated with TMD involving the TMJ structures, (3) Locking: jaw locking with a mouth opening of <35 mm, indicating limited mouth opening, as reported by the patient, (4) Muscle stiffness: stiffness, heaviness, or discomfort in muscles during function or at rest, (5) Bruxism: clenching or grinding of teeth while awake or sleeping, and (6) Tinnitus: a perception of various sounds, such as ringing or buzzing, without any corresponding external source. The three contributing factors included sleep disturbance, psychological distress, and a history of macro-trauma, with each factor assessed in a dichotomous manner.

STOP-bang and sleep time

The STOP-Bang questionnaire was used to evaluate factors associated obstructive sleep apnea (OSA). This validated screening tool is designed to identify individuals with a high likelihood of OSA. The STOP-Bang questionnaire consists of eight dichotomous (yes/no) questions related to the clinical features of sleep apnea. Each question, a response “yes” scores 1, a “no” response scores 0, and the total score ranges from 0 to 8. The exposure of interest was classified as either binary-low or high likelihood for OSA; the low likelihood of OSA: Yes to < 3 questions, moderate likelihood of OSA: Yes to ≥3 questions, and high likelihood of OSA: Yes to ≥5 questions17,18. In the logistic regression and DNN models predicting chronic TMD, individual STOP-Bang components were excluded to reduce the risk of overfitting due to high feature dimensionality. Instead, only binary indicators based on clinical cut-offs ( ≥ 3 and ≥5) were used. All patients were instructed to complete the STOP-Bang questionnaire. Additionally, we collected the average self-reported sleep time over the past two weeks in hours.

MR image acquisition

High-resolution MRIs were obtained using a 3 T MRI system (Signa™ Genesis, GE Healthcare, Chicago, IL, USA) with a 6-cm × 8-cm diameter surface coil. The MRI examinations were performed using the MR sequences and protocols of the Kyung Hee University Medical Center. All scans involved sagittal oblique sections (section thickness, ≤3 mm; field of view, 15 cm; matrix dimensions, 256 × 224), and spin-echo sagittal MRIs were obtained on axial localizer images. T2-weighted images (T2WIs) were obtained using a 650/14 repetition time (TR)/echo time (TE) and 2650/82 TR/TE sequences. Proton density (PD) images were obtained using a 2650/82 TR/TE sequence. MRI protocols for TMJ evaluation were performed as described previously2,19.

MRI abnormal findings

Anterior disc displacement (ADD), TMJ-OA, joint space narrowing, and effusion were coded dichotomously as positive or negative. The left and right sides of the patients with bilateral TMJ and ADD were evaluated separately using T2-weighted (T2WI) and proton density (PD) images. If an abnormal finding was present on either side, it was recorded as ‘positive. MRI indicators for assessing ADD in patients with TMD were defined as follows20: a positive finding of ADD was identified when the posterior band of the articular disc was displaced anteriorly beyond the normal range in the closed-mouth position. A positive finding of TMJ-OA was indicated by the presence of cortical erosion, subchondral cysts, osteophyte formation, flattening, or sclerosis, either alone or in combination21. However, the presence of flattening and sclerosis alone is insufficient to diagnose TMJ-OA. Joint space narrowing was defined as a distance of less than 1.5 mm between the outer lines of the condyle and temporal bone22. A positive finding of TMJ effusion was recorded when a high-intensity signal was observed in the superior or inferior joint space on closed-mouth sagittal T2WI or PD-weighted images23. Intra-examiner reproducibility yielded Cohen’s kappa values of 0.78 and 0.85, while inter-examiner kappa values for TMJ effusion diagnosis were 0.79 and 0.83, respectively. Any disagreements were resolved through discussion until a consensus was reached.

AI-based evaluation and visualization

The neural network model used in this study follows a multi-layer perceptron (MLP) architecture designed to perform a binary classification of chronic TMD against acute TMD. The dataset was divided into 60% training, 20% validation, and 20% for testing. The input layer received clinical data (features), which were standardized to ensure that the mean of all input features was zero, and the standard deviation was 1. This was followed by five hidden layers that were sequentially connected, with each layer consisting of 256, 64, 16, 4 and 2 neurons, respectively, gradually decreasing in size. Each hidden layer applies the ReLU activation function to introduce nonlinearity, thereby enhancing the model’s ability to learn complex patterns. To prevent overfitting, a dropout rate of 30% was applied to each hidden layer. The output layer consisted of two output nodes for binary classification. CrossEntropyLoss was used as the loss function, and softmax was applied to perform the classification. The activation function was applied in the final output layer. The model was trained for 100 epochs, and the training loss, validation loss, and accuracy were recorded for each epoch. The Adam optimizer was employed due to its ability to handle sparse gradients and noisy data efficiently, as well as its adaptive learning rate, which allows for faster convergence compared to standard stochastic gradient descent. The optimizer combines the advantages of momentum and RMSProp, making it particularly effective for deep-learning tasks. The model with the best performance, defined as the one with the lowest validation loss, was selected and evaluated using the test set to determine its final accuracy. The performance of the DNN model was compared to that of traditional machine learning methods to assess its effectiveness in predicting chronic TMD. The diagnostic performance of the DNN model was compared to that of conventional machine learning models (logistic regression model). To advance beyond an isolated understanding of the relationships between variables, we aimed to provide a comprehensive and intuitive understanding through 2D (two-dimensional) and 3D (three-dimensional) visualizations of the relationships between chronic TMD and related variables. During training, the model stored the weights at the point at which the validation loss was minimized, thereby ensuring optimal performance.

To improve the interpretability of the DNN model, we applied SHapley Additive exPlanations (SHAP), which quantify the contribution of each feature to the model’s prediction. Summary plots of mean SHAP values across 5-fold cross-validation were generated to visualize the most influential predictors of chronic TMD, including both the magnitude and direction of their impact. This approach provided transparent insights into the model’s decision-making process and enhanced the clinical relevance of the AI-based analysis.

Statistics and reproducibility

Data were analyzed using IBM SPSS Statistics for Windows (version 26.0; IBM Corp., Armonk, NY, USA). The 19 clinical characteristics and MRI findings used to predict chronic TMD, as opposed to acute TMD, included 15 clinical variables—sex, age, visual analog scale (VAS) score, TMJ noise, TMD-related pain, sleep time, joint locking, muscle stiffness, bruxism, macrotrauma, tinnitus, psychological distress, sleep problems, and STOP-Bang score thresholds ( ≥ 3 and ≥5)—and 4 MRI findings: ADD, TMJ-OA, effusion, and joint space narrowing. Descriptive statistics are reported as mean ± standard deviation or frequencies with percentages, as appropriate. The distributions of categorical data were assessed using the χ² test with Bonferroni correction for equality of proportions. Student’s t-tests were used to compare the mean values between the two TMD groups, with false discovery rate control applied to adjust for multiple comparisons among continuous variables. Cramer’s V analysis was also used to assess the strength of the associations between the two variables; the statistical values ranged from 0 to 1, with values closer to 1 indicating a stronger correlation. To predict chronic TMD compared to acute TMD, we identified variables with significant differences in means or percentages between acute and chronic TMD groups using t-tests, χ² tests, and Bonferroni correction. The contribution of each selected variable was further evaluated using a single logistic regression analysis, with the results expressed as odds ratios (ORs) and 95% confidence intervals (CIs). To assess the combined predictive power of these factors for chronic TMD, we performed multiple logistic regression analysis with backward selection. To evaluate the statistical significance of prediction differences between logistic regression and the DNN, McNemar’s test was applied to the paired classification outcomes. All analyses were considered statistically significant at a two-tailed p-value of <0.05. The performance of the prediction model for chronic TMD was evaluated by plotting the receiver operating characteristic (ROC) curve against the area under the ROC curve (AUC) calculated for each AI model. AUC values were interpreted as follows: AUC  =  0.5 (no discrimination), 0.6  ≥  AUC  >  0.5 (poor discrimination), 0.7  ≥  AUC  >  0.6 (acceptable discrimination), 0.8  ≥  AUC  >  0.7 (excellent discrimination); and AUC  >  0.9 (outstanding discrimination)24.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Demographics

Among the 239 patients with TMD, 117 (48.95%) were diagnosed with acute TMD, and 122 (51.05%) with chronic TMD. The female-to-male ratio across the sample was 161:78 (2.06:1). Notably, no statistically significant differences were observed in age or sex between the acute and chronic groups. The mean age of patients with acute TMD was 36.36 ± 17.69 years, while that of those with chronic TMD was 34.88 ± 18.20 years (p = 0.524). Similarly, the female-to-male ratio did not differ significantly between acute TMD (31.6% male vs. 68.4% female; 2.16:1) and chronic TMD (33.6% male vs. 66.4% female; 1.98:1, p = 0.783) (Table 1).

Table 1 Demographics and clinical characteristics of TMD patients
Full size table

Clinical characteristics

The duration of symptoms in acute TMD was 0.77 ± 0.67 months, while in chronic TMD it was 26.75 ± 36.37 months (p < 0.001). The VAS score, reflecting subjective pain intensity, was significantly higher in chronic TMD (4.82 ± 2.47) compared to acute TMD (3.64 ± 2.38, p < 0.001).

Among the six clinical symptoms evaluated, significant group differences between the acute and chronic TMD groups were observed for TMJ noise and bruxism. TMJ noise was more frequently reported in patients with chronic TMD than in those with acute TMD (52.1% vs. 70.5%; p = 0.005). Similarly, bruxism was significantly more frequent in patients with chronic TMD (31.1%) than in those with acute TMD (15.4%; p = 0.006). The most common clinical symptom was TMD pain, which was observed in 77.8% of the acute TMD cases and 78.7% of the chronic TMD cases; however, no significant differences were observed between the two groups (p = 0.877). Tinnitus was present in 33.3% of patients with acute TMD and 27.0% of patients with chronic TMD (p = 0.325). Notably, no significant differences were observed between the two TMD groups in terms of locking, muscle stiffness, or tinnitus. Regarding contributing factors, psychological distress was observed at similar frequencies in both groups (45.3% vs. 44.3%, p = 0.897), as was microtrauma (10.3% vs. 6.6%, p > 0.05) (Table 1).

Sleep-related factor

Sleep time was significantly shorter in chronic TMD (6.41 ± 2.37 h) compared to acute TMD (7.37 ± 2.17 h, p < 0.001). Sleep problems were significantly more frequent in the chronic TMD group (35.2%) than in the acute TMD group (22.2%; p = 0.032) (Table 1).

In the STOP-Bang questionnaire, significant differences were observed in the proportion of patients who answered “yes” to items 1 (Snoring), 3 (Observed apnea), and 5 (Body mass index ≥ 35 kg/m²). Specifically, patients with chronic TMD reported these symptoms more frequently than those with acute TMD: STOP-Bang 1 (20.5% vs. 39.3%, p = 0.002), STOP-Bang 3 (13.7% vs. 35.2%, p < 0.001), and STOP-Bang 5 (15.4% vs. 27.0%, p = 0.039). In contrast, no significant differences were observed between the two groups in the proportions of patients reporting the other items: STOP-Bang 2 (Tiredness), 4 (High blood pressure), 6 (Age ≥ 50 years), 7 (Neck circumference > 40 cm), and 8 (Male gender). The STOP-Bang total score (STOP-Bang score) was significantly higher in chronic TMD patients compared to acute TMD patients (3.02 ± 2.09 vs. 2.39 ± 1.73, p = 0.012). Additionally, the proportion of patients with a STOP-Bang score ≥ 3 was significantly higher in chronic TMD (59.8%) than in acute TMD (44.4%, p = 0.020). Similarly, the proportion of patients with a STOP-Bang score ≥ 5 was also significantly higher in chronic TMD (27.0%) compared to acute TMD (12.0%, p = 0.003) (Table 2).

Table 2 Results of the STOP-Bang Questionnaire
Full size table

MRI findings

In the MRI findings of patients with TMD, the most frequently observed features were as follows: in acute TMD, the order was joint space narrowing > effusion > ADD > TMJ-OA, whereas in chronic TMD, the order was joint space narrowing > ADD > TMJ-OA > effusion. The prevalences of ADD (60.7% vs. 86.9%, p < 0.001), TMJ-OA (58.1% vs. 82.0%, p < 0.001), and joint space narrowing (64.1% vs. 88.5%, p < 0.001) were significantly higher in patients with chronic TMD than in those with acute TMD (Fig. 1). Although effusion was more frequent in the chronic TMD group (71.3%) than in the acute TMD group (62.4%), the difference was not statistically significant (p = 0.169). Excluding effusion, the other three MRI findings showed statistically significant differences between the two groups (Table 3). Figure 2.

Fig. 1: MRI Findings in acute and chronic TMD patients.
figure 1

A Distribution of MRI findings, including anterior disc displacement (ADD), TMJ osteoarthritis (OA), effusion, and joint space narrowing, in acute and chronic TMD cases. B MRI of acute TMD (1 month) showing ADD with effusion. C MRI of acute TMD (2 months) displaying ADD. D MRI of chronic TMD (10 months) showing joint space narrowing along with TMJ-OA, ADD, and disc deformity. E MRI of chronic TMD (2 years) demonstrating both ADD and TMJ-OA. All images are proton density MRI scans, which provide detailed soft-tissue contrast essential for evaluating TMJ structural changes. TMJ temporomandibular joint, c mandibular condyle, d articular disc, e effusion. The results were obtained using the χ² test with Bonferroni correction. Analyses were conducted on data from all 239 patients. Statistical significance was set at p < 0.05. ***: p < 0.001.

Full size image
Fig. 2: Relationship between key clinical characteristics and MRI findings.
figure 2

TMD temporomandibular disorder, Chronicity: A chronic state in which TMD symptoms have persisted for more than six months, TMJ temporomandibular joint, ADD anterior disc displacement, TMJ-OA osteoarthritis of the temporomandibular joint. A Correlation between MRI findings across the entire dataset; B Correlation between MRI findings in patients with acute TMD; and C Correlation between MRI findings in patients with chronic TMD. The closer the absolute value of the correlation coefficient is to 1, the stronger the relationship between the two variables. Values near zero indicated little or no correlation. Visualization based on the full dataset (n = 239).

Full size image
Table 3 Comparison of MRI findings between acute and chronic TMD groups
Full size table

Logistic regression models for chronic TMD prediction

Single logistic regression analysis was used to examine whether each factor served as a significant predictor of chronic TMD compared to acute TMD. Joint space narrowing emerged as the most powerful predictor, increasing the likelihood of developing chronic TMD by 4.320 times compared with acute TMD (OR = 4.320, 95% CI: 2.204–8.466, p < 0.001). ADD (anterior disc displacement) significantly increased the likelihood of chronic TMD by 4.292 times (OR = 4.292, 95% CI: 2.256–8.168, p < 0.001). Similarly, TMJ-OA was associated with a 3.275-fold increase in the likelihood of chronic TMD compared to acute TMD (OR = 3.275, 95% CI: 1.816–5.908, p < 0.001). Other significant predictors for chronic TMD included bruxism (OR = 2.488, 95% CI: 1.323–4.680, p = 0.005), STOP-Bang score ≥5 (OR = 2.728, 95% CI: 1.373–5.420), TMJ noise (OR = 2.193, 95% CI: 1.288–3.733, p = 0.004), sleep problems (OR = 1.905, 95% CI: 1.075–3.733, p = 0.004), STOP-Bang score ≥5 (OR = 1.862, 95% CI: 1.114–3.113), and an increase in VAS score (OR = 1.220, 95% CI: 1.094–1.361, p < 0.001). Conversely, an increase in sleep time reduced the likelihood of developing chronic TMD by 0.826 times (OR = 0.826, 95% CI: 0.733–0.931, p = 0.002).

In the multiple logistic regression analysis with backward selection, all previously mentioned factors were included simultaneously to determine their collective impact on the development of chronic TMD and the extent to which each factor contributed. Among these predictors, bruxism was identified as the most powerful predictor, increasing the likelihood of chronic TMD by 4.048 times (OR = 4.048, 95% CI: 1.786–9.173, p = 0.01) (Table 4).

Table 4 Single and multiple regression analysis for predicting chronic TMD using the total sample (n = 239)
Full size table

Cut-off value for predicting chronic TMD

Figure 3 presents the cut-off values for sleep time, VAS, and STOP-Bang total scores in predicting chronic TMD. Among these, sleep time showed the strongest predictive power, with a cut-off value of 6.750 h (area under the curve [AUC] = 0.614, 95% confidence interval [CI]: 0.543–0.685, p = 0.002). Epidemiological data have demonstrated that short sleep duration ( < 7 h) is linked to a heightened risk of chronic musculoskeletal pain25. Nonetheless, empirical research directly evaluating sleep duration among individuals with TMD is currently scarce. This suggests that individuals who sleep less than 6.750 h are more likely to develop chronic TMD. Similarly, the VAS score was also a significant predictor of chronic TMD, with a cut-off value of 4.5 (AUC = 0.647, 95% CI: 0.567–0.707, p < 0.001). A VAS score higher than 4.5 indicates a significantly increased risk of developing chronic TMD. The STOP-Bang total score had a cut-off value of 2.50 (AUC = 0.594, 95% CI: 0.522–0.667, p = 0.012), with scores above this threshold significantly increasing the likelihood of chronic TMD.

Fig. 3: Cut-off value for predicting chronic TMD.
figure 3

VAS visual analog scale, AUC area under curve, STOP-Bang snoring, tiredness, observed apnea, high blood pressure (STOP)-BMI, age, neck circumference, and gender (Bang). When using acute TMD as a reference, the significant predictors for chronic TMD included the VAS score with a cut-off value of 4.50, STOP-Bang score with a cut-off value of 2.50, and sleep time with a cut-off value of 6.75 h. Model performance was evaluated using the full dataset (n = 239).

Full size image

Correlations between the clinical characteristics and MRI findings

When analyzing the entire dataset, the chronicity of TMD symptoms was associated with ADD (r = 0.30), joint space narrowing (r = 0.29), TMJ-OA (r = 0.26), and effusion (r = 0.09). Additionally, symptom chronicity was associated with clinical factors such as bruxism (r = 0.19), TMJ noise (r = 0.19), and TMD pain (r = 0.01). The factors that most correlated with symptom chronicity were ADD (r = 0.30), joint space narrowing (r = 0.29), TMJ-OA (r = 0.26), TMJ noise (r = 0.19), bruxism (r = 0.19), effusion (r = 0.09), and TMD pain (r = 0.01). Furthermore, ADD exhibited the strongest correlation with joint space narrowing (r = 0.82), along with positive correlations with TMJ-OA (r = 0.24) and effusion (r = 0.17). TMJ-OA showed a significant positive correlation with ADD (r = 0.24), effusion (r = 0.20), and joint space narrowing (r = 0.20). Effusion was also correlated with ADD (r = 0.17), TMJ-OA (r = 0.20), and joint space narrowing (r = 0.12).

A similar pattern of correlations was observed when focusing on acute TMD, consistent with the findings from the entire dataset. However, a stronger relationship was observed between ADD and joint space narrowing (r = 0.86). The correlation between effusion and ADD was also stronger than that observed in the whole dataset (r = 0.21), whereas the relationship between TMJ-OA and ADD was slightly weaker (r = 0.20) compared to the whole dataset.

For chronic TMD, the correlation between TMJ-OA and effusion was stronger than that for the entire dataset (r = 0.27). However, the interrelationships among ADD, effusion, and joint space narrowing were weaker in chronic TMD than in the entire dataset and acute TMD.

2D and 3D visualization of interrelationships

In the 2D visualization, factors significantly associated with chronic TMD (Euclidean distance ≤ 0.2) were identified. Among the MRI findings, the key factors were ADD, TMJ-OA, and joint space narrowing. For clinical characteristics, TMJ noise, bruxism, VAS, and sleep time were significantly associated with chronic TMD, with a STOP-Bang total score ≥ 5, also approaching a Euclidean distance of 0.2.

A 3D network was used to visualize the relationships among variables associated with chronic TMD. This visualization depicts the degree of interconnection between the factors associated with chronic TMD and their relationships with other variables. Among the clinical characteristics linked to chronic TMD, TMJ noise and bruxism are interrelated. Additionally, the VAS score was related to chronic TMD and was associated with effusion and sleep problems. MRI findings associated with chronic TMD include ADD, joint space narrowing of the TMJ, and TMJ-OA. The ADD demonstrated a strong relationship with joint space narrowing (Fig. 4).

Fig. 4: 2D and 3D plots (training set, n = 191) visualizing correlations among features associated with chronic temporomandibular disorders (TMD).
figure 4

A Two-dimensional force-directed plot illustrating the strength of correlation between chronic TMD and clinical, behavioral, and imaging-based features. Node size reflects variable importance, and edge thickness indicates correlation strength. B Three-dimensional force-directed network graph showing the interconnectivity and clustering of features related to chronic TMD. This visualization helps reveal latent patterns among variables. An interactive version is available at: https://htmlpreview.github.io/?https://github.com/SeonggwangJeon/Chronic-TMD/blob/main/Chronic%20TMD_3d_force_directed_correlation.html. TMJ temporomandibular disorder, TMJ OA osteoarthritis on temporomandibular disorder, ADD anterior disc displacement, VAS visual analogue scale, STOP-Bang snoring, tiredness, observed apnea, high blood pressure (STOP)-BMI, age, neck circumference, and gender (Bang).

Full size image

Machine learning algorithms for chronic TMD

We employed logistic regression model within machine learning to identify significant predictors of chronic TMD and extracted the weight values for each factor, as presented in Fig. 5. This figure shows the variables useful for predicting chronic TMD and highlights their respective weight values. Significant predictors with an absolute weight value ≥ 0.4 were identified in the following order: bruxism (0.6768), VAS (0.5812), sleep problem (0.5230), joint space narrowing (0.5116), TMJ noise (0.4457), ADD (0.4337), and STOP-Bang ≥ 5 (0.4082) (Fig. 5A). Thus, the presence or increase in these factors is associated with a higher likelihood of chronic TMD. Conversely, sleep duration was the only significant negative predictor. As sleep time decreases (weight = −0.5926), the likelihood of developing chronic TMD increases. SHAP analysis identified statistically significant predictors of chronic TMD based on 95% confidence intervals excluding zero. Significant contributors (SHAP > 0) included bruxism (mean SHAP = 0.008), TMJ noise (0.004), muscle stiffness (0.0035), macrotrauma (0.003), and tinnitus (0.0025), indicating increased risk (Fig. 5B).

Fig. 5: Predictors of Chronic TMD Identified through Machine Learning.
figure 5

A Feature weights in predicting chronic TMD. Feature weights derived from the machine learning model (logistic regression) indicating each variable’s direction and magnitude of influence on chronic TMD prediction. B Feature importance analysis using SHAP. SHAP summary plot showing the mean contribution and variability of each feature across cross-validation folds, highlighting the most influential predictors in the deep neural network model.

Full size image

Comparison of prediction performance between machine learning and deep neural network

Logistic regression, a commonly used machine learning model, was employed to predict chronic TMD based on 19 clinical characteristics and MRI findings. The model achieved an AUROC of 0.7550 (95% confidence interval [CI], 0.6550–0.8550). The prediction accuracy for chronic TMD was 0.7083, with higher sensitivity (73.4 %) compared to specificity (68.4%). When using Multi-Layer Perceptron (MLP), a type of deep neural networks, to predict chronic TMD, the AUROC improved to 0.7949 (95% CI: 0.6949–0.8949). However, the difference in the AUROC between logistic regression and MLP was not statistically significant (p = 0.3067). In other words, the prediction performance of deep learning (MLP) was 3.99% higher than that of logistic regression (79.49% vs. 75.50%); however, the difference was not statistically significant (p = 0.3067). The MLP model achieved a diagnostic accuracy of 77.08% for chronic TMD compared to acute TMD. Similar to the logistic regression, MLP demonstrated higher sensitivity (79.3%) than specificity (73.7%) (Fig. 6).

Fig. 6: Comparison of prediction performance between machine learning and deep neural network models.
figure 6

A Receiver operating characteristic (ROC) curves and performance metrics. B Confusion matrix comparing the prediction outcomes between logistic regression and DNN. The McNemar test comparing the two models yielded a chi-square value of χ² = 1.125 with a p-value of 0.289, indicating no statistically significant difference. Performance evaluation was based on a 20% held-out test set (n = 48). PPV positive predictive value, NPV negative predictive value, AUROC area under the receiver operating characteristic curve, CI confidence interval.

Full size image

To assess training stability and generalizability, loss curves from 5-fold cross-validation were analyzed (Supplementary Fig. 1). For logistic regression (Supplementary Fig. 1a–e), training loss consistently decreased, while validation loss declined initially and then slightly increased—a typical learning pattern. The dashed lines mark the epochs with minimum validation loss, which were used for model selection. The consistency across folds indicates stable training without underfitting or overfitting. For the deep neural network (Supplementary Fig. 1f–j), training loss steadily decreased, and validation loss showed an early drop followed by a mild rise, suggesting slight overfitting. Early stopping based on minimum validation loss helped prevent severe overfitting. These results support the robustness of both models.

Discussion

This study provides valuable insights into the distinguishing characteristics of acute and chronic TMD, focusing on clinical, sleep-related, and MRI-based predictors of chronicity. Our findings revealed that clinical and behavioral factors, including TMJ noise, bruxism, sleep problems, and elevated STOP-B Bang scores, were more prevalent in patients with chronic TMD than in those with acute TMD. Furthermore, structural abnormalities detected using MRI, such as anterior disc displacement (ADD), TMJ osteoarthritis, and joint space narrowing, highlight the relevance of joint-related changes in the chronicity of TMD symptoms. These findings underscore the multifaceted nature of chronic TMD and emphasize the importance of early intervention to prevent symptom persistence and deterioration. Identifying these key predictors will help clinicians proactively manage TMD and improve long-term patient outcomes.

Pain, sleep disturbance, and chronic symptoms are closely interconnected26,27. This study confirmed that patients with chronic TMD experience greater subjective pain, as reflected by higher VAS scores and shorter sleep durations, compared to those with acute TMD. These findings highlight that high pain intensity is significantly associated with chronic TMD, and sleep duration is notably shorter in patients with chronic TMD than in those with acute symptoms. These behavioral patterns suggest that chronic pain is closely linked to sleep disturbances, which is consistent with prior research showing that poor sleep quality exacerbates pain perception and impairs recovery from musculoskeletal conditions28. Elevated STOP-Bang scores in patients with chronic TMD further underscore the need to address sleep-related breathing disorders such as OSA as part of TMD management strategies. Although a complete consensus has not been reached, OSA is considered a potential risk factor for TMD29. Emerging studies have suggested that OSA and TMD share several overlapping mechanisms—such as sleep disturbances, bruxism, and inflammatory responses, which may contribute to the development or exacerbation of TMD symptoms30,31. Both conditions can disrupt normal sleep patterns, leading to increased pain sensitivity and psychological distress, which further complicates clinical outcomes32. In this study, machine learning revealed that a high risk of OSA, as indicated by a high STOP-Bang total score, is a significant predictor of chronic TMD. As research continues to explore this relationship, identifying OSA as a risk factor for TMD holds promise for integrated therapeutic approaches targeting both conditions. These findings are consistent with those of previous studies, suggesting that early intervention targeting joint health may prevent long-term damage and improve clinical outcomes.

The MRI findings indicate that structural changes in the TMJ play a significant role in the transition from acute to chronic TMD. The higher prevalence of ADD, TMJ-OA, and joint space narrowing in patients with chronic TMD supports the hypothesis that prolonged TMD contributes to joint deterioration and persistent symptom. TMJ noise is common in acute TMD and may be linked to ADD9,33. ADD plays a pivotal role in the progression of chronic TMD by initiating structural changes34. As ADD develops, it may gradually lead to joint space narrowing due to increased friction between the mandibular condyle, temporal bone, or articular disc35,36. This friction can further contribute to the development of TMJ-OA over time. In a recent MRI study, chronic TMJ inflammation has been associated with disc anomalies and bony erosive or deformative changes37.

In contrast, effusion, defined as fluid accumulation within the TMJ, was observed in a high proportion of patients (62.4% in acute TMD and 71.3% in chronic TMD). However, effusion did not emerge as a statistically significant predictor of chronicity in our model (p = 0.169). Despite its frequency, effusion demonstrated weak associations with other structural MRI abnormalities, such as ADD and joint space narrowing. This suggests that effusion may serve as a nonspecific marker of TMJ inflammation rather than a reliable indicator of structural progression. Prior studies have also reported that effusion can be present in both symptomatic and asymptomatic individuals, limiting its diagnostic specificity38,39. While effusion may reflect an active inflammatory response associated with TMD symptoms40, its inconsistent correlation with joint degeneration indicates that it might act more as a transient phenomenon than a predictor of chronic TMD. These findings highlight the need for a more nuanced interpretation of effusion in MRI analyses and support the inclusion of multiple structural and clinical variables when assessing chronicity.

From a diagnostic perspective, both the logistic regression and DNN models exhibited high performance in predicting chronic TMD, demonstrating their potential utility in clinical practice. Although the DNN achieved a slightly higher AUROC (0.7708) compared to logistic regression (0.7083), the difference was not statistically significant. Despite this modest performance gap, the inclusion of both models provides practical insights into their respective advantages. To better understand the relative strengths and limitations of different modeling approaches, we compared logistic regression and MLP models. Logistic regression provides a clinically interpretable and robust baseline, whereas the MLP model captures more intricate patterns in the data, illustrating the trade-off between clinical applicability and methodological innovation41,42. These complementary perspectives help minimize diagnostic errors and improve decision-making efficiency, ultimately enhancing patient outcomes43. Their strength lies in their ease of interpretation, which is crucial for clinicians making both time-sensitive and complex clinical decisions. Logistic regression, in particular, is highly interpretable, easily implemented, and well-aligned with clinical workflows44. In contrast, the DNN model excels in capturing complex nonlinear relationships and feature interactions that may be overlooked by traditional statistical methods45. It produces transparent coefficients that clinicians can readily interpret and use to guide patient care.

In contrast, the DNN captures non-linear relationships and complex interactions among features that may be overlooked by conventional statistical models46. The use of both models in this study reflects a deliberate trade-off between interpretability and predictive power—an issue increasingly emphasized in medical AI research. While DNNs may provide incremental gains in performance, particularly when larger and more complex datasets are available, the transparency and clinical familiarity of logistic regression remain valuable in real-world healthcare settings47,48. As deep learning methods continue to evolve and gain acceptance in clinical applications, traditional models like logistic regression still offer an optimal balance between performance and usability49. Additionally, this study incorporated advanced 2D and 3D visualizations along with AI-based analysis, which further enhance the understanding of the multifactorial mechanisms contributing to chronic TMD.

Bruxism, characterized by the involuntary clenching or grinding of teeth, is a significant contributing factor to the development and persistence of chronic TMD50,51,52. This condition can occur during sleep (sleep bruxism) or while awake (awake bruxism), and both forms are associated with excessive loading of the TMJ and masticatory muscles53,54. Repetitive mechanical stress from bruxism can exacerbate joint wear and tear, increasing the likelihood of structural abnormalities such as ADD and TMJ-OA55. These changes strongly correlate with symptom chronicity, as confirmed by the higher prevalence of both bruxism and MRI abnormalities in patients with chronic TMD than in those with acute TMD. Behaviorally, bruxism can disrupt sleep, contributing to poor sleep quality and increased pain sensitivity, further complicating TMD management27,30. This is consistent with the findings of the present study, in which patients with chronic TMD reported a higher frequency of bruxism and shorter sleep duration. Addressing bruxism through targeted interventions, such as behavioral therapy, occlusal appliances, or stress management, may help reduce joint strain, improve sleep quality56, and ultimately prevent the progression from acute to chronic TMD. This emphasizes the need for clinicians to proactively evaluate and manage bruxism in patients proactively, given its pivotal role in the persistence and progression of symptoms.

Although prior studies have consistently associated psychological distress and microtrauma with the development and persistence of chronic pain conditions57,58, these factors did not reach statistical significance in our analysis. Psychological stress or distress is widely recognized as a contributor to pain chronification through mechanisms such as central sensitization and dysregulation of the hypothalamic-pituitary-adrenal axis59. Similarly, microtrauma—defined as repetitive minor injuries to the temporomandibular joint or masticatory muscles—has been suggested to play a role in the onset and persistence of TMD9. One possible explanation is that the assessment of these variables relied solely on binary self-report items60, which may lack the sensitivity needed to capture the complexity and variability of individual experiences. Furthermore, while the use of a 6-month cut-off for symptom duration is consistent with clinical practice61, dichotomizing a continuous variable may have resulted in some loss of information and may not fully reflect the underlying spectrum of chronicity. These methodological limitations may have led to an underestimation of the true prevalence or clinical impact of these factors within our cohort. Moreover, subjective responses are inherently susceptible to recall bias and underreporting, particularly in the context of chronic or multifactorial conditions62. Future research should incorporate validated psychometric instruments and detailed clinical assessments to more accurately characterize the role of psychological and mechanical factors in the chronicity of TMD.

This study suggests that incorporating clinical symptoms, including bruxism, sleep patterns, and structural abnormalities, into treatment strategies may help mitigate the progression to chronic TMD. However, this study has certain limitations. First, the retrospective design may have introduced selection bias and limits the ability to establish causal relationships between identified predictors and chronic TMD. Future prospective studies are warranted to validate and extend these findings. Second, although the sample size met the required threshold for statistical analyses, it remains modest for training deep neural networks. To minimize the risk of overfitting and improve generalizability, several strategies were employed, including dropout regularization, L2 weight decay, 5-fold cross-validation, and the use of a separate validation set for model selection. Although internal cross-validation was performed to assess model stability, the generalizability of the findings could be further enhanced through external validation using independent, multi-center datasets. Future research should aim to replicate and extend these results across larger and more diverse populations. Third, while the STOP-Bang questionnaire was used to assess OSA risk, polysomnographic evaluation provided more accurate data on sleep disturbance. Fourth, although the study incorporated advanced AI models, the performance of the deep learning model was only marginally better than that of logistic regression, suggesting that further optimization and the inclusion of more comprehensive datasets may be required to enhance predictive accuracy. Finally, behavioral and psychological factors were assessed based on patient self-reports, which are subject to recall bias and may affect the reliability of the data. Future studies should consider longitudinal designs and include objective behavioral assessments to provide more robust evidence regarding the predictors of chronic TMD. Expanding to more diverse, high-dimensional datasets and exploring advanced AI architectures may help improve the predictive accuracy of chronic TMD models.

Conclusion

In this study, we identified key clinical and MRI features associated with the chronicity of TMD using both conventional and AI-based predictive models. Our results highlight the roles of TMJ noise, bruxism, pain intensity, sleep-related factors, and MRI findings such as anterior disc displacement and joint space narrowing in distinguishing chronic from acute TMD. Although the deep neural network demonstrated slightly higher predictive performance compared to logistic regression, the difference was not statistically significant. Nevertheless, incorporating both models provides complementary insights: logistic regression offers simplicity and clinical interpretability, while the DNN captures complex, nonlinear relationships among features. This trade-off underscores the need to balance model transparency and predictive power when applying AI to clinical decision-making. Future research with larger datasets and more advanced architectures may further refine the prediction of chronic TMD progression.