# Decoding adolescent TMJ osteoarthritis with multimodal machine learning

## TL;DR

This study uses machine learning to improve the diagnosis of adolescent TMJ osteoarthritis by combining clinical and imaging data.

## Contribution

The study introduces a multimodal machine learning approach to enhance diagnostic accuracy for adolescent TMJ-OA.

## Key findings

- Imaging-only models showed high specificity but low sensitivity for TMJ-OA diagnosis.
- Combining clinical and imaging data improved sensitivity and model performance.
- Panoramic radiography was the strongest predictor in the decision tree models.

## Abstract

Background: Early and accurate diagnosis of adolescent
temporomandibular joint (TMJ) osteoarthritis (OA) is critical, as degenerative
changes during growth can cause lifelong pain and deformity. This study aimed to
identify key clinical and imaging predictors of adolescent TMJ-OA and to evaluate
multimodal machine learning models. Methods: The diagnostic utility was
evaluated in 79 adolescents (10–18 years) with TMJ pain using panoramic
radiography (PR) and MRI. TMJ-OA was diagnosed based on the Diagnostic Criteria
for Temporomandibular Disorders (DC/TMD). Three decision tree models were
developed: Model 1 (clinical-only), Model 2 (imaging-only), and Model 3 (combined
clinical and imaging). Logistic regression was used for the comparisons.
Results: To ensure a robust evaluation with a small sample size (n =
79), the models were assessed using nested 5-fold cross-validation. Model 2
(imaging only) had the highest specificity (0.7714 ± 0.2321), accuracy
(0.5942 ± 0.0966), and AUROC (0.719 ± 0.101), but a low sensitivity
(0.4472 ± 0.2065). PR evidence of TMJ-OA (feature importance = 0.70; OR =
3.93) was the strongest predictor and root node in the decision tree. Model 3
(combined clinical and imaging data) showed improved sensitivity (0.6056 ±
0.1829), identifying PR_TMJ_OA, MRI_TMJ_ADD (anterior disc displacement),
Visual Analog Scale (VAS) score, and age as key nodes (AUROC = 0.6573 ±
0.0338; OR = 2.85 for PR_TMJ_OA). Model 1 (clinical-only) had limited
predictive performance (AUROC = 0.4859 ± 0.0894), with symptom duration
(importance = 0.64; OR = 1.40), VAS score, and joint locking (importance = 0.20)
contributing modestly. A model using PR_TMJ_OA alone achieved perfect
specificity (0.9714 ± 0.0571) but low sensitivity (0.3806 ± 0.1458).
Conclusions: Although PR is a meaningful screening tool for adolescent
TMJ-OA, it remains insufficient as a standalone diagnostic modality. Multimodal
integration of clinical and MRI findings improves diagnostic accuracy and
provides interpretable, clinically aligned decision-support tools for TMJ-OA.