TMD Diagnosis Using a Masked Self-Supervised Tabular Transformer.

Abstract

Temporomandibular disorders (TMDs) encompass a heterogeneous group of musculoskeletal conditions involving the temporomandibular joint (TMJ), masticatory muscles, and associated structures. Diagnosis remains challenging due to overlapping symptoms, multifactorial etiology, and variability across clinical settings. To address these limitations, we developed the Gated Attention Tabular Transformer (GATT), a novel deep-learning model that uses masked self-supervised learning and gated attention mechanisms, to classify TMD subgroups based on the diagnostic criteria for TMD (DC/TMD). A total of 4,644 structured clinical records from a university-based registry were analyzed, comprising 3,524 female and 1,120 male patients (mean age 36.9 ± 14.7 y), across 12 core TMD subgroups. GATT achieved robust diagnostic performance with area under the receiver-operating characteristic curve values ranging from 0.815 to 1.000, sensitivity from 0.652 to 1.000, and specificity from 0.773 to 1.000. The model significantly outperformed conventional machine-learning methods including logistic regression, random forest, support vector machine, and XGBoost as well as advanced tabular deep-learning models such as TabNet, TabTransformer, AutoGluon Tabular Predictor, and FT-Transformer. Shapley additive explanations (SHAP) analysis revealed "pain-free opening" (SHAP = 6.78, P < 0.001) and "current TMJ noise" (SHAP = 2.87, P = 0.003) as key features of mechanical TMJ disorders. Co-occurrence network analysis uncovered side-specific clustering and potential time-lagged progression between bilateral TMJs. These findings demonstrate the feasibility of using deep learning to classify heterogeneous TMD subgroups using only structured clinical data, without the need for imaging. The GATT model offers an accurate, explainable, and scalable tool to support clinician-assisted diagnosis and reduce variability in TMD management in real-world practice. These results support the integration of AI-driven tools such as GATT into clinical workflows for standardized, efficient, and patient-specific TMD diagnosis.