Osteoarthritis of the temporomandibular joint (TMJ OA) is the most common disorder of the TMJ. A clinical decision support (CDS) system designed to detect TMJ OA could function as a useful screening tool as part of regular check-ups to detect early onset. This study implements a CDS concept model based on Random Forest and dubbed RF+ to predict TMJ OA with the hypothesis that a model which leverages high-resolution radiological and biomarker data in training only can improve predictions compared with a baseline model which does not use privileged information. We found that the RF+ model can outperform the baseline model even when privileged features are not of gold standard quality. Additionally, we introduce a novel method for post-hoc feature analysis, finding shortRunHighGreyLevelEmphasis of the lateral condyles and joint distance to be the most important features from the privileged modalities for predicting TMJ OA.
Keywords: Multimodal learning; Privileged learning; Temporomandibular joint.