Beyond the Predicted Profile: Clinical Validity of Artificial Intelligence Based Facial Outcomes in Orthodontics

Introduction and Background: Artificial intelligence is increasingly being explored to predict facial soft tissue outcomes following orthodontic treatment. AI systems may aid in treatment expectations and profile prediction in patients. However, it is important to determine that the predicted outcome is not simply visually accurate.  Review: This narrative review discusses the differences between geometric accuracy, visual accuracy and clinical utility when using AI systems to predict post-orthodontic facial outcomes. Previous research has focused on using artificial neural networks, convolutional neural networks, deep neural networks and conditional generative adversarial networks. More recently, models have used machine learning with incisors and facial measurements, generative networks for lateral profile prediction and prediction from facial frontal photographs. Across all studies there has been varying success depending on the imaging type, region of interest, patient population and the type of measurement used. Difficulties were seen with predicting lower lip and chin areas. Additionally, even when there has been good success with numbers there has not been improved treatment decisions or outcomes.  Conclusions: Using AI to predict facial outcomes is an emerging field which may be beneficial. However, the clinical validity of the outcomes should be investigated. Future research should ensure independent validation occurs, measurement of error within each anatomical region, patient and clinician evaluation, clear communication of uncertainty and evidence of clinical decision making. Until this evidence is present the outcomes should be used as support.