Neural network models for predicting relapses after orthodontic treatment: a systematic review

The problem of predicting relapses after orthodontic treatment remains relevant due to their high incidence (20–40%) and the multitude of interconnected risk factors (morphological, periodontal, functional, dental, and behavioral factors). The human factor complicates manual analysis of all parameters, which creates a pressing need for digital solutions in modern orthodontics. Neural network models are capable of comprehensively processing heterogeneous clinical data (3D scans, cephalometric X-rays, medical histories) and identifying hidden risk patterns, thus improving prediction accuracy. Aims — to conduct a systematic review of recent scientific articles on the prediction of relapses after orthodontic treatment using neural network models. A systematic review of the effectiveness of using artificial intelligence and neural networks to predict the likelihood of recurrence after orthodontic treatment, conducted in accordance with the recommendations of PRISMA-2020, is presented. As a result of a systematic search, 32 literary sources were selected that fully correspond to the given topic and inclusion criteria. It was found that the highest predictive accuracy (89.2%) was achieved using deep neural networks (DNNs), which took into account the PAR and IOTN indices, patient age, and type of appliance. The capabilities of clinical software programs (ClinCheck, SimplyCeph 3DS, Avantis 3D) for indirect risk assessment through the analysis of morphology, occlusion, and tooth movement were analyzed. Limitations in the use of neural networks were identified, including a lack of representative training data, difficulties in interpreting results, and the high cost of implementation. Thus, irrefutable evidence has been obtained for the high effectiveness of various neural networks in predicting tooth movement patterns and assessing possible relapses. Neural network models have significant potential to improve the stability of orthodontic treatment, provided that existing barriers are overcome. Thus, the uniqueness of this review lies in its comprehensive, scientifically based approach to studying the use of neural network models in predicting orthodontic treatment.