ARTIFICIAL INTELLIGENCE-ASSISTED PREDICTION OF ORTHODONTIC TREATMENT OUTCOMES: A CLINICAL AND RADIOGRAPHIC STUDY

Abstract

Background: Accurate prediction of orthodontic treatment outcomes may facilitate individualized treatment planning and improve clinical decision-making. Artificial intelligence (AI) offers the potential to integrate clinical and radiographic variables to predict treatment response more objectively. Aim: To evaluate the performance of an AI-based model in predicting orthodontic treatment outcomes and to compare its predictive accuracy with conventional orthodontic assessment Materials and Methods: A total of 100 participants undergoing orthodontic treatment were evaluated. Clinical and radiographic/cephalometric parameters were assessed before and after treatment, and actual treatment outcomes were categorized as excellent, good, moderate, or poor. The AI model generated predicted treatment outcomes using the available clinical and radiographic parameters. Model performance was assessed using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, Cohen’s kappa, correlation analysis, intraclass correlation coefficients (ICC), and receiver operating characteristic (ROC) curve analysis. AI based prediction was also compared with conventional orthodontic assessment. Associations between baseline characteristics and prediction accuracy were evaluated using multivariable analysis. The study population had a mean age of 22.84 ± 3.76 years, with 54% females; fixed orthodontic appliances were used in 72% of participants and clear aligners in 28%. Results: Actual treatment outcomes were excellent in 28%, good in 45%, moderate in 22%, and poor in 5% of participants. The AI model correctly classified 83% of participants, with substantial agreement with actual outcomes (Cohen’s kappa = 0.743). For successful treatment, defined as an excellent or good outcome, the model demonstrated 95.0% accuracy, 95.9% sensitivity, and 92.6% specificity. ROC analysis showed excellent discrimination for treatment success (AUC = 0.92). Strong correlations were observed between AI-predicted and actual clinical parameters, while agreement was also high for clinical, radiographic, and combined outcomes. The AI approach demonstrated higher accuracy, sensitivity, specificity, and AUC than conventional orthodontic assessment. Baseline overjet and dental crowding were significantly associated with prediction accuracy, while age, sex, overbite, ANB, and FMA were not significant predictors. Cephalometric analysis demonstrated greater treatment related changes in dentoalveolar parameters than in skeletal measurements. Conclusion: The AI model demonstrated strong predictive and discriminatory performance for orthodontic treatment outcomes, particularly in distinguishing successful from less favourable treatment results. Its close agreement with clinical and radiographic findings and its superior performance compared with conventional assessment indicate that AI may serve as a useful adjunct to orthodontic treatment planning and outcome prediction. Further prospective studies involving larger and independently validated populations are warranted before routine clinical implementation.