Abstract
Abstract Introduction Artificial intelligence (AI) has the potential to redefine periodontal prognosis; however, its clinical readiness remains uncertain. This scoping review mapped and critically evaluated AI-based models for predicting periodontal disease progression and tooth loss. Methods Following PRISMA-ScR guidelines, PubMed, Scopus, and Web of Science were systematically searched without restriction by publication date. Of 567 records identified, 329 remained after deduplication. Three reviewers shared responsibility for title/abstract screening and full-text assessment. Each record was screened by one reviewer and independently verified by a second reviewer, while a senior reviewer audited the screening and data-charting processes. Twenty-eight full-text articles were assessed, of which eight studies met inclusion criteria. Data were synthesized descriptively according to study characteristics, model performance, predictors, validation methods, and clinical applicability. Results Seven of the eight studies used retrospective designs. Six studies predicted tooth loss, and two predicted periodontal disease progression. Random forest, neural networks, and logistic regression were among the most frequently evaluated approaches. Clinical periodontal variables were the most consistent predictors. Reported AUC or AUROC values generally ranged from approximately 0.71 to 0.91. One small, internally validated study reported near-perfect accuracy raising concerns about overfitting and limited generalizability. Only two studies performed external validation. One study directly comparing machine learning with a clinical prognostic system reported higher sensitivity but lower specificity for identifying unfavorable prognosis categories. None of the included studies reported prospective clinical implementation. Conclusion AI-based models show promise for predicting periodontal prognosis and tooth loss. However, methodological heterogeneity, limited external validation, and absence of prospective clinical implementation currently restrict their use in practice. Rigorous multicenter validation and direct comparisons with clinician-based prognostic assessments are needed before these models can be integrated into routine periodontal care.