Artificial intelligence models can predict the progression of periodontal disease and the loss of teeth with moderate to high accuracy, but almost none of them have been tested outside the data they were built on, and not a single one has been deployed in routine dental practice. That is the central conclusion of a scoping review published in Discover Artificial Intelligence, which systematically mapped the growing but still immature field of AI-based periodontal prognostication. The review, conducted according to PRISMA-ScR guidelines, searched PubMed, Scopus, and Web of Science without any restriction on publication date and identified 567 records.
After deduplication and rigorous screening by multiple reviewers, only eight studies met the eligibility criteria, a strikingly small number given the rapid expansion of artificial intelligence in dentistry more broadly. The eight included studies, published between 2021 and 2026, came from the United States, the United Kingdom, Lebanon, India, and multinational collaborations spanning Europe, North America, and Asia. Seven of the eight relied on retrospective designs, mining existing patient records rather than following patients forward in time.
Six studies focused on predicting tooth loss, with outcomes ranging from overall tooth loss and molar-specific loss to extraction need and tooth-loss phenotype, while two studies predicted periodontal disease progression, typically defined as changes in clinical attachment level over time. Prediction horizons varied enormously, from 12 months in the shortest study to 10 years in others, and one study examined a cohort with follow-up periods stretching from 10 to 50 years. The reported performance of these models ranged from moderate to high.
Area under the receiver operating characteristic curve values, the most commonly reported metric, generally fell between approximately 0.71 and 0.91. Among the disease-progression studies, one team reported an AUROC of 0.88 using a probabilistic graphical model that integrated clinical variables, demographic information, and salivary IL-1β biomarkers, while another achieved an AUROC of 0.82 with a random forest model predicting clinical attachment loss progression over at least two years of follow-up. In the tooth-loss category, a tuned random forest model reached an AUC of 0.9066 with 93 percent accuracy, and an externally validated logistic-regression and neural-network approach reported an AUC of 0.8943.
An externally validated ensemble model predicting molar loss came in at 0.726. One result stands out as a cautionary tale. A retrospective study of tooth loss prediction reported an AUC of 1.000 and an accuracy of 1.000, essentially perfect performance, using AdaBoost and neural-network models.
But those estimates came from a relatively small cohort of just 200 patients and were obtained through internal validation only, without any evaluation in an independent external dataset. The review's authors flag this as a red flag in itself: near-perfect performance is unusual in clinical prognostic modeling and raises concerns about overfitting, optimism bias, or other sources of performance inflation. Overfitting occurs when a model memorizes dataset-specific patterns or random noise rather than relationships that generalize to new patients, a risk that grows when model complexity is high relative to the number of patients and outcome events.
The algorithmic landscape of the field is broad. Neural networks, including multilayer perceptrons, logistic regression, support vector machines, Naïve Bayes classifiers, gradient boosting, decision trees, and ensemble stacking methods all made appearances. Some studies took more specialized routes, including RuleFit-based machine-learning pipelines designed to generate interpretable decision rules, and one study found that a probabilistic graphical model outperformed both logistic regression and neural networks on the same task.
Notably, the review found that increased model complexity did not consistently translate into superior performance, suggesting that data quality, feature relevance, and study design may matter more than algorithmic sophistication. What the models actually use to make predictions is reassuring from a biological standpoint. Clinical periodontal parameters were the most consistent predictors across all studies: probing depth, clinical attachment loss, tooth mobility, furcation involvement, bleeding on probing, plaque indices, baseline missing teeth, and periodontal staging or grading.
These align closely with established periodontal risk assessment frameworks, reinforcing the clinical plausibility of the models. Behavioral and systemic variables such as smoking, diabetes, age, hypertension, cardiovascular disease, and osteoporosis also featured prominently. One study went further, incorporating salivary biomarkers including IL-1β into its predictive models, hinting at a future of biologically informed, multimodal prognostication.
Yet the review exposes a deep methodological fault line: validation. Only two of the eight studies performed external validation using independent datasets from separate institutions or geographic regions, one drawing on cohorts from the United Kingdom and the United States and another spanning the United States, the United Kingdom, Germany, and China. The remaining six relied exclusively on internal validation techniques such as train-test splits, cross-validation, or bootstrapping.
Cross-validation can reduce bias in internal performance estimation, but it cannot answer whether a model remains accurate when clinical practices, patient characteristics, or measurement procedures differ. Tellingly, the externally validated studies reported more modest, and potentially more realistic, performance estimates than their internally validated counterparts, suggesting that published performance may decline when models are applied to independent populations. The comparison with conventional clinical practice is equally thin.
Only one study directly compared machine-learning models with a clinical tooth prognostic system, and the result was a tradeoff rather than a victory: the AI approach showed higher sensitivity for identifying teeth classified into unfavorable prognosis categories but lower specificity. In practical terms, the model would catch more high-risk teeth but would also incorrectly flag more healthy ones as doomed, potentially leading to unnecessary surveillance, additional treatment, or patient anxiety. The review emphasizes that comparing machine learning against statistical regression, which several studies did, is not the same as comparing it against clinician-assigned prognosis, which incorporates longitudinal context, treatment response, and patient preferences.
Reporting quality also varied considerably. Although every included study reported an AUC or AUROC value, additional metrics such as sensitivity, specificity, negative predictive value, F1-score, and confusion matrices were reported inconsistently. Only two studies reported calibration measures, which indicate whether predicted probabilities correspond to actual observed risks, and none reported confidence intervals for their performance estimates.
A model can discriminate well between patients who will and will not lose teeth while still systematically overestimating or underestimating absolute risk, a distinction that matters enormously when treatment decisions such as tooth extraction hang in the balance. The review also cautions that feature-importance and SHAP analyses describe how variables contribute to predictions but do not establish causal relationships or guarantee clinically appropriate reasoning. The bottom line, according to the review, is that artificial intelligence in periodontal prognostication is methodologically active but not yet clinically mature.
No included study reported prospective clinical implementation, evaluated clinician acceptance, assessed workflow burden, examined fairness across demographic groups, or measured the effect of model use on actual treatment decisions or patient outcomes. The authors call for prospective, multicenter studies with standardized outcomes, transparent reporting of discrimination and calibration, routine external validation in geographically and temporally distinct populations, and direct comparisons with clinician-based prognostic assessments. Until that evidence exists, AI models for predicting gum disease progression and tooth loss remain promising research tools rather than ready-made clinical instruments, and the gap between algorithm development and meaningful patient care remains wide open.
Subject of Research: Artificial intelligence models for predicting periodontal disease progression and tooth loss
Article Title: Artificial intelligence for predicting periodontal disease progression and tooth loss: a scoping review
Article References: Hung, M., Nelson, M., Chriss, H., Dhasi, S., Torp, K., & Mohajeri, A. (2026). Artificial intelligence for predicting periodontal disease progression and tooth loss: a scoping review. Discover Artificial Intelligence, 6(1), Article 987. https://doi.org/10.1007/s44163-026-02111-4
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02111-4
Keywords: artificial intelligence, machine learning, periodontitis, tooth loss, disease prediction, scoping review, dental prognosis, external validation, random forest, neural networks, overfitting, clinical implementation
News Source: Blake Davidson. (October 28, 2026). AI Shows Promise for Predicting Gum Disease and Tooth Loss, but Is Not Ready for the Clinic. Scienmag.
