Explainable Multimodal Artificial Intelligence for Early Prediction of Periodontitis Progression Using Clinical, Radiographic, and Salivary Biomarkers

Abstract

Periodontitis is a multifactorial inflammatory disease characterized by substantial interindividual variability in progression, making early identification of high-risk patients challenging. This study proposes the development and external validation of an explainable multimodal artificial intelligence model for predicting short-term periodontal disease progression by integrating routinely collected clinical parameters, dental radiographs, and salivary inflammatory biomarkers. Adults diagnosed with stage I–III periodontitis will be prospectively recruited from multiple dental centers and followed for 24 months. Baseline data will include probing pocket depth, clinical attachment level, bleeding on probing, plaque indices, smoking status, metabolic factors, standardized bitewing and periapical radiographs, and salivary concentrations of selected inflammatory and tissue-remodeling biomarkers. Deep convolutional neural networks will extract radiographic features, while gradient-boosting algorithms will model structured clinical and biomarker variables. A multimodal fusion architecture will subsequently combine these information streams to predict patient-level and site-level periodontal progression. Model discrimination, calibration, sensitivity, specificity, area under the receiver operating characteristic curve, and decision-curve net benefit will be evaluated through internal cross-validation and independent external validation. SHapley Additive exPlanations will quantify the relative contribution of individual predictors and identify clinically interpretable interactions underlying estimated risk. Performance will also be compared with conventional logistic regression and established periodontal risk assessment methods. Subgroup analyses will examine model fairness across age, sex, smoking, and diabetes categories. The proposed framework is expected to determine whether multimodal artificial intelligence offers clinically meaningful improvements over conventional periodontal risk stratification while maintaining transparency in algorithmic decision-making. By transforming heterogeneous diagnostic information into individualized progression probabilities, this study could support risk-adapted recall schedules, earlier preventive interventions, and more efficient allocation of periodontal care. The findings may provide a clinically applicable foundation for explainable precision dentistry rather than purely image-based automated diagnosis.