Saliva has emerged as the most extensively studied biological medium for AI applications in periodontology due to its noninvasive collection, stability, and rich molecular content. The development of ML models using salivary biomarkers has shown remarkable progress in recent years, with several studies demonstrating clinical-grade diagnostic accuracy.

Deng and colleagues developed a comprehensive multiclass screening tool that integrates nonclinical parameters with salivary biomarkers using multiple ML algorithms. Their model achieved area under the curve (AUC) values above 0.94 for discriminating among 3 classes (health, gingivitis, and periodontitis) and 6 classes (health, gingivitis, and stages I–IV periodontitis). This approach offers clinical value for population screening programs and early detection initiatives, as it requires no clinical examination or radiographic exposure while maintaining high diagnostic accuracy.

The polymicrobial nature of periodontal disease has been successfully addressed through AI analysis of salivary bacterial profiles. Kim and colleagues utilized salivary bacterial copy numbers in an ensemble ML approach to predict periodontitis severity, achieving 83.3% accuracy. Their model demonstrated that specific bacterial consortium patterns could reliably classify disease severity, providing insights into the microbial ecology underlying periodontal pathogenesis.

Beak and colleagues further advanced this field by developing a data-driven prediction model that achieved AUC value of 0.832 and 0.796 for internal and external validations, through correlational feature analysis of 16 risk factors. Their approach identified key biomarker combinations that significantly enhanced predictive performance compared to individual markers, demonstrating the value of comprehensive molecular profiling in periodontal diagnosis.

Shifting the focus from diagnosis to prognosis, Furquim and colleagues developed models to predict the future progression of periodontitis using baseline data from a longitudinal study. They integrated clinical and demographic data with the salivary levels of 10 analytes, finding that a Probabilistic Graphic Model (PGM) incorporating clinical data, salivary interleukin (IL)-1β, age, and sex provided the best performance with an area under the receiver operating characteristic curve (AUROC) of 0.88. This study highlights the potential of combining clinical assessments with salivary immunologic data to forecast disease advancement, supporting proactive and personalized treatment strategies.

GCF represents one of the most clinically relevant biological matrices for periodontal assessment due to its direct contact with diseased tissues and site-specificity. While AI applications using GCF have been less common than those for saliva, pioneering research demonstrates their significant potential. GCF contains concentrated inflammatory mediators, tissue breakdown products, and microbial components that directly reflect local periodontal conditions.

For instance, Huang and colleagues developed a custom antibody array to simultaneously measure 20 different proteins in GCF samples from healthy individuals and patients with severe periodontitis. They then used these data to compare different ML classifiers to distinguish between the 2 groups. The study identified 5 proteins, including IL-1β, IL-8, matrix metalloproteinase-13 (MMP-13), osteoprotegerin, and osteoactivin, as the most important features for classification. Among the models tested, the Linear Discriminant Analysis (LDA) classifier achieved the highest classification accuracy (96.94%–97.50%).

This study serves as a strong proof-of-concept, highlighting that AI models incorporating GCF biomarkers can provide enhanced diagnostic accuracy for periodontal disease. Future development in this area will depend on overcoming challenges such as standardizing GCF collection methods and addressing its limited volume. However, the high biological relevance of GCF markers makes it a promising frontier for AI in periodontology.

Beyond oral fluids, AI models have also been developed using biomarkers from more invasive sources like gingival tissue and peripheral blood, which can offer detailed insights into local gene expression and systemic immune responses.

Xiang and colleagues utilized publicly available gene expression data from gingival tissue biopsies to build a diagnostic model for periodontitis. They first applied a Random Forest algorithm to screen for key biomarkers among 153 differentially expressed genes, identifying a final set of 13. These biomarkers were then used to construct an artificial neural network (ANN) model, which demonstrated excellent diagnostic performance with an AUC of 0.945 in the training cohort and 0.900 in an independent validation cohort. Their analysis also provided insights into the immune landscape of periodontal tissue, highlighting the involvement of plasma cells.

The host immune response represents a critical component of periodontal pathogenesis, making immunologic markers attractive targets for AI-powered diagnostic systems. The application of ANNs to immunologic parameter analysis has shown promise for detecting aggressive forms of periodontal disease. Papantonopoulos and colleagues explored the use of systemic immunologic parameters from peripheral blood to differentiate between aggressive and chronic forms of periodontitis. They trained an ANN using various host immune markers, including leukocyte counts, interleukins, and antibody titers. The model demonstrated high accuracy, achieving between 90% and 98% in classifying patients. The best performance (98.1% accuracy) was achieved using a combination of monocyte, eosinophil, and neutrophil counts, along with the CD4/CD8 ratio. This study successfully showed that AI can leverage simple and routinely collected blood parameters to accurately classify different, complex periodontal phenotypes.

The development of automated diagnostic algorithms represents one of the most mature applications of AI in EDR analysis. These systems can analyze patterns in clinical measurements, treatment codes, and documentation to identify periodontal disease presence and severity without requiring additional clinical examination.

Patel and colleagues pioneered the development of automated computer algorithms to label diagnostic phenotypes of EDR data. Their approach utilized natural language processing to analyze clinical notes, procedure codes, and structured data fields of 27,138 EDR records to identify periodontitis cases. Building on the potential of such large-scale datasets, the same group developed a data-driven model to predict a patient’s risk for periodontal disease. Using an XGBoost ML model, they analyzed 74 distinct features, including not only dental findings but also demographics, medical history, and social determinants of health. The model classified patients as healthy, mild, or severe periodontitis with a weighted average AUC of 0.72. Significantly, their data-driven approach identified novel risk factors not typically included in traditional models—such as patient anxiety, specific systemic conditions, and recreational drug use—showcasing the power of EDRs to uncover more holistic insights into disease risk. Patel and colleagues further developed their automated diagnosis by utilizing a rule-based computational algorithm, which is a form of automation but is not ML, for periodontitis case definition.

Addressing the challenge of dental care accessibility, Zhu and colleagues developed an ML tool for the home-based assessment of periodontitis that does not require clinical examinations or radiographs. Using data from over 10,000 individuals in the National Health and Nutrition Examination Survey database, they trained several models on 27 nonradiographic features, including demographics, nutrition information, medical conditions, and self-reported oral health. The models were designed to classify individuals into “severe” (stage III/IV) or “nonsevere” periodontitis. Their best-performing model achieved an AUC of 0.81 and a precision of 0.80, with age, cotinine level, hemoglobin A1c (HbA1c), number of missing teeth, and gender identified as the most important predictive features. The study demonstrates the feasibility of using large public health datasets to create accessible screening tools, potentially empowering individuals with limited access to dental services to identify their risk and seek timely care.

One of the most clinically valuable applications of EDR-based AI is the development of personalized risk assessment models that enable precision periodontal medicine. Swinckels and colleagues created a personalized periodontitis risk assessment system using nonimage EDRs combined with ML algorithms. Their model analyzed patient demographics, medical history, previous dental treatments, and clinical measurements to generate individualized risk scores with AUC values ranging from 0.50 to 0.94. This personalized approach enables clinicians to tailor prevention strategies and treatment intensities based on individual patient risk profiles, optimizing resource allocation and improving cost-effectiveness of periodontal care.

The application of ML to large EDR datasets has also enabled population-level risk modeling that provides insights into periodontal disease epidemiology and public health implications. Enevold and colleagues demonstrated that ML models could effectively predict clinically defined stage III/IV periodontitis from questionnaires and demographic data in Danish cohorts (The Copenhagen Aging and Midlife Biobank and The Danish Health Examination Survey), achieving AUC values of 0.64 to 0.70. Their study validates the use of simple, easily collected data for accurate periodontal disease prediction at the population level, with important implications for public health screening programs and epidemiologic surveillance.

Similarly focused on population screening, Iwasaki and colleagues validated a 9 item self-report questionnaire in a Japanese population of 949 adults, using full-mouth clinical examinations as the gold standard. Their multivariable logistic regression models, which combined self-reported oral health with demographic and health-related variables, achieved AUC values ranging from 0.71 to 0.87 for detecting various categories of periodontitis. Notably, they developed a simplified and user-friendly screening score for severe periodontitis based on just 4 key questions (“have gum disease,” “loose tooth,” “lost bone,” and “bleeding gums”). This parsimonious model demonstrated strong diagnostic capability with an AUC of 0.82, a sensitivity of 73.1%, and a specificity of 74.3%. This study reinforces the utility of validated, population-specific questionnaires as a time-effective and cost-effective tool for epidemiologic surveillance where clinical examinations are not feasible.