Periodontitis is a chronic inflammatory disease of the tooth-supporting tissues that progresses through the destruction of the periodontal ligament and alveolar bone and, when untreated, results in attachment loss, tooth mobility, and tooth loss. It is among the most prevalent noncommunicable conditions of adulthood. A systematic review and meta-analysis of studies published between 2011 and 2020 estimated that periodontitis affects nearly two thirds of dentate adults worldwide, with severe disease concentrated in older age groups and in populations with limited access to care.(1) Estimates from the Global Burden of Disease study place oral disorders, of which periodontitis is a principal component, among the most common health conditions globally, affecting billions of people and generating substantial years lived with disability.(2) Because much of this burden is preventable when the disease is detected before irreversible tissue destruction has occurred, the accurate estimation of individual risk from information that is already available in routine care is a matter of practical consequence.

The epidemiology of periodontitis in the United States is well documented. National surveillance conducted within the National Health and Nutrition Examination Survey (NHANES) and interpreted with the surveillance case definitions of the Centers for Disease Control and Prevention and the American Academy of Periodontology (CDC/AAP) reported that almost half of adults aged 30 years and older had periodontitis in 2009 and 2010.(3) The condition does not occur in isolation: consensus documents link periodontitis bidirectionally with atherosclerotic cardiovascular disease, and it shares risk factors with diabetes and other metabolic disorders.(4) This comorbidity profile reinforces the value of risk prediction that can operate within medical as well as dental settings, where periodontal examination is not routinely performed.

The risk indicators of periodontitis are stable across populations and study designs. Older age, male sex, lower educational attainment, poverty, current smoking, and diabetes have been repeatedly associated with higher prevalence and greater severity in the NHANES adult population.( 5) The clinical case for risk prediction is reinforced by the burden of the reference examination itself. A full-mouth periodontal examination is time-consuming and is not performed during routine medical encounters, so many adults with undiagnosed periodontitis are seen only by physicians and other non-dental providers.( 3) A model that estimates periodontal risk from age, sex, socioeconomic indicators, and systemic conditions could extend case-finding into settings where periodontal probing is unavailable. This practical motivation makes the relative performance of statistical and machine learning methods a decision of consequence rather than a purely methodological exercise. These associations are conventionally quantified with multivariable logistic regression, which expresses each exposure as an additive and linear contribution to the log-odds of disease. The logistic model is transparent and familiar, but it cannot represent the nonlinearities and interactions that plausibly govern the relationship between continuous exposures such as age, body mass index, or the poverty-income ratio and periodontal risk unless those terms are specified by the analyst in advance.( 6) Specifying the correct functional form is rarely possible when the underlying biology is incompletely understood.

Tree ensemble learners relax this requirement by construction. Random forests aggregate many de-correlated decision trees and thereby approximate nonlinear decision boundaries and interactions without explicit specification.(7) Gradient boosting assembles trees sequentially, fitting each new tree to the residuals of the preceding ensemble, a strategy that tends to produce the most accurate models for tabular clinical data.( 8) The diffusion of these methods into periodontology has been rapid. Recent reviews describe a growing literature of artificial intelligence applied to periodontal diagnosis, staging, and risk prediction, spanning imaging, biological markers, and clinical records.(9-11) A systematic review and meta-analysis of patient-based studies concluded that artificial intelligence models achieve high accuracy for periodontitis diagnosis from non-invasive biological markers, while noting that the advantage over conventional statistical models is not uniform.(12) Whether machine learning actually outperforms logistic regression for periodontitis prediction from ordinary clinical variables remains unresolved. A systematic comparison of ten algorithms trained on Taiwanese and United States survey data found that gradient boosting and related ensembles reached the highest area under the receiver operating characteristic curve (AUROC), although the gain over logistic regression was modest and varied between datasets.(13) Models built on large electronic dental records and on the NHANES likewise reported competitive but not uniformly superior performance of machine learning relative to statistical models.(14,15) Longitudinal studies that predicted periodontitis progression with artificial intelligence also report ensemble methods ahead of, but not always statistically distinct from, simpler models.(16)

The resulting inconsistency leaves a specific question unanswered: given identical predictors, identical validation, and paired statistical testing, do gradient boosting and random forest predict periodontitis more accurately than logistic regression in a nationally representative US sample? The objective of this study was to compare the predictive performance of gradient boosting, random forest, and logistic regression for total periodontitis defined by the CDC/AAP case definition in adults from NHANES 2009-2014, and to rank the most influential predictors. Three hypotheses were tested. H1: gradient boosting achieves a higher AUROC than logistic regression. H2: random forest achieves a higher AUROC than logistic regression. H3: gradient boosting and random forest achieve comparable discrimination.

This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cross-sectional studies( 17) and the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement.( 6)

Gradient boosting predicted total periodontitis more accurately than logistic regression or random forest in this nationally representative sample of United States adults, whereas random forest did not improve on logistic regression. The findings resolve the three hypotheses in a consistent direction: boosting gained over the linear baseline, bagging did not, and the two tree ensembles were not equivalent. The results align with a broader pattern in which sequential boosting methods, but not bagging methods, extract a small and reliable advantage from a modest set of tabular clinical predictors.

H1 was supported. Gradient boosting exceeded logistic regression in discrimination, with an absolute difference of approximately two percentage points that was nonetheless statistically robust given the sample size. This direction agrees with prior work that identified gradient boosting or related boosting ensembles as the best-performing learners for periodontitis. In a systematic comparison of ten algorithms trained on Taiwanese and United States survey data, gradient boosting machines reached the highest AUROC, with a modest advantage over logistic regression that varied between datasets.(13) Risk prediction models for severe periodontitis in a Thai population likewise found that tree-based machine learning matched or slightly exceeded statistical models after feature engineering.(24) The magnitude of the gain here is consistent with those reports and suggests that boosting offers an incremental, not a large, improvement over a well-specified logistic model for this outcome.

H2 was rejected: random forest performed below logistic regression. The result is not anomalous. Studies that compared several learners on periodontitis outcomes have reported that random forests do not consistently outperform logistic regression, particularly when the predictor set is small and the underlying associations are largely monotonic.(25,26) A nomogram for identifying periodontitis cases in Denmark, for example, showed that a penalized linear model remained competitive with machine learning alternatives built on questionnaire data.( 26) The mechanism is instructive. A random forest averages many deep trees grown on bootstrap samples, which reduces variance but does not add the sequential bias reduction that boosting provides.(7,8) With only nine predominantly monotonic predictors, the extra flexibility of the forest offers little room for gain, and its tendency toward deeper, less regularized trees can slightly degrade discrimination on data of this kind.

H3 was also rejected: gradient boosting and random forest were not comparable, and boosting was clearly superior. The contrast between the two ensembles supports the interpretation that the advantage of gradient boosting arises from its sequential fitting strategy rather than from tree-based flexibility alone. Comparable patterns appear across the periodontitis machine learning literature. Explainable deep and gradient-boosted models outperformed simpler ensembles for periodontitis risk screening on NHANES data.(15) Models for periodontal disease built on electronic dental records favored boosting-type learners over random forests.(14) Machine learning analyses of periodontitis with metabolic syndrome and with systemic health indicators likewise reported ensembles ahead of random forests for classification.(27,28) A model that predicted coronary heart disease risk within a periodontitis population using the same NHANES cycles also relied on ensemble methods.(29)

Beyond prediction of disease presence, machine learning has been applied to periodontal outcomes with direct clinical consequence. A model that predicted the clinical response to periodontal treatment one year after baseline trained a random forest on demographic, clinical, and microbiological features and found that baseline disease severity dominated the prediction.(30) A longitudinal cohort study of periodontitis progression likewise reported that artificial intelligence models built on routine clinical variables could anticipate worsening, with gradient-boosted ensembles among the stronger learners.(16) These applications reinforce the general conclusion that, for tabular periodontal data, the choice of learner matters less than the predictive content of the variables, and that boosting consistently ranks at or near the top across tasks.

The systematic literature on the question is consistent with these findings. A systematic review and meta-analysis of patient-based studies reported that artificial intelligence models for periodontitis diagnosis from non-invasive biological markers achieve high pooled accuracy, although the included studies varied widely in design and reporting, and few offered head-to-head comparisons against logistic regression.(12) A scoping review of artificial intelligence in periodontology likewise observed that most studies report a single algorithm rather than a controlled comparison, which limits the strength of conclusions about relative performance.(10) Our design, which holds predictors and validation constant across learners, addresses exactly that gap.

Age was the most influential predictor in both ensembles, followed by the poverty-income ratio and body mass index. This ranking mirrors established epidemiology. Age is the strongest single risk indicator of periodontitis in United States adults, and socioeconomic position, indexed here by the poverty-income ratio, tracks lifetime exposure and access to care.( 5) Smoking and diabetes contributed less than their epidemiological prominence might predict, plausibly because part of their effect is captured through the age and socioeconomic gradients in this cross-sectional population. The prominence of sociodemographic over modifiable predictors is consistent with machine learning studies of periodontitis and metabolic syndrome, in which age and socioeconomic variables dominated the feature rankings.(27) Studies that incorporate biological markers or transcriptomic data have reported higher internal discrimination, but those inputs are not available in routine care, whereas the nine predictors used here are.(31) Analyses of the same NHANES periodontitis data through biological aging and osteoporosis lenses similarly confirm age and related measures as dominant.(32,33) The ranking also carries a practical message: the variables that best separate cases from non-cases are largely non-modifiable, which means that predictive models built on these inputs are better suited to risk stratification and targeted referral than to informing the specific behaviour changes that might prevent disease.

The absolute discrimination of all three models was moderate, with AUROC values near 0.70. This level is consistent with other periodontitis prediction models built from questionnaire and demographic data alone, which typically report AUROC values in the 0.60 to 0.75 range.(24,25) A value of this magnitude indicates that the models are informative for group-level risk stratification but are not sufficient for definitive individual diagnosis, which continues to require clinical examination. The value of the comparison therefore lies less in the absolute performance than in the consistent ordering, which shows that the small advantage of boosting is reproducible rather than an artefact of chance.

The choice of case definition deserves comment because it conditions every estimate in the study. The CDC/AAP definitions were designed for surveillance and are stricter in their thresholding than the 2018 staging and grading classification, so prevalence and model performance would shift if another definition were substituted.( 20-22) We retained the CDC/AAP definitions to remain comparable with national surveillance, but the sensitivity of the models to the outcome definition is a recognized boundary of the findings.

The study has several strengths. It used a large, nationally representative sample with a gold-standard clinical outcome defined by the CDC/AAP case definition.( 19) The three models shared the same predictors, the same cross-validation folds, and the same evaluation procedure, and differences in AUROC were tested with a paired DeLong test rather than compared informally.( 23) The analysis is deterministic and reproducible through a fixed seed and a self-contained script.

The fixed-hyperparameter design deserves emphasis. By refraining from tuning, the comparison measures the three learners under their standard defaults, which is the setting in which most clinical users would deploy them. It is possible that tuned random forests or tuned logistic models would close or reverse the observed differences, and the fixed design therefore answers a fairness question rather than a maximal-performance question. Studies that tune hyperparameters across a broad grid nevertheless reach the same ordering we observed, which supports the robustness of the conclusion.( 13)

The limitations should temper interpretation. The design is cross-sectional, so the models estimate prevalent rather than incident periodontitis and cannot establish temporal precedence between the predictors and disease. Complete-case analysis excluded participants with missing covariate data, which may reduce representativeness if missingness was not random. Survey weights were not applied, so the sample should not be used to estimate national prevalence; the weighting decision reflects the model-comparison aim of the study rather than an inferential target. Diabetes, hypertension, and smoking were self-reported, introducing measurement error that would bias the corresponding coefficients downward. Hyperparameters were fixed rather than tuned, which is transparent but may underestimate the achievable performance of each learner. The models were validated internally by cross-validation and were not tested on an external population, so their transportability to other settings is unknown. Calibration was not formally assessed, and the reported sensitivity and specificity depend on a threshold selected from the same data, so calibration curves and decision-curve analyses on independent data would be required before deployment. Each limitation suggests a specific next step: prospective or linked longitudinal data, multiple imputation or inverse-probability weighting for missingness, tuned pipelines, external validation in independent cohorts, and calibration assessment against observed outcomes.

For clinical and public health practice, the results support gradient boosting as the preferred learner when the goal is to estimate individual periodontitis risk from variables already collected in routine care. However, the modest absolute improvement over logistic regression also implies that a well-specified logistic model remains a defensible and more interpretable alternative when transparency and ease of implementation are the priority. The dominance of age and socioeconomic predictors is itself a finding with implications: risk stratification for periodontitis is inseparable from the social determinants of oral health, and screening tools that ignore these variables will systematically miscalibrate risk across populations.( 2) The choice between the two kinds of model also carries a transparency cost. Logistic regression reports coefficients with confidence intervals that a clinician can read directly, whereas the gradient boosting model requires a feature-importance or explanation layer to be interpreted. When the accuracy gain is on the order of two percentage points of AUROC, institutions may reasonably prefer the interpretable model and reserve the boosted model for settings in which the additional discrimination justifies the added explanation burden. The practical contribution of the study is therefore not a specific model to deploy but a clear and reproducible answer to a recurring methodological question, together with a ranking of predictors that clinicians can use to judge individual risk.

This study compared gradient boosting, random forest, and logistic regression for predicting total periodontitis in United States adults using nine sociodemographic and systemic predictors from NHANES 2009-2014. Gradient boosting achieved the highest discrimination and significantly outperformed logistic regression and random forest. Random forest did not improve on logistic regression, and the two tree ensembles were not equivalent. Age, poverty-income ratio, sex, race/ethnicity, and body mass index were the most influential predictors. The findings indicate that boosting methods offer a small but reliable advantage for periodontitis risk prediction from routine data, while logistic regression remains a reasonable and interpretable baseline. The concentration of predictive information in age and socioeconomic position also shows that accurate risk stratification requires attention to the social determinants of oral health as much as to the choice of algorithm. Future work should extend the comparison to longitudinal outcomes, tuned pipelines, and external validation to establish whether the observed ordering persists when the models are used for incident disease and in populations beyond the United States.