Periodontitis is among the most prevalent chronic inflammatory conditions in adults, and its distribution across the United States population is sharply unequal. The CDC/AAP surveillance case definition estimated that nearly half of adults aged 30 years and older had some form of periodontitis in the 2009-2012 survey, with prevalence climbing to 35% in Hispanic and 34% in non-Hispanic Black adults while remaining near 20% in non-Hispanic White adults.(1,2) The severe form of the disease ranks sixth among the most prevalent conditions worldwide, affecting an estimated 743 million people.( 3)

Against this backdrop, machine learning has been advanced as a tool for periodontal risk prediction, and a rapidly growing literature now reports models that predict periodontitis and its progression from clinical, demographic, radiographic, and molecular inputs with area-under-the-curve values that frequently exceed conventional regression-based risk scores.(4-9) Some of these models approach, and in a few tasks match, the accuracy of expert clinicians,(10) and systematic reviews describe the field as maturing toward clinical translation.(11)

The inputs to these models are heterogeneous, ranging from self-reported questionnaires and laboratory values through intraoral and panoramic radiographs to salivary and molecular biomarkers, and the reported performance varies with the richness of the input, so that multimodal and imaging-based models tend to outperform those built on survey variables alone.(8,9) Yet the published literature concentrates on the single-number summaries of these models, and the question of whether a given model performs equally well for the groups that bear the greatest burden of disease is almost never addressed.

The pattern this study anticipates is that a model trained without the sensitive attribute will nonetheless distribute its errors unevenly across racial and ethnic groups. The reason is that the variables available to a periodontitis model, such as education, household income, smoking, diabetes, and body mass index, are themselves structured by race and ethnicity, so a model that never sees race can recover it indirectly through these proxies. Algorithmic-fairness research has demonstrated this mechanism repeatedly in health care: algorithms that omit race still reproduce disparities by learning them from correlated variables, and the resulting differentials in sensitivity and false-positive rate are the operational meaning of unequal treatment.( 12-15)

The formal framework of equal opportunity and predictive equality makes the requirement precise: a fair model should attain similar sensitivity and similar false-positive rates across groups,( 16) while calibration, the agreement between predicted and observed risk, captures a separate axis along which models can fail equitably.( 17-19)

A model may be well calibrated within every group and yet assign systematically higher predicted probabilities to one group than another, which is the distinction between calibration and demographic parity, and a model may achieve equal sensitivity while retaining unequal false-positive rates, which is the distinction between equal opportunity and predictive equality. These conceptual distinctions matter because a fairness evaluation that reports only one metric can declare a model fair while it fails on the others, and the appropriate metric depends on the clinical use of the model and on which error is judged more costly in that use.

Whether race-blind models for periodontitis satisfy these fairness conditions has not been quantified in a population-based sample. Methodological reviews of periodontal prediction models flag the near-total absence of subgroup performance reporting as a recurring limitation, and they note that most studies report a single aggregate metric without examining how that metric varies by race or ethnicity.(20,21)

The consequence is that a clinician or health system cannot know, from the published literature, whether a periodontitis model will perform differently for the groups that already carry the highest burden of disease, and the choice of a decision threshold, which every deployed model requires, could silently distribute false negatives and false positives unevenly.

This gap is especially relevant because periodontitis is a preventable and treatable condition whose early detection depends on screening, and a screening tool that under-detects in any one group would reproduce, in algorithmic form, the very disparities that periodontal research has spent decades documenting. The objective of this study was to evaluate the fairness of a race-blind machine learning model for periodontitis by quantifying its discrimination and calibration within each racial and ethnic group, measuring disparities in sensitivity and false-positive rate at a common decision threshold, and assessing demographic parity in the predicted positive rate.

Discrimination differed only modestly across groups. The overall area under the receiver operating characteristic curve was 0.66, and the pattern across groups ranged from a low of 0.60 among non-Hispanic Black adults to a high of 0.68 in the Other group, with non-Hispanic White adults at 0.64 and Hispanic adults at 0.68 (Figure 1). The four curves overlapped substantially through most of the operating range, and the divergence was concentrated in the high-sensitivity region, where a fixed false-positive rate corresponded to different true-positive rates across the groups.

The spread of roughly 8 percentage points in the area under the curve is small relative to the spread in sensitivity and specificity reported next, a pattern indicating that discrimination and decision-level fairness are distinct properties of the model. In the middle of the operating range the four curves were nearly coincident, and only at the extreme ends, where few decisions are made in practice, did the groups separate, so the ranking of individuals by risk was similar across groups even though the threshold-based decisions that follow from that ranking were not.

The sensitivity and specificity pattern was far more pronounced than the discrimination pattern. At the common Youden threshold of 0.26, sensitivity climbed to 0.8% in Hispanic and 0.6% in non-Hispanic Black adults, dropped to 0.5% in the Other group, and reached its lowest value of 0.6% in non-Hispanic White adults (Figure 2). Specificity moved in the opposite direction: it dropped to 0.5% and 0.6% in Hispanic and non-Hispanic Black adults, while climbing to 0.6% in non-Hispanic White and 0.7% in the Other group.

The equal-opportunity gap, the difference between the highest and lowest sensitivity, was 0.3%, and the predictive-equality gap in false-positive rate was 0.3% (Table 1). In absolute terms, the model detected true periodontitis in nearly eight of every ten Hispanic cases but in fewer than six of every ten non-Hispanic White cases, while assigning positive labels to fewer than half of the disease-free Hispanic adults. The non-Hispanic Black group occupied an intermediate position on both axes, with sensitivity of 0.6% and specificity of 0.6%, and the Other group showed the inverse profile, with the second-highest specificity and the second-lowest sensitivity. The pattern therefore ordered the groups along a single gradient in which higher sensitivity was paired with lower specificity and vice versa, rather than producing a group that was simultaneously worse on both measures.

Calibration stayed within a narrow band across groups, in contrast to the decision-level disparities. The expected calibration error spread from 0.1% to 0.1%, a calibration gap of 0.1%, with the model slightly under-predicting risk in non-Hispanic Black adults and over-predicting risk in the Other group (Figure 3). The calibration curves tracked the diagonal closely for every group, which indicates that the predicted probabilities retained their meaning as absolute risks within each group even where sensitivity and specificity diverged sharply. The worst-calibrated group, non-Hispanic Black adults, showed an expected calibration error of 0.1%, still within a range that is widely considered clinically acceptable for risk stratification purposes, and the overall expected calibration error was 0.0%. The calibration curves for non-Hispanic White, Hispanic, and Other adults lay almost on the diagonal, while the curve for non-Hispanic Black adults fell slightly below it at intermediate predicted probabilities, which corresponds to the model assigning risks that were somewhat too low for that group in the middle of the risk range.

The predicted probability distribution confirmed a demographic-parity imbalance that tracked the underlying prevalence gradient. The median predicted risk clustered near the population prevalence in non-Hispanic White adults and shifted upward in Hispanic and non-Hispanic Black adults, where the model assigned higher probabilities to a larger share of participants (Figure 4).

The proportion of participants classified positive differed across groups by 0.3%, a spread that mirrors the group differences in periodontitis prevalence present in the source data, where prevalence reached 0.4% in Hispanic and 0.3% in non-Hispanic Black adults against 0.2% in non-Hispanic White adults (Table 2).

The overlap in the box plots was substantial, but the medians were ordered consistently with the prevalence gradient, which confirms that the demographic-parity gap is largely a consequence of different base rates rather than an independent distortion. The interquartile ranges also differed, being wider in the Hispanic and non-Hispanic Black groups, which received a broader spread of predicted probabilities, and narrower in the non-Hispanic White group, where the model concentrated its predictions around the lower prevalence of that group.

The principal finding is that a race-blind model for periodontitis produced a 25-percentage-point gap in sensitivity and a 28-point gap in false-positive rate across racial and ethnic groups while its calibration remained comparatively uniform. The pattern is the one anticipated from algorithmic-fairness research: excluding race from the predictors did not prevent disparities, because the model recovered group information from correlated variables such as education, income, smoking, and diabetes.(12,13) The disparities are large enough to matter clinically, since a 25-point difference in sensitivity means the model detects nearly three times as many true cases in the group with the highest sensitivity as in the group with the lowest, which would concentrate false-negative diagnoses in the lower-prevalence group and thereby delay care for the patients most likely to be missed.

The sensitivity and specificity pattern is the most consequential of the four figures. Sensitivity climbed to its highest value in Hispanic adults and dropped to its lowest in non-Hispanic White adults, while specificity moved in the opposite direction, the signature of a model whose predicted probabilities are shifted upward for higher-prevalence groups. The pattern coincides with the finding that blinding a model to sex still leaves discrimination and calibration uneven across groups,(17) and with the comparison of fairness methods in intensive-care risk prediction.(18) The mechanism here is prevalence rather than explicit encoding of the sensitive attribute: a common threshold applied to groups with different base rates mechanically trades sensitivity against specificity, the central point of the formalization of equal opportunity.(16) The imbalance also mirrors the observation that performance fairness degrades even in early-risk neural networks trained without demographic inputs.(22)

The calibration pattern was more reassuring than the decision-level pattern. The expected calibration error stayed below 12.2% in every group, and the calibration gap of 5.9 points was small relative to the sensitivity and specificity gaps. This dissociation means the model's predicted probabilities retained their meaning as absolute risks within each group, and the disparity was concentrated in the decision threshold rather than in the probability estimates themselves. The finding is consistent with the broader literature showing that calibration and discrimination can dissociate, and that fairness failures need not manifest as miscalibration.(19,23) For periodontitis specifically, the systematic reviews that catalog model performance have rarely reported calibration at all, so the near-diagonal calibration observed here cannot be assumed to hold for other models in the field.( 11,20)

The demographic-parity imbalance in the predicted probability distribution reflects the source-data prevalence gradient rather than an independent failure of the model. The distribution shifted upward in Hispanic and non-Hispanic Black adults, whose periodontitis prevalence in the sample was 35% and 34%, respectively, and downward in non-Hispanic White adults, whose prevalence was 20%. This is the classic ambiguity in demographic parity as a fairness criterion: a well-calibrated model will necessarily predict more positives in groups with higher underlying risk.(16) The policy-relevant question is therefore not whether predicted rates differ, but whether the sensitivity and specificity differences that follow are acceptable, and here the observed gaps argue for group-specific decision thresholds or a fairness-constrained training objective of the kind evaluated in the fairness-performance trade-off literature.(24,25)

What this study adds to the periodontitis literature is an explicit, population-based measurement of fairness where previous work has reported only aggregate accuracy. The methodological review of prognostic models for periodontitis found that subgroup analyses and calibration were systematically under-reported,(20) and the implementation review identified the same omission as a barrier to clinical translation.(21) By reporting sensitivity, specificity, calibration, and demographic parity within each racial and ethnic group, the present study supplies the subgroup evidence that those reviews called for, and it demonstrates that the omission is not benign, because the aggregate area under the curve conceals decision-level disparities of clinical magnitude. The result also connects the dental literature to the wider fairness literature in a concrete way: the equal-opportunity and predictive-equality gaps reported here are of the same order as those that have prompted intervention in other clinical prediction settings.( 12,14)

The model's overall discrimination, an area under the curve of 0.656, is comparable to, though slightly below, the values reported in recent periodontitis machine-learning studies, which typically span from the mid-0.60s to the high 0.80s depending on the predictors and imaging inputs available.(4,5,8,9) The lower discrimination here is expected for three reasons: the model was restricted to questionnaire and examination variables available in a population survey, without radiographic or salivary biomarkers; the race-blind specification deliberately discards a predictor empirically associated with periodontitis; and the binary outcome collapses a heterogeneous disease into a single class. more input-rich models can match expert clinicians in longitudinal periodontitis assessment,( 10) which suggests that the fairness problem documented here could recur, and potentially intensify, as model performance improves, since higher discrimination does not by itself equalize sensitivity across groups.

The clinical stakes of the sensitivity and specificity trade-off are concrete in periodontitis because the disease is chronic and slowly progressive, and a missed early case is unlikely to produce immediate harm but may nonetheless delay preventive care in exactly the populations that already carry the highest prevalence and the greatest burden of disease. A model deployed for population screening that misses nearly half of true cases in non-Hispanic White adults while flagging more than half of disease-free Hispanic adults as positive would generate both undertreatment and over-referral, and the two harms would be allocated asymmetrically across groups. This asymmetry is the operational content of unfairness in a screening tool, and it cannot be read off from the area under the curve, which differed by only a few points across groups. The dissociation between near-equal discrimination and unequal sensitivity is precisely why fairness evaluation must report decision-level metrics rather than ranking metrics alone, a point made in the surgical and intensive-care fairness literature(18,26) and reinforced by the governance reviews of clinical artificial intelligence.( 27)

The study has limitations that follow from its design, and each suggests a remedy. First, the outcome used the CDC/AAP case definition applied to survey measurements, which may not align with clinical diagnoses, and the model's absolute sensitivity and specificity are conditioned on that definition; future work could validate the fairness findings against adjudicated periodontal diagnoses. Second, the analysis used a single model class, logistic regression, and a single decision threshold, so the magnitude of the disparities may differ for tree-based or neural models; a comparison of model families would establish whether the pattern generalizes, and work on sex-specific clinical histories suggests that feature engineering can shift subgroup performance in both directions.(28) Third, the race and ethnicity categories collapse the Asian and multiracial population into an Other group whose sample size is the smallest of the four, limiting the precision of its subgroup estimates. Fourth, the model was evaluated internally on a single held-out split rather than by external validation, so the fairness metrics may be optimistic for new populations. Finally, the study measured statistical fairness criteria rather than the clinical or social consequences of the disparities, which require a decision-theoretic or utility-based evaluation of the kind advocated when moving beyond accuracy as the sole target.(29) Each of these limitations is addressable, and together they define a research program in which fairness is evaluated across model families, validated externally, and ultimately linked to the downstream periodontal outcomes that matter to patients.

The practical implication for developers and reviewers of periodontal machine-learning models is that a race-blind specification is not a fairness guarantee. The disparities documented here arise through proxy variables, so the remedies must operate on the decision rule or the training objective rather than on the exclusion of race, and the surgical-risk literature reaches the same conclusion for algorithmic bias in risk prediction.(26) Group-specific thresholds calibrated to a chosen cost ratio, post-processing that equalizes sensitivity across groups, and the routine reporting of subgroup sensitivity, specificity, and calibration all follow from these results and are low-cost to implement.(14,30,31) The consistency of these recommendations with the broader fairness literature(12,13) indicates that periodontal models face the same equity challenge as algorithms in other clinical domains, and the review of bias and oversight in clinical decision support makes clear that regulatory and governance expectations are moving in the same direction.(27) For a field in which the burden of disease is already concentrated in Hispanic and non-Hispanic Black adults, the failure to measure and report these disparities would mean that the very tool introduced to improve periodontal care could, if deployed at a single fixed threshold, allocate its errors in a way that reinforces rather than reduces existing inequities.

Looking forward, the most direct path to fairer periodontal prediction is to move fairness from a post hoc audit to a design constraint. The fairness methods evaluated in adjacent clinical domains, including pre-processing that reweights training observations, in-processing that adds a fairness penalty to the objective, and post-processing that selects group-specific thresholds, are all applicable to the logistic regression used here.(18,24) Because the model remained well calibrated within groups, a group-specific threshold calibrated to equalize sensitivity across groups would be a minimal intervention that preserves the probabilistic output while eliminating the largest disparity observed. Future work should evaluate whether such interventions reduce the gaps documented here without degrading calibration, and whether the same pattern holds for the imaging-based and biomarker-based periodontitis models that now dominate the literature,(8,9) since those richer models may either amplify or attenuate the disparities through their additional inputs.

A race-blind machine learning model for periodontitis discriminated the outcome with an overall area under the curve of 0.656 and remained adequately calibrated in every racial and ethnic group examined, with expected calibration error below 12.2% throughout.

Sensitivity differed across groups by 25.1 percentage points and false-positive rate by 27.9 points at a common decision threshold, with the largest sensitivity in Hispanic adults and the smallest in non-Hispanic White adults, so that false negatives and false positives were allocated unevenly across the four groups.

The predicted positive rate differed across groups by 29.5 percentage points, a gap that tracked underlying prevalence differences and persisted despite the exclusion of race from the predictors, which shows that proxy variables alone reproduced the disparities documented here.