Beyond Accuracy: Machine Learning Models for Predicting the Presence of Permanent Molar Caries in U.S. Children and Adolescents With Fairness Consideration

Abstract

Introduction

Although predictors of dental caries have been previously explored, a comprehensive understanding of factors influencing permanent‐molar decay in U.S. children and adolescents, especially with respect to racial and ethnic biases, remains limited. This study aims to develop and evaluate machine learning models incorporating algorithmic fairness to predict caries in permanent molars.

Methods

Data from the National Health and Nutrition Examination Survey were analyzed using the 2011–2014 cycles for training and validation and the 2015–2016 cycle for testing. The primary outcome was decayed, missing, and filled teeth in at least 1 permanent molar, dichotomized to represent the presence or absence of caries. Candidate predictors included demographic characteristics, dietary factors, laboratory and examination measures, behaviors, and lifestyle indicators. Three machine learning models (logistic regression, random forest, and XGBoost) were implemented and compared using the area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity. Fairness was quantified by comparing group-specific true-positive rates and false-positive rates across race/ethnicity over 10 cross-validation folds, using Shapiro–Wilk tests and paired tests to assess subgroup gaps and pre/postmitigation changes. To mitigate disparities, an in-processing equalized-odds constraint was applied through an exponentiated-gradient reduction framework. Sensitivity analyses using National Health and Nutrition Examination Survey weights were also conducted.

Results

Among 5,710 participants aged 6–19 years, 37.30% had caries in at least 1 permanent molar. Logistic regression achieved the highest premitigation area under the receiver operating characteristic curve (78.50%). SHAP (SHapley Additive exPlanations) analyses identified age, family income, BMI, and household education as key contributors. Fairness evaluation indicated minimal equal-opportunity gaps but showed evidence of equalized-odds violations across race/ethnicity. After mitigation, race/ethnicity-related disparities decreased across models, with varying performance trade-offs. XGBoost and random forest maintained stable area under the receiver operating characteristic curve values in both unweighted (∼77%–78%) and survey-weighted (∼75%–77%) evaluations. In contrast, logistic regression showed a substantial decline in unweighted area under the receiver operating characteristic curve (78.50%–65.53%) and a marked reduction in recall and F1 score, whereas changes under weighted evaluation were more modest (77.91%–76.54%).

Conclusions

Machine learning models effectively predicted dental caries in children and adolescents using National Health and Nutrition Examination Survey data. Applying equalized-odds constraints reduces disparities across models, although with varying fairness-performance trade-offs. These findings highlight the importance of tailoring bias mitigation strategies to clinical priorities, balancing equitable outcomes and predictive utility, and they can be used to inform policy and better design-targeted interventions to improve oral health.

Highlights

• •Machine learning models accurately predicted permanent-molar caries in U.S. youth.
• •Age, SES, diet, and dental-care behaviors were top predictors.
• •Fairness testing revealed race/ethnicity differences in false positive rates.
• •Tree-based machine learning models balanced fairness and accuracy best.
• •Findings support equitable, data-driven approaches for pediatric oral-health policy.

Graphical abstract

INTRODUCTION

Dental decay (caries) is a significant public health problem, impacting more than half of adolescents aged 12–19 years. In the U.S., approximately 17% of children aged 6–11 years and 57% of adolescents aged 12–19 years have experienced caries, with 5.2% and 16.6% having untreated decay in those age groups, respectively.¹ Untreated caries often lead to pain and can negatively impact school attendance and learning.^2,3 Reducing caries experience for children and adolescents has been a national health objective for decades, as reflected in the recently published Healthy People 2030. Caries is preventable, most notably reduced by receipt of preventive oral health services, such as dental sealants and fluoride. It is well known that there are disparities in caries by race and ethnicity, family income level, and insurance coverage.^4,5 Prior research has also found a relationship between caries and access to care, including living in a dental healthcare shortage area and other social determinants of health.⁶

Artificial intelligence technology is showing a promising role in transforming dental caries management. With high performance metrics across detection, diagnosis, and prediction tasks, artificial intelligence offers significant potential to improve diagnostic accuracy, identify important predictors, and improve dental patient care.⁷ Recent systematic reviews have assessed the performance of machine learning (ML) algorithms, both in diagnosing and in predicting the prognosis of dental caries; convolutional neural networks have shown promise in image-based diagnosis,⁸ whereas support-vector machines and generalized regression neural networks have performed well in prognostic prediction.⁹

Building on these findings, subsequent studies have applied both classical ML and deep learning methods to caries prediction across diverse populations and outcomes.10, 11, 12, 13, 14, 15, 16, 17, 18 For instance, multivariate models and deep artificial neural networks have been used to classify caries status in deciduous teeth across a broad population (ages 0–80 years), relying on demographic and dietary factors from the nationally representative National Health and Nutrition Examination Survey (NHANES) 2013–2014 data.^10,11 Similarly, an automated ML classifier has been employed using data from an epidemiologic study of North Carolina to address early childhood caries in preschoolers aged 36–71 months, incorporating age and parent-reported oral health status.¹⁵ A recent study used the Iowa Fluoride Study data set to build a dental caries prediction model for children and young adults.¹⁶ Other studies include using historical cohort data from Japan with patients from the age group 20–64 years using mostly clinical predictors,¹⁴ a sample of elementary school children (aged 5–8 years) from Japan,¹⁷ and elderly people (aged 65–74 years) with participants from China using demographic and lifestyle predictors.¹⁸ However, the limited sample sizes and specific geographic characteristics of these studies require further validation, especially considering potential race and ethnicity bias, through additional research with a broader national-level population.

The HHS requires the detection and mitigation of sources of bias and discriminatory outputs in ML algorithms used for clinical decision making and healthcare service delivery¹⁹ in light of growing evidence showing that ML systems can unintentionally encode bias, leading to decisions that put underrepresented or marginalized groups to disadvantage.20, 21, 22 As a result, ensuring fairness in healthcare-focused ML is critical to preventing exacerbating existing disparities. Although some recent studies have evaluated bias and provided mitigation techniques in Alzheimer disease,²³ foregoing preventive dental care²⁴ and anxiety disorder prediction,²⁵ a significant research gap remains in applying fairness-focused ML models for predicting dental caries. Specifically, to the best of the authors’ knowledge, no prior model exists for predicting dental caries in permanent molar teeth that considers algorithmic fairness among children and adolescents aged 6–19 years, a demographic particularly vulnerable to oral health issues. This study addresses this gap by developing ML models for permanent molar caries prediction that explicitly assess and mitigate race/ethnicity bias. The objective of this research is to ascertain which input variables among clinical, behavioral, and sociodemographic predictors have the most significant impact on dental caries in permanent molars using a variety of ML tools. By adjusting ML models on the basis of fairness metrics, this study provides evidence to inform policy and better design-targeted interventions to improve oral health.

METHODS

Study Population

Publicly available NHANES data were used for the analysis. NHANES uses stratified multistage probability samples of the civilian, non-institutionalized U.S. population. It contains information on sociodemographic data, dietary data, questionnaires, and laboratory data, in addition to detailed dental examinations.

NHANES 2011–2014 data were used to develop the predictive models, and the 2015–2016 cycle was reserved as an independent temporal holdout test set. Combining the 2011–2014 cycles increased the sample size and improved estimate stability, whereas the later cycle provided a more stringent evaluation of generalizability than a random split by assessing performance in new, unseen data that could differ modestly in sample composition, access to care, and population behaviors. The analytic sample was restricted to children and adolescents aged 6–19 years. Variables with >10% missing data were excluded, and candidate predictors were selected on the basis of prior studies. After preprocessing, the numbers of observations retained for training, validation, and testing are shown in Figure 1.

Figure 1.

Model flowchart.

There is no widely accepted, survey-authorized guideline for how to integrate NHANES sampling weights into a modern ML framework beyond descriptive or inferential analyses. NHANES-based ML papers typically do not apply sampling weights when fitting ML models¹²; instead, they use weights mainly for descriptive summaries (e.g., survey-weighted proportions) without explicitly weighting the ML training process.¹³ A recent methodologic work further suggested that the impact of incorporating survey weights into ML algorithms on predictive performance is context dependent and does not necessarily improve out-of-sample accuracy.²⁶ Accordingly, the primary analysis was conducted without survey weights, and NHANES sampling weights were incorporated into a sensitivity analysis of the test data set.

This study was exempt from IRB review because only publicly available data were used in the analysis. The analytical codes used to perform the analyses presented in this paper can be made available upon reasonable request to the corresponding author.

Measures

Sixteen predictor variables and 1 target variable were selected to evaluate the determinants of dental health outcomes. The binary outcome variable was the presence of dental caries in at least 1 of the 8 permanent molars, defined as at least 1 decayed, missing, or filled permanent molar (>0). In this study, the assessment focused specifically on the 8 permanent molars, providing a targeted evaluation of caries prevalence in these key teeth because they are the most vulnerable to caries in children and adolescents aged 6–19 years.²⁷

Predictor selection was based on previous NHANES-based ML studies and established oral-health disparity frameworks. Prior work on dental caries and tooth loss prediction has demonstrated the importance of demographic and socioeconomic variables (e.g., age, race/ethnicity, education, income), dental-care utilization measures (e.g., dental insurance status and time since last dental visit), and behavioral or lifestyle proxies (e.g., smoking exposure, sedentary behaviors, and dietary intake) in improving predictive performance.^12,13 Additional insights were drawn from oral-health disparity frameworks that emphasize socioeconomic context, access to care, and health behaviors as key drivers of population-level oral health outcomes.²⁸ In addition, a univariate screening step was conducted to corroborate predictor relevance using chi-square tests for categorical variables and correlation analysis for continuous variables, which served as a filter-based feature selection check before finalizing the 16 predictors for model training.

Therefore, all predictors were organized into demographic/SES context, dietary intake (carbohydrate and sugar intake rates), examination-based clinical measures (BMI), and questionnaire-derived behaviors/access variables (insurance status, dental visit status, sedentary screen-time measures, household smoking, and food expenditures), consistent with prior evidence in building predictive models in oral health. The predictor variables in this study are summarized in Table 1.

Table 1.

List of Predictors and Target Variable, NHANES 2011–2014 Training Data Set (Weighted Frequencies, N=3,866)

Variable/category	n	Weighted % (or mean±SD)
Age (years), mean (SD)	3,866	12.39±3.90
BMI, mean (SD)	3,866	21.77±6.19
Income-to-poverty ratio, mean (SD)	3,866	2.45±1.63
Total carbohydrate and sugar intake (g/day), mean (SD)	3,866	394.02±165.75
Food expenditure, at home ($/month), mean (SD)	3,866	183.65±212.66
Food expenditure, away from home ($/month), mean (SD)	3,866	33.50±77.11
Sex
Male	1,973	51.4%
Female	1,893	48.6%
Race/ethnicity
Mexican American	819	15.2%
Other Hispanic	407	7.1%
Non-Hispanic White	1,015	56.0%
Non-Hispanic Black	1,038	13.5%
Non-Hispanic Asian	361	3.9%
Other/multiracial	226	4.3%
Household education level
Less than 9th grade	409	8.3%
9th–11th grade	567	11.5%
High-school graduate/GED	860	21.0%
Some college or AA degree	1,145	30.6%
College graduate or above	885	28.6%
Household marital status
Married	2,360	68.6%
Widowed	100	2.1%
Divorced	363	10.1%
Separated	221	3.9%
Never married	503	8.6%
Living with partner	319	6.8%
Household size
1–2 persons	173	4.6%
3 persons	579	15.6%
4 persons	1,069	31.5%
5 persons	976	25.5%
6 persons	569	12.8%
≥7 persons	500	9.9%
Dental visit status
6 months to 1 year	725	16.6%
Within last 6 months	2,426	66.3%
>1 year ago	715	17.1%
Health insurance status
Uninsured	413	9.5%
Private insurance	1,667	54.7%
Public insurance	1,786	35.7%
Household smoker
No	730	17.8%
Yes	3,136	82.2%
Hours watching TV per day
<1 hour	803	22.9%
1 hour	1,172	30.1%
2–3 hours	648	16.6%
4–5 hours	314	7.3%
≥6 hours	382	8.8%
Do not watch TV	85	2.3%
Hours using computer per day
<1 hour	906	24.5%
1 hour	659	15.9%
2–3 hours	318	9.0%
4–5 hours	142	3.2%
≥6 hours	263	6.4%
Do not use computer	582	13.4%
Dental caries status (DMFT)
No caries (DMFT=0)	2,423	62.67%
Any caries (DMFT>0)	1,443	37.32%

Note: Weighted percentages are calculated using NHANES examination sample weights to produce nationally representative estimates. Continuous variables are presented as weighted mean±weighted SD. Categorical variables are presented as unweighted sample size (n) and weighted population percentage (%). Reference categories are indicated in parentheses.

AA, Associate of Arts; DMFT, Decayed, Missing, and Filled Teeth; NHANES, National Health and Nutrition Examination Survey.

Statistical Analysis

Three widely used approaches were evaluated for dental caries prediction: logistic regression, random forest, and XGBoost. All categorical variables were 1-hot encoded, and continuous variables were standardized to reduce scale-related distortion. The Synthetic Minority Over-sampling Technique was explored initially, but because it did not materially improve accuracy metrics, those results are not reported.

To optimize hyperparameters while minimizing the risk of overfitting, 10-fold cross-validation with Bayesian optimization (hyperopt) was applied to the training data. After the selection of the hyperparameter combination with the highest area under the receiver operating characteristic curve (AUC) for each model, final algorithms were fitted to the training set and evaluated on the test set. An optimal decision threshold was selected by maximizing the F1 score on the validation set. Using that threshold, AUC, accuracy, precision, recall, and F1 score were calculated for each model on the test set. All reported performance metrics are based on the test set (NHANES 2015–2016).

Race/ethnicity was selected a priori as the protected attribute because persistent and statistically significant caries disparities by race/ethnicity have been documented in U.S. children, even after adjusting for socioeconomic and clinical factors.⁵ The true positive rate (TPR) is the proportion of actual positives correctly predicted as positive. The false positive rate (FPR) is the proportion of actual negatives incorrectly predicted as positive. To quantify algorithmic fairness, group-specific TPR (equal opportunity), FPR (equalized odds), and overall prediction rates (demographic parity) were compared across self-reported race and ethnicity categories, which were designated a priori as the protected attribute, using data from 10 cross-validation folds on the test data set. Equal opportunity, equalized odds, and demographic parity are standard and widely used algorithmic fairness metrics. Shapiro–Wilk tests were conducted to assess the normality of the fold-wise differences. Next, paired 2-tailed t-tests (or Wilcoxon signed-rank tests when normality was violated) were conducted to determine whether the fairness metrics differed significantly between each racial or ethnic subgroup and all other subgroups combined. After that, an in-processing bias mitigation technique (equalized-odds constraint through an exponentiated-gradient reduction framework) was applied to reduce fairness gaps across the subgroups. Demographic parity was excluded from the fairness adjustment because it could be a misleading criterion, masking real disparity when dental caries prevalence differs across racial groups. Paired tests were applied to evaluate differences in fairness metrics between base models and bias-mitigated (both weighted and unweighted) models using a statistical significance level of p<0.05. Table 2^29,30 presents the definitions of the fairness metrics used in this study.

Table 2.

Mathematical Definitions of 3 Common Fairness Metrics

Fairness metric	Definitiona	Explanation with sample use in dental caries prediction
Equal opportunity	True positive rates are the same across groups: $P\left(\widehat{Y}=1|A=0,Y=1\right)=P\left(\widehat{Y}=1|A=1,Y=1\right)$	The probability of correctly predicting dental caries in permanent molars is the same across groups defined by race or ethnicity.29
Equalized odds	True positive rates and false positive rates are the same across groups: $P\left(\widehat{Y}=1|A=0,Y=y\right)=P\left(\widehat{Y}=1|A=1,Y=y\right),y\in \left\{0,1\right\}$	The probabilities of correctly and incorrectly predicting dental caries in permanent molars are the same for groups defined by a protective attribute such as race or ethnicity.29
Demographic parity	Equal probability of being classified with the positive label: $P\left(\widehat{Y}=1|A=0\right)=P\left(\widehat{Y}=1|A=1\right)$	The proportion of individuals predicted to have dental caries in permanent molars is the same across groups defined by race or ethnicity.30

^aY indicates the observed outcome, Ŷ is a prediction of Y, and A represents a binary protected attribute.

RESULTS

The study included a total of 5,710 children and adolescents aged 6–19 years. Among them, 2,130 individuals (37.30%) had dental caries in at least 1 permanent molar. The target variable of this model, dental caries, exhibited a class distribution of 62.70% majority (no caries) vs 37.30% minority (with caries), indicating a slight imbalance. In data sets with such imbalances, ML algorithms may exhibit a predisposition to favor the majority class. Given the modest degree of imbalance in the data set, regular class weights were incorporated into model fitting rather than relying on additional resampling techniques.

The performance of the ML algorithms without sampling weights for predicting dental caries both before and after fairness adjustment is summarized in Table 3. Before adjustment, all models showed comparable performance, with AUC values consistently ranging between 77.10% and 78.45% and accuracy between 67.95% and 69.41%, which indicates that model choice has a limited impact on baseline performance. Logistic regression is favored for its higher sensitivity/recall in detecting caries, whereas the tree-based models provide a more balanced precision-recall trade-off. After fairness adjustment, the AUC and accuracy were maintained for the tree-based models, whereas logistic regression showed a substantial reduction in recall, indicating that the trade-off between performance and fairness varies by algorithm.

Table 3.

Performance Metrics of ML Algorithms Before and After Fairness Adjustment (%) Without Sampling Weights

Model	AUC (%)		Accuracy (%)		Precision (%)		Recall (%)		F1 score (%)
	Pre	Post	Pre	Post	Pre	Post	Pre	Post	Pre	Post
XGBoost	77.94	77.85	69.41	66.14	55.61	52.12	77.73	78.62	64.84	62.63
Logistic regression	78.45	65.53	67.95	65.36	53.90	56.67	80.57	18.73	64.59	28.12
Random forest	77.10	76.95	69.31	70.52	55.91	57.82	72.80	69.12	63.25	62.97

AUC, area under the receiver operating characteristic curve; ML, machine learning.

Feature importance, quantified with SHAP (SHapley Additive exPlanations), showed age as the highest-influence predictor in all 3 models (XGBoost, random forest, logistic regression). Among non-age variables, household education level, income–poverty ratio, and insurance coverage category consistently ranked high. Lifestyle-related variables, particularly total carbohydrate and sugar intake, money spent eating out, and frequency of dental visits, also demonstrated substantial contributions. BMI appeared among the top predictors specifically in the random forest model. Figure 2 shows the SHAP values for the ML models.

Figure 2.

SHAP feature importance plots of logistic regression, random forest, and XGBoost.

(A) Logistic regression. (B) Random forest. (C) XGBoost.

SHAP, SHapley Additive exPlanations.

Potential bias was evaluated in predictive performance using 3 ML models (XGBoost, logistic regression, and random forest), considering race/ethnicity as the sensitive factor. Initial bias evaluation revealed no statistically significant disparities in equal opportunity (TPR) across racial groups for any of the models. However, there was a significant difference in equalized odds (FPR) when comparing all other groups with non-Hispanic White, the largest demographic group in the data set. Thus, in the fairness gap analysis, non-Hispanic White was used as the reference group.

The application of the equalized-odds–based mitigation reduced intergroup TPR and FPR disparities across all 3 models (Figure 3). The maximum FPR difference (max ΔFPR) declined from 0.142 to 0.121 for XGBoost, from 0.117 to 0.064 for logistic regression, and from 0.143 to 0.110 for random forest, whereas reductions in the maximum TPR difference (max ΔTPR) were more modest (XGBoost: 0.187–0.175; logistic regression: 0.192–0.165; random forest: 0.178–0.156). Relative to the non-Hispanic White reference group, FPR gaps that were statistically significant before mitigation for XGBoost (p=0.008) and random forest (p=0.006) became nonsignificant after mitigation (XGBoost p=0.604; random forest p=0.804). TPR gaps were not statistically different from zero before or after mitigation. Complete TPR and FPR estimates with 95% CIs are provided in Appendix Tables 1–9 (available online).

Figure 3.

Pre–post fairness gaps across algorithms.

(A) TPR, mean |Δ| versus NH-White. (B) FPR, mean |Δ| versus NH-White. (C) TPR, worst-case Δ. (D) FPR, worst-case Δ.

Note: Error bars show bootstrap 95% CIs. Worst-case gap was defined as the maximum pairwise absolute difference in group-wise TPR or FPR across racial/ethnic groups within each resample; the resample mean with percentile bootstrap 95% CIs is reported.

FPR, false positive rate; NH, non-Hispanic; TPR, true positive rate.

In sensitivity analyses using NHANES sampling weights, all models demonstrated stable results after bias mitigation, with only modest reductions in AUC (≤1.4 percentage points) and small declines in accuracy and F1 score (Table 4). XGBoost and random forest showed minimal performance changes, whereas logistic regression exhibited a slightly larger decrease in recall and F1 score, consistent with the greater sensitivity of linear models to threshold adjustments under weighted evaluation. Overall, the weighted results and metrics were consistent with those of the primary analyses, supporting the robustness of the findings when accounting for the complex survey design.

Table 4.

Performance Metrics of ML Algorithms Before and After Weighted Fairness Adjustment (%)

Model	AUC (%)		Accuracy (%)		Precision (%)		Recall (%)		F1 score (%)
	Pre	Post	Pre	Post	Pre	Post	Pre	Post	Pre	Post
XGBoost	76.92	76.45	70.99	70.12	58.93	58.03	58.60	55.35	58.77	56.65
Logistic regression	77.91	76.54	72.22	68.12	61.67	55.59	56.17	47.87	58.79	51.44
Random forest	75.49	75.04	69.29	68.64	56.59	55.40	56.94	55.57	56.16	56.07

AUC, area under the receiver operating characteristic curve; ML, machine learning.

Being consistent with the unweighted analysis, non-Hispanic White was used as the reference group. Residual error-rate disparities varied by algorithm but consistently showed larger gaps for TPR than for FPR (Figure 4). Collectively, these findings indicate that under survey-weighted evaluation, equal opportunity mitigation yielded tighter alignment in FPR than in TPR across race/ethnicity subgroups, consistent with the unweighted analysis.

Figure 4.

Survey-weighted pre–post fairness gaps across algorithms.

(A) TPR, mean |Δ| versus NH-White. (B) FPR, mean |Δ| versus NH-White. (C) TPR, worst-case Δ. (D) FPR, worst-case Δ.

Note: Error bars show 95% CIs. Worst-case gap was defined as the maximum pairwise absolute difference in group-wise TPR (or FPR) across racial/ethnic groups within each resample; the resample mean with percentile bootstrap 95% CIs is reported.

FPR, false positive rate; NH, non-Hispanic; TPR, true positive rate.

DISCUSSION

Using NHANES data, this study developed and evaluated 3 ML algorithms—XGBoost, logistic regression, and random forest—for the prediction of permanent molar caries in U.S. children and adolescents, while explicitly considering fairness evaluation and mitigation. All models showed good discrimination, with AUC values >77% in the primary unweighted analysis and >75% in the secondary weighted analysis.

The predictor set combined socioeconomic, dietary, clinical, and questionnaire-derived variables, allowing the models to capture multiple dimensions of caries risk. All comprehensive base models and secondary models showed better predictive performance than previous works (AUC<75%), which used smaller data sets from historical cohort studies for predicting general dental caries.^14,17,18

An observational analysis of common features across the 3 models reveals age, household education level, income-related measures, and dietary habits as critical factors influencing dental caries risk. The consistency in these top predictors provides actionable insights from a public health perspective, underscoring the importance of comprehensive approaches that address both socioeconomic conditions and individual lifestyle behaviors to improve oral health outcomes. All findings of this study are consistent with those of previous studies, where age was considered the most dominant factor for different dental disease conditions, including caries and tooth loss.^12,13,16 As children age, more permanent molars erupt and remain exposed to the oral environment for longer periods, thus increasing the cumulative risk for caries. In this context, age likely captures both dentition stage and duration of exposure to caries-related factors. This finding is especially relevant given the wide age range included in the analysis (6–19 years), during which permanent molars erupt. The prominence of age highlights the importance of accounting for developmental stage in caries risk prediction. Future work could further isolate the age effect by developing age-stratified prediction models.

Dietary habits, such as sugar and carbohydrate intake rates, were influential, which supports a recent study that analyzed the relationship between untreated decay and high sugar intake rates in U.S. adults using NHANES data.³¹ Socioeconomic factors (FPL and household education level), behavior and lifestyle factors, and last dental visit status are also significant in predicting caries among children and adolescents. This agrees with conventional knowledge that both SES and healthcare behavior impact oral health outcomes.^32,33 Children from lower-income households or with parents with limited education levels are more likely to face barriers to accessing dental care, either owing to financial constraints or health literacy. Similarly, irregular dental visits prevent early caries detection and can lead to subsequent untreated dental issues.³⁴

BMI emerged as one of the dominant factors for the random forest model. Although BMI is not a direct cause of dental caries, it suggests that body weight may serve as a potential early indicator for underlying dietary and lifestyle patterns that affect oral health. As noted in some previous research, elevated BMI among children and adolescents is often associated with behaviors such as increased frequency of snacking, higher consumption of carbohydrates, and lower participation in health-promoting activities, all of which contribute to an increased risk of dental caries.^35,36 However, other studies indicate that the evidence of an association between BMI and dental caries is inconsistent.³⁷ Further investigation is necessary to understand the nature of this relationship. Similarly, the primary analysis identified additional behavior and lifestyle activities (i.e., hours watching TV, hours using the computer, and food expenditure) as influential variables in predicting oral health outcomes. These factors may serve as proxies for unhealthy routines and dietary choices that contribute to the development and progression of dental caries in the younger population.

Fairness analyses highlighted that model choice can meaningfully affect subgroup-level equity in dental caries prediction. These findings support previous studies on bias evaluation for healthcare data sets^23,25 and highlight the importance of systematically evaluating and addressing algorithmic bias to ensure equitable healthcare decision making. Demographic parity was assessed descriptively but was not used as the mitigation target because forcing equal positive prediction rates can be misleading when the underlying disease prevalence differs across groups.

For the logistic regression model, applying the exponentiated-gradient in-processing algorithm with an equalized odds constraint forced the model to equalize sensitivity (TPR) and false-alarm rates (FPR) across racial/ethnic groups.³⁸ Because a single-boundary logistic model cannot express group-specific thresholds, the optimizer met the constraint by shifting the boundary toward the majority (nondisease) class, sharply reducing the proportion of positive predictions. This adjustment cut recall and, consequently, the F1 score, exemplifying the well-documented accuracy–fairness trade-off that arises when strict parity is enforced on low-capacity classifiers.³⁹ In summary, if fairness is paramount, logistic regression is the standout model; if balanced accuracy and moderate fairness improvement are more desirable, XGBoost or random forest may be better choices.

The outcome variable in the test set of this study was modestly imbalanced (62.70% no caries vs 37.30% caries); standard classifiers may preferentially predict the majority no caries class, which can yield apparently strong overall accuracy while reducing sensitivity (missed caries) and thereby limiting clinical usefulness for screening and triage. Class weights and F1-based threshold selection were therefore used to support a more clinically meaningful balance between false negatives and false positives (rather than relying solely on default thresholds). In real-world public health implementation, prioritizing higher sensitivity can reduce missed opportunities for preventive care and referral, whereas prioritizing higher specificity can reduce unnecessary follow-up burden when capacity is constrained; accordingly, the operating point should be chosen for the intended workflow, and fairness adjustment should be interpreted in light of its impact on the sensitivity–specificity trade-off.

Fairness adjustment is important when predictive models are intended for real-world clinical use and policy implications, where they can influence dental screening decisions, referrals, and preventive care. Models that do not account for fairness outcomes may perform well on average but miss certain demographic groups, which further impacts oral health disparities. By incorporating fairness adjustments, models are able to deliver more consistent performance across every subpopulation and better identify individuals who may benefit from preventive oral health interventions. Although these adjustments may lead to trade-offs in some performance metrics, they are essential to ensure responsible policies in oral health decision making.

Finally, paired within-group analyses (Appendix Tables 1–9, available online) showed significant pre-to-post improvements in fairness-related error rates, most notably a reduction in non-Hispanic White FPR across all 3 models (p<0.05 by paired t-test). These findings support calls from the health-equity literature to embed fairness methods directly into clinical risk tools, ensuring subgroup-level equity without unduly sacrificing model utility.^23,25,40 ML by itself cannot overcome structural, behavioral, or access-related determinants of oral health.^6,41 These predictive models should therefore be viewed as one component of a broader equity strategy. Truly, equitable dental care will require interventions such as reconfiguring school and community-based preventive programs^42,43 and strengthening provider training on implicit bias,⁴⁴ which are beyond the scope of this study.

Limitations

This study has a few limitations. First, these models rely solely on NHANES data. Even though it is a comprehensive data source, the analysis and results are limited to the variables and outcomes captured by this survey data. Therefore, some confounding factors, such as longitudinally obtained dietary patterns, fluoride exposure, and genetic factors, may influence dental caries but cannot be included owing to the absence of these variables in the NHANES data sets. Second, the temporal scope of the 2011–2016 cycle may not capture more recent dental health trends or changes in care delivery. Relatedly, temporal trends (e.g., secular shifts in caries risk factors, access to care, or survey measurements across cycles) were not modeled explicitly; therefore, model performance and calibrated risk estimates may not fully generalize to future cohorts if the underlying data-generating process changes over time. Future work can enhance robustness by integrating multisource longitudinal data, such as electronic health records and community-level social/environmental indicators, to better capture time-varying risk and strengthen external validity.

NHANES 2011–2016 data were used because concerns have been raised about the reduced detection sensitivity of the NHANES oral examination in the later cycles (including 2017–2018) on the basis of inter-reference examiner reliability analyses⁴⁵ and because the 2019–March 2020 cycle is not nationally representative as a standalone cycle owing to the coronavirus disease 2019 (COVID-19) suspension of field operations.⁴⁶ Future research should validate these ML models using updated NHANES data sets.

Bias in ML can also arise from skewed sampling and from cultural assumptions embedded in the data.⁴⁷ In health applications, some inequities stem from underlying biological or systemic factors that may be difficult or impossible to remove.⁴⁸

CONCLUSIONS

This study developed and validated the first set of ML models to predict dental caries in permanent molars among children and adolescents aged 6–19 years using publicly available NHANES data. All 3 ML algorithms demonstrated strong predictive performance, with AUC values exceeding 75%. Besides age, socioeconomic factors, including household education level, family income, poverty ratio, and insurance status, and lifestyle-related variables, such as carbohydrate/sugar intake and screen time, also significantly influenced model predictions.

In fairness testing, no significant gaps were observed in true-positive rates across racial/ethnic groups, but non-Hispanic Whites showed higher false-positive and overall prediction rates (p<0.05). Targeted bias-mitigation strategies (exponentiated gradient in processing) substantially reduced these disparities while preserving good accuracy.

These findings provide actionable insights for policymakers and clinicians seeking to allocate preventive resources more equitably and underscore that best models in health care must be judged by both accuracy and fairness.