Related Experiment Videos
Machine Learning for Predicting Dental Caries Presence and Severity: An Age-Cohort Perspective from a Nationally
Introduction:
Although preventable, dental caries remains a global public health challenge due to its multifactorial nature and the deep influence of socio-structural determinants. This study evaluates the performance of Artificial Intelligence (AI) models as decision-support frameworks for population-level risk stratification, screening, and public health planning.
Methods:
We analyzed a representative sample (N=23,726) from the SB Brasil 2023 national survey, stratified into three age cohorts: children (12y), adolescents (15-19y), and adults (35-44y). Eleven predictors, encompassing socioeconomic (wealth score via PCA, household density, race, water access), psychosocial (self-rated oral health status), behavioral, and clinical domains, were used to predict caries presence (classification: D>=1) and severity (DMFT index regression). To account for overdispersion and the high frequency of zero counts, we utilized advanced count-based regressions, including Zero-Inflated Negative Binomial (ZINB) and Negative Binomial (NB) models. Performance was assessed using AUC, Mean Absolute Error (MAE), and Decision Curve Analysis (DCA) via 5-fold cross-validation.
Results:
Caries prevalence increased progressively with age (36.8% to 52.7%). The non-clinical triage models demonstrated significant Net Benefit in DCA across all groups, effectively identifying high-risk individuals for prioritization without requiring clinical examination. For severity prediction, the ZINB model identified structural signals of inequity, particularly in late adolescence (15-19y), where socio-structural determinants explained a meaningful proportion of clinical variance (R2=0.1010; MAE=2.79). Across all stages, SHAP analysis reinforced that wealth score and household density are the primary structural predictors of caries risk.
Conclusion:
AI provides a feasible framework for population-level risk stratification and public health planning. The findings suggest that AI-driven frameworks should be utilized for managerial triage and resource allocation, enabling proportional universalism by identifying high-risk clusters through structural determinants rather than individual-level clinical diagnostics.