AI Models Map Cavity Risk Across Brazilian Population Groups
An international research team has used artificial intelligence to predict cavity risk from population-level social and behavioral data, moving beyond the usual clinic-based view of tooth decay. The work, published in Caries Research, points to a model of oral health surveillance built around who is most likely to need prevention before disease becomes visible in the chair.
Dental caries remain the most common health condition worldwide, and their distribution is rarely random. Diet, sugar exposure, water fluoridation, access to dental care, income, housing density and the timing of dental visits all shape risk long before a lesion is diagnosed. In that setting, machine learning becomes more than a statistical exercise; it can sort through overlapping social forces and identify patterns that conventional analyses may flatten into averages.
The dataset came from the Brazilian Oral Health Survey, a national epidemiological study coordinated by the Brazilian Ministry of Health. Investigators analyzed information from 23,726 people who participated in the 2023 survey, including children, adolescents and adults. Participants contributed demographic data such as age, sex and race, along with social determinants of health and clinical examinations measuring the prevalence and severity of tooth decay.
Wealth and household density emerged as the strongest predictors of cavity risk in traditional analyses. Non-white populations had higher risk across all age groups, while private dental insurance and more recent dental visits were protective. One result stood out immediately: among 12-year-olds, 62.5% of those without access to fluoridated water had cavities, nearly double the 35.5% seen in those served by fluoridated systems.
Social Determinants Carved the Strongest Risk Signal
Those findings fit the biology of caries. Tooth decay develops where bacterial fermentation, sugar exposure and enamel vulnerability intersect over time, but the upstream exposures are social as much as clinical. A crowded household can amplify shared dietary patterns and access barriers. Lower wealth can delay routine care, reduce the likelihood of preventive fluoride exposure and increase the chance that disease advances before treatment begins.
Access to fluoridated water remains one of the clearest structural protections in oral health. Fluoride strengthens enamel and shifts the balance away from demineralization, especially in children whose teeth are still developing and whose exposures can set the course for future disease. The Brazilian data make that effect visible at population scale rather than in isolated case studies.
The race-associated risk difference adds a harder public health dimension. It signals that oral disease is tracking the broader distribution of structural disadvantage, not just individual hygiene habits or appointment attendance. In practical terms, the clinic sees the lesion, but the system creates the risk.
Machine Learning Separated Clusters That Ordinary Averages Blur
Investigators then applied several machine learning algorithms to the survey data and found that the models could accurately predict cavity risk and severity by age group. The algorithms evaluated the relative contribution of the different variables and detected patterns that reflect the multidimensional nature of caries risk. In other words, the models did not simply sort patients into high and low risk; they helped identify population clusters whose combined social profile carried a sharply elevated burden of disease.
Fernando Neves Hugo, chair of the Department of Epidemiology and Health Promotion at NYU College of Dentistry and senior author of the study, framed the approach as a public health tool rather than a replacement for clinical diagnosis. He described AI as a way to address the global burden of cavities through surveillance and risk stratification, with the goal of targeting preventive care to groups most likely to benefit.
That framing fits the structure of the data. Traditional analysis can identify which factors matter, but machine learning can rank how those factors combine within age-specific profiles. A child with poor access to fluoridated water, limited dental insurance coverage and a high-density household may sit in a far different risk category from a peer of the same age in a more advantaged setting, even if both appear similar on a cursory clinic intake form.
Targeted Prevention May Reach Farther Than Universal Messaging Alone
The public health implication is straightforward: prevention resources can be concentrated where they are most likely to interrupt disease early. That could mean directing fluoride programs, school-based prevention, outreach and surveillance toward specific communities rather than distributing the same intervention intensity everywhere. The logic is not to abandon broad prevention, but to refine how limited resources are deployed.
Hugo said the goal is to use country-specific data to build tailored prevention strategies and make better use of available resources. The researchers also plan to replicate the approach with U.S. data from the National Health and Nutrition Examination Survey, which would test whether the same variables carry the same weight in a different health system. Local insurance structures, water access, income gradients and care-seeking behavior may shift the relative importance of each predictor.
The findings may travel beyond Brazil, but not untouched. Middle-income countries can share broad patterns of social inequity while still differing in fluoridation coverage, dental benefit structures and access to routine care. Any model exported to a new setting will need local calibration, because the variables that dominate one country’s risk profile may weaken in another.
A more targeted approach also raises a practical bedside question: what should clinicians do when a model flags risk before the patient has a cavity? For now, the answer is earlier prevention, closer follow-up and more deliberate attention to social history. Machine learning will not restore enamel already lost, but it may help public health teams decide where to intervene before disease becomes irreversible.
Additional authors on the study included Jessye Melgarejo do Amaral Giordani, Orlando Luiz do Amaral Júnior and Maria Laura Braccini Fagundes of the Federal University of Santa Maria, along with Juliana Balbinot Hilgert of the Federal University of Rio Grande do Sul. The central limitation is the same one faced by any model built from national survey data: its value depends on how well the local predictors match the next population in which it is used. Until those replications are complete, the strongest claim is narrower—AI can sort cavity risk with surprising precision when the social context is measured carefully enough.