AI can help identify people most at risk for tooth decay

A dental examination can reveal a cavity. Identifying communities that need more examinations is a different challenge, especially when household circumstances shape who receives care. An international research team has tested whether artificial intelligence can help make those needs more visible using national survey data from Brazil.

The study, published in Caries Research, analyzes social, economic and health information alongside dental findings. Senior author Fernando Neves Hugo chairs epidemiology and health promotion at NYU College of Dentistry. The proposed use is public health planning, rather than an automated replacement for a dentist’s diagnosis.

The approach connects cavity prevention with the conditions surrounding daily life. Income, crowded housing and access to services can help identify unmet needs. AI cavity risk models aim to distinguish higher-risk profiles that might disappear inside a community average, giving planners a more detailed view of whom to reach.

“These are powerful tools that could be helpful for targeting resources and delivering preventive care to specific groups that will benefit the most,” Hugo said. The work complements research into oral health, while leaving substantial questions about prediction, fairness and implementation unanswered.

Researchers used the SB Brasil 2023 survey, coordinated by Brazil’s Ministry of Health. Their sample included 6,698 children aged 12, 8,041 adolescents aged 15–19, and 8,987 adults aged 35–44. The survey covered urban residents in permanent private households, an important boundary when interpreting its national scope.

Trained teams collected questionnaire responses and examined participants’ teeth. The analysis incorporated wealth, residents per bedroom, self-reported race or color, piped-water access, dental insurance and time since the last dental visit. Other predictors included perceived oral health, orthodontic appliance use, gum bleeding and dental calculus, or hardened dental plaque.

The team assessed two outcomes. One recorded whether someone had at least one untreated decayed permanent tooth. The other counted teeth that were decayed, missing because of decay or filled. This DMFT index captures accumulated disease experience and treatment, including tooth loss, rather than simply counting active cavities.

The authors include Jessye Melgarejo do Amaral Giordani, Orlando Luiz do Amaral Júnior and Maria Laura Braccini Fagundes of the Federal University of Santa Maria, and Juliana Balbinot Hilgert of the Federal University of Rio Grande do Sul. They compared several statistical and machine-learning approaches across the three age groups.

Untreated decay affected about 36.9% of children, 43.7% of adolescents and 52.7% of adults. Average DMFT scores also rose across these separate groups, from 1.67 in children to 10.70 in adults. Because participants were not followed over time, those comparisons do not describe the same people aging.

Wealth and household density stood out in analyses of predictor importance. Lower wealth and greater crowding were associated with higher disease burden. Non-White participants also experienced disparities. These associations concern social conditions and inequities; they do not establish race as a biological cause of decay or determine an individual’s outcome.

Among 12-year-olds, untreated decay prevalence was 62.5% without piped-water access, compared with 35.5% among those with access. The paper measures piped water, so this comparison cannot establish an effect of fluoridation. Unlike metabolic factors, infrastructure captures part of the wider environment surrounding dental health.

Recent dental visits and private insurance were associated with lower untreated-decay prevalence in reported comparisons. Those findings support examining barriers to preventive care. However, a cross-sectional survey cannot show that insurance, a visit or a water connection caused the differences observed between groups.

The researchers compared a fuller model with a triage version that removed gum bleeding and calculus measurements requiring an examination. The latter relied on demographic, socioeconomic and self-reported information. That distinction matters for outreach: a screening tool is more accessible if it can operate before a dental appointment.

As with other forms of risk prediction, identifying existing disease differs from forecasting future illness. This study classified dental findings recorded in the same survey. It did not demonstrate that a questionnaire could predict who would develop a new cavity over the following months or years.

Severity estimates were modest. The strongest adolescent model explained about 10% of variation in DMFT scores, with an average prediction error of roughly 2.8 teeth. Explained variation was lower in children and adults. Some more complex algorithms performed worse than predicting the average score, showing that additional complexity did not guarantee improvement.

The published classification results also contain discrepancies between the narrative and table, requiring clarification before quoting one headline accuracy rate. Table entries show that some adult models identified most cases while poorly distinguishing people without disease. This illustrates the need for careful evaluation beyond a model’s overall accuracy.

The researchers used five-fold cross-validation, repeatedly testing models on portions of the survey data separated from their training data. They also examined calibration and decision curves. These assess agreement between estimated and observed risks, and whether a scoring approach might offer advantages over default screening strategies under specified assumptions.

Such analyses provide an internal assessment, rather than evidence that deploying the system actually improved dental health. Testing on independent populations remains necessary. Experience with AI screening highlights why performance and fairness must be evaluated for the people and setting in which a tool will operate.

The authors emphasize models whose contributions can be understood, including simpler regression approaches. Knowing which circumstances influence a score could help planners connect predictions to services. It would not prove that changing the highest-ranked predictor produces the largest health improvement, since predictive importance and causal influence are different questions.

Including race and socioeconomic information also demands attention to algorithmic bias. The intended purpose is to reveal underserved groups and direct support toward them. Testing who is missed or incorrectly flagged would be essential before using scores to influence access to examinations or preventive services.

AI cavity risk models might eventually help public health teams prioritize outreach or invitations for dental assessment. They could complement established dental treatments by identifying where access is weakest. The study did not show that its models reduced cavities, saved money or improved outcomes after deployment.

Several limitations shape that next step. The data exclude rural populations, lack detailed dietary and microbiological profiles, and provide only a snapshot in time. Local insurance arrangements, water systems and patterns of care may change how predictors behave elsewhere. A model developed in Brazil cannot simply be assumed to transfer unchanged.

The researchers plan a related analysis using the U.S. National Health and Nutrition Examination Survey. That would test the approach in another context, rather than establish international reliability automatically. Alongside new dental technology, the challenge remains demonstrating practical benefits under real conditions.

The study’s contribution is a framework for connecting oral disease with social circumstances at scale. Its value will depend on whether transparent, independently tested predictions lead to better access and prevention. An algorithm can help describe unmet needs, but people and health systems must still deliver the care.

This article is for informational purposes and does not provide medical advice. Consult a qualified dental professional about oral health concerns.

These resources explain oral health inequalities and standards for evaluating prediction tools. They exclude the featured paper and appear newest first.

PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods
Guidance for assessing model development, evaluation and applicability, including potential sources of bias. (BMJ, March 2025)

Oral health
An overview of oral diseases, preventable risk factors and barriers to affordable care worldwide. (World Health Organization, March 2025)

Global strategy and action plan on oral health 2023–2030
The global policy framework connects oral health services, prevention and progress toward universal health coverage. (World Health Organization, May 2024)

TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods
Reporting recommendations help readers assess how prediction models were developed and evaluated, and what evidence supports their use. (BMJ, April 2024)

Global oral health status report: towards universal health coverage for oral health by 2030
Country profiles and global evidence describe disease burden, unequal access and opportunities to strengthen oral health systems. (World Health Organization, November 2022)

Research findings are available online in the journal Caries Research.

The original story "AI can help identify people most at risk for tooth decay" is published in The Brighter Side of News.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.