Artificial Intelligence in Caries Risk Assessment: Evaluating the Current Status of CAMBRA and Cariogram with Large Language Models

BACKGROUND: Caries risk assessment (CRA) is essential for individualized caries management in early childhood, but existing tools are difficult to standardize in routine practice. The ability of large language model (LLM)-based artificial intelligence (AI) systems to recognize and correctly apply validated CRA tools remains unclear. This study aimed to evaluate the ability of ChatGPT and Google Gemini, LLM-based AI systems, to apply and interpret CRA tools in early childhood under different prompting conditions. METHODS: Thirty standardized clinical vignettes of children aged 0-5 years were evaluated using Caries Management by Risk Assessment (CAMBRA) and the Cariogram by ChatGPT-5.1 (OpenAI, San Francisco, CA, USA) and Google Gemini-3.0 Pro (Google LLC, Mountain View, CA, USA) under three prompting conditions: unguided, guideline-informed, and guideline-only. An expert panel provided reference classifications. AI outputs were assessed for categorical agreement, mean absolute error (MAE), quality, accuracy, and readability. MAE differences were analyzed using the Friedman test and mixed-design repeated-measures ANOVA, while quality, accuracy, and readability were analyzed using repeated-measures analysis of variance (ANOVA). Reliability was assessed using intraclass correlation coefficients (ICC). RESULTS: Inter- and intra-rater reliability were high (ICC = 0.89-0.93). Reference CAMBRA and Cariogram classifications showed a moderate correlation (ρ = 0.509; n = 30). MAE differed significantly across AI model-prompting condition combinations for both tools (p < 0.001). No significant difference was observed between AI models for CAMBRA, whereas ChatGPT showed lower MAE than Gemini for the Cariogram (p = 0.002). Guideline-informed and guideline-only prompting significantly reduced MAE and improved quality and accuracy compared with the unguided condition (p < 0.001). CONCLUSIONS: LLM-based AI systems can support CRA, particularly when guideline-based prompting is used. However, performance depends on the CRA tool and AI model, and numerical or algorithmic components, such as those in the Cariogram, remain challenging.