Large language models for early childhood caries risk triage: A validation study using standardized pediatric dental vignettes.

PURPOSE: To compare the diagnostic accuracy and public health utility of Gemini 2.5 Flash and GPT-4 for early childhood caries (ECC) risk triage in classifying presentations as high risk (warranting immediate referral) or low risk (amenable to preventive counseling). MATERIALS AND METHODS: A standardized dataset of 50 clinical vignettes (25 high risk and 25 low risk) was evaluated against both large language models via official application programming interfaces using a blinded, prospective comparative design (CONSORT-AI; TRIPOD+AI). This constitutes a vignette-based internal validation study and does not represent a clinical diagnostic accuracy study. Ground truth was established by two consultant pediatric dentists (intraclass correlation coefficient = 0.92). Performance was quantified using sensitivity, specificity, positive predictive value, negative predictive value, and false negative rate (FNR). Readability, actionability, and multilingual competence in Hindi and Tamil were additionally assessed. RESULTS: Both models exceeded the 80% accuracy threshold. GPT-4 achieved significantly higher sensitivity (92.0% vs. 80.0%; P = 0.032) and a lower FNR (8.0% vs. 20.0%), with superior readability (Flesch-Kincaid Grade Level 5.9 vs. 8.1), actionability (4.8 vs. 4.2), and cross-lingual consistency. Both models achieved 100% safety disclaimer compliance in English. CONCLUSIONS: GPT-4 demonstrates a statistically superior safety profile for ECC risk triage. As a Level 1 adjunct, it can assist pediatric dentists by prescreening caregiver-reported presentations, reducing diagnosis lag, and directing specialist capacity toward confirmed high-risk pathology without replacing clinical judgment. CLINICAL RELEVANCE: This study establishes that GPT-4 is a safer and more accessible AI triage tool than Gemini 2.5 Flash for pediatric dental screening, particularly in linguistically diverse, resource-constrained settings where specialist access is limited.