Population-Level Oral Disease Surveillance Using Large Language Models on Clinical and Public Health Data

Population-level oral disease surveillance is critical for guiding public health interventions, yet traditional systems relying solely on structured data often fail to capture contextual, behavioral, and access-to-care determinants embedded in unstructured clinical narratives. In this study, we developed a hybrid large language model (LLM) framework that integrates structured epidemiological features with embeddings derived from examination notes and survey text to improve the detection and monitoring of dental caries, periodontal disease, and tooth loss. Using the publicly available NHANES Oral Health Dataset, we compared the performance of traditional machine learning models, text-only LLM models, and our proposed hybrid approach. The hybrid model consistently outperformed all baselines, achieving higher accuracy, precision, recall, F1-score, and calibration, while maintaining equitable performance across demographic and socioeconomic subgroups. Explainability analyses revealed that combining structured and unstructured features captured clinically meaningful patterns, including behavioral risk factors and care access barriers. Our findings suggest that hybrid LLM-based surveillance can enhance real-time population-level monitoring, identify high-risk communities, and inform preventive strategies within the U.S. public healthcare system, offering a scalable, interpretable, and equitable approach to oral health monitoring.