Abstract
Patients who use the emergency department (ED) for non-traumatic dental-related conditions are likely to return for multiple visits, accounting for a disproportionate share of healthcare cost. Reliance on the ED for receiving dental care continues to rise, particularly among Medicaid beneficiaries. We sought to predict future "ED users" for non-traumatic dental-related conditions using machine learning algorithms and identify associated risk factors to help Medicaid outreach programs better target populations in most need of improvement in access to dental care. Among 6.4 million Medicaid beneficiaries aged 19-64 years, the state-specific ensemble prediction models demonstrated high performance across four states, with C-statistics ranging from 0.70 to 0.73, accuracy from 0.90 to 0.95, specificity from 0.93 to 0.97, and sensitivity from 0.17 to 0.23, consistently outperformingtraditional logistic regression models. Individual-level variables, such as age, race/ethnicity, and musculoskeletal complications, and area-level variables, such as racial/ethnic composition, indices of food access, and proportion of US-citizenship, explained a significant proportion of the variation in dental-related ED use. Using a machine learning-based risk model to monitor for risk factors of dental-related ED use can assist in developing more targeted care management interventions, addressing unmet dental care and social needs while reducing unnecessary dental-related ED visits.