Predicting Dental General Anesthesia Use among Children with Behavioral Health Conditions.

Abstract

To evaluate how different data sources affect the performance of machine learning algorithms that predict dental general anesthesia use among children with behavioral health conditions. Observational study using claims data. Using Medicaid claims from Partners For Kids (2013-2019), electronic medical record data, and the Ohio Child Opportunity Index, we conducted a retrospective cohort study of 12,410 children with behavioral health diagnoses. Four lasso-regularized logistic regression models were developed to predict dental general anesthesia use, each incorporating different data sources. Lift scores, or the ratio of positive predictive value to base case prevalence, were used to compare models, and a lift score of 2.5 was considered minimally acceptable for risk prediction. Dental general anesthesia use ranged from 3.2% to 3.9% across models, which made it difficult for the machine learning models to achieve high positive predictive value. Model performance was best when either the electronic medical record (lift = 2.59) or Ohio Child Opportunity Index (lift = 2.56), but not both (lift = 2.34) or neither (lift = 1.87), was used. Incorporating additional data sources improved machine learning model performance, and 2 models achieved satisfactory performance. The model using electronic medical record data could be applied in hospital-based settings, and the model using the Ohio Child Opportunity Index could be more valuable in community-based settings. Machine learning was applied to satisfactorily predict which children with behavioral health diagnoses would require dental treatment under general anesthesia. Incorporating electronic medical record data or area-level social determinants of health data, but not both, improved the performance of the machine learning predictions. The 2 highest performing models could be applied by hospitals using medical record data or by organizations using area-level social determinants of health data to risk stratify the pediatric behavioral health population.