Background: Healthcare costs, especially those associated with emergency department (ED) visits, are increasing at an unsustainable rate. Often, ED visits for certain conditions can be prevented through patients utilizing their primary care physician. We consider two of these conditions: diabetes for adults (n=342,286 patient quarters), and ear, nose, and throat (ENT) conditions for children (n=2,660,733 patient quarters). Being able to identify patients at risk of an avoidable admission to the ED could lead to dramatically reduced costs for both patients and healthcare systems. We develop models to predict avoidable admissions (defined as visits to the ED for either of these ambulatory care sensitive conditions) and reduce healthcare costs.Methods: Patients with the chosen conditions (adult diabetes and juvenile ENT) were drawn from a major hospital system. The training set includes 10 total quarters, spanning from the third quarter of 2016 to the last quarter of 2018. The test set, where all models were compared, includes the first quarter of 2019.Logistic regression has commonly been used to identify high-risk patients, but more recently other statistical and machine learning techniques have been employed. We use a variety of models, including the lasso, a mixed model, random forest, and XGBoost to determine which model best predicts avoidable ED visits. All available predictors were included in the full model and compared. We also include novel predictors, such as how far a patient lives from the ED and a patient's family's tendencies to visit the ED. The predictors are compared using multiple methods (including LASSO, P values, and boosting). Results:We find the XGBoost model generally outperforms the other models in the validation sample (C-index of 0.80 in both the diabetes and ENT cohorts). Among the best predictors of future ED visits are past ED visits; a patient's age and weight; and, for patients with diabetes, the amount of time since their initial diabetes diagnosis. Conclusions:If implemented, this model can identify 50 patients who would have gone to the ED unnecessarily by only contacting 600 patients. Or, by contacting 5,500 patients, identify (and potentially prevent) 170 unnecessary ED visits.