Frequent emergency department (ED) use has been operationalized in research, clinical practice, and policy as number of visits to the ED, despite the fact that this definition lacks empirical evidence and theoretical foundation. To date, there are no studies that have attempted to understand ED use empirically, without arbitrary use of "cut-points." This study was conducted to identify the best-performing, empirically grounded definition of frequent ED use. The performance of machine learning supervised clustering algorithms based on the most common definitions of frequent ED use in peer-reviewed literature (i.e., 3+, 4+, 5+ visits per year) were compared to unsupervised clustering algorithms that take into account numerous systemic factors associated with patients' ED use. All ED visits for the State of Florida, 2011-2015, including more than 100 clinical and payment-related variables per visit were employed in the model. Supervised algorithms using number of visits to the ED, alone, were unable to differentiate patients into clusters, while unsupervised models using all patient data formed clusters in which patients within a given cluster were alike, and patients between clusters were different. Cluster size and characteristics were stable across years. The results of this study indicate that mean number of ED visits by patients differ between patient clusters, but this does not allow for accurate identification of ED patients. Machine learning algorithms using all systemic and biopsychosocial patient data can be used to identify and group patients for the purpose of developing and testing integrated, whole health interventions. (PsycInfo Database Record (c) 2021 APA, all rights reserved).