On October 19, 2017, the research team queried PubMed and Google Search (Google LLC, Mountain View, CA) to identify literature containing studies or programs that had applied super-utilizer criteria to health care populations. In order to capture all available definitions that could be applied to Medicaid, the team included criteria applied to non-Medicaid populations. PubMed searches included “super-utilizers,” “super utilizers,” “‘high need’ AND ‘high cost,”’ “frequent fliers,” and “hot spotters.” These terms were taken from the team’s experience and a preliminary review of the literature. Google searches included “Medicaid super utilizers”; the team only retained papers and articles in the first 40 results as this was a supplementary search meant to capture top government reports and professional papers from the gray literature. The team found additional documents from published systematic reviews identified in the PubMed search and from links, footnotes, or citations in the Google search results.

The research team included only official reports (defined as formal documents from academia, government, non-profits, or the private sector) that applied ≥1 super-utilizer criteria (as defined by CMS¹) to original individual-level population data. Two reviewers independently reviewed document abstracts from PubMed results and title of citations from the Google search to exclude work that did not meet this inclusion criterion. Disagreements between reviewers were resolved by consensus. All remaining documents were reviewed in detail (Figure 1).

Criteria included were categorized by the CMS super-utilizer targeting approach¹ and logical subcategories (determined by the reviewers) within each approach. Specific criteria within these subcategories were tabulated for evaluation of whether they were commonly used. Combinations of criteria in the same document were separated into component parts for categorization (eg, 2 emergency department [ED] visits or 2 inpatient admissions would be tabulated under both categories). Criteria from documents containing multiple studies or multiple sets of criteria were likewise tabulated separately. Because this analysis was intended for retrospective data, the team excluded referral-based criteria, which would not be available in the data set and would be difficult to simulate, because of inconsistent application in practice.⁸ Criteria that included rates over time were prorated to a 12-month period. For documents with multiple gradations of utilization, the more inclusive criterion was used (eg, some documents distinguished between high utilizers and super-utilizers). Criteria that used the term behavioral health without further explanation were categorized under both mental health and substance use disorders. Individual criteria that were applied in at least 5 documents and accounted for at least 10% of all criteria that fell in a category subset were considered commonly used. Additionally, documents were reviewed for any specific methods used to determine which criteria would be applied, and criteria developed using clearly defined data-driven methods were noted.

The research team obtained Utah Medicaid claims data from the Division of Medicaid and Health Financing for all beneficiaries with claims in Utah during State Fiscal Year (SFY) 2017 (July 1, 2016–June 30, 2017). Cluster analysis was used to group these beneficiaries by super-utilizer criteria identified from the literature review. Two of 21 criteria that were commonly identified in the literature were excluded because Utah Medicaid lacked data on homelessness and social needs. Additionally, 2 criteria identified from the literature review were excluded because of overlap with other criteria. Jenks Natural Breaks analysis was used to identify top cost percentage break points based on methodology described in one of the literature review documents,6 which brought the total number of applied criteria for this analysis to 20. For Jenks Natural Breaks, the team used the classInt package⁹ in R 3.4.2¹⁰ (R Foundation for Statistical Computing, Vienna, Austria) with N = 3 clusters and considered the top cluster (3.1%) to be the break point of interest.

Dummy variables for each criterion were assigned based on claims data during the 12-month period before each beneficiary’s last claim of SFY 2017. This included the amounts paid by Medicaid, International Classification of Diseases, Tenth Revision (ICD-10) codes,¹¹ and revenue codes. Out-of-pocket patient costs were not included in the analysis because of the study’s focus on institutional costs. Chronic health conditions were assigned using Elixhauser Comorbidity Index¹² categories based on the beneficiary’s ICD-10 codes. Top cost percentage variables were based on the population of beneficiaries with claims. Excluded from the study were enrollees in the Children’s Health Insurance Program, refugees, juvenile justice cases, and custody medical cases because of shared stewardship with other agencies. No other subpopulations or groupings were excluded in order to achieve a largely agnostic top-level view of potentially impactable institutional costs.

The research team used the CLARA (clustering large applications) implementation of k-medoids from the cluster package¹³ in R 3.4.2 to conduct the clustering analysis. The “Manhattan” distance¹⁴ was used to calculate dissimilarities between super-utilizer criteria dummy variables. Eligible Utah Medicaid beneficiaries who filed no claims during SFY 2017 were excluded and treated as a separate cluster with no associated costs. The number of clusters was selected using the average silhouette method¹⁵ for up to 30 clusters. The minimum number of clusters was confirmed visually by applying the elbow method, an approach commonly used in cluster analysis that selects the optimum number of clusters such that adding an additional cluster does not significantly reduce the objective function (ie, where the curve bends).¹⁶ The presence or absence of each criterion among Medicaid beneficiaries (ie, the percentage of beneficiaries in a cluster who met the specified criteria) was used to characterize each cluster.

This study consisted of a methodological review of research methods used in the literature and application of review findings to claims data from the Utah Medicaid population, and entailed secondary analysis of data collected routinely by Utah Medicaid. The Institutional Review Boards at the Utah Department of Health (Approval #533) and the Centers for Disease Control and Prevention (HSR #2018–00189) determined that this study did not constitute human subjects research.

Searches of PubMed and Google Search resulted in 845 documents for review (Figure 1). Of these, 136 were from PubMed searches (N = 96) and Google searches (N = 40). Of the PubMed search results, 32 were from the “frequent fliers” search, 31 from “‘high need’ AND ‘high cost,”’ 28 from “super-utilizers,” and 5 from “hot spotters.” Of 709 documents identified from other sources, 641 records were identified from links, footnotes, or citations from the Google search results and 68 were from systematic reviews found among the PubMed results. Of these documents, 613 were removed after deduplication, 443 were excluded, and fulltext review was conducted for the 170 remaining documents. After full review, a further 59 documents were excluded, resulting in 111 documents included in the synthesis (Supplementary Data).

The studies included produced 180 super-utilizer criteria: 112 from the academic literature, 51 (39 state and 12 federal) from government reports, and 17 from nonprofit organizations. Fifty-three criteria (29%) were specifically applied to Medicaid populations. Of the 180 super-utilizer criteria, 89% identified super-utilizers based on very high levels of utilization, most commonly ED visits or number of inpatient admission (Table 1). Half of criteria identified super-utilizers based on risk factors associated with high preventable costs, mostly chronic health conditions or behavioral health conditions (Table 1).

Altogether, 21 definitions met the minimum criteria to be considered commonly used (Table 2). Only the following categories met the inclusion criteria: targeting very high levels of utilization; targeting based on risk factors associated with high, preventable costs; and targeting based on referrals and follow-up investigation. Five subcategories were apparent within the utilization-based targeting category (number of ED visits, number of inpatient admissions, top predicted risk score, number of prescribed medications, and top cost percentage). Top predicted risk score was excluded because of substantial variation and lacking detail in the approaches used to determine the risk score among sites. Number of prescribed medications also was excluded because a time frame for determining the number of medications was available in only 3 documents. Four subcategories were apparent within the risk factor-based targeting category (chronic health conditions, number of chronic health conditions, behavioral health conditions, and social determinants of health).

Seventy-seven documents (69%) stated their super-utilizer criteria, allowing them to be categorized, but did not describe how their criteria were selected. Twenty-three documents (21%) based their criteria on previous criteria found in the literature. Four documents (4%) referred only generally to deriving their criteria from previous work the authors had completed. Only 7 documents (6%) included detailed descriptions of data-driven approaches for developing their criteria. Of those documents with detailed descriptions, 3 used the mean number of ED visits plus 2 standard deviations as a criterion,^7,17–18 1 used the mean number of inpatient admissions plus 2 standard deviations,¹⁹ 2 used other combinations of the mean and standard deviation of the number of visits,^20–21 and 1 used natural break points to identify the top cost percentages.⁶ Certain documents acknowledged that the criteria used to identify super-utilizers were arbitrary.^22–24

Cluster analysis identified an optimum number of 5 distinct clusters of super-utilizer criteria among the 309,921 Utah Medicaid beneficiaries with claims during SFY 2017 (Figure 2). The average silhouette width at 5 clusters was exceeded slightly at 29 clusters, but the elbow method confirmed 5 clusters as the optimum number to use (Figure 3). These clusters included the following: Cluster 1: Beneficiaries who were not super-utilizers (<10% met any super-utilizer criteria; n = 163,118); Cluster 2: Beneficiaries with multiple chronic or mental health conditions (breakdown of chronic health conditions and mental health disorders shown in Table 3; n = 68,054); Cluster 3: Beneficiaries with a single chronic health condition (breakdown of chronic health conditions and mental health disorders shown in Table 3; n = 43,939); Cluster 4: Beneficiaries with chronic or mental health conditions who are ED super-utilizers (breakdown of chronic health conditions, mental health disorders, ED use, and costs shown in Table 3; n = 7809); and Cluster 5: Beneficiaries with ≥1 hospitalizations and no chronic or mental health conditions (breakdown of admission, chronic health conditions, and mental health disorders shown in Table 3; n = 27,001).

Only Clusters 2 and 4 had ≥4 criteria that applied to at least 10% of beneficiaries in the cluster. Cluster 1 did not exceed 10% for any criterion. Clusters 3 and 5 only exceeded 25% on 1 criterion each. Although 100% of beneficiaries in Cluster 3 had ≥1 chronic conditions, 0% had ≥2 chronic conditions, indicating that this cluster was completely composed of beneficiaries with a single chronic condition. Similarly, 100% of beneficiaries in Cluster 5 had ≥1 inpatient admissions, but only 8% had ≥2 inpatient admissions. Closer inspection (data not shown) revealed that many inpatient admissions in this cluster were for infant deliveries and other routine hospitalizations. Meanwhile, Clusters 2 and 4 had more than 10% of the cluster populations meet a majority of the super-utilizer criteria (11 and 18 criteria, respectively). These clusters included beneficiaries with complex needs and ED super-utilizers, who are more traditional super-utilizer populations.¹