Understanding regional variations in the veterans health administration’s clinical resource hub program using hierarchical agglomerative clustering

Abstract Background Large healthcare systems such as the Veterans Health Administration (VHA) have potential to optimize resource allocation and improve efficiency through regionalization. VHA’s Clinical Resource Hubs (CRH) is a regionally based program with one “hub” based in each of the 18 VHA administrative regions; hubs fill temporary staffing gaps across each region’s “spoke” ambulatory care sites. Understanding clusters of pre-existing regional characteristics in relationship to observed program outcomes could inform program design and evaluation. Existing clustering approaches, however, are difficult to interpret and apply. We aimed to test, as a proof-of-concept, whether unsupervised hierarchical agglomerative clustering (HAC) can (a) reveal interpretable regional typologies and (b) show face validity for learning about variations in CRH program outcomes. Methods We first developed a set of characteristics with hypothesized links to regional CRH variations. Based on these, we created regional variables based on Veteran characteristics by aggregating their relevant data. We also included relevant pre-existing regionally defined VHA data variables. We used principal component analysis (PCA) for dimensionality reduction and for developing interpretive feature loadings. We applied HAC to identify regional clusters among the retained variables and iteratively tested cluster face validity. We compared access pre-CRH and post-CRH (FY19 and FY23) and utilization (FY23) across HAC-based clusters. Results HAC identified four clusters, or regional profiles (e.g., older/sicker/rural vs. younger/growing/more female). Analysis of access and utilization outcome variations showed apparent differences by cluster; e.g., one cluster with the lowest variation in access scores also had the lowest access score, suggesting regional access challenges. Conclusions PCA-informed HAC produced interpretable regional typologies that aligned with operational leadership insights and corresponded with illustrative variations in CRH program access and utilization results. Future work, however, is needed to confirm the usefulness of this method more generally as an approach for understanding outcome variations across regionally based programs.