Background: Healthcare systems are increasingly turning to data-driven approaches, such as clustering techniques, to inform interventions for medically complex older adults. However, patients seeking care in multiple healthcare systems may have missing diagnoses across systems, leading to misclassification of resulting groups. We evaluated the impact of multi-system use on the accuracy and composition of multimorbidity groups among older adults in the Veterans Health Administration (VA).
Methods: Eligible patients were VA primary care users aged ≥65 years and in the top decile of predicted 1-year hospitalization risk in 2018 (n = 558,864). Diagnoses of 26 chronic conditions were coded using a 24-month lookback period and input into latent class analysis (LCA) models. In a random 10% sample (n = 56,008), we compared the resulting model fit, class profiles, and patient assignments from models using only VA system data versus VA with Medicare data.
Results: LCA identified six patient comorbidity groups using VA system data. We labeled groups based on diagnoses with higher within-group prevalence relative to the average: Substance Use Disorders (7% of patients), Mental Health (15%), Heart Disease (22%), Diabetes (16%), Tumor (14%), and High Complexity (10%). VA with Medicare data showed improved model fit and assigned more patients with high accuracy. Over 70% of patients assigned to the Substance, Mental Health, High Complexity, and Tumor groups using VA data were assigned to the same group in VA with Medicare data. However, 41.9% of the Heart Disease group and 14.7% of the Diabetes group were reassigned to a new group characterized by multiple cardiometabolic conditions.
Conclusions: The addition of Medicare data to VA data for older high-risk adults improved clustering model accuracy and altered the clinical profiles of groups. Accessing or accounting for multi-system data is key to the success of interventions based on empiric grouping in populations with dual-system use.
Keywords: Veterans health; chronic disease; electronic health records; latent class analysis; patient care management.
Published 2023. This article is a U.S. Government work and is in the public domain in the USA.