1079 — Comparing Methods for Clustering Comorbidity Data
Goulet JL (VA Connecticut Healthcare System), Fultz SL
(Office of Public Health and Environmental Hazards, VACO ), Justice AC
(VA Connecticut Healthcare System)
Objectives:
Many veterans suffer from multiple chronic conditions. This multi-morbidity may lead to morbidity, mortality, and difficulties in coordinating health care. We compared two methods of identifying multi-morbidity patterns using statistical clustering methods.
Methods:
The sample consisted of 100,260 patients in the VACS Virtual Cohort. HIV-infected veterans were identified from VA administrative data, and a 2:1 age-race-gender matched sample of HIV-uninfected comparators was identified. High-prevalence and high-impact conditions (n=27) were chosen for analysis (e.g. coronary artery disease, depression). Conditions were identified using validated ICD-9 codes.
Two analytic methods were employed to cluster conditions:
1. A hierarchical nesting algorithm, with dendrograms to depict clusters. The agglomerative coefficient (AC) assessed the degree of structure in the data.
2. Latent class analysis (LCA) was used to identify individuals with similar multi-morbidity profiles and the prevalence of each class. Multinomial logistic regression was used to assess the association of patient characteristics with class membership.
Results:
Patients were a median 45 years of age, 98% were male, 43% were black, and 46% had 2 or more conditions (mean= 1.8, range 0-15).
The hierarchical model revealed six clusters: the cluster with the highest degree of association included alcohol and drug disorders. Diabetes, HTN, and hyperlipidemia formed the next cluster. Substance disorders and psychiatric disorders combined to form one larger cluster. There was a moderate degree of clustering in the data (AC=0.62). The LCA identified five classes: Metabolic, containing older, male, white veterans with few other conditions; HIV-Related, the class with the highest number of deaths; Substance Abuse; Psychiatric Disorders; and Healthy, containing patients with a low probability of any comorbid conditions.
Implications:
Both clustering methods revealed specific patterns of multi-morbidity for the conditions assessed. However, the methods and their results are not equivalent. LCA may be more informative, as the prevalence of each class, the patients falling within that class, and the probability of additional comorbidities can be assessed more readily.
Impacts:
VA health services must be organized and delivered to meet the needs of the large number of veterans with multi-morbidity. Identification of clusters may be useful in coordinating resources and designing more effective and accessible treatment strategies.
