PROBABILISTIC MODEL-BASED PATTERN ANALYSIS OF HEALTH RESOURCE USE AMONG PEOPLE LIVING WITH HIV/AIDS

Purpose: Identifying patterns of health resource utilization (HRU) of people living with HIV/AIDS (PLHIV) can help in the development of targeted interventions on health outcomes and costs.

Method: We conducted a population-level analysis of HRU for individuals having received a CD4 test after HIV diagnosis. All individuals in British-Columbia in the modern antiretroviral treatment-era (post-September 2006) were included. We derived from the first year following linkage ten categorical and binary indicators capturing HRU from linked comprehensive administrative health databases. Using a probabilistic model-based clustering analysis for mixed data, parameters were estimated by the method of maximum likelihood (ML) using the expectation maximization (EM) algorithm. Individuals with estimated parameters maximizing the posterior probability of belonging to a similar cluster are classified with each other, and the optimal number of clusters was estimated by the Bayesian Information Criterion (BIC). The analysis was conducted across CD4 count stratification (>200cells/mm³; <200cells/mm³). 

Result: Our study included 941 individuals with at least one year follow-up (median age 40, 21% female) and with a CD4 count obtained between September 1^st, 2006 and March 31^st, 2011. The 215 individuals with CD4<200 clustered in 2 HRU patterns. The high cost cluster (N=58; mean $23,691 [SD: $25,443]) had costs more than four times the low cost cluster (N=157; $5,494 [$9,066]). Driving the difference in costs were lengthy HIV-related hospitalizations (62.1% with >7 days in the high cost cluster) and more frequent non-HIV-related physician visits (mean visits 92 vs. 24). The 726 individuals with CD4>200 were best classified in 3 clusters. The high cost cluster (N=146; $11,981 [$16,490]) was characterized by numerous non-HIV physician visits and ER hospitalizations (87.8% of all individuals with ≥1 day) as well as a high prevalence of mental health issues. Mean costs were more than triple that of the medium cost cluster (N=428; $2,723 [$4,170]). The low cost cluster (N=152; $1,391 [$8,360]) had almost no hospitalizations (98.7% with 0 days) and relatively few total non-HIV medication days.

Conclusion: Electronic medical records can be used to characterize heterogeneous HRU patterns in the interest of designing public health interventions to optimize clinical response and improve efficiency in medical care delivery.