We built a first-order stochastic decision-analysis model to determine the costs of treating a hypothetical cohort of patients identified due to a Rapid Testing screening program vs. costs of treating this same cohort of patients were they identified due to symptoms of disease (Usual Care). Cohort size indicates the number of HIV-positive persons identified by the screening programs in the ED. The ED was chosen as the screening site due to its ability to access an otherwise difficult-to-reach population and its predicted higher prevalence of HIV infection compared to primary care locations (11
). This analysis used cost and economic data from an urban VA ED and Infectious Disease clinic located in a major metropolitan area. The VA was chosen due to the extensive data available from its cost databases and electronic medical records. Institutional Review Board approval was obtained from the VA system.
Due to the fact that the VA generally keeps patients for life, this analysis assumed all patients would be identified at some point in the system due to their symptoms. Therefore, sizes for the Rapid Test and Usual Care cohorts were identical. We modeled a variety of cohort sizes to reflect uncertainty in the number of patients offered testing, the percentage of patients accepting testing, and the prevalence of HIV (). This analysis evaluated the financial impact of three different offer rates: five per business day, 10 per business day, or universal. The former two rates were used to reflect realistic levels of testing that can occur in environments with existing capacity constraints; the latter was used in accordance with CDC recommendations. The model assumes HIV prevalence is independent of test acceptance.
Costs of Rapid Testing include both program implementation and treatment of disease. Costs of Usual Care include disease-treatment costs only, but health care utilization is more intensive in this population due to late identification of disease. Treatment costs were calculated for the following budgets: inpatient, outpatient, pharmacy, and global (where global indicates the sum of the former three sub-budgets). All treatment costs were assumed to be dependent upon disease severity at diagnosis (measured by CD4+ count). Disease severity was noted by the following four categories: CD4+ < 50, CD4+ 50–199, CD4+ 200–350, and CD4+ > 350 cells/mm3 at diagnosis (where low CD4+ values indicate higher disease severity). Costs were projected over a time horizon of 7 years. A horizon of 7 years is longer than usual for BIAs; however, we felt this longer horizon was necessary to capture the delay between when patients are diagnosed through screening and when they begin accruing costs of treatment.
Treatment cost was a function of cost per disease-severity category and the proportion of cohort members in each disease-severity category. Estimated costs were calculated on an annual basis for each budget using 2007 dollars. The following equation was used to model the total estimated costs of treating disease:
- Cj is the average cost per case in sub-budget category j
- i is the disease-severity category (CD4+) assigned at time of diagnosis
- t is the year in which costs are accrued
- C is the average cost per case of disease
- N is the number of people
Cost per disease-severity category was determined by examining utilization patterns of HIV-positive patients and allocating VA-specific direct medical cost values to this utilization. VA-specific costs were used, following best practice recommendations for conducting BIAs. (10
). Utilization data came from chart review of all treatment-naïve patients diagnosed and treated at this site from Fiscal Year (FY) 2000 to FY 2007 (n = 112). Although this number is small, the analysis required the use of these actual data, due to the dearth of literature available on treatment-naïve patients and the fact that data from treatment-experienced patients would substantially overestimate treatment costs. Modeling HIV-related utilization over time also poses significant difficulties, given that drug treatment efficacy decreases over time, is influenced by patient adherence, and that the relationship between adherence and consequences of much highly active anti-retroviral therapy (HAART) is not fully understood (13
). For example, partial adherence increases the risk of drug resistance more so than complete non-adherence (13
Only HIV-related utilization was included in the model. All inpatient stays with HIV as a primary or secondary diagnosis were reviewed by two infectious disease physicians who had treated the majority of these patients; only those inpatient stays unanimously deemed to be due to HIV were included in this analysis.
Once the cost per disease-severity category was calculated, we populated the model with the program-specific CD4+ distributions. As Usual Care patients are diagnosed at later, more advanced stages of disease than are patients identified through Rapid Testing, the disease-severity distributions differed across the programs; a larger proportion of Usual Case patients had lower CD4+ counts at diagnosis. The CD4+ distribution for the Usual Care program was determined using patients from the national Veterans Aging Cohort Study (VACS) Virtual Cohort, which contains clinical data on all HIV-infected veterans in the country (14
). Disease severity at diagnosis was obtained for VACS Virtual Cohort members who had ED visits in a VA facility before their HIV diagnosis (n = 3355), which represents the Usual Care patients in this model.
We used a back-calculation scheme to determine the CD4+ distribution for the Rapid Testing program. Annual CD4+ decline in the absence of treatment can be estimated by the following equation: CD4+^½ =− .0584 – 0.918 [log(vRNA/1000)] (15
). Entering patients’ viral load at diagnosis, CD4+ count at diagnosis, as well as the length of time elapsed between first ED visit and HIV diagnosis into the equation yields an estimate of patients’ CD4+ count had they been diagnosed with HIV at the time of first ED visit (a Rapid Testing approach). For example, a patient presenting 2 years after his first ED visit with an initial CD4 count of 150 cells/mm3
and viral load of 100,000 copies would be back-calculated to have had a CD4 of 292 cells/mm3
had he been diagnosed at the time of initial ED visit ().
Model Inputs: Percentage of Patients in Each CD4 Category
As identified HIV-positive patients, persons in the Rapid Testing arm begin accruing treatment costs immediately. However, it takes some time for Usual Care patients to display symptoms of disease, be diagnosed, and therefore start incurring treatment costs. Data from the VACS Virtual Cohort indicate a mean 1.25-year delay between the first ED visit and diagnosis of HIV. All estimated costs in the Usual Care program were adjusted back 1.25 years to reflect this delay to diagnosis. Over the entire analysis, the Rapid Test program therefore had 7 years of data, whereas the Usual Care program had 5.75 years of data.
In the outpatient and pharmacy models, cost values associated with the Usual Care and Rapid Test programs were identical. The model assumed drug costs and office visits were not driven by method of HIV diagnosis, but simply by the severity of patient illness. However, cost values for the inpatient and thus global models differed across the Usual Care and Rapid Testing programs; inpatient utilization, and cost, was dependent upon whether a patient was diagnosed through Rapid Testing or Usual Care. Inpatient stays within 3 months of HIV diagnosis were excluded from the Rapid Testing costs. These hospitalizations represent patients who were diagnosed due to opportunistic infections; their exclusion was more indicative of the clinical scenario of Rapid Testing patients, who are not diagnosed due to opportunistic infections and therefore would not have experienced such inpatient stays.
Once the model was populated with the cost per CD4+ category and the distributions of CD4+ counts, we ran the analysis and employed statistical techniques to mitigate the noise associated with small samples. Probabilistic sensitivity analyses (with 10,000 Monte Carlo iterations) were run to gain a more accurate estimate of cost per CD4+ category, with each iteration randomly sampling from each distribution of cost per CD4+ category. This re-sampling increases the likelihood that outlier cost inputs would not have an undue influence on cost estimates. Model analyses were conducted using TreeAge Pro 2008 software (TreeAge Software Inc., Williamstown MA).
Usual Care program costs were directly obtained from model results. Rapid Testing program costs were obtained by adding program implementation costs to model results. Implementation costs were determined by micro-cost estimates of resources used in offering and conducting HIV tests, including testing kits and labor. Time-and-motion data were obtained by direct observation of HIV rapid testing in the ED. An applied rate of 85% was used for all salary calculations. Two-tailed t-tests of significance were run on the total costs of each program to determine if the overall 7-year global costs of the Rapid Test program were significantly different from the Usual Care program. Tests of significance were run using STATA version 9.2 (StataCorp LP, College Station, TX).
To summarize, the following assumptions were made in conducting these analyses:
- Patients can be properly stratified into four CD4+ count categories.
- All health care utilization, and therefore cost, is contingent upon CD4+ count at diagnosis.
- Utilization patterns of patients identified through these programs can be modeled using utilization data from past HIV-positive patients treated at the same location.
- All persons diagnosed are immediately linked to treatment.
- All patients identified by Rapid Test would be diagnosed an average of 1.25 years later in the Usual Care program.