|
BACKGROUND/RATIONALE:
Improving access to care was identified by the VA Undersecretary for Health as a major VA goal in 2002. System-level studies indicating areas that can be improved to promote better access were designated as high priority for VA-funded research. Such projects would ensure that all enrolled veterans are linked to the services that they need. This project identified system-wide errors in identification of HIV patients and demonstrated putative access deficits for a group of patients who may not be receiving the care that they need.
OBJECTIVE(S):
The specific aims of this study were to (1) develop, test, and validate procedures for HIV casefinding by analyzing HIV testing, and using algorithms for more complete, accurate, and timely identification of HIV cases onto the Immunology Case Registry (ICR, which is the VA s HIV disease registry), as compared to past and current versions of the ICR; (2) evaluate the effects of enrollment onto the ICR in determining access to quality HIV care and outcomes.
METHODS:
We conducted retrospective analyses of VA patient data from October 1997 through May 2005, inclusive. To address the first objective, we used a subset of the data from June 2004 to May 2005 to model disease status using logistic regression (LR), decision tree (DT), and artificial neural networks
(NN) approaches. Predictive performance of the candidate models were quantitatively compared with respect to sensitivity, specificity, and Receiver Operating Characteristic area-under-the-curve (AUC) to a reference model (RM) that only used HIV-specific ICD9 diagnostic codes. Fully merged data were derived from National Patient Care Database (NPCD), Decision Support System (DSS), and Pharmacy Benefits Management (PBM) databases.
FINDINGS/RESULTS:
For the algorithm development phase, we found that our best three models outperformed the reference model (RM) both in terms of lower false negative
(FN) rate and higher AUC index. Adjusted FN rates (# variables selected at
p 0.5) were: Stepwise LR=0.012 (19); Boosted DT with 500 trees=0.011 (53); NN with 3 neurons=0.010 (54); compared to RM=0.016 (1). All AUC indices for our models were significantly higher than the RM (p 0.001). DT, LR, and NN models predicted more patients than were entered on the HIV registry for the study year.
IMPACT:
Use of disease registries in management of chronic care ensures regular follow-up and timely provision of evidence-based care by providers who have the appropriate expertise. Additionally, the simplicity of the current ICR s capture formula and the limited range of codes used result in significant amounts of false positives. Our models have the unique features of adding probabilistic weights, more indicators of infection, and antiretroviral prescriptions to the capture mechanism. This resulted in improved sensitivity and specificity of the mechanism which has potential to reduce the problem of false identifications and eliminate the need for extensive manual confirmation.
PUBLICATIONS:
None at this time.
DRA:
Chronic Diseases, Health Services and Systems
DRE:
Technology Development and Assessment, Diagnosis and Prognosis
Keywords:
HIV/AIDS, Quality assurance, improvement, Research method
MeSH Terms:
none
|
