144-2008: In Search for a Golden Algorithm

This paper describes an application of data mining methods for development of an HIVcasefinding algorithm with SAS Enterprise Miner (EM) 5.2. Access to HIV care depends on accurate identification of all infected persons. The Veterans Health Administration (VHA) provides care to ~20,000 HIV-infected veterans. The current algorithm for patient identification into the Registry is based only on HIV-specific diagnostic codes. We built logistic regression (LR), decision tree (DT), and neural network (NN) models to predict a binary outcome variable HIV status. We applied these models to the VHA population to identify patients with high predicted probability of disease. In addition to the diagnostic codes we were using demographic, geographic, laboratory, pharmacy and service utilization variables. False Negative (FN) rates and Area Under the Curve (AUC) indices were used for model comparisons. Our best models outperformed the reference model (RM) both in terms of lower FN rate and higher AUC index. The lowest FN rate (0.010% vs. 0.016% for the RM) was demonstrated by the NN model, while the highest AUC index was observed for the LR model (0.995 [0.994, 0.996] vs. 0.974 [0.971, 0.977] for the RM). Non-HIV-specific variables selected by our models included age, race/ethnicity, marital status, service-connected disability, number of days hospitalized, number of primary care and social work visits, number of total and lipids lab tests, blood pressure and liver co-morbidities. Apart from those already on the registry, new algorithms have identified additional 5% new cases. Using EM, our approach can be applied to other disease registries where electronic clinical data are available.