Stroke prediction in a sample of HIV/AIDS patients: Logistic regression, Bayesian networks or a combination of both?
Conference
Gutierrez, J, Yoo, C. (2011). Stroke prediction in a sample of HIV/AIDS patients: Logistic regression, Bayesian networks or a combination of both?
. 1030-1036.
Gutierrez, J, Yoo, C. (2011). Stroke prediction in a sample of HIV/AIDS patients: Logistic regression, Bayesian networks or a combination of both?
. 1030-1036.
Background: Cardiovascular disease is an increasingly frequent diagnosis in patient with HIV/AIDS. Predictive models of stroke in these populations can helps planning targeted strategies to reduced morbidity and mortality in HIV/AIDS populations. Methods: In a hospital-based sample, we explored associations with stroke. Three different statistical models were used: multivariate logistic regression (LR), Bayesian networks (BN) and a combination of both. Goodness of fit was evaluated with the area under the curve and reliability was tested with three-cross validation. Results: One hundred and nine patients with HIV/AIDS were included in the analysis. The mean age was 46.87 (SD± 11.6), 56.8% were men, 77% were black, 79% came from low income areas (less than $40,000 year median income). Stroke was associated with high CD4 count, lack of insurance and the presence of cardiovascular risk factors. Using as inputs the presence of cardiovascular risk factors, race, insurance status, HAART administration, CD4 count and medical comorbidities in a multivariate LR, 76 % of the stroke cases were accurately predicted. Using BN analysis, 75% of stroke cases were accurately predicted. Using a combination of LR and BN, only 72% of the cases were correctly classified, although this model had the best discriminant function as evidenced by good positive and negative predictive value. Conclusions: Stroke is associated with cardiovascular risk factors, lack of medical insurance and higher CD4 counts. Using a combination of LR and graphical probabilistic models improved the discriminant function of the predictive model. Longitudinal, population based studies are needed to confirm these associations.
