Logistic Regression Models for Predicting Resistance to HIV Protease Inhibitor Nelfinavir

Resumo

The development of models to predict the resistance to the antiretroviral drugs can be useful in making a decision regarding the best therapy for HIV+ individuals. This study developed predictive models of resistance to the protease inhibitor Nelfinavir using logistic regression. The data comprises a total of 625 patients for which HIV-1 genotype was available, with 130 resistants to Nelfinavir in the last regimen. Feature selection was carried out using a combination of bootstrap resampling procedure with the stepwise selection technique. Additionally, due to the unbalanced nature of the dataset, we develop four balanced final models. The accuracies of the models ranged from 70.40 to 76.80% and areas under the ROC curve (AUC) ranged from 0.657 to 0.687. The best model had AUC equal to 0.687, accuracy of 76.80%, specificity of 84.21% and sensitivity of 53.33%. The agreement between this model and the known resistance level was fair, Kappa index of 0.3712.