Artificial neural networks to predict virological and immunological success in HIV patients under antiretroviral therapy from a nationwide cohort in Colombia, using the SISCAC database.

ObjectiveThis study aimed to develop predictive models both for viral suppression and immunological reconstitution using a standard set of reported variables in a nationwide database system (SISCAC) from a cohort of patients living with HIV in Colombia. Materials and MethodsWe included 2.182 patients with no missing data related to the outcomes of interest, during a 12 month follow up period. We randomly assigned a 0,7 proportion of this cohort to de training dataset for 2 different predictive models (logistic regression, artificial neural networks). The AUC/ROC results were compared with those obtained through the construction of artificial neural networks with the specified parameters. ResultsFrom a cohort of 2182 patients, 85,79% were male and at HIV diagnosis, the mean value of the CD4 count was 342 x mm3. The logistic regression models obtained AUC/ROC accuracy for the outcomes "suppressed viral load" 0,7, "undetectable viral load" of 0,66 and "immunological reconstitution" 0,83; whereas the artificial neural network perceptron multilayer obtained AUC/ROC of 0,77, 0.69 and 0,87 for the same outcomes. ConclusionsThe selection of specific variables from a nationwide database in Colombia with quality control purposes allowed us to generate predictive models with an initial evaluation of performance regarding three predefined outcomes for virological and immunological success.