Neural networks morbidity and mortality modeling during loss of HIV T-cell homeostasis

Abstract

Despite the proven clinical benefits of HAART, mortality may still occur; particularly in those with less than 50 CD4+ cells/mL and, in some cases, with a viral burden below detectable plasma levels of HIV-1 RNA. Multiple factors may predict mortality including initial response to therapy, viral factors and host immune parameters. Due to the complexity of this problem, we developed Artificial Intelligence based tools/Neural Network (NN) to optimally evaluate outcomes of therapy and predict morbidity and mortality. To further validate the accuracy of these tools, we challenged their performance with that of Cox regression modeling (RM). Our study population involved 116 HIV+ individuals who consistently maintained CD4+ count < 50 cells/mL for over 6 months. All patients were treated with antiretrovirals. To assess clinical outcomes, we developed a feedforward back-propagation Neural Network. We then compared the performance of this network to a Cox regression model. The Neural Network outscored the Cox regression model in the ROC curve areas: 0.888 vs 0.760 (HIV+ first Seropositivity to AIDS), 0.901 vs 0.758 (HIV+ first Seropositivity to Last Assessment incl. death) and 0.832 vs 0.799 (AIDS to Last Assessment incl. death), for the NN & Cox, respectively. In patients with a history of AIDS defining events and with severe T-Cell depletion, mortality occurs despite therapy. Although Neural Networks and Cox modeling were successful in predicting mortality, the Neural Network was superior in assessing risk in this population.