Neural networks in the assessment of HIV immunopathology

Surrogate markers are by definition quantifiable laboratory variables that have clinical and biological relevance to disease outcomes. Virologic and immunologic surrogate markers have proven useful in following HIV-associated viral burden, immune dysregulation, dysfunction and deficiency. Monitoring of sequential changes in these markers and their interrelationships may provide significant information about viral-host-drug dynamics. The complexity and fluidity of these changes necessitates that an efficient means be developed for their monitoring. We therefore generated a neural network-based model for assessing host dynamics over time and compared its performance with that of a multiple regression model. Both modeling approaches were applied to the actual, non-filtered, clinical observations on 58 HIV-infected individuals treated consistently with Highly Active Anti-Retroviral Therapy (HAART), for a period of over-52 weeks resulting in an average of 16 observations per patient throughout this time span. Results demonstrated that the neural network was at least as accurate as a multi-regression model. Since our dataset was modest in size we also believe that neural networks warrant further consideration for modeling the complexity of HIV-host dynamics on larger datasets.