Computational intelligence methods for risk assessment of HIV

Demographic and medical history information obtained from annual South African antenatal surveys is used to estimate the risk of acquiring HIV. Biomedical individualism refers to the factors that place an individual at risk of acquiring an infectious disease, which affects the risk profile of that individual. The design of the estimation system consists of two stages, the first of which is a neural network trained to perform binary classification, using supervised learning with the survey data. The survey information containing discrete variables such as age, gravidity and parity, as well as the quantitative variables race and location, make up the input to the neural network, and the HIV status as the output. A multilayer perceptron with a logistic function is trained with a cross entropy error function, which allows for a probabilistic interpretation of the output. Predictive and classification performance, sensitivity and specificity are measured, and the Receiver Operating Characteristic is derived. This curve illustrates the ability of the classifier to produce a binary output based on varying thresholds. In the second part of the system, the trained neural network produces the inferred risk probability, using Bayesian classification methods to estimate the class conditional densities. The predicted posterior probability is adjusted using Bayes Theorem to account for differing prior probabilities of the training and testing data. One of the difficulties encountered in the survey data is missing data, and this presents a significant problem with neural networks, as they are unable to make predictions using partially complete inputs. An auto-associative neural network is trained on complete datasets, and when presented with partial data, global optimization methods are used to approximate the missing entries. The effect of the imputed data on the network prediction is investigated.