USE OF NEURAL NETWORKS TO MODEL COMPLEX IMMUNOGENETIC ASSOCIATIONS OFDISEASE - HUMAN-LEUKOCYTE ANTIGEN IMPACT ON THE PROGRESSION OF HUMAN-IMMUNODEFICIENCY-VIRUS INFECTION
Jpa. Ioannidis et al., USE OF NEURAL NETWORKS TO MODEL COMPLEX IMMUNOGENETIC ASSOCIATIONS OFDISEASE - HUMAN-LEUKOCYTE ANTIGEN IMPACT ON THE PROGRESSION OF HUMAN-IMMUNODEFICIENCY-VIRUS INFECTION, American journal of epidemiology, 147(5), 1998, pp. 464-471
Complex immunogenetic associations of disease involving a large number
of gene products are difficult to evaluate with traditional statistic
al methods and may require complex modeling. The authors evaluated the
performance of feed-forward backpropagation neural networks in predic
ting rapid progression to acquired immunodeficiency syndrome (AIDS) fo
r patients with human immunodeficiency virus (HIV) infection on the ba
sis of major histocompatibility complex variables. Networks were train
ed on data from patients from the Multicenter AIDS Cohort Study (n = 1
39) and then validated on patients from the DC Gay cohort (n = 102). T
he outcome of interest was rapid disease progression, defined as progr
ession to AIDS in <6 years from seroconversion. Human leukocyte antige
n (HLA) variables were selected as network inputs with multivariate re
gression and a previously described algorithm selecting markers with e
xtreme point estimates for progression risk. Network performance was c
ompared with that of logistic regression. Networks with 15 HLA inputs
and a single hidden layer of five nodes achieved a sensitivity of 87.5
% and specificity of 95.6% in the training set, vs. 77.0% and 76.9%, r
espectively, achieved by logistic regression. When validated on the DC
Gay cohort, networks averaged a sensitivity of 59.1% and specificity
of 74.3%, vs, 53.1% and 61.4%, respectively, for logistic regression.
Neural networks offer further support to the notion that HIV disease p
rogression may be dependent on complex interactions between different
class I and class II alleles and transporters associated with antigen
processing variants, The effect in the current models is of moderate m
agnitude, and more data as well as other host and pathogen variables m
ay need to be considered to improve the performance of the models. Art
ificial intelligence methods may complement linear statistical methods
for evaluating immunogenetic associations of disease.