THE USE OF NEURAL NETWORKS TO PREDICT VIROLOGICAL RESPONSE IN HIV-POSITIVE PATIENTS

randombroadAI and Robotics

Oct 15, 2013 (3 years and 11 months ago)

72 views

THE USE OF NEURAL NETWORKS TO PREDICT VIROLOGICAL RESPONSE IN
HIV
-
POSITIVE PATIENTS

Y. Singh

UKZN, Durban, South Africa

Introduction
:

In 2009, 30 million of the 33.3 million people infected with HIV live in low
-

and middle
-
income countries. HIV infection c
an be effectively managed with standard
antiretroviral therapy (ART). Close monitoring of the patient is vital in

order to determine the
optimal time to start the ART and to measure the success of the ART. One of the best
available surrogate markers for HI
V progression is the use of CD4 lymphocyte cell counts.
There have been some attempts to measure the progression and other aspects of HIV using
machine learning. Many studies have shown machine learning may predict current drug
resistance, future drug resi
stance, evolution of the HIV in a patient, and predicting future CD4
count. Most of these techniques however have a major limitation in its use in developing
countries: they make use of part of the patient genome as an input into the machine
-
learning
model
. Determining the GAG and POL genes is still an expensive and laborious task in
environments where there are limited recourses. AIM The aim of this pilot study was to apply
a machine
-
learning technique to investigate if it is possible to forecast CD4 count

change by
using readily available data, and specifically without the use of genomes.

Methodology
:

Separate patient datasets containing treatment information and disease
progression were obtained from the Stanford HIV drug resistance database
(http://hivd
b.stanford.edu/). These datasets consisted of information regarding the ARV drugs
the patients have been exposed to, CD4 cell counts, viral loads and the number of weeks from
the baseline CD4 cell count measurement. A neural network was created that takes
as input
the ARV drugs the patients have been exposed to, viral loads and the number of weeks from
the baseline CD4 cell count measurement, and tries to determine if the change in CD4 count >
25%.

RESULTS
:

Preliminary results indicate an accuracy of 81%.
The neural network produced a
sensitivity of 72%, specificity of 71%, a positive predictive value of 71% and a negative
predicative value of 73%.

Conclusion
:

Preliminary results indicate that it is possible to create a machine learning
technique that can
predict a change in CD4 count. This type of tool is beneficial in CD4 count
guided treatment.

Keywords: bioinformatics, decision
-
support system, computer
-
software, AI