Skip to main content

Coronavirus (COVID-19): Visitor Restrictions, Resources, and Updates

Explore URMC
menu
URMC / Del Monte Institute for Neuroscience / NeURoscience Blog / May 2019 / Using Machine Learning to Diagnose HIV-associated Neurocognitive Disorders

Using Machine Learning to Diagnose HIV-associated Neurocognitive Disorders

Udaysankar ChockanathanMore than 25 percent of individuals with HIV experience a set of complications known as HIV-associated neurocognitive disorders (HAND). Symptoms are often subtle and dynamic, which renders diagnosis of the condition challenging. A new study published in Computers in Biology and Medicine explores automated approaches to detect biomarkers of HAND. In the study, led by third-year URMC MSTP student Udaysankar Chockanathan, researchers trained a machine learning model using brain network properties derived from functional MRI (fMRI) data. They then applied the learned model to predict HIV-status and cognitive performance at the level of individual subjects.

“To perform this study, we integrated expertise from many different disciplines,” said Chockanathan. “For instance, one of the study’s authors is Dr. Giovanni Schifitto, a neurologist with experience in the management of patients with HAND. The senior author, Dr. Axel Wismüller, is a radiologist who runs a research lab on computational methods for medical image analysis.”

lsGC top regions

The image shows the top 10 brain regions that were most informative at distinguishing HIV-negative from HIV-positive subjects.

Researchers took raw fMRI data and built brain connectivity maps using two different analysis methods: Pearson correlation, a widely used technique for comparing the similarity between two variables, and large-scale Granger causality (lsGC), a novel method developed from an algorithm initially deployed in economics.  The brain connectivity maps generated from each method were analyzed with a machine learning model to predict (1) HIV-status and (2) cognitive performance. Prediction of both variables was better with lsGC than with Pearson correlation.

“The fact that a technique originally developed for economics might help predict brain scores better was surprising to me,” said Chockanathan. “No one has taken these brain networks and thrown them at a machine learning algorithm to use them to predict HIV status, so this is new.”

Down the line, Chockanathan would like to understand how biological and socioeconomic variables, such as age, sex, education level, drug use, and antiretroviral medications, affect the brain networks of individuals with HAND, and, ultimately, to find biomarkers that can be used to identify HIV-positive individuals at risk for HAND, before neurocognitive decline begins.

Additional co-authors Dr. Adora DSouza (PhD ’19) and Dr. Anas Abidin (PhD ’19) provided substantial insight on the design and implementation of the models used in the study.

An excerpt of this article appeared in the Spring 2019 issue of NeURoscience.

Samantha Jean | 5/24/2019

You may also like

No related posts found.