A New Research Examines Alternate Viral Suppression Prediction Models to Minimize HIV Transmission
Jingxin Liu, a recent graduate of the Ph.D. in Epidemiology program in the Department of Public Health Sciences (DPHS) at the University of Miami Miller School of Medicine is the primary author of a new study that evaluated alternative viral suppression prediction models to reduce HIV transmission.
The new study, titled “Strategies of Managing Repeated Measures: Using Synthetic Random Forest to Predict HIV Viral Suppression Status Among Hospitalized Persons with HIV” was published in the AIDS and Behavior journal in February 2023.
Since the 1980s, the HIV/AIDS epidemic has been a significant public health issue. Researchers are concerned about the consequences of untreated HIV infection on the quality of life, including increased risks of heart disease, physical and mental health risks, and diminished socioeconomic standing, according to the authors.
Despite the significant progress that has been accomplished, major obstacles still exist in the HIV diagnosis and care continuum. In this study, researchers targeted vulnerable populations of people living with HIV (PLWH) and assessed various prediction models for viral suppression utilizing longitudinal or recurring assessments in order to improve patient outcomes and decrease HIV transmission.
Jingxin Liu, Ph.D.
For this study, researchers included two-parent studies, CTN-0049 and CTN-0064, both recruited by the National Drug Abuse Treatment Clinical Trials Network (CTN).
Parent study CTN-0049 was carried out between July 2012 and January 2014. It evaluated how a structured patient navigation intervention, both with and without financial incentives, affected the HIV viral suppression rates among HIV patients who were originally hospitalized and who also used drugs. There were 801 individuals altogether, and they were divided into three groups at random: patient navigation alone, those who received standard care, and those who received standard care with financial incentives.
When compared to standard care, the study indicated that neither additional financial incentives nor patient guidance had a positive impact on HIV viral suppression at the 12-month follow-up (6 months after the intervention was finished).
The same cohort was used for the parent research CTN-0064, which was started two years after CTN-0049's recruitment process was complete. It evaluated the effectiveness of a Hepatitis C Virus (HCV) care intervention for HIV/HCV co-infected people who use drugs and looked at the long-term results of the CTN-0049 study, documenting HCV prevalence among the HIV-infected CTN-0049 sample.
Among the 801 patients who participated in CTN-0049, 422 of them completed a baseline evaluation in CTN-0064 and are included in this study.
“We used repeated assessments of HIV-positive patients who were first enrolled while in the hospital and who subsequently took part in two trials from CTN-0049 and CTN-0064, to compare various approaches of feature engineering,” said Dr. Liu.
Random Forest VS Synthetic Random Forest
“Random Forest (RF) is a machine learning approach for classification and regression, which involves constructing a number of decision trees by bootstrapping samples from the training dataset and selecting random features (predictors) in tree induction,” explained Dr. Liu.
A synthetic random forest (SRF) is defined as a secondary random forest, which uses a collection of both all the original features as well as new inputs of synthetic features, she added.
These new synthetic features are the anticipated results of numerous random forests with different tuning parameters. Researchers stated that the SRF approach outperforms traditional RF and can enhance random forest prediction without necessitating the additional work of locating the ideal tuning parameters, therefore they used SRF to compare the predictive power (overall error rates) of four strategies for incorporating repeated measures as predictors of HIV viral suppression.
Predicting viral suppression
Little research has compared different organizational methods for predicting HIV health outcomes. “Only a few studies have specifically looked at using longitudinal or repeated measurements to predict viral suppression among PLWH,” said Dr. Liu. The availability of longitudinal data allows for incorporating trajectories of change, an approach common in behavioral research, she added.
“Representing repeated measurements as trajectories of change lowers the number and correlation of characteristics and may make it easier to understand the findings,” said Dr. Liu.
Participants who enrolled in CTN-0064 on average had better perceived health and higher CTN-0049 baseline counts of CD4 (T cells that circulate immune cells that search for and eliminate bacteria, viruses, and other foreign pathogens; HIV damages the immune system by targeting CD4 cells).
Except for perceived health and baseline CD4 cell counts, there were no statistically significant differences in the majority of demographic factors, including age, gender, education, and income, between those who participated in CTN-0064 and those who did not.
When comparing out-of-bag (OOB) error rates (method of measuring the prediction error of machine learning models) of the training dataset, person-specific trajectories performed best, yielding a 32.0% OOB error rate.
“This study has a unique advantage in combining data from both CTN-0049 and CTN-0064. This allowed us to have more options in selecting predictors from three distinct time points across two years and to explore the predictive power of longitudinal data and repeated measures for long-run outcomes after two years,” explained Dr. Liu.
Dr. Liu assures the model with person-specific trajectories had the best predictive power as compared to other models. “The findings from this study provide evidence that incorporating not just levels of predictors but also their change over time improves the predictive performance of our models,” she said.
Using person-specific intercepts and slopes provides a novel and useful approach to creating predictive models using repeated measurements. It also suggests the possibility of incorporating these types of modeling efforts into ongoing clinical monitoring using medical records.
Researchers hope that this study will be useful in implementing fresh and innovative perspectives on the development of longitudinal features to forecast outcomes along the HIV care continuum and in determining the most effective predictive methods by employing repeated measurements.
Other DPHS collaborators include Yue Pan, Ph.D., Assistant Professor of Biostatistics, and Daniel J. Feaster, Ph.D., Professor of Biostatistics.
