Poster BB19 - Predictive Modeling for Bacterial Vaginosis in a Tanzanian Cohort of Women Living with Human Immunodeficiency Virus( HIV)

Saturday, October 26, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Wambui Gachunga

Graduate Research Assistant
University of Florida
Gainesville, Florida, United States

Co-Author(s)

DO

Diandra Ojo

Postdoc
University of Florida, United States

Primary Investigator(s)

IP

Ivana Parker

Assistant Professor
University of Florida, United States

Co-Author(s)

CC

Carleigh Coffin

Graduate Research Assistant
University of Florida, United States

Introduction: Women of African descent are disproportionately affected by the HIV epidemic. One of the major drivers of the HIV epidemic is bacterial vaginosis (BV), a prevalent vaginal syndrome that is often associated with a multitude of other sexual and obstetric complications. AI and machine learning present opportunities for analyzing data for diagnosing BV, which has been demonstrated in prior works focusing on female cohorts in the United States. In this work, we implement four machine learning models for predicting BV in a cohort of women from Tanzania, Africa living with HIV
using 16S rRNA sequencing data. We find that model performance for indicating BV in this African cohort was lower compared to prior works using cohorts from the United States. Results indicated that intermediate Nugent scores posed challenges for model discrimination, often leading to misclassification. Interestingly, we find Lactobacillus iners, typically found to be less resistant to BV, to be a significant predictor of BV among this cohort. These findings point to new areas of investigation for improving the feasibility and equity of predictive diagnosis of BV involving women with vaginal and sexual disorders and conditions.

Materials and

Methods: This study analyzed the microbial communities in 272 samples from 118 Tanzanian women with HIV, aged 18-45, to explore the relationship with bacterial vaginosis (BV) diagnosis. The dataset included operational taxonomic unit (OTU) variables, total reads, Amsel’s criteria, and Nugent score. OTU variables were derived from 16S rRNA gene sequences. Predictive models for BV were developed using random forest, logistic regression, support vector machine, and multi-layer perceptron. The performance of these models was compared with two non-HIV cohorts: a symptomatic cohort of 220 women from a Seattle study and an asymptomatic cohort of 396 women from Baltimore and Atlanta. Model development followed TRIPOD+AI guidelines, employing four-fold nested cross-validation to minimize overfitting, with stratification based on BV outcome and ethnicity. Model performance was assessed using balanced accuracy, precision, recall, false positive rate, and false negative rate. BV was diagnosed using the Nugent score, with a score of seven or greater indicating BV positive. Calibration, sensitivity, and specificity were evaluated for each model. Analyses were conducted in Python. Study sponsors did not influence the study design, data collection, analysis, interpretation, writing, or publication decisions.

Results, Conclusions, and Discussions: Results Summary
The study analyzed 272 samples from 118 HIV-positive Tanzanian women, revealing overlaps between BV-positive and BV-negative diagnosis groups, particularly in samples with intermediate Nugent scores (4-6). These intermediate samples complicate model discrimination, often leading to misclassification.

Performance of Predictive Models
Four machine learning models (RF, LR, SVM, MLP) were evaluated for predicting BV outcomes. Logistic regression (LR) and multi-layer perceptron (MLP) showed the highest average AUROCs (0.89), while SVM had the lowest balanced accuracy (0.77). Random forest (RF) excelled in precision (0.77) and had the lowest false positive rate (0.19). SVM performed best in recall (0.78) and false negative rate (0.22). However, all models struggled with intermediate Nugent scores, often misclassifying them as BV-positive.

Nugent Score-Based Performance
When predicting BV outcomes by Nugent scores, RF and MLP were more accurate in classifying intermediate samples as BV-negative, but overall performance remained suboptimal. LR was most effective in predicting BV-negative for low Nugent scores, while SVM excelled in predicting BV-positive for high scores.

Key Predictors
The most significant predictors across all models were Lactobacillus iners, Prevotella timonensis, and Prevotella amniotica. L. iners, identified as the most significant variable in the HIV cohort, contributed to overlaps in BV diagnosis due to its coexistence with both healthy and disruptive bacteria. In HIV-negative cohorts, anaerobic bacteria like Parvimonas micra and Gardnerella were top predictors.

Comparison with HIV-Negative Cohorts
Models performed better in HIV-negative cohorts compared to the HIV-positive cohort. There were no shared significant predictors among all cohorts, highlighting differences between HIV and non-HIV populations. The study found that predictive models for BV were less accurate for the Tanzanian HIV cohort than for Black women in the symptomatic and asymptomatic HIV-negative cohorts.

Conclusion
Machine learning models demonstrated potential in predicting BV in high-risk, HIV-positive populations, offering an advancement over traditional diagnostic methods. However, intermediate Nugent scores pose a challenge for accurate diagnosis. The study underscores the need for tailored diagnostic tools considering specific biological and epidemiological contexts to improve health outcomes and reduce HIV transmission, thereby addressing global health disparities.

Acknowledgements (Optional):