Predicting Incident Bacterial Vaginosis with Machine Learning

Longitudinal microbiome modeling to predict bacterial vaginosis before clinical onset.

Overview

Bacterial vaginosis (BV) is a common vaginal condition with high recurrence rates and limited ability to predict disease before symptoms appear. This project developed machine-learning models that identify incident BV weeks prior to clinical diagnosis using longitudinal vaginal microbiome data.

The focus was early risk stratification in real patients, rather than retrospective classification.

Approach

  • Longitudinal cohort with dense, patient-level sampling before BV onset
  • Vaginal microbiome profiling using 16S rRNA sequencing
  • Artificial neural networks trained on temporal microbial features
  • Built-in explainability using SHAP-based feature attribution
  • Parallel analysis of cohort-level trends and individual patient trajectories

Models were designed to learn how microbiomes evolve over time leading into disease.

Key Findings

  • Predictive signal emerged well before clinical BV diagnosis
  • Model performance remained robust across heterogeneous baseline microbiomes
  • Feature attribution highlighted specific taxa and temporal shifts driving risk
  • Patient-level explanations revealed multiple mechanistic paths to BV

Outputs

  • U.S. Provisional Patent Application No. 63/778,989
    Predicting Bacterial Vaginosis Development Using Artificial Neural Networks
    Inventors: Jacob Elnaggar, Christopher Taylor, John Lammons, Christina Muzny
    Filed March 27, 2025

  • First-author manuscripts in preparation
  • Conference presentations and invited talks
  • Modeling framework reused in downstream dysbiosis and cancer-related projects

Clinical Relevance

Early identification of BV risk enables preventive interventions, personalized monitoring strategies, and a generalizable framework for predicting other microbiome-associated diseases. This work established a foundation for translating longitudinal microbiome data into clinically actionable tools.

Tools & Methods

Python · TensorFlow · SHAP · scikit-learn · pandas · longitudinal modeling · reproducible pipelines