Predicting Bacterial Vaginosis Development using Artificial Neural Networks

Area of Science:

• Microbiology
• Computational Biology
• Gynecology

Background:

• Bacterial vaginosis (BV) is a vaginal microbiome dysbiosis linked to depleted *Lactobacillus* species and increased anaerobes.
• Early detection of incident BV (iBV) is crucial for timely intervention and management.
• Artificial neural network (ANN) modeling presents a novel approach for analyzing complex microbial community data.

Purpose of the Study:

• To develop and validate an ANN model for the early detection of incident BV (iBV) using vaginal microbial data.
• To identify key vaginal bacterial taxa contributing to BV prediction.
• To assess the impact of race-stratified data on model performance.

Main Methods:

• 16S rRNA gene sequencing and quantitative PCR were used to determine the inferred absolute abundance (IAA) of vaginal bacterial taxa.
• ANN models were trained using IAA data from 420 vaginal specimens to classify samples as pre-iBV or Healthy.
• Feature importance analysis was conducted to identify significant microbial contributors to model predictions.

Main Results:

• ANN models achieved high accuracy (>97%), sensitivity (>96%), and specificity (>98%) in classifying pre-iBV and Healthy specimens using 20 taxa.
• Excellent predictive performance (>97% accuracy) was maintained even with models trained on only the top five most important features.
• Race-stratified models demonstrated improved accuracy, with three-feature models achieving >96% accuracy for both White and Black participants.

Conclusions:

• ANN modeling is a highly accurate method for early detection of incident BV (iBV) using vaginal microbiome composition.
• A small set of key vaginal taxa can effectively predict BV status, simplifying diagnostic approaches.
• Stratifying models by race may enhance predictive accuracy, highlighting the need to consider population-specific microbial patterns.