Predicting bacterial vaginosis incidence using artificial neural networks

Bacterial vaginosis (BV) is a vaginal dysbiosis associated with adverse reproductive and infectious outcomes. Current diagnostics identify BV after symptom onset. We evaluated whether artificial neural network (ANN) models of the vaginal microbiome could predict incident BV (iBV) up to 14 days before clinical diagnosis.

ANNs were trained using 16S rRNA gene sequencing data from 1,201 longitudinal vaginal specimens collected from 58 women across two prospective cohorts. Models classified individual specimens as pre-iBV or healthy using the relative abundance of vaginal bacterial taxa. Model performance was assessed using a held-out participant-level test set, participant-level cross-validation, and external validation. SHAP analysis was used to identify taxa associated with model predictions.

On the held-out participant-level test set, the ANN achieved 93% accuracy (AUC = 0.97, sensitivity = 95%, specificity = 92%). Models using only five taxa (Lactobacillus crispatus, Gardnerella spp., L. iners, L. mulieris, and Megamonas spp.) maintained >91% accuracy, sensitivity, and specificity. SHAP analysis identified Lactobacillus spp. and Gardnerella spp. as the taxa most strongly associated with model predictions. Participant-level cross-validation produced lower and more variable performance (AUC = 0.826 ± 0.076; accuracy = 71.5% ± 8.2%), while external validation achieved approximately 80% balanced accuracy.

Vaginal microbiome composition contains predictive information that can identify women at risk of iBV before clinical onset. Lower performance during participant-level cross-validation and external validation indicates that additional validation in larger, more diverse cohorts is required before clinical implementation.