(L-468) Feature selection in ML identifies unique bacteria to consider for fair BV diagnosis among ethnicities

Introduction:: Bacterial Vaginosis (BV) is a common condition affecting 20-60% of women globally and is caused by an imbalance of bacteria in the vaginal microbiome. Due to its diverse composition, it is difficult to characterize what bacteria compose a “healthy” vaginal microbiome. As the vaginal microbiome has been shown to vary based on race and ethnicity, there can be misdiagnosis through Nugent score and Amsel criteria, which are currently the main methods for diagnosing BV. Feature selection can be used with 16S sequencing data to identify microbes important for diagnosis across ethnicities. These methods can be paired with machine learning methods to accurately predict BV in all women.


Materials and Methods:: The dataset used from Srinivasan includes 221 patients categorized as White, Black, Asian, American Indian or Alaska Native, and Native Hawaiian and Pacific Islander. Ethnicities were categorized as White, Black, and Asian and Other to produce a large enough sample set. The dataset includes 16S sequencing data of each patient with their respective Nugent score and Amsel criteria. Feature selection was done using the F-test, Point Biserial Correlation, T-Test, and Gini impurity to condense the original 159 features to a smaller set of significant features. Feature selection was performed to find significant features of importance including all ethnicities, and then used to find significant features based on each individual ethnicity within the dataset (White, Black, Asian and Other). In order to assess fairness of prediction using significant features, a Support Vector Machine (SVM) was used to classify each patient (BV +/-). The SVM was then used to classify each patient with the T-Test with each individual ethnicities specific feature sets. The accuracy and false positive rates (FPR) and false negative rates (FNR) were used as metrics to assess accurate predictions between ethnicities.


Results, Conclusions, and Discussions:: Top features found by T-Test between Nugent score and Amsel criteria were similar. Both feature sets identified pH, Eggerthella, Gardnerella Vaginalis, Leptotrichia amnionii, Lactobacillus Crispatus, and Sneathia as the top features. The T-Test and F-Test achieved the highest accuracies (0.929 and 0.857) when predicting BV status based on Nugent score and Amsel criteria, respectively. Reducing features to ~40 had comparable accuracy to using all features. Overall, Nugent scoring had a higher accuracy than Amsel criteria. The White subset had the highest accuracy (0.948) with the Black subset resulting in the lowest accuracy (0.884). When testing how ethnicity specific features influence accuracy, we found that the Black feature set most accurately predicted BV across all ethnicities for the Nugent scoring (0.931). Conversely, when predicting BV using Amsel criteria, the White feature set was the most accurate across all ethnicities (0.858). In both tests, the Asian and Other feature set did produce the lowest accuracies. When investigating how the selected features affect false positive rates and false negative rates, we found the White feature set produced the highest false positive rates (0.072) when testing for Nugent score. Nugent scoring had lower overall false positive rates and false negative rates compared to the Amsel criteria.

Nugent score and Amsel criteria had similar significant features for accurate prediction. When features were reduced to ~40, we achieved similar accuracy in prediction overall, showing these are important identifiers for BV diagnosis. Amsel criteria is a syndromic diagnosis, whereas Nugent score is based on bacterium types which may need to be taken into account when assessing BV status across ethnicities. The Black feature set overall performed the best showing it may better contain the complexity of bacteria in other ethnicities.

This study emphasizes the significance of understanding the complexity of the vaginal microbiome. Machine learning and feature selection can aid in identifying disparities of the vaginal microbiome between ethnicities leading to personalized approaches for women’s health


Acknowledgements (Optional): : University of Florida
J. Crayton Pruitt Family Department of Biomedical Engineering
Department of Computer & Information Science & Engineering

References (Optional): : Srinivasan, Sujatha, et al. “Bacterial Communities in Women with Bacterial Vaginosis: High Resolution Phylogenetic Analyses Reveal Relationships of Microbiota to Clinical Criteria.” PLOS ONE, Public Library of Science, https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0037818#s5.