Predicting preterm birth through vaginal microbiota, cervical length, and WBC using a machine learning model

Abstract

An association between the vaginal microbiome and preterm birth has been reported. However, in practice, it is difficult to predict premature birth using the microbiome because the vaginal microbial community varies highly among samples depending on the individual, and the prediction rate is very low. The purpose of this study was to select markers that improve predictive power through machine learning among various vaginal microbiota and develop a prediction algorithm with better predictive power that combines clinical information. As a multicenter case-control study with 150 Korean pregnant women with 54 preterm delivery group and 96 full-term delivery group, cervicovaginal fluid was collected from pregnant women during mid-pregnancy. Their demographic profiles (age, BMI, education level, and PTB history), white blood cell count, and cervical length were recorded, and the microbiome profiles of the cervicovaginal fluid were analyzed. The subjects were randomly divided into a training (n = 101) and a test set (n = 49) in a two-to-one ratio. When training ML models using selected markers, five-fold cross-validation was performed on the training set. A univariate analysis was performed to select markers using seven statistical tests, including the Wilcoxon rank-sum test. Using the selected markers, including Lactobacillus spp., Gardnerella vaginalis, Ureaplasma parvum, Atopobium vaginae, Prevotella timonensis, and Peptoniphilus grossensis, machine learning models (logistic regression, random forest, extreme gradient boosting, support vector machine, and GUIDE) were used to build prediction models. The test area under the curve of the logistic regression model was 0.72 when it was trained with the 17 selected markers. When analyzed by combining white blood cell count and cervical length with the seven vaginal microbiome markers, the random forest model showed the highest test area under the curve of 0.84. The GUIDE, the single tree model, provided a more reasonable biological interpretation, using the 10 selected markers (A. vaginae, G. vaginalis, Lactobacillus crispatus, Lactobacillus fornicalis, Lactobacillus gasseri, Lactobacillus iners, Lactobacillus jensenii, Peptoniphilus grossensis, P. timonensis, and U. parvum), and the covariates produced a tree with a test area under the curve of 0.77. It was confirmed that the association with preterm birth increased when P. timonensis and U. parvum increased (AUC = 0.77), which could also be explained by the fact that as the number of Peptoniphilus lacrimalis increased, the association with preterm birth was high (AUC = 0.77). Our study demonstrates that several candidate bacteria could be used as potential predictors for preterm birth, and that the predictive rate can be increased through a machine learning model employing a combination of cervical length and white blood cell count information.

Keywords: 16s ribosomal RNA metagenome sequencing; cervicovaginal fluid; machine learning; microbial-marker; pregnancy; preterm birth; vaginal microbiome.

Figure 1

Flowchart of the study. CVF, cervicovaginal fluid; rRNA, ribosomal ribonucleic acid; OTUs, operational taxonomic units.

Figure 2

Flowchart of marker selection and evaluation in exhaustive search. The data were split to a training set and test set in a two-to-one ratio. Markers with frequency more than 25% and mean proportion more than 0.001% were selected. Then, markers, showing significant p values in two or more statistical tests, were selected. Venn diagram of significant markers (p < 0.05) after seven statistical methods (ZIG, ZIBSeq, ANCOM, CLR permutation, Wilcoxon rank-sum test, DESeq2, and edgeR) is shown. Additional filtering steps were applied to the selected markers to finalize the set of 10 and 17 markers. For the given set of markers, exhaustive search was applied to every possible combination of markers using LR. Best marker sets for each number of combinations were selected using AUC from the training set. The global best marker set among these selected sets was chosen as the one that showed the highest AUC from the five-fold CV. Then, the final marker set was select based on the test set. Lastly, the final marker sets were used in building machine learning (ML) models.

Figure 3

Differences in alpha- and beta-diversity between PTB and TB groups. (A) Shannon’s alpha diversity was significantly higher in the PTB group (PTB, n = 54; TB, n = 96). (B) Multidimensional scaling plot. Boxes show median and interquartile ranges, and whiskers extend from minimum to maximum values.

Figure 4

ROC curve and feature importance plot of the Random Forest (RF) models using covariates and selected markers. (A) RF model’s ROC curve on test data using 10 selected markers and WBC (B) RF model’s feature importance plot using 10 selected markers and WBC. (C) RF model’s ROC curve on a test using forward-selected markers, WBC and cervical length. (D) RF model’s Feature Importance plot using forward-selected markers, WBC and cervical length.

Figure 5

Decision trees made from GUIDE algorithm using covariates, cervix length and WBC with (A) ten pre-selected markers and (B) seven markers forward selected from the total markers. GUIDE v.38.0 classification tree for predicting Y using estimated priors and unit misclassification costs. Tree constructed with 109 observations. Pruning parameter α was 0.02 for A and 0.03 for B. At each split, an observation goes to the left branch if and only if the condition is satisfied. Predicted classes and sample sizes printed below terminal nodes; class sample proportion for Y = Preterm beside nodes. In (A), V1 stands for Prevotella timonensis. In (B) V1 stands for Peptoniphilus lacrimalis.