Introduction

Preterm birth (PTB), defined as delivery of newborns before 37 weeks of gestation, is a leading cause of morbidity and mortality in newborns1. Established risk factors for PTB include genitourinary tract infections, short cervical length, and multiple pregnancies2. However, there is still disagreement regarding the magnitude of these factors’ effects on birth outcomes. Early identification of pregnant women at high risk of PTB can facilitate the implementation of strategies to prolong gestation and improve birth outcomes3.

Despite increased understanding of risk factors contributing to PTB, there remains a considerable lack of sensitivity in predictive models that can serve as a framework for intervention strategies4. Numerous attempts have been made to predict PTB using machine learning techniques combined with data from health records, inflammatory markers, and vaginal microbiome5. Fetal fibronectin is widely used clinically due to its affordability and simplicity. However, it has a low prediction rate with a sensitivity of only 56%6. Cervical length measurement also has limitations due to the hassle and inaccuracy of the procedure and the need for a skilled specialist7.

Approximately 70% of PTBs result from spontaneous onset of preterm labor and preterm pre-labor rupture of membranes due to intrauterine infection and inflammation8. However, the mechanism of PTB cannot be fully explained by inflammatory and infectious pathways as anti-inflammatory and antibiotic treatments could not reduce PTB rates9. With advancements in molecular genetic technology, studies on maternal microbiomes using 16S ribosomal RNA (rRNA) sequencing have emerged to explore unknown pathways of PTB10.

Microorganisms associated with PTB have been postulated to originate from one of two places: the reproductive or genitourinary tract ascending through the cervix or a hematogenous route11. Recent research has identified vaginal microbial signatures in women who later experience PTB and attempted to predict PTB using cervicovaginal fluid12. Although existing reports have verified a potential relationship of vaginal microbiome with PTB, they can only explain an ascending route.

Decades of epidemiological research studies have suggested that periodontitis is an independent risk factor for various adverse birth outcomes, including PTB13. Based on these precedents, it is expected that the oral microbiome can explain another hematogenous route. However, prenatal oral microbiome is not well understood14.

Thus, this study aimed to compare oral microbiome compositions between a PTB group and a full-term birth group, to identify oral microbiome associated with PTB, and to develop a machine learning prediction model of PTB based on oral microbiome compositions.

Methods

Study design and participants

This study was conducted on singleton pregnant women admitted for delivery at Jeonbuk National University Hospital between 2019 and 2021. This study received approval from the Ethical Research Committee (IRB file No. 2019-01-024). All participants provided written informed consent. Eligible participants included women admitted for induction delivery, elective cesarean section, and those who were hospitalized due to symptoms of preterm labor or preterm pre-labor rupture of membranes.

Data collection and grouping

Data on current and historical pregnancy outcomes were collected from questionnaires and electronic medical records. This information encompassed demographic factors (gestational age, birth weight, sex) and maternal risk factors (maternal age at delivery, cesarean section, preterm pre-labor rupture of membranes, previous preterm delivery history, gestational or overt diabetes mellitus, pregnancy-induced or chronic hypertension, and pre-pregnancy overweight or gestational weight gain). All subjects were divided into a PTB group or a full-term birth (FTB) group, with PTB defined as delivery before 37 weeks of gestation.

Oral microbiome sample collection

Oral microbiome samples were collected using mouthwash within 24 h before delivery. Standard sterile techniques were employed. Medical staff supervised all sample collection procedures. Participants were instructed to avoid brushing their teeth, eating, or drinking 30 min before sampling. Saliva samples were obtained by rinsing the mouth with 12 mL of a gargle solution (E-zen Gargle, JN Pharm, Pyeongtaek, Korea) for 30 s. Sample were labeled with the subject’s anonymous ID and stored at 4 °C until further processing. The resuspended sample was transferred to a microcentrifuge tube. Genomic DNA was extracted using an Exgene™ Clinic SV kit (GeneAll Biotechnology, Seoul, Korea) according to the manufacturer’s instructions and stored at − 20 °C.

16S rRNA gene sequencing and taxonomy assignment

Specimens were sent to Department of Biomedical Engineering, Ulsan National Institute of Science and Technology for taxonomy assignment. Then 16S rRNA sequencing was performed using an Illumina MiSeq Reagent Kit v3 (Illumina, San Diego, CA, USA) commissioned by Macrogen (Macrogen, Seoul, Korea). Library protocols for amplifying V3 and V4 regions were used. Pooled library was sequenced using a v3 600 cycle chemistry after diluting the sample to a final concentration of 6 pM with 20% PhiX control to generate 300 bp paired end reads.

Bioinformatics and statistical analysis

Independent t-test and chi-square test were used to compare differences between the PTB group and the FTB group. Statistical Package for Social Sciences version 20.0 was used for all data analyses. Statistical significance was considered at \(p<0.05\).

16S rRNA sequences from study subjects were imported with Quantitative Insights Into Microbial Ecology (version 2022.2) for further processing14. Sequences were filtered with DADA215. Amplicon sequence variants were assigned taxonomy with the Human Oral Microbiome Database (version 15.22)16.

To measure richness of microbiomes, two diversity indices were calculated. Alpha diversity, a measure of species within a particular community, was calculated using the Faith’s phylogenetic diversity index within the Quantitative Insights Into Microbial Ecology platform. Communities numerically dominated by a few species will exhibit a low Faith’s phylogenetic diversity index, whereas communities in which abundance is distributed equally among species will exhibit a high evenness. Mann–Whitney–Wilcoxon test was used to find statistically significant differences in Faith’s phylogenetic diversity index. Beta diversity measures dissimilarity between pairs of communities. It was calculated using the Hamming diversity index. PERMANOVA multivariate test was used to calculate statistically significant differences in the Hamming diversity index.

To identify differentially abundant taxa (DAT) with distinct abundance differences between the PTB group and the FTB group, DESeq2 was implemented using “DESeqDataSetFromMatrix” method as described in the package tutorial17. Taxa with \(\left|{\mathrm{log}}_{2}{\text{FoldChange}}\right|>1\) and \(\text{adjusted }{\text{p}}<0.05\) were considered as significantly different.

Machine learning prediction model development

Following qualitative and quantitative analyses of associations between the PTB group and each bacterium, a random forest classifier was used to find the criteria to predict PTB with oral microbiome data. Random forest classification is an ensemble machine learning algorithm that summarizes many decision trees to improve classification evaluations and robustness18. Random forest classifier was implemented to predict PTB based on oral microbiome compositions. To assure consistence and reliable classification results, we performed stratified \(k\)-fold cross-validation (\(k=5\)). Moreover, to decide the best features that could maximize classification evaluations, we performed random forest classification evaluations only with some DAT selected by their importances. Evaluations for classification included accuracy, balanced accuracy, precision, sensitivity, and specificity.

Ethics approval and consent to participate

The research protocol was approved by the Institutional Review Board of Jeonbuk National University Hospital (No. 2019-01-024) and was performed according to the Declaration of Helsinki.

Results

Study participant demographics

In this study, a total of 69 volunteer mothers were initially recruited. However, one participant with incomplete data and nine individuals with twin pregnancies were excluded from the study cohort. As a result, 59 women (30 in the PTB group and 29 in the FTB group) were included in the final analysis. Demographic and clinical characteristics of subjects in the PTB group and the FTB group are summarized in Table 1. Because prelabor rupture of membrane (PROM) is a major cause of preterm birth, it was significantly higher in PTB group. There was no significant difference in other maternal clinical characteristics between the PTB group and the FTB group. There were no cases with a history of smoking or concurrent periodontal disease in both groups.

Table 1 Baseline clinical characteristics of study subjects.
Full size table

Comparison of oral microbiomes

The oral microbiome comprised 13,953,804 sequences from 59 oral microbiome samples, with 102,305.95 ± 19,095.60 and 64,823.41 ± 15,841.65 reads per sample before and after filtering low-quality reads and trimming extra-long tails, respectively. After filtering low-quality reads and trimming extra-long tails, remaining representative reads were clustered in amplicon sequence variants with their exact sequence match. There were no significant differences in measures of alpha diversity (Faith’s phylogenetic diversity index) and beta diversity (Hamming distance) indices for samples between the PTB group and the FTB group (Supplementary Fig. 1).

Of 465 genera and species analyzed, 32 DAT between the PTB group and the FTB group were selected by DESeq2, including 26 FTB-enriched DAT and six PTB-enriched DAT. In order to mitigate the confounding effect of PROM, we excluded 7 PROM-related DATs from these 32 DAT (Supplementary Fig. 2). There were a total of 25 DATs between the PTB group and the FTB group, with 22 DATs enriched in the FTB group and 3 in the PTB group, as depicted in Fig. 1.

Figure 1
figure 1

DAT volcano plots. Volcano plot shows DAT, with PTB-enriched DAT shown as red dots and FTB-enriched DAT shown as green dots. Taxa with \(\left|{\mathrm{log}}_{2}{\text{FoldChange}}\right|>1\) and \(\text{adjusted p}<0.05\) are considered as significantly different. DAT, Differentially abundant taxa; PTB, Preterm birth; FTB: Full-term birth.

Full size image

Figure 2 displays DAT volcano plot in the oral microbiome sorted by differences between FTB-enriched DAT and PTB-enriched DAT proportions, indicating a decrease in gestational age for FTB-enriched DAT proportions. Pearson correlation analysis revealed a strong negative correlation (\(r=-0.542\)) between gestational age and difference between PTB-enriched DAT and FTB-enriched DAT proportions.

Figure 2
figure 2

Oral microbiome compositions over DAT. (a) Proportions of DAT of study subjects. Samples are sorted by difference between PTB-enriched DAT proportion and FTB-enriched DAT proportion. GA of samples are shown, matched with order of upper panel. PTB: red bar, FTB: green bar. PROM: arrow head. (b) Correlation plot with GA and difference between PTB-enriched DAT proportion and FTB-enriched DAT proportion. Pearson correlation shows strong negative coefficient (r = − 0.542) between GA and difference between PTB-enriched DAT proportion and FTB-enriched DAT proportion. DAT, Differentially abundant taxa; PTB, Preterm birth; FTB: Full-term birth; PROM: prelabor rupture of membrane; GA, Gestational age.

Full size image

Random forest classification

Random forest classifiers were established to classify PTB based on DAT. The best balanced accuracy (0.765 ± 0.071) was achieved using 9 most important taxa (Fig. 3a). The Random forest model calculated the importance of each DAT (Fig. 3b). Overall accuracy, precision, sensitivity, and specificity were 0.714 ± 0.061, 0.700 ± 0.194, 0.728 ± 0.058, and 0.743 ± 0.138, respectively. In order to validate the performance of our machine learning prediction model, we conducted a validation test on 9 twin pregnancies that were excluded from the paper (Supplementary Fig. 3). On PTB subjects of these twin samples, the machine learning classifications have 87.5% accuracy, comparable to the machine learning classification on the singleton study subjects.

Figure 3
figure 3

Machine learning evaluations over DAT. (a) Machine learning evaluation upon number of features (DAT). Random Forest classifier has the best balanced accuracy (0.765 ± 0.071) with the 9 most important DAT. (b) Importance of DAT is shown. Note that 0 ≤  importance  ≤ 1 and \(\sum {\text{importance}}=1\). The 20 most important DAT that give the best balanced accuracy is marked with blue. DAT, Differentially abundant taxa; ACC, Accuracy; BA, balanced accuracy; SEN, Sensitivity; SPE, Specificity; PRE, precision.

Full size image

Discussion

In this study, we developed a method for predicting PTB based on random forest classifier using oral microbiome compositions. Recently, several sporadic reports have suggested a bidirectional relationship between oral microbiome and pregnancy11. However, prenatal oral microbiome is not well understood yet. Some research has shown that oral microbial dysbiosis combined with gingival inflammation can lead to adverse pregnancy outcomes, including low birth weight, PTB, preeclampsia, and miscarriages19. Nevertheless, these results have been inconsistent due to methodologies employed in studies that only target known pathogens.

Fusobacterium nucleatum is the most prevalent oral microbiome studied. Fusobacterium nucleatum is a Gram-negative, anaerobic, filamentous oral microbiome. It is considered one of the most abundant species in the oral microbiome. It can also be isolated from vaginal microbiome20,21. Intra-amniotic Fusobacterium nucleatum infection leading to PTB has been reported in human and animal studies22. Other studies have shown that other oral pathogens including Porphyromonas gingivalis and intrauterine Bergeyella spp. can be isolated from the placenta of women who deliver prematurely23,24. In the present study, although Bergeyella spp. was overrepresented in the PTB group with PROM, it was excluded in the finally selected 25 DATs. Furthermore, Campylobacter gracilis was one of the FTB-enriched DAT that can aid colonization by periodontal pathogens including Porpyhromonas gingivalis in subgingival microbiome25. Lactobacillus gasseri was also one of the FTB-enriched DAT. It is known that Lactobacillus gasseri in vaginal microbiome can decrease early PTB risk21,26,27.

We found that decisive species differentiating between two groups were mainly abundant in the FTB group, with DAT consisting of 22 FTB-enriched DAT and 3 PTB-enriched DAT. We hypothesize that deficiency of species having a protective impact might have triggered the pathophysiology of PTB. Two different mechanisms have been proposed to explain the relationship between unhealthy microbiota composition and adverse pregnancy outcomes. The first mechanism proposed that periodontal bacteria originating in the gingival biofilm could translocate from the unhealthy oral cavity and cross the placenta, reach the intra-amniotic fluid and fetal circulation and directly affect the fetoplacental unit, resulting in bacteremia28. The second mechanisms proposed that systemic dissemination of endotoxins and/or inflammatory mediators derived from periodontal plaque and secreted by the subgingival inflammatory site could be carried to the fetoplacental unit29,30. Although certain microbiota has the same species, their subgroups can have both positive and negative influences on pregnancy outcomes. Following this line of thought, we believe that composition or dysbiosis of the oral microbiome is more important than the presence of specific microbiota.

It is worth nothing that microbial changes occurring during pregnancy might be nature consequences of a healthy pregnancy. Three reasons can explain the susceptibility to oral diseases such as periodontitis during pregnancy. These diseases are common in pregnant women due to hormonally driven hyper-reactivity of the gingiva to bacteria in the subgingival biofilm. Other factors that increase the risk of poor oral health during pregnancy include changes in dietary habits (frequent snacking or increased consumption of carbohydrate-rich foods), stomach acids from nausea and vomiting that contribute to the breakdown of tooth enamel, and a decreased likelihood of seeking dental care during pregnancy. We plan to implement pathway analyses to investigate direct link between the microbiome and PTB.

Even though there was limited power resulting from a small number of participants and restricted validation sample size, our study verified that oral microbiota might provide potential biomarkers for predicting pregnancy complications using machine learning methods including random forest classification. Additionally, the fact that the entire microbiome was not analyzed was a limitation of this study because our analysis only used relative values measured by 16S rRNA sequencing, not 16S metagenome sequencing. We did not investigate other factors could impact the oral microbiome, such as participant’s diet and socioeconomic status.

Despite these limitations, this prospective study demonstrated the potential of a PTB prediction model using oral microbiome in mouthwash. Further multi-center and larger-scale studies are needed to confirm out results before applying techniques developed in this study in the clinical field.