327 - Longitudinal Airway Microbiota Dynamics and Machine Learning-Based Prediction of Bronchopulmonary Dysplasia in Extremely Preterm Twins

Publication Number: 1311.327

mengmeng Ge, Children's Hospital, Shanghai, Shanghai, China (People's Republic)

Background: Bronchopulmonary dysplasia (BPD), a leading chronic lung disease in preterm infants, is closely associated with dysregulated airway microbiota. However, the temporal dynamics of microbial communities and their predictive value in BPD progression remain underexplored.

Objective: This study aimed to longitudinally characterize the airway microbiota dynamics in extremely preterm twins and evaluate the potential of microbial biomarkers for early prediction of bronchopulmonary dysplasia (BPD) using machine learning approaches.

Design/Methods: This prospective cohort study enrolled 10 twin pairs (20 infants) born at < 32 weeks of gestation age. Airway secretions were longitudinally collected at six timepoints for 16S rRNA sequencing and cytokine profiling. Microbial diversity, community structure, and clinical correlations were analyzed using bioinformatics and multivariate statistics. LASSO regression was applied to identify microbial biomarkers for BPD prediction.

Results: Severe BPD (sBPD) infants exhibited prolonged respiratory support and hospitalization (p < 0.001). By leveraging twin-pair stratification, microbial diversity significantly differed between concordant and discordant twin pairs (p < 0.05, PERMANOVA), with Staphylococcus dominating late-stage BPD. Longitudinal analysis revealed dynamic shifts: Ralstonia predominated in early stages (T1-T2), transitioning to Staphylococcus monodominance by T5-T6. Key taxa (Pseudomonas ASV3140, Cutibacterium ASV1143) correlated with IL-6 and IL-17A levels (p < 0.01). LASSO models achieved high predictive accuracy at T1, with the top6 model balancing complexity and performance in test set (AUC=0.87).

Conclusion(s): This study delineates stage-specific airway microbiota signatures in BPD and establishes machine learning models for early prediction. The findings underscore the potential of microbiome-driven diagnostics and targeted interventions in high-risk preterm neonates.

Flowchart of the study design.


Microbial diversity and immune responses at different stages of BPD in very preterm twins.
(A) Overview of clinical data for study participants in different BPD stages (T1-T6). (B) The saturation curve of the relationship between sample size and microbial community feature richness. (C) Expression levels of IL-6, IL-17a cytokines at different time stages after adjusting for the effects of group and birth weight. (D) Alpha diversity of airway microorganisms across stages in Concordant and Discordant twins. The vertical coordinate represents the Shannon index, with larger values indicating a higher response biodiversity. (E) NMDS analysis between the concordant and discordant groups. (F) The effect of clinical and environmental confounders (e.g., antibiotics, birth weight, sex, IL-6 levels) on differences in microbial community structure at different stages. (G) Beta-diversity variance (R²) of the airway microbiota across different clinical variables in very preterm twins with BPD. The y-axis represents the percentage of explained variance (R2) in beta-diversity, and the x-axis represents clinical variables. Asterisks (*) indicate clinical variables with significant differences in beta-diversity between the two groups (*p < 0.05, **p < 0.01).

LASSO regression and prediction model of airway microbiota in BPD. (A) Optimal parameter selection in Lasso regression analysis.
(A) Optimal parameter selection for the LASSO regression model. The top-left corner shows the cross-validation curve of binomial deviance for optimal 𝜆 selection, with green and red dashed lines marking 𝜆. min and 𝜆.1se, respectively. Eight ASVs were identified using the optimal 𝜆.min. The top-right figure displayed how regression coefficients of microbial variables change with 𝜆. The bottom-left figure showed regression coefficient trends under varying L1 Norms. The bottom-right figure revealed the relationship between deviance explained and regression coefficients: larger absolute coefficients mean greater deviance and better model fit. (B) A correlation plot of ASVs from the LASSO model, with circle size and color depth reflecting correlation strength, and asterisks denoting significance (p < 0.05). (C) The LASSO formula coef of ASVs, where red and green indicated positive and negative coefficients, respectively. (D) ROC curves for LASSO model evaluation. Three models were built using the top3, top6, and all ASVs based on regression coefficient magnitudes. Samples were split 6:4 into training and testing sets to compute AUC and 95% CI for ROC curves. (E) ROC curve assessments for samples at different stages.