RCOG World Congress 2025

The Role of Artificial Intelligence in Predicting Pre-Term Birth: A Meta-Analysis

Miss Fathima Amanah Badurdeen (presenting) - f.badurdeen1@uni.bsms.ac.uk

BACKGROUND

•Defined as delivery before 37 weeks of gestation, Preterm birth (PTB) remains a leading cause of neonatal morbidity and mortality worldwide.

•Despite decades of research, the ability to accurately predict PTB remains limited, with current clinical tools offering modest sensitivity and specificity. The early identification of at-risk pregnancies is crucial. It's a matter of life and death for both mothers and infants, and it's a challenge we must urgently address.

•Artificial intelligence (AI) and machine learning (ML) offer promising avenues for enhancing PTB prediction. Their ability to process complex, multidimensional datasets—including clinical history, genomics, and biomarkers—with greater precision than traditional statistical models is a testament to their power and potential healthcare.

•However, existing research varies considerably in the AI models employed, data sources used, and outcome definitions applied. This data heterogeneity—including clinical, methodological, and statistical differences across studies—poses a significant challenge to synthesising the evidence.

•This meta-analysis aims to fill this gap by systematically reviewing and evaluating the performance of various AI models in predicting PTB.

METHOD

•Design: Systematic review and meta-analysis conducted according to PRISMA guidelines.

•Databases Searched: PubMed, Scopus, EBSCO (2015 – 2025).

•Search Strategy: Used Boolean operators to combine AI and ML terms with PTB and prediction-related keywords.

•Inclusion Criteria:

oStudies using AI/ML to predict PTB.

oReported performance metrics (e.g. accuracy, AUC)

oHuman subjects; English language

•Exclusion Criteria: Non-AI models, insufficient methodology, or non-peer-reviewed literature.

•Screening:

o724 records identified

o147 after filters applied

o75 full texts reviewed

o20 studies eligible

•Data Extraction:

oModel type, input features, sample size

oPrediction outcomes and performance metrics

•Analysis: Subgroup and sensitivity analyses to explore heterogeneity

Figure 1: PRISMA Diagram

RESULTS

1. AI Models and Performance:

•Neural CDEs (Controlled Differential Equations) achieved the highest performance (AUC: 0.82, Accuracy: 75%) by analysing vaginal microbiota dynamics in weeks 9–24 of gestation.

•LSTM (Long Short-Term Memory) models using routine clinical data (e.g., blood pressure, metabolic markers) showed moderate accuracy (73.9%) but lower sensitivity (40.7%).

•XGBoost consistently outperformed other models (AUC: 0.75–0.89) in studies leveraging biomarkers (e.g., ALP, AFP) and EHR data.

•RNNs with word embedding predicted extreme preterm birth (birth at <28 weeks gestation) with high sensitivity (96.5%) but low PPV (3.3%).

2. Key Predictive Features

•Microbiome: Low Lactobacillus and high Gardnerella Prevotella abundance has shown to be linked to PTB.

•Clinical/EHR Data:

oEarly predictors: Short cervical length, hypertensive disorders, and twin pregnancy.

oMetabolic markers: Triglycerides, glucose, uric acid.

•Biomarkers:

oLipid mediators (e.g., resolvin D1, LTB4) and inflammatory markers (AUC: 0.78–0.84).

oRoutine blood tests (ALP, AFP, hemoglobin) in women under 35 (AUC: 0.89).

•Electrohysterography (EHG): Phase synchronization in low-frequency bands (0.08–1.0 Hz) has shown to predict PTB (Accuracy: 90.5–97%).

3. Clinical Utility

•Early Prediction: Neural CDEs and EHG-based models enabled risk assessment as early as 9–24 weeks.

•Non-Invasive Tools: Vaginal swabs, EHG signals, and cfDNA analysis (e.g., PTerm classifier, AUC: 0.85) offered scalable alternatives to cervical length ultrasounds.

•Race-Specific Insights: Chronic stressors (e.g., low education, violence) drove PTB risk in Black women; biomedical factors (e.g., BMI) were key for White women (AUC: 0.75–0.76).

4. Limitations & Future Directions

•Challenges: Low PPV, overfitting in ML models, and lack of diverse validation cohorts.

•Opportunities:

oIntegrate multi-omics data (microbiome, biomarkers, imaging).

oDevelop hybrid models (e.g., neural CDEs + XGBoost) for improved generalizability.

CONCLUSIONS

•Artificial intelligence demonstrates transformative potential for PTB prediction by integrating multi-source data.

•AI models (e.g. Neural CDEs, XGBoost) demonstrate superior accuracy. They achieve AUCs of 0.75–0.97, outperforming traditional methods (e.g. fetal fibronectin, cervical length).

•Reliable risk stratification as early as weeks 9–24 using vaginal microbiota, biomarkers (ALP/AFP), or EHG signals, allows for early detection.

•Equity & Scalability:

oRace-specific models uncover disparities (e.g. chronic stressors)

oNon-invasive tools (e.g. vaginal swabs, cfDNA) enable low-resource deployment.

•Low PPV in sensitive models (e.g., RNNs: PPV 3.3%) risks false alarms and is a critical gap to be addressed.

•There is an urgent need for:

oHybrid models (e.g., Neural CDEs + XGBoost) integrating microbiome, biomarkers, and socio-demographics.

oClinical-AI partnerships to translate predictions into interventions (e.g., targeted care pathways).

"AI is not a replacement—but rather a multiplier—for obstetric care. Prioritising diverse data and equitable implementation will unlock its potential to predict PTB and reduce global disparities"

REFERENCES:

•This work is based on the authors’ own meta-analysis of 20 studies conducted in 2025.

•References of all included studies are available on request.

RCOG World Congress 2025

The Role of Artificial Intelligence in Predicting Preterm Birth: A Meta-Analysis

Poster Presenter

Authors

Affiliations

Part of Topic

Audio