Integrating Ai-Powered Multiomics for Personalized Prediction and Management of Pregnancy Complications In 2025

Abstract

Background: Early and reliable prediction of preterm birth remains challenging. Prior work often relies on single-modality signals (e.g., EHG or clinical variables) or a narrow subset of ‘omics.

Objective: To evaluate whether integrating multi-omics (genomics, transcriptomics, proteomics, metabolomics, microbiome) with routine clinical variables and electrohysterography (EHG) improves prediction of preterm birth compared with single-modality baselines.

Methods: We conducted a multicenter study across [number] hospitals with [N] pregnancies. After standardized QC and normalization, modality-specific encoders fed an attention-based fusion layer and a stacked meta-learner. The primary endpoint was preterm birth (<37 weeks). Internal performance used stratified cross-validation with calibration assessment; external validation used a held-out site. We compared against EHG-only, clinical-only, and the best single-omics models.

Results: The integrated model outperformed all single-modality baselines in both internal and external evaluations. In our cohort, the fused model achieved an AUC of 0.91 internally and 0.89 on an external site (with improved calibration and net benefit on decision-curve analysis). Step-up ablations indicated additive value from proteomics/metabolomics and microbiome features beyond clinical and EHG inputs. Model explanations highlighted biologically plausible pathways and inflammatory signatures associated with risk.

Conclusions: Integrating diverse ‘omics with clinical and EHG data yields materially better discrimination and clinical utility than single-modality models. While related multimodal studies exist, to our knowledge this work is among the first to unify full multi-omics, clinical data, and EHG within a single, validated pipeline. Prospective, cost-aware implementation studies are warranted to establish impact on care.