Placenta–pulmonary coupling–guided multimodal AI for fetal lung maturity staging and individualized glucocorticoid therapy

TL;DR

A new AI framework uses placenta-lung interactions and multimodal data to assess fetal lung maturity and guide glucocorticoid therapy, improving interpretability and clinical alignment.

Contribution

A physiology-informed hybrid AI model that integrates multimodal data and provides interpretable fetal lung maturity staging and individualized glucocorticoid dosing.

Findings

The AI model clustered fetal lung maturity into four stages with a silhouette score of 0.72, aligning with biochemical benchmarks.

The model predicted glucocorticoid doses within ±0.5 mg of clinical regimens and reduced projected RDS risk by 27%.

The framework maintained neurotoxicity index below thresholds while improving decision traceability and clinical alignment.

Abstract

Deep learning has improved medical image analysis but often produces opaque decisions and correlation-driven predictions that may diverge from clinical reasoning. We hypothesize that a physiology-informed hybrid framework, which explicitly models placenta–pulmonary interactions and integrates multimodal data, could provide interpretable and reliable guidance for assessing fetal lung maturity (FLM) and optimizing antenatal glucocorticoids (GCs).

In a prospective cohort study involving 320 pregnancies—including 160 with hypertensive disorders of pregnancy (HDP)—each with weekly acquisitions from 28 to 36 weeks, we combined 2D/3D ultrasound, shear-wave elastography, Doppler, and maternal plasma metabolomics. A biophysical placenta–pulmonary coupling model used the umbilical artery pulsatility index (PI) and a metabolomic hypoxia–steroid score to represent placental reserve, while backscatter integrals and elastography were used to characterize fetal lung properties. Constrained by this model, a dual-branch network was developed: (i) a cross-modal attention Transformer with self-supervised contrastive learning to generate unsupervised FLM stages from fused representations and (ii) a spatiotemporal convolution–LSTM network to predict individualized GC dosing and the optimal administration window. A composite loss penalized both projected respiratory distress syndrome (RDS) risk and the biomarker-derived neurotoxicity index.

The cross-modal representations clustered into four distinct maturity stages matching biochemical benchmarks, with an inter-stage silhouette score of 0.72. A downstream classifier achieved 92.3% accuracy in discriminating early from late maturity. The dosing branch predicted the GC dose within ±0.5 mg of clinically prescribed regimens and reduced projected RDS risk by 27% compared to standard dosing, while maintaining the biomarker-derived neurotoxicity index below the prespecified threshold.

A mechanism-guided, multimodal AI framework constrained by placenta–pulmonary physiology transforms imaging features into traceable decision pathways that align with clinical cognition. This interpretable framework may enable non-invasive FLM staging and individualized GC therapy, providing hypothesis-generating decision support that warrants external validation and prospective trials.

Figures