Long-term Cardiac Autonomic Effects of Prenatal Steroid Exposure: A Machine Learning Approach Integrating Heart Rate Variability and ECG Foundation Models

Background: Prenatal glucocorticoid administration is standard care for threatened preterm birth, but long-term cardiac autonomic effects remain incompletely understood. We investigated whether children exposed to prenatal steroids exhibit persistent differences in cardiac autonomic function at age 8 years using comprehensive heart rate variability (HRV) analysis and deep learning-based ECG foundation models. Methods: We analyzed Holter ECG recordings performed in a well-controlled laboratory environment from 49 children (24 prenatal steroid-exposed, 25 controls) at age 8 years in this exploratory study. Children were exposed to prenatal glucocorticoids in the context of maternal multiple sclerosis treatment. We employed ensemble R-peak detection, computed 112 HRV metrics with PCA dimensionality reduction (7 components, 88% variance), and extracted 512-dimensional embeddings using a pre-trained ECG foundation model. Statistical analysis used linear mixed-effects models (LMM) with Bonferroni correction and covariate adjustment for sex and gestational age to assess confounding. Results: Before covariate adjustment, 3/7 HRV principal components and 334/1024 FM dimensions showed significant group differences (FDR-corrected). After covariate adjustment, traditional HRV findings lost significance (0/3 HRV PCs remained significant, p>0.13), while 11 foundation model dimensions remained robust (adjusted p<0.05, |Cohen's d|>0.8), suggesting confounding of HRV by sex/gestational age but biologically robust FM differences. The study was severely underpowered for small effects (10-17% power, n=24/group); detected large effects (d[≥]0.8) likely reflect genuine biological differences requiring validation. Conclusions: Deep learning-based ECG foundation models detect robust cardiac effects of prenatal steroid exposure independent of demographic confounders, while traditional HRV metrics show confounded group differences. This exploratory study demonstrates proof-of-concept for transfer learning in pediatric cardiology and underscores the critical importance of covariate adjustment in small observational studies. Independent replication in larger cohorts (n>175/group) is essential before clinical translation.