Artificial Intelligence in the Analysis of the Fetal Genome in Utero: A Critical Review of Current Paradigms, Clinical Utility and Future Horizons

Abstract

Prenatal genomic medicine now generates data at a scale that exceeds what unaided human interpretation can process within the time available to a continuing pregnancy. Cell-free DNA screening, chromosomal microarray, exome sequencing and, increasingly, genome sequencing produce millions of candidate observations per case, and computational methods described as artificial intelligence have been proposed at almost every step of the resulting pipeline. This critical narrative review evaluates the strength, consistency and clinical relevance of the evidence supporting those methods when they are applied specifically to the genome of a fetus in utero. Literature was identified through structured searching of biomedical and multidisciplinary scholarly sources from 1997 to 20 June 2026, prioritised by methodological quality and by direct relevance to prenatal application rather than by citation count. The evidence divides sharply. Analytical performance for tasks that are essentially signal-detection problems, including trisomy classification from low-coverage plasma sequencing, fetal fraction estimation and small-variant calling, is supported by internally consistent and technically credible work. Evidence for tasks that require inference about meaning, including missense and splice-effect prediction, phenotype-driven gene prioritisation and automated candidate diagnosis, is substantially weaker in the prenatal context, because the models were trained and calibrated on postnatal cohorts whose phenotypes were fully observable. Fetal phenotypes evolve during gestation, are captured through imaging rather than examination, and map imperfectly onto the ontologies that phenotype-driven algorithms consume. Several studies show that ontology-driven gene selection in fetuses misses a substantial minority of causative genes that curated prenatal panels retain. Almost no prospective study has tested whether an artificial intelligence component changes a prenatal decision, and reporting quality across the field falls short of contemporary prediction-model standards. Structural inequities in reference genomic resources compound these problems. Priorities include prenatally calibrated variant-effect evidence, gestational-age-aware phenotype representation, prospective evaluation with decision-relevant endpoints, and governance that treats uncertainty as a reportable output rather than a defect to be smoothed away.