Artificial Intelligence for Detecting Fetal Genetic Variants Across the Prenatal Testing Pathway: A Critical Narrative Review

Abstract

Background: Prenatal genetic testing now spans screening of cell-free DNA in maternal plasma, sonographic phenotyping, cytogenetic analysis and exome or genome sequencing of fetal samples. Artificial intelligence, including machine learning and deep learning, has been applied at each point, yet the literature is fragmented across laboratory genomics, imaging science and clinical genetics, and its evidential maturity has rarely been compared across domains. Purpose and scope: This critical narrative review examines how artificial intelligence has been used to detect or interpret fetal aneuploidy, copy number variants, single-nucleotide variants and syndromic phenotypes, and evaluates the strength, validity and limitations of that evidence. Embryo selection in assisted reproduction and prediction of non-genetic obstetric outcomes were excluded. Approach: Literature from January 1995 to July 2026 was identified through bibliographic databases, a trial registry and citation searching, and was appraised against reporting and validation principles for diagnostic artificial intelligence. Principal findings: Evidence is strongest for supportive analytical tasks, notably fetal DNA fraction estimation, automated chromosome classification and prioritisation of candidate variants, where algorithms operate under human oversight against measurable reference standards. Machine-learning refinement of aneuploidy calls from cell-free DNA has improved positive predictive value in a large retrospective series, but the incremental benefit over established screening is bounded by placental mosaicism and maternal confounders rather than by computation. Genome-wide non-invasive genotyping of monogenic variants has been demonstrated only in small family cohorts. Image-based prediction of chromosomal or syndromic disease from ultrasound remains developmental, relying on case-control sampling, single-centre data and scarce external validation. Prenatal sequencing interpretation is constrained by incomplete fetal phenotypes, ancestry bias in reference data and unverified large language model outputs. Unresolved questions: Prospective, multicentre evaluations with appropriate comparators, fetal rather than placental reference standards and assessment of counselling and decision outcomes are largely absent. Conclusion: Artificial intelligence is best regarded as an adjunct that improves efficiency and consistency within existing testing pathways rather than as an autonomous detector of fetal genetic disease. Safe expansion depends on transparent reporting, equitable datasets, calibrated uncertainty and governance that preserves informed reproductive choice.