Performance of AI-based models in prenatal detection of trisomy 21 via ultrasonography: a systematic review.

Authors

Abstract

To systematically evaluate the performance, strengths, and limitations of artificial intelligence (AI) models for trisomy 21 classification and risk prediction, and for enabling ultrasound tasks, using fetal ultrasound. PubMed, Scopus, and Web of Science were searched from database inception through October 2026, supplemented by manual screening of reference lists. Original, peer-reviewed studies that developed or validated AI models for trisomy 21 classification or risk prediction, or for enabling ultrasound tasks (nuchal translucency and facial marker measurement, segmentation, and plane recognition), using human fetal ultrasound images or ultrasound-derived measurements and reporting quantitative performance metrics were eligible. Two reviewers independently selected studies, extracted data, and assessed risk of bias with PROBAST (prediction-model studies) and QUADAS-2 (enabling-task studies), resolving disagreements by consensus. Studies were categorized by primary task and synthesized narratively, without meta-analysis. Nineteen studies published between 2015 and 2026 were included: eight image-based classifiers, seven marker-based risk models, and four enabling-task studies. Image-based classifiers reported accuracies of 84-98.3% and AUCs of 0.95-0.985. Marker-based models reported AUCs of 0.84-0.983; detection rates of 90-100% at false-positive rates of about 3.6-3.9% came from one research group, whereas the largest cohort reported a sensitivity of 84% at a false-positive rate of 2.28%. Enabling-task studies were evaluated with task-specific metrics and were not compared with classification performance. Ten of fifteen prediction-model studies were at high risk of bias, mainly in the analysis domain; seven reported only image counts, and leakage between training and test data was likely in six. Most studies were retrospective with limited external validation. Reported performance of AI models for trisomy 21 is high but comes mostly from retrospective, internally validated studies, and benefits for screening performance, invasive testing, or operator dependence have not been demonstrated. No study compared AI with cell-free DNA screening, so the role of these tools within current screening pathways remains undefined. Prospective, multicenter validation with karyotype-confirmed outcomes and standardized reporting is needed before clinical integration can be supported.

Tags

Topics