The role of artificial intelligence in predicting late-onset intrauterine growth restriction in singleton pregnancies: a systematic review and meta-analysis

Purpose: Late-onset intrauterine growth restriction (IUGR) is difficult to detect and is associated with adverse perinatal outcomes.This systematic review and meta-analysis evaluated the performance of artificial intelligence (AI) models, including machine learning (ML) and deep learning, for predicting late-onset IUGR in singleton pregnancies.Methods: PubMed, Scopus, and Web of Science were searched for studies published during the past decade that used AI to predict IUGR.Nineteen studies including 13,379 pregnancies were analyzed.Sensitivity, specificity, likelihood ratios, and diagnostic odds ratios were pooled using DerSimonian-Laird random-effects models; the area under the summary receiver operating characteristic curve (AUC) was estimated in Stata.Results: In the primary analysis (one model per study, k=17), pooled sensitivity was 0.83 (95% confidence interval [CI], 0.74 to 0.89), specificity was 0.89 (95% CI, 0.85 to 0.91), the positive and negative likelihood ratios were 6.6 (95% CI, 5.3 to 8.1) and 0.22 (95% CI, 0.16 to 0.30), and the AUC was 0.93 (95% CI, 0.91 to 0.95).Estimates were unchanged when all model entries were included (k=51).No threshold effect or funnel-plot asymmetry was detected.Subgroup estimates for support vector machines, random forests, and logistic regression were similar but reflected indirect comparisons.Conclusion: AI/ML models showed high pooled discriminative accuracy for late-onset IUGR.However, the evidence derives mainly from retrospective, single-center case-control studies without external validation; calibration was reported in six of 19 studies and decision-curve analysis in none.These estimates are an optimistic upper bound, not evidence of superiority over current ultrasound practice.