AI-based live birth prediction in IVF cycles: a systematic review without meta-analysis of model performance and validation

Abstract Background Artificial intelligence and machine learning have been identified as promising avenues for improving embryo selection and predicting the outcome of in vitro fertilization (IVF). However, there is much heterogeneity in terms of model architectures and validation methodologies used for these models, as well as their actual clinical utility in terms of the end outcome of achieving a live birth. Objective To evaluate the diagnostic performance and validation methodologies of artificial intelligence-based predictive models for the end outcome of achieving a live birth following in vitro fertilization cycles. Methods This systematic review without meta-analysis was carried out following the PRISMA 2020 guidelines, and the study was registered on the PROSPERO database (CRD420261298569). An extensive literature search was carried out on PubMed/MEDLINE, Scopus, EMBASE, Web of Science, and CENTRAL databases for original studies published between January 2010 and January 2026. Original studies on the development or validation of AI models for embryo evaluation or prediction of IVF outcomes, where live birth was the primary or secondary outcome, were considered for inclusion. The methodological study quality and risk of bias assessment were done using the QUADAS-2 tool, while the certainty of the evidence was assessed using the modified GRADE approach. Results A total of 23 primary clinical studies (20 retrospective cohort studies, 2 prospective cohort studies, and 1 randomized controlled trial) were included. In total, model evaluation on the validation cohorts comprised around 45,000 embryos/cycles, while the training datasets exceeded 200,000 embryos/cycles. Diagnostic accuracy was highly variable depending on model architecture. Multimodal models combining embryo image data and patient-level clinical variables demonstrated the highest model performance. These models achieved up to 0.97 Area Under the Curve (AUC) values and accuracy ranging from 74% to 82%. In contrast, Time-Lapse models achieved 0.64 to 0.97 AUC values (64% to 78% accuracy), while models using clinical variables alone achieved 0.70 to 0.80 AUC values (76% to 78% accuracy), and models using static images demonstrated the lowest performance (62% to 69% accuracy). Center-based models always outperformed national registry-based models. It is also interesting to note that the single prospective, double-blind randomized controlled trial evaluating an AI algorithm (iDAScore) against traditional morphology did not demonstrate noninferiority with clinical pregnancy rates of 46.5% vs. 48.2%, respectively. Conclusions In terms of the diagnostic potential of the AI-based predictive models for the prediction of live birth following IVF, these are highly effective when utilizing multimodal prediction approaches to assess embryonic competence and the systemic reproductive environment. The existing body of evidence is heavily based on retrospective cohorts and is subject to a number of limitations due to the presence of risks for selection bias and internal validation overfitting. As prospective RCT data have not been used to prove the independent clinical superiority or non-inferiority of AI-based prediction for IVF outcomes, currently, it should be used as a supplementary tool for embryologists only.