A comprehensive systematic review synthesizing clinical studies across international fertility research indicates that artificial intelligence algorithms predicting live birth following in vitro fertilization perform most effectively when combining embryo imaging data with patient-level clinical variables. Evaluating evidence from 23 primary clinical studies that spanned model evaluations across approximately 45,000 validation embryos or treatment cycles and training datasets exceeding 200,000 embryos or cycles, researchers determined that diagnostic accuracy varies substantially depending on model design. Across the analyzed literature, multimodal architectures that paired embryo visuals with clinical characteristics achieved area under the receiver operating characteristic curve values up to 0.97 and overall accuracies ranging between 74% and 82%, outperforming predictive approaches that relied on isolated data modalities.
The review investigated the persistent clinical bottleneck in reproductive medicine where success rates remain modest, with live birth rates generally ranging between 20% and 50% depending on maternal age and specific clinical factors. Standard laboratory methods in in vitro fertilization depend predominantly on visual microscopic evaluations conducted at specific developmental phases, most frequently blastocyst grading at day five or six using the Gardner and Schoolcraft classification system. This grading framework appraises blastocoele expansion along with distinct ratings for the inner cell mass and the trophectoderm. Although universally employed, microscopic morphology remains subjective and prone to substantial inter-observer and intra-observer variability even among veteran embryologists, with roughly 40% of morphologically normal embryos failing to implant after uterine transfer.
To rigorously quantify how algorithmic frameworks address these limitations, the systematic review gathered original development and validation studies indexed across PubMed, Scopus, EMBASE, Web of Science, and the Cochrane Central Register of Controlled Trials published between January 2010 and January 2026. The authors adhered to Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines and registered their methodology prospectively in the PROSPERO database under identification CRD420261298569. Eligible investigations evaluated artificial intelligence, deep learning, convolutional neural networks, or related predictive systems applied to patients undergoing in vitro fertilization or intracytoplasmic sperm injection where live birth served as a primary or secondary clinical outcome.
The synthesized evidence comprised 20 retrospective cohort studies representing 87% of the included literature, two prospective cohort studies accounting for 9%, and a single multicenter randomized controlled trial representing 4%. Geographically, 12 investigations originated in Asia, five in Europe, two in North America, one in Australia, and three involved international multicenter collaborations. Methodological study quality and potential risk of bias were scrutinized using the revised Quality Assessment of Diagnostic Accuracy Studies tool, while certainty was appraised through a modified Grading of Recommendations Assessment, Development and Evaluation framework. Because model designs, validation techniques, input modalities, and outcome reporting exhibited pronounced methodological heterogeneity, the researchers conducted a qualitative narrative synthesis rather than a quantitative pooled meta-analysis.
Model architectures separated into distinct diagnostic tiers based on input configurations. Systems trained exclusively on static microscopy photographs of blastocysts demonstrated the lowest predictive capability, generating diagnostic accuracies ranging from 62% to 69%. Time-lapse monitoring platforms incorporating morphokinetic analyses achieved accuracies between 64% and 78%, accompanied by area under the curve metrics spanning 0.64 to 0.97 across eleven clinical investigations. Within this time-lapse category, performance varied widely among specific models; for instance, a ResNet-like convolutional neural network achieved an area under the curve of 0.968 in a single-blastocyst cohort of 15,434 transfers, whereas the annotation-free iDAScore version 1.0 recorded an area under the curve of approximately 0.70 across 3,010 vitrified-warmed transfers, and dropped to 0.64 in preimplantation genetic testing cycles.
Machine learning models built solely upon preprocedural female clinical and demographic variables without any image data achieved intermediate diagnostic accuracies between 76% and 78%, alongside area under the curve values of 0.70 to 0.80 using gradient boosting, random forest, and support vector machines. Crucially, investigations comparing data sources revealed that center-specific models developed from individual clinic populations consistently outperformed algorithms trained on aggregated national registry databases. The highest diagnostic performance across all literature belonged to multimodal deep learning architectures that integrated both time-lapse embryo imaging and systemic patient features, exemplified by transformer-based networks reaching 81.5% accuracy by evaluating embryo competence alongside recipient clinical factors.
Biologically, the heightened performance of multimodal architectures reflects the dual physiological requirements governing reproductive success. Blastocyst morphology and morphokinetic parameters captured through continuous time-lapse photography serve as direct indicators of intrinsic embryonic developmental competence and underlying genetic quality, with deep neural networks frequently concentrating attention on the inner cell mass and trophectoderm. Conversely, clinical predictors including maternal age, ovarian reserve assessments, and endometrial factors reflect the receptive state of the systemic maternal environment into which the embryo is transferred. While clinical parameters help compute absolute population probabilities of live birth, the review emphasized that identical maternal features generate a theoretical within-cohort area under the curve of 0.5, demonstrating that maternal variables cannot help embryologists rank sibling embryos from the same patient.
The review highlighted a critical divergence between retrospective computational metrics and prospective clinical efficacy. In the only double-blinded, multicenter randomized controlled trial included in the synthesis, investigators randomized 1,066 patients to compare the iDAScore artificial intelligence algorithm against standard human morphological grading. That trial failed to establish noninferiority, recording clinical pregnancy rates of 46.5% in the algorithmic selection group compared with 48.2% in the traditional embryologist assessment group. This discrepancy underscores that highly optimistic discrimination metrics reported in retrospective cohorts often stem from selection bias, internal validation overfitting, or testing confined exclusively to embryos already selected for transfer by human practitioners.
Methodological quality assessments using the QUADAS-2 framework exposed persistent vulnerabilities across the underlying literature. Patient selection represented the primary source of bias, with 26% of studies categorized as high risk and 48% deemed unclear risk due to retrospective designs and selective cohort inclusion. Index test interpretation presented low risk in 61% of studies, while reference standard ascertainment had low risk in 78%. Most investigations relied entirely on internal validation routines such as k-fold cross-validation or isolated holdout test sets, with very few conducting external validation across independent, multicenter clinics. Consequently, the certainty of evidence for live birth prediction was rated as low under modified GRADE criteria due to serious risk of bias, inconsistency, and imprecision.
Because prospective randomized controlled trials have not demonstrated algorithmic superiority or noninferiority over conventional assessment, the review concluded that artificial intelligence cannot currently replace human embryologists. Instead, computational models offer immediate clinical utility as supplementary decision-support tools capable of standardizing morphological grading, reducing observer variability, expediting workflow efficiency, and acting as automated quality assurance alerts. Moving forward, the authors underscored that clinical translation requires prospective randomized trials powered for live birth endpoints, rigorous external validation across heterogeneous laboratory environments, model explainability techniques, and standardized protocol reporting conforming to the updated TRIPOD-AI statement.
The systematic search and screening sequence yielded specific study screening metrics that illustrate the broader state of artificial intelligence research in reproductive medicine. From an initial pool of 2,862 unique database records, the investigators retrieved 316 full-text publications for detailed eligibility appraisal, ultimately selecting 23 primary clinical studies that met the stringent inclusion criteria. Reviewers established high inter-rater agreement during this screening process, achieving Cohen’s kappa values of 0.84 during title and abstract screening and 0.79 during full-text review. Although the initial literature query spanned research published through January 2026, the included primary trials concentrated between 2019 and 2026, tracking the rapid transition from proof-of-concept modeling to clinical testing.
Study quality appraisals revealed additional methodological nuances across specific evaluation domains. In the flow and timing domain of the QUADAS-2 framework, 12 studies demonstrated low risk of bias, nine showed unclear risk, and two exhibited high risk. When examining comparative clinical performances among the included investigations, specific studies exposed distinct operational boundaries between algorithmic scoring and manual assessment. In one direct comparison encompassing 429 blastocysts, the manual morphokinetic scoring system KIDScore D5 achieved an area under the curve of 0.695 for live birth, whereas the algorithmic iDAScore system yielded an area under the curve of 0.657. In the largest single dataset analyzed, the iDAScore version 2.0 architecture attained a 77.5% internal validation accuracy across 181,428 embryos, contrasting with earlier multicenter models that achieved 64.3% accuracy predicting clinical pregnancy across 11 fertility clinics.
Addressing the technical gap between research claims and routine clinical deployment, the review highlighted several critical research priorities for future investigations. The authors stressed that algorithmic tools often operate as opaque computational black boxes, pointing out that very few existing studies incorporate explainability methods such as saliency maps, attention visualizations, or feature importance evaluations to clarify which specific biological structures drive algorithmic predictions. Furthermore, current literature lacks dedicated evaluation data for crucial clinical subpopulations, including patients with advanced maternal age, poor ovarian responders, and couples experiencing severe male factor infertility. To facilitate meaningful cross-study comparisons, the researchers recommended that future trials prospectively evaluate whether algorithmic scores reflect chromosomal euploidy or metabolomic profiles, while adhering to CONSORT-AI and TRIPOD-AI reporting standards.
Subject of Research: Artificial intelligence predictive models for live birth outcomes in IVF cycles
Article Title: AI-based live birth prediction in IVF cycles: a systematic review without meta-analysis of model performance and validation
Article References: Yadav, R., Banerjee, D., & K, J. R. (2026). AI-based live birth prediction in IVF cycles: a systematic review without meta-analysis of model performance and validation. Middle East Fertility Society Journal, 31(1), Article 44. https://doi.org/10.1186/s43043-026-00334-0
Image Credits: AI Generated
DOI: 10.1186/s43043-026-00334-0
Keywords: in vitro fertilization, artificial intelligence, machine learning, live birth prediction, embryo selection, deep learning, multimodal models, time-lapse imaging, blastocyst assessment, systematic review, clinical pregnancy, iDAScore, QUADAS-2, reproductive medicine, TRIPOD-AI, morphokinetics
Cite Scienmag News
Blake Davidson. (November 20, 2026). Systematic Review Finds Artificial Intelligence IVF Predictions Require Multimodal Data and Validation. Scienmag. https://scienmag.com/systematic-review-finds-artificial-intelligence-ivf-predictions-require-multimodal-data-and-validation/
Blake Davidson. "Systematic Review Finds Artificial Intelligence IVF Predictions Require Multimodal Data and Validation." Scienmag, 20 November 2026, https://scienmag.com/systematic-review-finds-artificial-intelligence-ivf-predictions-require-multimodal-data-and-validation/. Accessed 20 November 2026.
Blake Davidson. "Systematic Review Finds Artificial Intelligence IVF Predictions Require Multimodal Data and Validation." Scienmag. November 20, 2026. https://scienmag.com/systematic-review-finds-artificial-intelligence-ivf-predictions-require-multimodal-data-and-validation/








