Time-lapse imaging and deep learning have rapidly become widespread with the promise of reducing interobserver variability in embryo selection. So does this technology really increase the live birth rate? We consider the evidence honestly—both pros and cons.
⚕️ Target audience: This content is aimed at healthcare professionals and is for educational/current-information purposes. AI tools offer clinical decision support; It does not replace clinician responsibility and approved clinical workflows.
One of the most critical decisions in the in vitro fertilization (IVF) process is selecting the most competent embryo to transfer within a cohort. For decades, this selection has relied on morphological grading (e.g., the Gardner system for the blastocyst stage), in which embryos are evaluated at specific time points under a light microscope. However, morphological evaluation shows substantial variability both between embryologists and across readings by the same embryologist at different times; it also poorly predicts the chromosomal status of the embryo (euploidy/aneuploidy). Artificial intelligence (AI) entered the field with the promise of making this subjectivity objective and standardized.
This article honestly addresses the technical underpinnings of AI in embryo selection, the actual performance of commercial tools, and—most importantly—the contradictory picture revealed by randomized controlled trials (RCTs). Because this field is extremely open to excessive optimism fueled by impressive retrospective results.
The backbone of modern AI-based embryo assessment is time-lapse imaging (TLM — time-lapse microscopy that allows the embryo to be photographed at continuous intervals within the incubator). Unlike traditional assessment, TLM monitors the preimplantation development of the embryo without interruption, revealing morphokinetic parameters (dynamic features such as the exact timing of cell divisions) that are missed in discrete observations.
These image sequences are analyzed by convolutional neural networks (CNN — a layered deep learning architecture used in image recognition). Models follow one of two basic approaches: (1) Explainable feature extraction—automatically producing measurements that the embryologist can understand and verify, such as fertilization status, degree of fragmentation, symmetry; (2) “Black box” viability scoring—predicting implantation/live birth outcome directly from raw images. Although the second approach promises stronger predictive power, it compromises interpretability and generalizability.
💡 Important distinction: Consistent classification of morphological quality by an AI model and predicting clinical outcome (pregnancy, live birth) are entirely different levels of difficulty. A high morphology-classification accuracy does not reflect actual pregnancy prediction performance.
This distinction is best exemplified by the STORK study. Khosravi and colleagues (npj Digital Medicine, 2019) trained a deep neural network called STORK, based on Google's Inception model, with approximately 50,000 embryo time-lapse images. The model graded blastocyst quality with AUC >0.98 (area under the curve—1.0 indicating perfect discrimination) and outperformed individual embryologists. However, this metric belongs to predicting the morphological quality category of the embryo — not pregnancy outcome. Indeed, the authors had to build a separate decision tree combining embryo quality with patient age to capture clinical significance (pregnancy chance varied between 13.8% and 66.3% depending on age and quality). STORK was an early warning of why caution should be exercised in interpreting the field's evidence.
The most common commercial product in the field is Vitrolife's fully automatic and "annotation-free" deep learning-based embryo scoring module called iDAScore (intelligent Data Analysis Score). iDAScore received CE marking under the European Medical Device Directive in early 2020 and ranks embryos according to their probability of implantation, 1.0–9.9, without manual input from the embryologist.
The development and validation of iDAScore relies on large data sets. Berntsen and colleagues (PLoS One, 2022) developed the model with 115,832 embryos from 18 IVF centers and tested external-center generalizability; AUC for embryos with known implantation results was 0.67, and for new clinics it was in the range of 0.60–0.75. The updated v2.0 version was trained with 181,428 embryos from 22 clinics (Lassen et al., Scientific Reports, 2023) and reported AUC 0.62–0.71 by day of transfer. In retrospective single-center cohorts, higher scores on iDAScore were associated with increased live births (adjusted OR 1.81; 95% CI 1.67–1.98) and reduced miscarriage rate ( Ueno et al., 2022 ).
These results are consistent but modest: pregnancy prediction AUCs are typically in the 0.66–0.74 band—that is, at most on par with or slightly better than, but not dramatically superior to, manual assessment by embryologists. The main advantage of iDAScore appears to be not the leap in accuracy, but objectivity, repeatability and speed: the time to evaluate an embryo is reduced almost tenfold, from approximately 208 seconds to 21 seconds.
⚠️ Beware of over-optimism: Impressive AUC values in retrospective studies do not imply prospective clinical benefit. RCT evidence in this area has not confirmed the retrospective promise.
The strongest evidence measuring the true clinical value of AI in embryo selection is the multi-center, randomized, double-blind, non-inferiority (comparable efficacy) study published in 2024 (Illingworth et al., Nature Medicine, 2024). At 14 IVF clinics in Australia and Europe, 1,066 women under 42 years of age with at least two early-stage blastocysts on day 5 were randomized to either standard morphological assessment or selection by iDAScore (v1).
The result was striking: the clinical pregnancy rate was 46.5% (248 of 533) in the iDAScore arm and 48.2% (257 of 533) in the morphology arm (risk difference −1.7%; 95% CI −7.7 to +4.3; p=0.62). Given the predefined 5% non-inferiority limit, the study did NOT show that deep learning was non-inferior to standard morphology. So, not only could AI not surpass morphology in clinical pregnancy rate, but it also could not prove to be its equal.
This finding caused such a stir in the field that an accompanying commentary in the journal Human Reproduction (Sakkas, 2025) summarized the study with the headline: “The golden fleece of embryology is slipping away once again.” The author emphasizes that good blastocyst morphology has again proven itself as “an insurmountably high bar” in predicting live birth—but adds that research should not be abandoned to find the viability characteristics that morphology cannot capture.
This table is consistent with the broader literature. Pooled meta-analyses evaluating time-lapse imaging systems have consistently reported “uncertainty of benefit”; The individual-participant data (IPD) meta-analysis protocol (Bhide et al., BMJ Open, 2025) was designed to address precisely this uncertainty—despite the clinics' significant investment of resources.
It is also illuminating to what extent AI and embryologist decisions overlap. In a small retrospective cohort (Harir et al., 2025; 82 IVF cycles) embryologist and iDAScore choices agreed in only 64.6% of cases; Moreover, there was no difference in pregnancy rate between compatible and incompatible choices (45.2% vs. 44.8%; p>0.05). This suggests that AI is “closely but not identically” aligned with human judgment and should be positioned as a decision support tool—not a substitute—at the current level of evidence.
The most ambitious — and most controversial — promise around AI is that it can noninvasively predict the euploidy state (chromosomal normality) of the embryo from images. This would be an inexpensive alternative to preimplantation genetic testing for aneuploidies (PGT-A—the costly, invasive, and ethically controversial gold standard that requires trophectoderm biopsy).
However, the evidence does not support this claim. A systematic review and meta-analysis (Xin et al., EClinicalMedicine, 2024) examining twenty studies (12 in meta-analysis, totaling 6,879 embryos) found the pooled sensitivity of AI for euploidy prediction to be 0.71, specificity to be 0.75, and AUC to be 0.80 (95% CI 0.76–0.83). Performance is even lower when looking at a single commercial score: iDAScore's euploidy prediction AUC alone is only 0.60–0.61, only rising to ~0.69 when clinical and embryonic variables are added ( Cimadomo et al., 2023 ; Ma et al., 2024 ). Although this performance provides an auxiliary signal for sorting embryos, it is far from replacing NGS (next-generation sequencing)-based genetic diagnosis. The authors of the meta-analysis clearly conclude that: AI models cannot completely replace invasive methods in determining ploidy; They can be a helpful tool for patients who cannot undergo PGT-A.
In light of the available evidence, a balanced position can be summarized as follows:
In summary, AI in embryo selection is a real addition to the IVF laboratory's standardization and efficiency toolkit; but the “golden fleece of embryology”—the ability to select the most competent embryo to transfer, significantly better than morphology—has yet to be achieved. Ongoing IPD meta-analyses and new RCTs will clarify which patient subgroups, if any, benefit from this technology in the coming years.
Last updated: . Sources have been confirmed with their publication dates and current versions.