JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):27
Poster Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263549

P-015. Artificial Intelligence in Embryo Quality Assessment: Potential for Optimizing Embryo Selection in In Vitro Fertilization

Júlia Pereira de Matos¹, Ana Beatriz Carvalho Rocha¹, Luiza Faria Vasques¹, Bruna Barbosa Coimbra Corlaite¹

¹Faculdade Ciências Médicas de Minas Gerais - Belo Horizonte - MG - Brasil

Objective: Evaluate the use of Artificial Intelligence (AI) models, particularly deep learning (DL) techniques, in the analysis of human embryo images in the context of in vitro fertilization (IVF), aiming to predict viability, optimize embryo selection for transfer, assess implantation potential, and ultimately reduce the number of treatment cycles required to achieve pregnancy.
Methods: A narrative review was conducted, analyzing six studies published between 2020 and 2025, comprising retrospective studies, randomized clinical trials, methodological reviews, and scoping reviews on AI applied to embryo assessment. The models incorporated convolutional neural networks trained with static images of blastocysts (Day 5) obtained through optical microscopy, as well as dynamic image sequences acquired via time-lapse incubator systems. Databases ranged from hundreds to more than 249,000 embryos, sourced from multiple fertility clinics across different countries. Performance metrics for the AI models were compared with conventional morphological grading performed by embryologists, focusing on metrics such as accuracy, sensitivity, specificity, and area under the curve (AUC).
Results: The Life Whisperer model achieved a sensitivity of 70.1%, specificity of 60.5%, and an overall accuracy of 64.3% for predicting clinical pregnancy, outperforming embryologists by up to 24.7% in binary viability classification. The iDAScore v1.0, trained with over 115,000 embryos, obtained an AUC of 0.67 for embryos with known implantation outcomes and maintained consistent performance across new clinics settings, with AUC values ranging from 0.60 to 0.75). While static images proved effective, time-lapse systems integrating morphokinetic data provided additional predictive gains for implantation and live birth outcomes (AUC up to 0.74), without requiring manual annotations of developmental milestones. Despite these promising metrics, multicenter randomized clinical trials did not demonstrate statistically significant difference between iDAScore-based selection and conventional morphology-based evaluation in terms of clinical pregnancy rates (46.5% vs. 48.2%). Reviews highlight the rapid expansion of AI applications in reproductive medicine over the past decade, with potential to standardize embryo selection, reduce subjectivity, and enhance IVF laboratory efficiency. Along with technical potential, ethical and operational considerations remain, including risks of depersonalizing patient care, exacerbating inequalities in access to advanced technologies, the lack of data standardization, and active involvement of healthcare professionals in model development and validation.
Conclusion: AI-based embryo analysis demonstrates consistent and promising performance for improving embryo selection in IVF, with gains in accuracy and reduced observer-dependent variability. DL models, particularly when trained on large datasets and integrated with time-lapse systems, demonstrate high predictive capacity and good generalization across diverse clinical environments. However, clinical outcome data remain heterogeneous, underscoring the need for well-designed, prospective and multicenter trials to validate the benefits of these tools in reducing time to pregnancy and increasing live birth rates. The safe, transparent, and ethical integration of AI mechanisms into reproductive clinical practice has the potential to accelerate and standardize, enhance decision-making, and ultimately serve as a valuable complementary resource in the future of assisted reproduction.