JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):26
Poster Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263548

P-14. Artificial Intelligence for Oocyte Selection: A Multicenter Experience from Two Brazilian IVF Clinics

Ricardo Azambuja¹, Fabiana Mariani Wingert¹, Luiz Mauro Oliveira Gomes², Amanda Cerquearo Rodrigues dos Santos², José Fernando de Macedo², Carolina Comissoli Fernandes³, Sophia Abur Said³, Alvaro Petracco¹, Marta Ribeiro Hentschke¹, Mariangela Badalotti¹

¹Fertilitat - Porto Alegre - RS - Brasil
²Reproferty, São José dos Campos - SP - Brasil
³Pontificia Universidade Católica - Porto Alegre - RS - Brasil

Objective: To evaluate the clinical performance of an artificial intelligence tool (Magenta^®) for oocyte selection in IVF cycles across two assisted reproduction centers.
Methods: This is a multicenter, retrospective, case-control study conducted at two assisted reproduction centers in Brazil, with data collected from January to December 2024. The AI system used was Magenta^®, an AI tool that analyzes 2D images of mature oocytes to predict blastocyst formation for each oocyte, using normal semen parameters. Each oocyte image is scored on a scale of 0 to 10, and the blastulation rate (percentage of blastocysts formed) is given by groups divided as follows: G1) 0 to 2.5 (18.3%); G2) 2.6 to 5 (38.20%); G3) 5.1 to 7.5 (48.80%) and G4) 7.6 to 10 (56.80%). In the present study, the sample consisted of 4,395 oocyte images obtained from a time-lapse incubator (Embryoscope^®), and was divided into four groups according to the AI score. Semen parameters were not controlled, and cycles with varying sperm concentrations were included. Descriptive statistics were presented as absolute values and percentages. The two-proportion Z-test was applied to compare the percentage of blastocysts formed using eggs from the clinics versus AI, considering p<0.05.
Results: When comparing the blastulation rate from the clinics with the rate predicted by AI, the following results were observed: G1) 354/1.132 (31.2%) vs. 18.30%, p<0.001; G2) 495/1.000 (49.5%) vs. 38.20%, p<0.001; G3) 592/1.053 (53.3%) vs. 48.80%, p<0.001; and G4) 746/1.210 (61.6%) vs. 56.80%, p<0.001.
Conclusion: Clinical outcomes with the AI tool aligned with its original predictive model, reinforcing its clinical applicability. Observed blastulation rates were higher than AI estimates across all score groups, and followed the expected trend: lower rates in low-scoring oocytes and higher rates in high-scoring ones. Importantly, while the AI was trained on samples with normal semen parameters, our study included all profiles, suggesting that, had altered semen samples been excluded, the differences observed might have been even greater. Further research is needed to validate these findings across broader populations.