P-137 Validation of a novel artificial intelligence (AI) model assessing retrospective oocyte images to predict blastocyst PGT-A outcomes

Abstract

Study question Can a non-invasive image analysis AI (Ploidy) model predict the likelihood that a mature oocyte will develop into a euploid blastocyst from previously unseen data? Summary answer The Ploidy AI model assesses oocyte images from an external Spanish dataset, predicting chromosomal ploidy status of the resulting blastocyst with an AUC of 0.68. What is known already Embryo chromosomal integrity is crucial to implantation and IVF success. PGT-A testing has been adopted to assess the genetic status of embryos and aid in blastocyst selection for transfer. However, high costs, invasiveness, and technical challenges introduce barriers for patient access. Most embryonic chromosomal abnormalities originate from maternal meiotic errors, making oocyte assessment a crucial opportunity to gain early genetic insights. However, non-invasive methods to evaluate the impact of oocyte quality on embryonic genetic integrity remain unavailable. Here, we validate the performance of a non-invasive, AI-powered oocyte assessment tool in predicting blastocyst ploidy (euploid/aneuploid) outcomes from the mature oocyte stage. Study design, size, duration Images of 925 mature oocytes (153 patients, 30-48 years old) undergoing IVF-ICSI at a Spanish clinic using GERI time-lapse incubators were retrospectively analyzed by the Ploidy AI model to predict the probability (0-100%) of each oocyte developing into a euploid blastocyst. Within the dataset, 418 oocytes did not develop into a blastocyst, whereas 507 oocytes developed into a blastocyst (235 aneuploid, 149 euploid, 26 mosaics, 97 untested). Mosaic and untested blastocysts were excluded. Participants/materials, setting, methods The euploidy-prediction probabilities generated by the Ploidy AI model for each oocyte were analyzed for model performance in predicting true outcomes by Area-under-the-curve (AUC), positive predictive value (PPV), and negative predictive value (NPV). Correlations of the probabilities with key clinical parameters, including PGT-A outcomes of the resulting blastocysts, blastocyst morphology, and patient age, were also assessed by Welch’s t-test, One-Way Analysis of Variance (ANOVA) with post-hoc pairwise comparisons, or Two Proportions z-test. Main results and the role of chance All oocyte cohorts included had at least one oocyte that developed into an aneuploid or euploid blastocyst. The oocytes that did not develop into blastocysts(n = 418) or those that became aneuploid blastocysts(n = 235) were considered as a negative outcome, whereas those that became euploid blastocysts were the positive outcome(n = 149) when assessing the performance of the model. The Ploidy AI model displayed AUC=0.68, sensitivity=0.46, specificity=0.76, PPV=0.3, and NPV=0.86. Oocytes that developed into euploid blastocysts had significantly higher mean euploidy-predicted-probabilities than those that developed into aneuploid blastocysts (25% vs. 20%; p < 0.0001). This significant difference persisted for patients ≥35 (euploid (n = 126): 23% vs. aneuploid (n = 217):19%, p < 0.0001), but not for those <35 in subgroup analysis (euploid (n = 23): 34% vs. aneuploid (n = 18): 34%, p > 0.05). Subgroup analysis for blastocyst quality revealed significantly higher euploidy-predicted-probabilities among euploid blastocysts compared to aneuploid in both high- (Gardner expansion 4-6, ICM/TE=A/B) and low-quality (Gardner expansion 1-6, ICM/TE=C/D) groups: High-quality euploid (n = 85) 25% vs. aneuploid (n = 94) 20%,p<0.01; Low-quality euploid (n = 64) 25% vs. aneuploid (n = 137) 20%,p<0.01). Lastly, when divided into equal-sized quartiles (Q) based on model-predicted probabilities, a stepwise increase in the proportion of PGT-A-tested euploid blastocysts was observed (∼200 oocytes/quartile; Q1:4.5%, Q2:17%, Q3:21%, Q4:32%) with significant difference between Q1/Q2(p < 0.001) and Q3/Q4(p < 0.05). Limitations, reasons for caution The retrospective nature of this study prevents definitive determination of aneuploidy origin in PGT-A-tested blastocysts. While the model may indicate meiotic origin, prospective studies are needed to distinguish between meiotic and mitotic aneuploidy origins. Additionally, subgroup analyses were limited by small sample sizes for <35, and medium-quality blastocysts (n = 3, excluded). Wider implications of the findings The novel AI Ploidy model predicts the ploidy status (euploid/aneuploid) of blastocysts from the oocyte stage. The model’s predictions align with true PGT-A outcomes, even among older patients and different-quality blastocysts, providing a versatile, non-invasive, cost-effective method to assess blastocyst genetic potential from the earliest stage possible–the oocyte. Trial registration number No