O-005 Advancement of a novel image-analysis artificial intelligence (AI) model to predict blastocyst ploidy development of mature oocytes

Study question

Leveraging analysis by MAGENTA, can an additional AI model be developed to specifically predict the ploidy status (euploid/aneuploid) of individual mature oocytes in a cohort?

Summary answer

A non-invasive image-analysis Ploidy-AI model was developed to predict the likelihood of euploid blastocyst development of mature oocytes with an AUC of 0.71.

What is known already

MAGENTA is an AI-based model that assesses images of mature oocytes and provides a score (0-10) correlated to its likelihood of developing to a blastocyst-stage embryo, however research also shows correlation to blastocyst-ploidy (euploid/aneuploid) outcomes. Unlike sperm, the oocyte is responsible for most blastocyst-stage aneuploidies, as chromosome segregation during female meiosis is highly susceptible to errors, which increases with age. As such, a model trained to specifically predict the likelihood of a mature oocyte to develop into a euploid blastocyst, from a patient’s cohort of oocytes, could indicate greater clinical value regarding the quality and developmental potential of these oocytes.

Study design, size, duration

Retrospective study including 15,849 mature oocyte images (3746 patients, 4168 cycles) from 3 clinics (Canada, USA, Spain) obtained from EmbryoScope Time-Lapse incubators (Vitrolife), immediately post-ICSI. Oocyte images were acquired from cohorts with at least one blastocyst that underwent PGT-A. Oocytes that failed to develop into a blastocyst or became an aneuploid or euploid blastocyst were included and split into 60/20/20 for training, validation, and test subsets for model development. Mosaic or untested blastocysts were excluded.

Participants/materials, setting, methods

MAGENTA assessed 15,849 images providing a score and probability of blastocyst development. A Ploidy-AI model was then developed utilizing the images plus additional features (oocyte age, MAGENTA score and probability), to predict blastocyst-ploidy outcomes. The negative class were oocytes that failed blastulation/became an aneuploid blastocyst; positive class were those that became a euploid blastocyst. Once trained, the model’s probabilities were calibrated to adjust the prediction threshold from 0.50 to 0.28 –better reflecting true euploid development.

Main results and the role of chance

On the test set of 3,484 mature oocytes, the Ploidy-AI model predicted blastocyst-ploidy development outcomes with an AUC of 0.71, sensitivity 0.59, specificity 0.70.

Subgroup analysis by clinic revealed similar performances across locations; Clinic 1 (n = 1779)—AUC 0.67, sensitivity 0.72, specificity 0.52; Clinic 2 (n = 885)—AUC 0.70, sensitivity 0.60, specificity 0.66; and Clinic 3 (n = 820)—AUC 0.76, sensitivity 0.60, specificity 0.76. Comparing model performance between clinics displayed similar performance on Clinic 1 and Clinic 2 (p = 0.1267, DeLong’s test); however, significantly higher performance on Clinic 3 compared to Clinic 1 and Clinic 2 (p < 0.001, p < 0.05, respectively).

Oocytes that developed into euploid blastocysts (n = 1483) had significantly higher median model-predicted euploid probabilities than those that developed into aneuploid blastocysts or failed to develop into a blastocyst (n = 2001) by Mann-Whitney U-test (0.30 vs. 0.19, p < 0.001).

Model probabilities in the test set were divided into quartiles (Q1-4) according to model probability distribution. A significant, stepwise increase in true euploid development rate from the mature oocyte stage was observed between each quartile group; Q1 – 156/871 (18%), Q2 – 339/871 (39%), Q3 – 415/871 (48%), Q4 – 573/871 (66%). The proportion of true euploid development within all quartiles was significantly different by pairwise-proportions test with Bonferroni correction (all p < 0.01).

Limitations, reasons for caution

This dataset contains retrospective data from 3 clinics. Additional data from diverse clinic geographies would potentially add to model generalizability. The model displayed significantly higher performance on Clinic 3, which should be further investigated. A prospective validation study is needed to further confirm the clinical utility of the model.

Wider implications of the findings

Utilizing analysis from the existing MAGENTA model, a Ploidy AI-model was developed to specifically predict euploid blastocyst development from a mature oocyte. Robust non-invasive analysis of oocytes is achieved by these two models, providing valuable insights into oocyte quality, improving expectation management, and presenting an alternative assessment of genetic integrity.

Trial registration number

No