O-004 Oocyte image assessment by novel artificial intelligence (AI) model predicts blastocyst PGT-A results, outperforming age as a predictor

Abstract

Abstract Study question Can a non-invasive, image analysis artificial intelligence (AI) model be built to predict the likelihood that a mature oocyte will develop into a euploid blastocyst? Summary answer A novel AI model was developed to assess images of mature oocytes, predicting chromosomal ploidy status of the developed blastocyst with an AUC of 0.71. What is known already The genetic status of blastocysts greatly impacts implantation and cycle success; therefore, PGT-A has become widely adopted as routine to aid in blastocyst selection for transfer and aims to decrease the time to pregnancy. Most chromosomal abnormalities arising in a blastocyst embryo are of maternal meiotic origin–indicating that the oocyte may have valuable information to be harnessed at the earliest stage in the process. Gaining insights into the potential genetic status of a blastocyst at the oocyte stage in a non-invasive manner would provide critical information to the overall fertility cycle. Study design, size, duration This study included a labeled dataset—9,260 images of mature denuded oocytes, immediately post-ICSI—from 3 clinics (Canada, United States (US), Spain) with euploid or aneuploid trophectoderm biopsy results (excluding mosaic/inconclusive embryos, n = 980). The labeled dataset was split into training (60%), validation (20%), and test (20%). An unlabeled dataset was also included—34,709 oocyte images (2 clinics; Canada, US)—and utilized in training. An external dataset (n = 473) from a separate US clinic was assessed for model generalization. Participants/materials, setting, methods Fully-supervised and semi-supervised machine learning models were experimented to predict blastocyst PGT-A from oocyte images. The labeled data were utilized in both methods. The semi-supervised method integrates unlabeled data using the well-known Mixmatch algorithm—learning important features from augmented images (with pseudo labels) that were then utilized in the supervised phase to predict outcomes. Age was incorporated into the models, therefore, a model only using age to predict ploidy was developed to act as a comparison. Main results and the role of chance Following various experiments, the fully-supervised model displayed an AUC 0.61, sensitivity 0.41, specificity 0.76, and accuracy 0.62 on the test set. The semi-supervised model displayed a performance of AUC 0.71, sensitivity 0.63, specificity 0.67, and accuracy 0.65–notably higher AUC and balanced performance on the positive/negative classes. The semi-supervised method benefits from both labeled and augmented unlabeled data for enhanced learning efficacy. The model utilizing age only to predict ploidy was extremely imbalanced in performance on the test set with AUC 0.63, sensitivity 1, specificity 0, and accuracy 0.48, confirming the importance of image features. Performance of the semi-supervised model was assessed by clinic. On Clinic 1 (n = 364; 40% euploid) it displayed AUC 0.61, sensitivity 0.60, specificity 0.56, and accuracy 0.58; on Clinic 2 (n = 675; 52% euploid) it displayed AUC 0.64, sensitivity 0.56, specificity 0.65, and accuracy of 0.60; on Clinic 3 (n = 466; 50% euploid) it displayed significantly better performance with AUC 0.85, sensitivity 0.76, specificity 0.79, and accuracy 0.77–warranting further investigation into model generalizability. On the external dataset (n = 473, 58% euploid), the semi-supervised model displayed improved performance (AUC 0.62, sensitivity 0.61, specificity 0.61, accuracy 0.60) in comparison to the age only model (AUC 0.59, sensitivity 1, specificity 0, accuracy 0.58). Limitations, reasons for caution This dataset includes oocytes that developed into blastocysts and underwent PGT-A. Datasets including oocytes that did not reach the blastocyst stage must also be assessed to validate the model’s clinical utility. Despite utilizing diverse data, advancing the generalizability of the model requires additional investigations to ensure balanced performance across clinics. Wider implications of the findings Utilizing images of mature oocytes, an AI model predicts the ploidy status of resulting blastocysts, providing a powerful tool for assessing oocyte quality and genetic integrity. These preliminary findings are promising for those undergoing oocyte cryopreservation by striving to elucidate genetic potential at the earliest phase of the IVF process. Trial registration number Not applicable.