Table 6.

Embryo assessment studies using artificial intelligence

Study	ART process	Outcomes of interest	Dataset	AI methods	Results
Khosravi et al. (2019)85	Prediction of blastocyst quality (poor vs. good).	-Classification of blastocyst quality at 110 hrs. post insemination.	Retrospective dataset consisting of 12,001 time-lapse images at 110h post insemination.	Deep learning—CNN	-Development of AI model (STORK) to predict blastocyst quality. -Predicted blastocyst quality with AUC above 0.98. -AUC of 0.90 and 0.76 achieved on validation with two external datasets.
Dimitriadis et al. (2019)81	Determination of normal fertilization (2PN vs. non-2PN embryos).	-Categorization of embryos based on fertilization outcomes.	Retrospective dataset of 3469 embryos (2893 2PN; 576 non-2PN).	Deep learning—CNN	-AUC of 0.90, with PPV of 96.2% and NPV of 78.1%. -Trained CNN capable of automated fertilization check with high accuracy.
Fukunaga et al. (2020)82	Pronuclei determination.	-Categorization of oocytes based on pronuclei status.	Retrospective dataset of 900 embryos (300 each 0PN, 1PN, and 2PN).	Deep learning—CNN	-Precision of machine learning equivalent to that of expert embryologist. -Sensitivity for detection of 0PN, 1PN, and 2PN: 99%, 82%, and 99%, respectively.
Coticchio et al. (2021)83	Cytoplasmic movement to predict blastocyst development.	-Deep learning methods based on cytoplasmic movements at early cleavage stage to predict development to blastocyst.	Retrospective analysis of 230 embryo time-lapse sequential images.	Deep learning ANN extended by k-NN.	-Combination of blind operator assessment and deep learning models led to prediction accuracy of 82.6%, 79.4% sensitivity and 85.7% specificity. -Highlights importance of cytoplasm dynamics as novel source of data.
Zhao et al. (2021)84	Labeling of segmented day-1 embryos.	-CNN labeling of zona pellucida, cytoplasm, and pronuclei performance compared with manual labeling by a clinical embryologist.	1218 images from 24 day-one embryos of 14 subjects.	Deep learning—CNN	-Good precision in measurement of cytoplasm, pronuclei, and zona pellucida (97%, 84%, and 80% accuracy respectively) and comparable with morphometrics reported in literature. -Rapid labeling of all images: 130 hrs. for manual labeling against 12.18 s for CNN.
Thirumalaraju et al. (2021)86	Blastocyst classification based on morphological data.	-Classifying blastocysts based on morphological data in eight different neural network architectures.	742 embryo images used for validation.	Deep learning—CNN	-XCeption CNN architecture correctly classified > 99.5% of the highest quality blastocysts as good embryos. -Accuracy of Xception model in categorizing blastocyst and non-blastocyst was 90.9%.
Berntsen et al. (2022)87	Embryo selection for transfer.	-Prediction of implantation outcome with fully automated deep learning tool.	115,832 embryo time-lapse sequences (validation set of 17,249 embryos, 2212 with known outcomes).	Deep learning—CNN (iDAScore v1).	-AUC of 0.95 in predicting implantation when all embryos are considered together (including 1510 embryos labeled as discarded due to manual deselection by embryologist or aneuploidy). -Inclusion of discarded embryos in model training aids deep learning.
Hickman et al. (2022)95*	Embryo selection for transfer.	-CHLOE EQ™ score based on embryo bioinformatics and relation to expert embryologist grading, implantation, and live birth.	799 day-5 embryo time-lapse videos	Not disclosed	-CHLOE EQ™ score was directly related to embryologist ranking of morphology. -CHLOE EQ™ score differentiated between embryos that implanted and those that did not. -Strong correlation between human and AI-determined morphokinetic labeling. -Was not predictive of live birth.
Diakiw et al. (2023)89	Embryo selection for transfer.	-AI model using deep CNN and Grad-CAM++ mapping.	-9359 day-5 blastocyst images from 4709 women who underwent IVF.	Deep learning—CNN	-Heat maps generated for regions relating to viable and nonviable embryo classification and AI score generated. -Positive linear correlation of AI scores with pregnancy outcomes were found, leading to 12.2% reduction in time to pregnancy in comparison with standard morphological grading methods. -AI scores significantly correlated with Gardner morphological score and associated with embryo ploidy status.
Meseguer Escriva et al. (2022)99*	Aneuploidy assessment	-AI model using 5 feature extraction models to predict ploidy status (abnormal morphokinetic patterns, an embryo grading classification algorithm, differential cell division activity, mitochondrial DNA content, and quantification of blastocoelic contractions).	Retrospective dataset of 2502 embryo time lapse sequences with known ploidy status.	Deep learning—CNN	-Integration of all 5 features led to 90% accuracy in prediction of ploidy status. -Non-invasive AI-guided PGT triage could be a useful adjunct to conventional embryo selection or recommendation for PGT.
Barnes et al. (2023)100	Aneuploidy assessment	-Prediction of ploidy status based on static images, morphokinetic parameters, morphological assessments, and maternal age.	Retrospective dataset of 10,378 annotated blastocysts from 1385 patients with known ploidy status.	Deep learning—CNN	-‘STORK-A’ automated embryo evaluation predicted aneuploid versus euploid embryos with an accuracy of 69.3% (AUC 0.761) when using images, maternal age, morphokinetics, and blastocyst score. -Accuracy increased to 77.6% in prediction of complex aneuploidy vs. euploidy. -Two external test datasets, achieved an accuracy of 63.4% and 65.7%, showing generalizability.

Summary of studies using artificial intelligence (AI) and machine learning (ML) methods for embryo assessment, prediction, and selection. The asterisk (*) indicates studies from conference proceedings. PN pronuclear, AUC area under curve, CNN convolutional neural network, ANN artificial neural network, k-NN k-nearest neighbor.