Embryo selection in assisted reproduction has traditionally relied on subjective morphological grading, which is prone to inter-observer variability and correlates poorly with ploidy, the primary cause of miscarriage. Artificial intelligence (AI) models utilizing time-lapse imaging have emerged; however, many rely on human-defined morphokinetic markers or static blastocyst images. We developed two deep learning models that non-invasively predict the likelihood of clinical pregnancy and embryo ploidy status, based solely on time-lapse videos and maternal age. In this retrospective study, 2,436 embryos were analyzed for pregnancy prediction and 1,645 for ploidy prediction, with external validation performed using data from independent IVF centers. Both models were implemented using a spatiotemporal convolutional neural network. The pregnancy prediction model achieved an area under the receiver operating characteristic curve (AUROC) of 0.799 in the validation set, 0.717 in the internal test set, and 0.746 in the external test set. The ploidy prediction model yielded AUROCs of 0.802, 0.738, and 0.759 in the validation, internal, and external test sets, respectively. These results suggest that AI models can serve as non-invasive decision support tools that leverage developmental dynamics associated with implantation potential and chromosomal status. These models may supplement conventional morphology-based assessment and contribute to more objective embryo selection.
MAEKAWA et al. (Thu,) studied this question.