Delineating the heterogeneity of preimplantation development via unsupervised clustering of embryo candidates for transfer using automated, accurate and standardized morphokinetic annotation

Abstract

The majority of human embryos, whether naturally or in vitro fertilized (IVF), do not poses the capacity to implant within the uterus and reach live birth. Hence, selecting the embryos with the highest developmental potential to implant is imperative for improving pregnancy rates without prolonging time to pregnancy. The developmental potential of embryos can be assessed based on temporal profiling of the discrete morphokinetic events of preimplantation development. However, manual morphokinetic annotation introduces intra- and inter-observer variation and is time-consuming. Using a large clinically-labeled multicenter dataset of video recordings of preimplantation embryo development by time-lapse incubators, we trained a convolutional neural network and developed a classifier that performs fully automated, robust, and standardized annotation of the morphokinetic events with R-square 0.994 accuracy. To delineate the morphokinetic heterogeneity of preimplantation development, we performed unsupervised clustering of high-quality embryo candidates for transfer, which was independent of maternal age and blastulation rate. Retrospective comparative analysis of transfer versus implantation rates reveals differences between embryo clusters that are distinctively marked by poor synchronization of the third meiotic cell-cleavage cycle. We expect this work to advance the integration of morphokinetic-based decision support tools in IVF treatments and deepen our understanding of preimplantation heterogeneity.

Embryo classification beyond pregnancy: early prediction of first trimester miscarriage using machine learning

“…Feature screening as reported here is not a deterministic process and the selection of a different feature subset that is both predictive and non-redundant cannot be excluded. While the reported features were extracted using semi-automated algorithms, clinical implementation would be further supported by developing fully automated tools for the extraction of the reported features or of an equivalent subset of predictive and non-redundant features [47].…”

Improving the assessment of embryo developmental potential via morphokinetic forecasting of future events using language modeling

“…To maximize the dataset size, automatic morphokinetic annotation was performed to additional 47,454 embryos. We used an automatic annotation algorithm that we recently developed with near-perfect accuracy, R-square 0.994 4 . The embryos were divided into a train-validation set (65376 embryos) and test set (2331 embryos).…”

“…To eliminate outlier contributions, we set the upper limits of these time windows to be the 97.5 percentile of the temporal distributions of the morphokinetic events. We have previously generated the most accurate statistical description of embryo preimplantation potential using 20,253 annotated embryos 4 and used it here to calculate the temporal distributions of each event. In this manner, we were able to label arrested events as follows: An embryo is arrested at morphokinetic event tN if it failed to morphokinetically advance by the time of the 97.5 percentile of the t(N+1) event.…”

“…With respect to feature size, morphokinetic annotation effectively allows massive and interpretable dimensional reduction of the original video raw data (~200 Mb per embryo). In addition, morphokinetic prediction models a-priori offer non-invasive real time developmental assessment, thus alleviating potential risks of biopsy collection and allowing fresh transfers 4 . Indeed, morphokinetic embryo assessment has been clinically tested via prospective studies and was approved for clinical applications 5 .…”

A hybrid artificial intelligence model leverages multi-centric clinical data to improve fetal heart rate pregnancy prediction across time-lapse systems

“…Other studies use AI to predict the chances an embryo has to develop to a blastocyst ( Wong et al , 2010 ; Liao et al , 2021 ). Another approach consists of feeding the timing of key biological events to machine learning algorithms, in the hopes that they can use this information to predict the chances of pregnancy or live birth ( Zabari et al , 2022 ) better than embryologists. In contrast, a more objective approach consists of training models that either analyze a single image ( VerMilyea et al , 2020 ) or the entire embryonic development ( Tran et al , 2019 ; Berntsen et al , 2022 ; Lassen et al , 2022 ) using pregnancy outcome as a label.…”