Introduction: Current methods fail to evaluate the developmental competence of pre-implanted embryos. As a result, the practice is to transfer more than one embryo to obtain reasonable pregnancy rates. The prediction of embryogenic developmental competence at early stages of pre-implantation development will enable identifying the highest quality embryos for transfer, thus significantly increase live-birth rates and reduce time-to-pregnancy and multiple pregnancy. The current state of the art in IVF clinics is based on time-lapse incubators with annotating the morphokinetics events to use an algorithm to predict the embryo potential (like KIDScoreTM). Unfortunately, this is a non-standardized and time consuming process with limited reproducible predictive power.
Aim: Revisiting morphokinetic annotation schemes to define the potential predictive power of the first three days of preimplantation development.
Material & Methods: Clinical data (age, ET results) and embryonic development videos were collected from four IVF units. We utilize three state-of-the-art unbiased statistical and machine learning models to define the information contained in the video and clinical data: Monotonic regression, Logistic regression and convolutional neural network (CNN).
Results: Using a large database, we analyze the first three days of preimplantation development. We exclude embryos that failed to reach the four-cell stage by day-3. The analysis of transferred and non-transferred embryos showed limited performances in discriminating the high-performing from the low-performing embryos. All the models utilized here share these results.
Conclusions: We define the limited capacity of morphological and morphokinetic analyses in identifying high-potential embryos by day-3 of preimplantation development. Our results motivate analyses of data that extend beyond the discrete description of morphokinetic events and are based on raw time-lapse images using machine-learning methods.