P-144 Utilizing artificial intelligence to uncover Day 5 morphokinetic features associated with single blastocyst transfer implantation outcome

Abstract

Abstract Study question Can AI aid in studing embryos’ continuous development in Day5 and identifying morphokinetic features (blastocyst growth and spontaneous collapse etc.), to improve implantation potential prediction? Summary answer The AI system utilized time-lapse images to measure newly identified morphokinetic features, establishing a significant correlation with clinical pregnancy outcomes. What is known already In Single Blastocyst Transfer (SBT), the conventional Gardner evaluation system may inadequately differentiate between embryos of comparable morphology scores in specific cases. To address this, integrating new morphokinetic features is essential for enhanced blastocyst evaluation. Spontaneous collapse (SC) during blastocyst expansion is a phenomenon that was revealed relatively recently following the clinical application of time-lapse monitoring in IVF laboratories. Previous studies link embryo size and SC to quality, yet these correlations require further validation. In the critical selection process utilizing time-lapse imaging (TLI) systems, AI technology is indispensable for developing automated and interpretable tools to ensure precise quantitative measurements. Study design, size, duration 255 patients underwent IVF with Day5 single blastocyst transfer in fresh cycles at the Center for Reproductive Medicine, Nanjing Drum Tower Hospital, from April 2021 to August 2022, and were analyzed in this retrospective study. Embryos were cultured in time-lapse incubators (CCM-iBIS, ASTEC). On Day 3 embryos were transferred to sequential media (G2, Vitrolife, Sweden) and cultured in the same incubator until Day5. In this cohort, there were 167 clinical pregnancies and 88 non-pregnant cases. Participants/materials, setting, methods An AI system was developed to calculate the development curves of areas in each part of the embryo (ZP,blastocel, ICM and TE) based on TLI images from 92 to 116 phi(±2). T-test and correlation analysis were used to select effective features. The dataset was divided into training and testing subset in an 80:20 ratio. The performance of different prediction models were compared by Delong test. Main results and the role of chance In this study, an AI system automatically generated blastocyst development curves, from which 35 morphokinetic features were extracted. These features were carefully defined to comprehensively capture the dynamics of embryo size development and spontaneous collapse (SC). Six parameters, including maximum size, max/min ratio of blastocyst area (with and without ZP region), speed of expansion after tSB, and number of minor SC, exhibited statistical correlation with clinical pregnancy outcomes. Using these parameters, an XGBoost model achieved an AUC of 0.71 for pregnancy prediction, surpassing (p = 0.0307, p < 0.05) the AUC of 0.48 from the model trained with manual Gardner evaluation. These findings underscore the potential of AI-generated morphokinetic features of D5 in enhancing blastocyst evaluation systems. Limitations, reasons for caution The limited sample size used in this retrospective experiment may lead to an underrepresentation of certain significant features. Wider implications of the findings In contrast to prior findings, most SC-based features in our study did not significantly correlate with clinical pregnancy outcome. Further research is needed to understand their causes and implications. Integrating AI with TLI offers promise for improving single blastocyst transfer outcomes, emphasizing quantifiability and interpretability in approach. Trial registration number NA