O-293 An Artificial Intelligence Based Approach for Selecting the Optimal Day for Triggering in Antagonist Cycles

Abstract

Abstract Study question Can a machine learning (ML) algorithm suggest an optimal trigger day for maximizing the number of mature (MII) oocytes retrieved during an antagonist protocol cycle? Summary answer Using ML algorithm for trigger day selection may increase the mean number of MII oocytes and usable embryos by 3.6 and 1.1 per cycle respectively. What is known already ML has been increasingly used in the field of reproductive medicine, mainly in lab related technologies such as embryo selection. Only a few studies, aimed at improving clinical decision making, were published. A recent study used ML to optimize trigger day selection based on a retrospective analysis, showing an increase of 2.3 MII oocytes and 1.0 usable embryo. Study design, size, duration A retrospective cohort study including data of 9,622 antagonist protocol cycles performed between August 2017 to November 2022. The evaluation of the ML algorithm was conducted using a test dataset including the following quality groups: 1. “Freeze all- oocytes” cycles – a unique population of patients mostly for social fertility preservation. 2. “ICSI only” cycles – for maturation rate evaluation 3. “Fertilize all” cycles including IVF and ICSI, for evaluation of the number of embryos. Participants/materials, setting, methods The ML algorithm suggested optimal trigger days for maximizing the number of MII oocytes retrieved by considering the MII prediction, prediction errors and outlier detection results. The model suggested one, two, or three days as trigger options, depending on the difference in potential outcomes. It recommended the days that have a 10% higher prediction of MII oocytes with a confidence level of over 50% compared to less optimal options if they exist. Main results and the role of chance To evaluate the performance of the management algorithm, it was applied to cycles in the test sets. For each cycle, the algorithm provided a suggestion for trigger days, and this was compared to the actual trigger day chosen by the physician. When the day chosen by the physician was one of the algorithm’s suggestions, the result was labeled as “correct”, otherwise it was labeled as “incorrect”. Comparing the “correct” and “incorrect” groups, using the trigger management algorithm resulted in a higher number of total and MII oocytes retrieved, 2PN and usable embryos. Specifically, when using the algorithm in the “Freeze All” test set, an average increase of 4.8 oocytes and 3.4 MII oocytes retrieved in the “correct” group (consisting 36.2% of the quality group subset) compared to the “incorrect group” (63.8%). In the “ICSI-only” test set, the algorithm resulted in an average increase of 4.5 oocytes, 3.8 MII oocytes, 2.4 2PN and 1.1 usable embryos in the “correct” group (26.8%) compared to the “incorrect group” (73.2%). Lastly, in the “Fertilize all” test set, the algorithm resulted in an average increase of 3.6 oocytes, 2.1 2PN, and 0.9 usable embryos in the “correct” group (25.1%) compared to the “incorrect group” (74.9%). Limitations, reasons for caution The trigger management algorithms' decision-making is based solely on the predicted number of MII oocytes to be retrieved. Moreover, it was developed and applied exclusively for antagonist cycles. Wider implications of the findings The trigger management algorithm may improve oocyte yield and IVF outcomes in antagonist cycles. Moreover, provision of two or three trigger options allows for more flexibility when choosing the trigger day and enables taking into account other factors without negatively affecting the outcome. Trial registration number NA