P-645 A machine learning model for optimal trigger day recommendation

Abstract

Abstract Study question Could AI be used for trigger day recommendations to improve the number of M2 oocytes retrieved in IVF processes? Summary answer The use of AI for trigger day recommendations can improve the number of M2 oocytes retrieved in IVF processes by 1 on average. What is known already One of the most important decisions during ovarian stimulation during IVF cycles is when to trigger follicular maturation. Optimally timed trigger injection should maximize the number of retrieved high-quality oocytes. While these decisions are commonly made based on follicle sizes, hormone levels, and patient parameters, they are subjective and the criteria vary across medical centers and physicians. Numerous research studies are exploring how follicle sizes and other parameters translate to stimulation outcomes, but only in recent years, the use of artificial intelligence in decisions about the day of the trigger has become one of the main fields of interest. Study design, size, duration A retrospective study was made on 1085 IVF cycles conducted between 2019 and 2021 at 6 IVF centers in Poland. Processes were selected to have at least one set of ultrasound cine-loop videos for both ovaries from a maximum of three days before trigger day. Additionally, data on age, AMH, estradiol level, endometrium thickness, and the process outcome (MII oocytes number) was collected. Follicles were measured using the FOLLISCAN system. Participants/materials, setting, methods The recommendation algorithm uses a model predicting the number of MII oocytes retrieved if the trigger was applied on each of the three days following examination. The day with a maximum expected number of oocytes is recommended. The average number of retrieved MII oocytes was compared between two groups: cycles where the recommendation was the same as the actual physician decision, and cycles where it was different. Ten-fold cross-validation and propensity score matching were performed. Main results and the role of chance The model predicting MII oocyte numbers had an R2 of 0.56 and errors of 3.5 (RMSE), and 2.54 (MAE) overall. For the trigger on the day of the USG examination, it was RMSE: 3.33, MAE: 2.39, and later (+1, +2, +3 days respectively) RMSE: 3.6, 3.67, 3.44, MAE: 2.57, 2.71, 2.56. On average, 1.98 more MII oocytes (95% Confidence Interval: 1.22-2.77) were retrieved from processes where the model recommendation was the same as the actual physician decision (on time) compared to those where the model recommendation differed (not on time). Because this was not a randomized trial, and patient parameters vary across compared groups, propensity matching was used to ensure their homogeneity. After this, there were 1.11 more MII oocytes (CI: 0.22-1.99) retrieved on average in the “on time” group than in the “not on time” group. For patients with less than 10 follicles, there were on average 0.3 more MII oocytes in the “on time” group. For patients with 10-20 follicles, on average 0.72 more MII oocytes were observed, and for those with at least 20 follicles 1.32 more on average. Limitations, reasons for caution The study was limited to a few IVF centers in Poland, and its retrospective findings do not necessarily generalize to different centers. Additional evaluation of the model on data from multiple centers should be performed. Also, randomized control trials could be used to evaluate its effectiveness. Wider implications of the findings Results show the potential of using machine learning to optimize ovarian stimulation outcomes. An increased number of MII oocytes results in a higher success rate of in vitro fertilization. Trial registration number Project support was provided by the Polish National Center for Research and Development no. POIR.01.01.01-00-1634/20-00 and ERC Consolidator Grant TUgbOAT no. 772346.