O-121 Temporal development of scores from a time-lapse based artificial intelligence provides no additional benefit compared to the latest score

Abstract

Abstract Study question Does the temporal development of scores from an artificial intelligence (AI) based embryo selection model provide additional predictive power compared to the latest score? Summary answer Earlier predictions provide no additional predictive power given the latest prediction. Knowing the development of AI scores is thus not beneficial for embryo selection. What is known already Traditionally, embryo grading takes into consideration the final morphology score together with the history of embryo development events. Recent publications have shown that time-lapse based AI models trained on clinical outcomes can automatically rank embryos by the likelihood of implantation that equal or surpass traditional methods, without the necessity for manual evaluation. AI-based methods based on time-lapse data may account for the impact of the full temporal development on likelihood of implantation. In this study, we examine if the addition of previous scores adds to the performance of an embryo selection model. Study design, size, duration A retrospective multicenter study of transferred embryos (n = 2422) with known implantation data (KID) between 2012-2020 from 21 international clinics. Embryos were cultured in EmbryoScope time-lapse incubators for at least 5 days. Both single and multi-embryo transfers in fresh and warmed cycles were included. Embryos were classified as positive (KIDp) or negative (KIDn) by the presence of fetal heartbeat. Implantation likelihood was evaluated at cleavage- and blastocyst stages by a 3D CNN model; iDAScore v2.0. Participants/materials, setting, methods A kernel-based conditional independence test was used to evaluate if a score at x hours post insemination (hpi) provide extra information compared to a score at y hpi. The test estimates the likelihood that a prediction after x hpi does not provide any additional information with regards to predicting KIDp when a prediction after y hpi is available. These likelihoods (p-values) were Bonferroni corrected to adjust for multiple comparisons. Main results and the role of chance KIDp predictions were computed for all embryos at 40, 44, 64, 68, 112, and 116 hpi corresponding to early and median time of transfer for day 2, 3, and 5. For all time points, the conditional independence test showed that there was no added information from knowing predictions at an earlier time point (p > 0.21). Similarly, it was tested if a known later prediction will improve model performance for an earlier prediction. This was always the case (p < 0.01) except for predictions close in time (68 vs 64 hpi; p = 1.0) and (44 vs 40 hpi; p = 1.0). This may reflect that there was little additional embryo development information during a 4-hour period at this development stage. The above shows that knowledge of the temporal score development does not improve model performance compared to using latest available score. This means that an AI model based on time-lapse videos likely considers earlier embryo development events in the context of the entire development history. Limitations, reasons for caution This study is a retrospective study that only looks at AI based on time-lapse videos. There is a potential bias as the embryos were likely of higher quality having been chosen for transfer. It is possible that the results do not generalize to simpler non-3D CNNs. Wider implications of the findings Knowing the temporal score development can introduce subjectivity that might lower implantation rates as the earlier scores were shown to not provide any significant prediction improvement. Trial registration number not applicable