P-113 The performance of pregnancy prediction is improved in an updated deep-learning based embryo selection model: A retrospective validation study using 3960 single blastocyst transfer cycles

Abstract

Abstract Study question Does increasing the size of a training dataset improve the performance of pregnancy (fetal heartbeat) prediction in a deep-learning model? Summary answer Yes. Increasing the training data for a deep-learning model for embryo selection improved prediction of the likelihood of implantation after single vitrified-warmed blastocyst transfer (SVBT). What is known already Recently, artificial intelligence (AI) for implantation prediction after blastocyst transfer has been extensively studied. AI can address the issue of subjective assessment for the selection of blastocyst transfer. Generally, the performance of deep learning models is said to improve by increasing the training dataset. However, to the best of our knowledge, there exists no study on the effect of increasing training data on the performance of pregnancy prediction on the same deep-learning model. Study design, size, duration A total of 3,960 SVBT cycles (1 patient,1 cycle) were retrospectively analyzed. Embryos were stratified according to SART age groups. The quality and scoring of embryos were assessed by iDAScore v1.0 (iDAV1; Vitrolife, Sweden), v2.0 (iDAV2; Vitrolife, Sweden), and Gardner grading. The discriminative performance of the pregnancy prediction for each embryo scoring model was compared using the area under the curve (AUC) of the receiver operating characteristic curve for each maternal age group. Participants/materials, setting, methods Embryos were cultured in the EmbryoScope+ and/or EmbryoScope Flex (Vitrolife). iDAV2 and iDAV1 were based on an identical deep-learning architecture, but the training data for iDAV2 has been increased with 15% more data. ICM and TE were annotated according to the Gardner grading system. The degree of blastocyst expansion was Grade 4 due to our freezing policy. Furthermore, Gardner's grading (GG) was stratified into four grades (A: AA, B: AB BA, C: BB, D: others). Main results and the role of chance The AUCs of the < 35 years age group (n = 757) for pregnancy prediction were 0.718 for iDAV1, 0.733 for iDAV2, and 0.694 for GG. The AUC of iDAV2 was significantly higher compared to GG (P < 0.05). For the 35–37 years age group (n = 821) the AUCs were 0.696, 0.712, and 0.695 for iDAV1, iDAV2, and GS, respectively, and were significantly different between iDAV1 and iDAV2 (P < 0.05). The AUCs of the 38–40 years age group (n = 1,007) were 0.698 for iDASV1, 0.706 for iDAV2, and 0.700 for GG and there was no significant difference. The AUCs of the 41–42 years age group (n = 715) were 0.734, 0.745, and 0.696 for iDAV1, iDAV2, and GS, respectively, and the AUC of iDAV2 was not significantly different compared to iDAV1 (P = 0.174) but significantly different compared to GG (P < 0.05). For the > 42 years age group (n = 660) AUCs were 0.685 for iDAV1, 0.698 for iDAV2, and 0.682 for GS and were not significantly different among groups. The AUCs of iDAV1, iDAV2, and GG in all ages were 0.736, 0.720, and 0.702, respectively. iDAV2 were significantly higher than iDAV1 and GG (p < 0.05). Limitations, reasons for caution This study was based on minimal stimulation and natural cycle IVF treatment, and a freeze-all strategy whereby all transferred blastocysts had previously been vitrified. Therefore, we had only a few cycles with elective blastocyst transfer. In addition, this study was retrospective in nature. Wider implications of the findings For all age groups, iDAV2 had a higher AUC than iDAV1, although a significant difference was only observed for the young age group. This result suggests that the increased dataset used for the development of iDAV2 improved the performance of pregnancy prediction for SVBT in the deep-learning model. Trial registration number not applicable