O-106 Bringing Transparency to Oocyte Assessment: the importance of including confounders when building Artificial Intelligence (AI) based support tools to quantify oocyte viability

Abstract

Abstract Study question Which confounders (sperm quality, oocyte dysmorphism, culture time, images pre or post-ICSI, age) affect the ability of AI to predict blastulation based on oocyte images? Summary answer Sperm quality, oocyte dysmorphism, pre or post-ICSI image should be controlled for when building AI algorithms to predict blastulation based on oocyte images. What is known already Previous studies reporting on the use of AI to predict blastulation based on oocyte images have: (i) not accounted for confounders affecting blastulation (i.e. sperm quality, culture time), and (ii) used post-ICSI images; without assessing whether the ICSI procedure affects the oocyte image as assessed by AI. Therefore, there is a risk of mislabeling viable oocytes as non-viable due to external factors, which could cause uncontrolled bias and failure to generalize when used in clinical practice. The objective was to assess how these confounders affect efficacy of prediction of blastulation from oocyte images by an AI-based oocyte assessment tool: CHLOE-OQ(Fairtility). Study design, size, duration Cohort study. Images of 1281 oocytes (February to June 2022) were taken pre and post ICSI using the Embryoscope, and the embryos cultured until day 7. Oocyte donor source and age, oocyte dysmorphias and sperm quality were documented. CHLOE-OQ algorithm was trained, validated and tested in a diverse data set, accounting for pre and post ICSI image datasets, quality of oocytes, quality of sperm and patient age. Participants/materials, setting, methods The primary endpoint was blastulation. Sperm quality data was classified into 4 groups: (A)All (n = 1281), (B)donor sperm only (n = 51), (C)donor sperm and normospermic samples from men not diagnosed with male factor infertility (n = 557), (D)abnormal sperm samples and other diagnosed male factor cycles (n = 747). Eggs were classified by source (own/donor), and by dysmorphisms: enlarged perivitelline space, abnormal Zona pellucida, cytoplasmic abnormalities, dark, enlarged oocytes. Main results and the role of chance Post-ICSI images had higher mean CHLOE-OQ score than pre ICSI images (0.28±0.1 vs 0.33±0.1, p < 0.001, paired t-test). Discrepancies were particularly identified in oocytes that degenerated following ICSI, and scored 0 by CHLOE-OQ despite having higher scores pre-ICSI. Using Post-ICSI images (AUC=0.66, 95% confidence interval, CI: 0.63-0.69, n = 1281) improved the efficacy of prediction of blastulation compared to pre-ICSI images (AUC=0.57: 0.53-0.60, n = 1281, p < 0.001), suggesting that ICSI affected the quality morphology of the oocyte, and how an oocyte responds to ICSI, as assessed by AI, contributes to prediction of blastulation. Efficacy of prediction (AUC) was not affected by the quality of the sperm: (A-OVERALL 0.658 [CI(95%): 0.626-0.687]; B-Donor 0.586 [CI(95%): 0.449-0.728]; C-normospermic 0.645 [CI(95%): 0.600-0.688], D male factor 0.678 [CI(95%): 0.639-0.715]). Oocyte features associated with low CHLOE-OQ scores were: enlarged perivitelline space, dysmorphic oocytes, abnormal Zona pellucida, cytoplasmic abnormalities and dark and enlarged oocytes. Whilst spherical oocytes with normal zona and perivitelline space were characterized as being more likely to form a blastocyst. Limitations, reasons for caution This single-clinic study is retrospective. A multi-center study is underway. External factors affecting blastulation must be accounted for to avoid mislabeling of good oocytes as non-viable. There is also a need to understand oocyte dysmorphias identified by the AI algorithm to ensure biological transparency in clinical decision making. Wider implications of the findings Taking into account clinical and gamete confounders when building AI algorithms is a necessary strategy to ensure AI algorithms are generalized when incorporated into clinical practice, whilst reducing bias and promoting transparency in clinical decision making. The risk of not considering confounders leads to mislabeling, bias and inaccurate predictions. Trial registration number not applicable