Abstract Study question Do AI-based and manual embryo selection strategies based on morphological and kinetic evaluation favour male over female embryos? Could the differences enable sex selection? Summary answer Male embryos receive higher-quality scores using both AI-based models and manual grading, but not using deep-learning approaches. The differences, although present, preclude reliable sex selection. What is known already A key aspect of IVF, embryo selection, involves methods executed manually via direct embryo morphological assessment, or by artificial intelligence algorithms which assess parameters such as tPNf, t2, t3, t4, t5, t8, tB and ICM and TE grades using time-lapse imaging. However, all current embryo selection models do not take embryo sex into account despite reported differences in developmental timings between XX and XY embryos. If the predictions of embryo selection algorithms themselves are affected by embryo sex, their accuracy and fairness may be impacted, leading to sex disparities in populations heavily reliant on IVF. Study design, size, duration A retrospective study was conducted on 1411 embryos with known sex information following PGT-A, at a single centre between 2018-202, making this the largest study to date interrogating morphological and morphokinetic sex differences using time-lapse and PGT-A data. Three embryo assessment methods were interrogated: manual morphological grading (Gardener system), KIDScore D3 (VitroLife) and CHLOE (Fairtility). These algorithms are representative of the current embryo grading landscape, encompassing manual selection, traditional machine-learning and modern deep-learning approaches, respectively. Participants/materials, setting, methods Kinetic parameters were annotated using CHLOE on incubator time-lapses. KIDScore computation was implemented according to Petersen et al.,2016 making use of kinetic annotations from CHLOE. Manual morphological grading was done at the blastocyst stage using a modified Gardner system. Mann-Whitney U and chi-squared tests compared XX and XY gradings, with detailed analyses on morphokinetics (tPNf,t2-t9+,cc2,cc3,tM,tSB-tEB,ICM and TE grades). To evaluate the ability of morphokinetic differences to predict embryo sex, four machine learning models were built. Main results and the role of chance Overall, XY embryos (4.182 ± 1.353, N = 692) were more likely to be assigned higher scores than XX embryos (4.022 ± 1.420, N = 642) by the machine learning-based KIDScore (U = 207604, p = 0.0182). Similarly, XY (462/668) blastocysts were more likely to receive good grades than XX (351/614) blastocysts under manual morphological grading (χ2 =19.843, df = 1, p < 0.00001). We pinpoint the source of sex disparities in conventional morphological grading to variations in trophectoderm morphology (male embryos receiving higher TE grades). Nevertheless, the observed differences, although present, were of marginal magnitude, and the overlap between the sexes substantial enough to preclude reliable sex selection. No significant difference was found in the scores assigned by more advanced deep learning methods such as CHLOE EQ scores between XX (0.787 ± 0.276, N = 628) and XY (0.802 ± 0.254, N = 679) embryos (U = 204621, p = 0.208). Limitations, reasons for caution This retrospective, single-centre study encounters limitations such as selection bias and constrained generalisability beyond the specific centre. It focuses on blastocyst-stage sex analysis, which, while statistically-powerful, may not fully represent the societal implications of embryo selection on sex ratios. The use of automated kinetic annotations, while efficient, might introduce inaccuracies. Wider implications of the findings Our findings raise significant questions surrounding the fairness of widely-used embryo assessment methods. We highlight the urgent need for the reproductive health community to recognise and address algorithmic bias, ensuring equitable and ethical treatment of all embryos, especially in the context of increasing global reliance on assisted reproduction. Trial registration number not applicable
Read full abstract