Embryo selection tools in IVF can favour XY embryos: implications for equitable reproductive AI

Embryo selection tools in IVF can favour XY embryos: implications for equitable reproductive AI | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Embryo selection tools in IVF can favour XY embryos: implications for equitable reproductive AI Teodora Popa, Chloe He, Christian S. Ottolini, Colin Davis, Matthew Lau, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8230063/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Assisted reproductive technologies (ART) are increasingly used worldwide, with up to 8.5% of babies in some countries conceived via in vitro fertilisation (IVF). Embryo selection, a critical step in IVF, is shifting from manual morphology-based to artificial intelligence (AI)-based assessment. While these systems use outcome data to select embryos with the highest likelihood of pregnancy, their potential to introduce unintended bias has not been evaluated at scale. We analysed 1,334 human embryos with known sex (from preimplantation genetic testing), graded by two widely-used AI models (KIDScore D3™, CHLOE EQ™) and standard manual morphology-based Gardner grading. Manual grading and the traditional AI model (KIDScore D3™) assigned higher scores to XY embryos; the deep-learning model (CHLOE EQ™) did not. Simulations suggest such biases could modestly skew sex ratios at birth in high-IVF-utilisation settings, which is of global concern. These findings highlight the need for systematic bias evaluation in embryo grading and demonstrate that algorithmic selection without sex bias is achievable with appropriate design. Biological sciences/Developmental biology Health sciences/Medical research Figures Figure 1 Figure 2 Main text Global fertility rates are at an all-time low and have declined sharply over recent decades, with many individuals delaying parenthood into ages associated with reduced natural fertility. As a result, reliance on assisted reproductive technologies (ART) such as in vitro fertilisation (IVF) has increased, accounting for up to 8.5% of all live births in Greece, 7.9% in Spain and around 9% in Denmark, among the highest rates worldwide ( 1 ) . A critical stage in IVF is embryo selection: identifying the embryo(s) with the highest chance of implantation and live birth. Traditionally, this has been performed manually by embryologists using morphological criteria, most often the Gardner grading system ( 2 ) , assessing the inner cell mass (ICM) and trophectoderm (TE) quality at the blastocyst stage ( 2 ) . More recently, time-lapse incubators and AI-based grading algorithms have been introduced to improve selection accuracy, reduce inter-operator variability, and increase efficiency ( 3 – 5 ) . These systems integrate developmental timings (known as kinetics ) with morphological data, a combined approach referred to as morphokinetics ( 6 ) . However, as in other domains, AI in reproductive medicine may encode and propagate biases present in training data or arising from model design. Algorithmic bias in medical AI has been documented across sex, race, and socioeconomic lines in diagnostics and treatment recommendations ( 7 ) , but its potential role in embryo selection has not been investigated systematically. This gap is important: if grading tools systematically favour one sex over the other, the cumulative effect in high-IVF-utilisation populations could alter the sex ratio at birth, with biological, social, and ethical implications. It is well established that natural conception yields a modest male bias at birth (≈ 51–52% XY; sex ratio ≈ 1.05) ( 8 , 9 ) , and IVF may amplify this baseline skew. Some studies have found that XY embryos cleave and reach the blastocyst stage faster than XX embryos ( 10 ) , potentially due to the additional developmental demands of X-chromosome inactivation in XX embryos ( 11 ) . Extended in vitro culture may also impose metabolic and oxidative stress that disproportionately affects XX embryos ( 12 ) . Together, these morphokinetic differences may favour XY embryos and be inadvertently reinforced by embryo selection systems designed under the assumption of sex-neutral development, even in the absence of deliberate sex selection. To investigate this, we retrospectively evaluated three widely used embryo grading approaches: the Gardner system (a manual grading system based on blastocyst morphology) ( 2 , 13 ) , KIDScore D3™(Vitrolife) (a classical machine learning model based on annotated morphokinetic events) ( 14 ) , and CHLOE EQ™ (a deep learning model that analyses time-lapse videos without manual feature selection) ( 15 ) . Together, these three systems represent the spectrum of current clinical practice, spanning traditional human-led techniques to modern automated approaches. We assessed whether each tool assigned different scores to XX vs XY embryos, and modelled the potential demographic impact if such biases were applied at scale (Fig. 1 ). Sex Bias in Embryo Grading - In the full cohort of 1,334 embryos, manual morphological grading assigned ‘good’ grades more often to XY embryos (462/668) than XX embryos (351/614; χ²=19.843, p < 0.001). KIDScore D3™ also scored XY embryos (M ± SD: 4.182 ± 1.353, n = 642) higher than XX (4.022 ± 1.420, n = 692; U = 207604, p = 0.0182). CHLOE EQ™ scores showed no difference between XX (0.787 ± p.276, n = 628) and XY embryos (0.802 ± 0.254, n = 679; U = 204621, p = 0.208) (Fig. 2 a). When restricted to euploid embryos, manual grading again favoured XY embryos (χ²=15.839, p < 0.001), but KIDScore D3™ and CHLOE EQ™ showed no statistically significant difference (Fig. 2 b). Drivers of Bias - Post-hoc analyses indicated that the disparity in manual grading was driven largely by higher TE grades in XY embryos (2.809 ± 0.628) than XX embryos (2.656 ± 0.623, n = 616; U = 179584, p < 0.001), with no significant differences in ICM scores or other morphokinetic parameters (Fig. 2 c, Supplementary Table 1). Aneuploidy rates between XX (73/642) and XY embryos (75/692) did not differ ( X 2 (1, N = 1334) = 0.0958, p = 0.757). Predicting Sex from Embryo Grades - Machine learning models trained to predict embryo sex from TE grade and KIDScore D3™achieved low predictive performance (Fig. 2 d), confirming that while sex differences in grades exist, grades cannot be used to reliably infer embryo sex for individual embryos. Modelling of Potential Population Impacts - Monte Carlo simulations (Fig. 1 e) suggest that in high-IVF-utilisation countries, even small grading biases could modestly skew the sex ratio at birth. In simulations of IVF birth populations that assumed correlation between grades and live birth (Fig. 1 f), embryo selection skewed the sex ratio toward XY for TE grade (XX/XY = 0.886, 95% CI [0.658,0.904]), KIDScore D3™ (XX/XY = 0.900 [0.881,0.920]), and CHLOE EQ™ (XX/XY = 0.940 [0.924, 0.959]). When correlation was not assumed (Fig. 1 g), the skew persisted: TE grade (XX/XY = 0.951, 95% CI [0.930, 0.971]), KIDScore D3™ (XX/XY = 0.950, 95% CI [0.929, 0.972]), and CHLOE EQ™ (XX/XY = 0.984, 95% CI [0.963, 1.005]). To our knowledge, this analysis of over 1,300 human embryos with known sex, graded using both manual and AI-based tools, is the largest study to date assessing sex bias in both manual and AI-based embryo selection methods. We show that manual Gardner grading and the KIDScore D3™ algorithm both assign higher scores to XY embryos, whereas the CHLOE EQ ™ deep learning model does not. These findings highlight two key points. First, bias is not unique to AI: manual grading by experienced embryologists showed stronger sex bias than either AI model. Second, sex bias in embryo selection is not inevitable and can be avoided. An obvious starting point for interpreting these results is biology, as the differences observed may reflect intrinsic sex-based developmental dynamics. Some works suggest that XY embryos develop more quickly and may be more resilient to extended culture, traits that correlate with implantation and live birth potential ( 10 ) . Clinical data support this, with male-skewed birth ratios more pronounced after blastocyst-stage than cleavage-stage transfer (~ 54–58% vs ~ 50–52%) (16) , suggesting that sex-related factors become more influential beyond the cleavage stage and are mirrored in real outcomes. However, the absence of bias in CHLOE EQ™ shows that, although biological factors may contribute, sex-neutral embryo selection is achievable. The reason for this may lie in algorithm design. Conventional morphological grading relies on coarse visual appraisal of blastocyst structures. While blastocyst morphology and grading do correlate with live birth rates ( 17 , 18 ) , the correlation is imperfect, leaving room for confounding. Following this reasoning, the skew observed in manual grading may have emerged unintentionally over decades of human-guided refinement of grading practices, rather than from deliberate sex selection. In contrast, CHLOE EQ™ employs a convolutional neural network explicitly optimised for the prediction of implantation. This setup enables CHLOE EQ™ to discover its own features, focusing on direct predictive value rather than those that are simply accessible to human observers. In doing so, it may have learnt sex-invariant markers of embryo viability or, if sex-linked developmental differences exist, developed parallel strategies for each sex. Interpretability tools could then reveal the features driving these patterns and generate hypotheses on the role of sex differences in embryo development. Similarly, the KIDScore D3™algorithm relies on human-provided developmental timings, a feature that, as previously discussed, may introduce sex bias. Yet, as it was also optimised directly for implantation prediction, it acts as a middle way between manual grading and CHLOE EQ™, consistent with the intermediate level of bias observed. As ART becomes more common, particularly in countries with high IVF utilisation rate, even modest biases in selection tools could translate into measurable demographic shifts. Unlike overt sex selection, which is regulated in most jurisdictions, these biases are unintended and therefore largely invisible to clinicians and regulators. This underscores the need for systematic bias evaluation not only of AI systems but also of long-established manual grading practices. Our study has several limitations. It was conducted in a single centre, and most embryos were cryopreserved, leaving us with limited follow-up on live births. While simulations suggest that biases in trophectoderm grading could modestly skew sex ratios at birth, our modelling was necessarily simplistic and did not account for higher-order pregnancies, potential sex differences in embryo viability, or the influence of clinician judgement in transfer decisions. Larger, multi-centre datasets with live birth outcomes will be needed to confirm and extend these findings. A further limitation is that KIDScore D3™only incorporates morphokinetic events up to day 3 and does not capture blastocyst-stage features, unlike Gardner grading or CHLOE EQ™. More advanced algorithms such as KIDScore D5™(Vitrolife) or iDAScore® (Vitrolife) could not be assessed, as they require access to proprietary incubator software and their methodologies are not publicly available. This restricted our ability to benchmark across the full range of current embryo selection systems. Despite these caveats, our study provides the first large-scale evidence that embryo selection tools used in IVF, whether manual or AI-based, can systematically favour XY embryos, and the first attempt to model the possible demographic consequences of such bias. These findings highlight the need to routinely assess risk of bias in embryo grading systems and to report clinical prediction models with greater transparency in their design and training datasets ( 19 ) . Incorporating fairness constraints during model development offers a pathway to more equitable selection. Ultimately, ensuring that embryo grading methods are both effective and equitable will be essential to prevent unintended demographic shifts and to uphold the fairness of ART practice worldwide. For panels a-c , slight variations in sample size reflect missing annotations, and distributions are normalised such that the area under the curve for each sex equals 1. Kinetic features include: tX , time to reach the X-cell stage (e.g., t2 = time to 2-cell stage); tM , time to morula; tSB , time to start of blastulation; tB , time to blastocyst; tEB , time to expanded blastocyst; and cc1-3 , first to third cell cycles. For panels f and g , error bars represent 95% confidence intervals. Online Methods Ethics statement This study was conducted in accordance with relevant guidelines and regulations. An electronic ethics application was submitted to the UK Integrated Research Application System (IRAS project ID: 328309). The Committee confirmed that formal review was not required, as the project involved secondary analysis of anonymised data collected during routine clinical care, in which individual patients could not be identified. All patients undergoing IVF at the clinic provided written informed consent for the use of anonymised data in research, quality control, and clinical process development. All procedures were undertaken under the clinic’s HFEA licence, in line with UK regulatory requirements and best practice for assisted reproduction. Embryo imaging was performed solely as part of standard clinical care. Data were anonymised such that patient identities could not be ascertained directly or indirectly, and investigators neither contacted nor attempted to re-identify participants. Importantly, the clinical team did not have access to embryo sex information at any stage, ensuring that no intentional bias could have influenced clinical decision-making. Study design and dataset We retrospectively analysed 1,411 diploid embryos from 398 IVF/ICSI cycles at a single UK centre (2018–2021). Embryos with other karyotypes (n = 49) and those not derived from normally fertilised zygotes (n = 28) were excluded, leaving 1,334 embryos. The mean maternal age at oocyte retrieval was 38.4 ± 3.2 years. Most cycles (n = 947) used intracytoplasmic sperm injection (ICSI); the remainder used conventional IVF. All followed an antagonist stimulation protocol with oocyte retrieval 36 h after trigger. Embryos were cultured in Geri time-lapse incubators (Genea Biomedx), enabling continuous imaging without removal from culture. Time-lapse videos were exported for manual and AI-based grading. Embryos were biopsied on day 5, 6, or 7 for preimplantation genetic testing for aneuploidy (PGT-A). Biopsied cells were analysed by next-generation sequencing (PGTai 1.0 and PGTai 2.0; CooperSurgical Inc). In line with standard clinical reporting, results disclosed to the clinical team included only chromosomes 1–22. Sex chromosome data were not disclosed to clinicians or patients at any point. Instead, anonymised identifiers were generated, and only the research team had restricted access to sex chromosome information for research purposes. The key linking patient identities to research codes was stored separately in a password-protected, access-restricted file on the clinic’s secure drive. Access was strictly limited to authorised personnel. Embryo grading systems For the manual morphological grading, blastocysts were graded for inner cell mass (ICM) and trophectoderm (TE) quality using a modified Gardner system (2, 13). Grades were binarised into “good” and “poor” following McCoy et al. (2023) (20). The computation of the KIDScore D3™ decision tree algorithm was implemented according to ​Petersen et al., 2016​ (14) making use of kinetic annotations from CHLOE EQ™. Code for our implementation can be found at https://github.com/chlohe/embryo-sex . The deep learning-based CHLOE EQ score was assessed using CHLOE (Fairtility). Statistical analysis Comparisons between XX and XY embryos for each grading method were performed using two-tailed Mann–Whitney U tests for continuous scores and χ² tests for binary grade categories. Post-hoc analyses examined morphokinetic variables (timings of pronuclear fading, cleavage stages, morulation, blastulation) and ploidy status. ICM and TE grades were converted to numerical scales (D = 1 to A = 4) for analysis. For each test, embryos missing all required morphokinetic annotations were excluded. For example, we did not consider any embryos that did not have both a TE and ICM grade when analysing the Gardner gradings. All statistical analysis was carried out using Python (v3.8.10) with the Pandas (v1.3.3) and SciPy (v1.7.1) packages. Sex prediction from embryo grades We trained and evaluated four widely-used machine learning models on the task of predicting whether an embryo is XX or XY from variables showing significant associations (TE grade and KIDScore D3™). The models included a decision tree, a random forest, multinomial logistic regression and a multilayer perceptron. The models were compared against two simple baseline models based on simple cut-offs for TE grade and KIDScore D3™. These baseline models took the form: H ( x ) = XX (if x ⊕ X ) XY (otherwise) where x is the embryo score, X is some cutoff value and ⊕ is a comparison operator. We treat both X and ⊕ as hyperparameters. The dataset was split into training (N = 875) and test (N = 375) sets. Hyperparameters for each model were computed by grid search and 5-fold cross validation on the training set (a full list of hyperparameters considered can be found in Supplementary Table 2). Each model was evaluated on the test set using the accuracy, sensitivity, specificity and area under the receiver operating characteristic curve (AUC) metrics, with XY being the ‘positive’ outcome. All predictive modelling was carried out using Python (v3.8.10) with the Sci-kit Learn (v1.0) package. Monte Carlo Simulations We modeled a population of N individuals undergoing IVF. We assumed that each individual obtains M ~ p (M) usable (euploid) embryos, and that XX and XY embryos are equally likely to develop to be usable. Here, p represents an arbitrary probability distribution. We assigned each embryo its sex s i from the set { XX, XY } with equal probability. Each embryo is also assigned a score xi ~ p(x i | s i ) conditioned on its sex. In practice, p(x i | s i = XX) and p(x i | s i = XY) are derived empirically from the observed distributions of scores for each sex given a specific grading system. We then simulate the transfer of embryos sequentially, in descending order of score until a live birth is achieved. This is in keeping with the strategy of elective single embryo transfer (eSET), which is widely used across IVF clinics. When transferred, each embryo has a fate f i ∈ { Live Birth, No Live Birth } sampled from the bimodal distribution p(f i | x i ) conditioned on x i . It is important to note that this modelling approach assumes that embryos with a certain grade all have the same probability of live birth, regardless of sex, and that higher-order pregnancies do not happen. We repeat these steps R times, simulating R populations, and calculate the mean and 95% confidence intervals for the ratio of XX to XY (that is, XX births per XY birth) births using the percentile method. In our experiments, N = 50000, M ~ Binom(8, 0.3) and R = 1000. Live birth probability distributions were obtained from previously published work mapping scores to implantation rates (Hill et al. 2013 for TE grade (17), Petersen at al. 2023 (14) for KIDScore D3, CHLOE EQ score used directly, as per Erlich et al. 2022 (15)). Moreover, an additional batch of experiments were carried out under the assumption that all embryos have the same probability of live birth, regardless of sex or grade. Declarations Conflicts of Interest T.P. and C.H. are employees of Nuevo Healthcare Ltd, an AI-enabled fertility clinic. H.C.O. holds shares in Hertility Health Ltd, though this entity is unrelated to the work described here and is not judged by the authors to pose a competing interest. C.D. is a shareholder in The Evewell (Harley Street) Ltd. F.V, M.L. and C.O. declare no competing financial or non-financial interests. Competing Interests T.P. and C.H. are employees of Nuevo Healthcare Ltd, an AI-enabled fertility clinic. H.C.O. holds shares in Hertility Health Ltd, though this entity is unrelated to the work described here and is not judged by the authors to pose a competing interest. C.D. is a shareholder in The Evewell (Harley Street) Ltd. F.V, M.L. and C.O. declare no competing financial or non-financial interests. Funding Statement This work was supported by a PhD studentship awarded to T.P. from the British Heart Foundation, United Kingdom (FS/19/63/34902D). Author Contribution T.P., C.H. and H.C.O. conceived the study. C.H. and M.L. conceived the method and designed the algorithmic techniques. C.H. wrote the codes and performed the computational analysis with input from M.L. and F.V. C.D. and C.S.O. provided the Evewell time-lapse and PGT-A datasets. T.P and C.H. drafted the manuscript with input from H.C.O., C.S.O., C.D., M.L. and F.V. All the authors read the paper and suggested edits. H.C.O. supervised the project. Acknowledgement We thank the patients who participated in this study. We are grateful to Hertility Ltd. for providing the illustrations in Fig. 1 and to Fairtility Inc. for providing CHLOE EQ scores for research purposes. We thank Alexis Gkantiragas for helpful comments. Data Availability The embryo genetics and imaging datasets were not collected as part of this study, and were analysed retrospectively. The embryo-imaging datasets are available under restricted access owing to reasonable privacy and security concerns. Researchers can request access to the data which will be evaluated on a case-by-case basis. Any requests should be sent to the corresponding author ( [email protected] ). References Smeenk, J. et al. ART in Europe, 2020: results generated from European registries by ESHRE. Hum. Reprod. (2025). Gardner, D. K. & Schoolcraft, W. B. Culture and transfer of human blastocysts. Curr. Opin. Obstet. Gynecol. 11 (3), 307–311 (1999). Jiang, V. S. & Bormann, C. L. Artificial intelligence in the in vitro fertilization laboratory: a review of advancements over the last decade. Fertil. Steril. 120 (1), 17–23 (2023). Fernandez, E. I. et al. Artificial intelligence in the IVF laboratory: overview through the application of different types of algorithms for the classification of reproductive data. J. Assist. Reprod. Genet. 37 (10), 2359–2376 (2020). Rosenwaks, Z. Artificial intelligence in reproductive medicine: a fleeting concept or the wave of the future? Fertil. Steril. 114 (5), 905–907 (2020). Meseguer, M. et al. The use of morphokinetics as a predictor of embryo implantation. Hum. Reprod. 26 (10), 2658–2671 (2011). Mittermaier, M., Raza, M. M. & Kvedar, J. C. Bias in AI-based models for medical applications: challenges and mitigation strategies. NPJ Digit. Med. 6 (1), 113 (2023). Pergament, E., Toydemir, P. B. & Fiddler, M. Sex ratio: a biological perspective of 'Sex and the City'. Reprod. Biomed. Online . 5 (1), 43–46 (2002). D'Alfonso, A. et al. Sex ratio at birth: causes of variation and narrative review of literature. Minerva Obstet. Gynecol. 75 (2), 189–200 (2023). Alfarawati, S. et al. The relationship between blastocyst morphology, chromosomal abnormality, and embryo gender. Fertil. Steril. 95 (2), 520–524 (2011). Lyon, M. F. Sex chromatin and gene action in the mammalian X-chromosome. Am. J. Hum. Genet. 14 (2), 135–148 (1962). Dallemagne, M. et al. Oxidative stress differentially impacts male and female bovine embryos depending on the culture medium and the stress condition. Theriogenology 117 , 49–56 (2018). Gardner, D. K., Lane, M., Stevens, J., Schlenker, T. & Schoolcraft, W. B. Blastocyst score affects implantation and pregnancy outcome: towards a single blastocyst transfer. Fertil. Steril. 73 (6), 1155–1158 (2000). Petersen, B. M., Boel, M., Montag, M. & Gardner, D. K. Development of a generally applicable morphokinetic algorithm capable of predicting the implantation potential of embryos transferred on Day 3. Hum. Reprod. 31 (10), 2231–2244 (2016). Erlich, I. et al. Pseudo contrastive labeling for predicting IVF embryo developmental potential. Sci. Rep. 12 (1), 2488 (2022). Perlman, B. E. et al. Increased male live-birth rates after blastocyst-stage frozen-thawed embryo transfers compared with cleavage-stage frozen-thawed embryo transfers: a SART registry study. F S Rep. 2 (2), 161–165 (2021). Hill, M. J. et al. Trophectoderm grade predicts outcomes of single-blastocyst transfers. Fertil. Steril. 99 (5), 1283–9e1 (2013). Awadalla, M., Kim, A., Vestal, N., Ho, J. & Bendikson, K. Effect of Age and Embryo Morphology on Live Birth Rate After Transfer of Unbiopsied Blastocysts. JBRA Assist. Reprod. 25 (3), 373–382 (2021). Collins, G. S. & Moons, K. G. M. Reporting of artificial intelligence prediction models. Lancet 393 (10181), 1577–1579 (2019). McCoy, R. C. et al. Meiotic and mitotic aneuploidies drive arrest of in vitro fertilized human preimplantation embryos. Genome Med. 15 (1), 77 (2023). Additional Declarations Competing interest reported. T.P. and C.H. are employees of Nuevo Healthcare Ltd, an AI-enabled fertility clinic. H.C.O. holds shares in Hertility Health Ltd, though this entity is unrelated to the work described here and is not judged by the authors to pose a competing interest. C.D. is a shareholder in The Evewell (Harley Street) Ltd. F.V, M.L. and C.O. declare no competing financial or non-financial interests. Supplementary Files SupplementaryMaterials.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 10 Mar, 2026 Reviews received at journal 05 Mar, 2026 Reviews received at journal 09 Feb, 2026 Reviewers agreed at journal 08 Feb, 2026 Reviewers agreed at journal 05 Feb, 2026 Reviewers agreed at journal 01 Feb, 2026 Reviewers invited by journal 18 Dec, 2025 Editor invited by journal 04 Dec, 2025 Editor assigned by journal 02 Dec, 2025 Submission checks completed at journal 02 Dec, 2025 First submitted to journal 28 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8230063","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":563490069,"identity":"f9723739-cb2d-4bb5-9737-2da58ac00034","order_by":0,"name":"Teodora Popa","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Teodora","middleName":"","lastName":"Popa","suffix":""},{"id":563490070,"identity":"043266fb-ac04-409e-b889-5ba1d8eabacb","order_by":1,"name":"Chloe He","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYDACCeYGIHmAgZ+BgfEAWISHoBZGiBbJBhBJkhaDA8Rq4Z/d2PjgA8MdOeMbyQ8OMNTYMRicOUDAkjsHmw1nMDwzNruRBrToWDKDwdkG/FoMJBLbpHkYDiduu5EDdBgb0IXnCTgMrOUPw+H6zTNAWv4Rq4WB4XCCgQRQC2PbAcIOk7iR2GzYY3DYcMaZZwYHEvuSeSQJeZ9/RvLBBz8qDsvztyc/fPDhm50c35kEAi6DOA9KJxARkaNgFIyCUTAKiAAAQ25HRLEvzhoAAAAASUVORK5CYII=","orcid":"","institution":"University College London","correspondingAuthor":true,"prefix":"","firstName":"Chloe","middleName":"","lastName":"He","suffix":""},{"id":563490071,"identity":"09d8aa18-b9cd-45fd-b11a-251a2bbe4f8e","order_by":2,"name":"Christian S. Ottolini","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Christian","middleName":"S.","lastName":"Ottolini","suffix":""},{"id":563490072,"identity":"e23e0a8c-989e-4e19-9607-82920bab827a","order_by":3,"name":"Colin Davis","email":"","orcid":"","institution":"The Evewell","correspondingAuthor":false,"prefix":"","firstName":"Colin","middleName":"","lastName":"Davis","suffix":""},{"id":563490073,"identity":"dda31c71-e56d-492e-958f-13a53833c781","order_by":4,"name":"Matthew Lau","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Matthew","middleName":"","lastName":"Lau","suffix":""},{"id":563490074,"identity":"a7e9e3ee-0608-47fd-bd1b-c7c49dae9c03","order_by":5,"name":"Francisco Vasconcelos","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Francisco","middleName":"","lastName":"Vasconcelos","suffix":""},{"id":563490075,"identity":"6ddc726f-287b-431b-a2d4-f30628f486d6","order_by":6,"name":"Helen C. O'Neill","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Helen","middleName":"C.","lastName":"O'Neill","suffix":""}],"badges":[],"createdAt":"2025-11-28 11:53:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8230063/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8230063/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98819225,"identity":"4e11915a-3bd2-45c4-ae91-85f3d3fb7f35","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":764787,"visible":true,"origin":"","legend":"","description":"","filename":"SUBMISSION2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/da31d981c37e7f4554b6a12d.docx"},{"id":98819217,"identity":"ad4a57da-b2f5-4b8c-83fa-4290a8f7cf92","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9136,"visible":true,"origin":"","legend":"","description":"","filename":"e80d8a59c1e44a85b69de0fb6bf84d2c.json","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/ed70b7f6b3083147dff47104.json"},{"id":99307607,"identity":"fc212aaa-aae9-4057-9bcb-9543f20d0b50","added_by":"auto","created_at":"2025-12-31 16:06:26","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18626,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/0a0187205f4c0e575f867dd5.docx"},{"id":98819222,"identity":"35cb3750-82b3-4d17-8fe7-07736c5f6228","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62298,"visible":true,"origin":"","legend":"","description":"","filename":"e80d8a59c1e44a85b69de0fb6bf84d2c1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/640b2353337e78becb61bdd4.xml"},{"id":99307460,"identity":"68939779-6894-48d0-9dec-ac1d1799f5a7","added_by":"auto","created_at":"2025-12-31 16:06:17","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":86117,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/c57b9d9fe1411bfb74e51360.png"},{"id":98819221,"identity":"f5ae7608-e8fa-44ab-954a-d8d743b10fba","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":58023,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/9d3fae6a7bad088006eab765.png"},{"id":98819223,"identity":"7afff779-a9c4-40c3-bb73-b717e51e0688","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60152,"visible":true,"origin":"","legend":"","description":"","filename":"e80d8a59c1e44a85b69de0fb6bf84d2c1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/a2f84f45551f56919484c67c.xml"},{"id":98819227,"identity":"043ca6f2-2b3f-4956-9406-cc130adddd79","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"html","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72367,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/c5b75675aa942cd863a7adce.html"},{"id":99307647,"identity":"43353350-d965-4f91-be95-1b92faf43d98","added_by":"auto","created_at":"2025-12-31 16:06:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":402837,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSex bias in embryo grading systems and potential demographic impact. \u003c/strong\u003eSchematic overview showing how three widely-used embryo grading approaches: the Gardner system (a manual grading system based on blastocyst morphology)\u003csup\u003e(2, 13)\u003c/sup\u003e, KIDScore D3™(Vitrolife) (a classical machine learning model based on annotated morphokinetic events)\u003csup\u003e(14)\u003c/sup\u003e, and CHLOE EQ\u003csup\u003eTM\u003c/sup\u003e (Fairtility) (a deep learning model that analyses time-lapse videos without manual feature selection)\u003csup\u003e(15)\u003c/sup\u003e differentially score XX and XY embryos. Manual grading and KIDScore D3™(assigned higher scores to XY embryos, while CHLOE EQ\u003csup\u003eTM\u003c/sup\u003e showed no sex bias. Such sex biases in embryo grading could skew the sex ratio at birth towards males, especially when applied at scale in high-IVF-utilisation populations.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/609f64c11f856cab4c12081e.png"},{"id":99307571,"identity":"cdf1f545-fe4b-49c8-b9b8-c4712837c42b","added_by":"auto","created_at":"2025-12-31 16:06:25","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":303944,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEmbryo score distributions, model performance, and simulation workflow.\u003cbr\u003e\na)\u003c/strong\u003e Distributions of embryo scores across all embryos, colour-coded by sex, for CHLOE EQ\u003csup\u003eTM\u003c/sup\u003e (deep learning; top left), KIDScore D3™ (classical machine learning; top right), and manual grading (bottom).\u003cbr\u003e\n\u003cstrong\u003eb)\u003c/strong\u003e Equivalent score distributions restricted to euploid embryos.\u003cbr\u003e\n\u003cstrong\u003ec)\u003c/strong\u003e Distributions of kinetic parameters (measured in hours post insemination, HPI) by sex.\u003cbr\u003e\n\u003cstrong\u003ed)\u003c/strong\u003e Comparison of test set performance of machine learning models in predicting embryo sex from grading scores.\u003cbr\u003e\n\u003cstrong\u003ee)\u003c/strong\u003e Monte Carlo simulation process for a representative patient with 4 embryos. We first assign sexes to each embryo. Embryos are then assigned scores, sorted and sequentially transferred from highest to lowest score until a live birth is achieved (here, with the second embryo).\u003cbr\u003e\n\u003cstrong\u003ef)\u003c/strong\u003e Results from Monte Carlo simulations assuming correlation between grades and live birth.\u003cbr\u003e\n\u003cstrong\u003eg)\u003c/strong\u003e Results from Monte Carlo simulations assuming no correlation between grades and live birth.\u0026nbsp;\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/a75dab05bf31103645e15d43.png"},{"id":99788047,"identity":"21fec8fe-6de4-43d7-a26d-b992fbec4f4d","added_by":"auto","created_at":"2026-01-08 12:43:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1329716,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/4a5d4d7a-b450-460e-b37e-a8d6e2e63be2.pdf"},{"id":98819219,"identity":"508b0d3b-3ebb-4743-981f-fd601ac1977c","added_by":"auto","created_at":"2025-12-22 17:07:07","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":18626,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8230063/v1/cc563a08f4ba26ca387886c7.docx"}],"financialInterests":"Competing interest reported. T.P. and C.H. are employees of Nuevo Healthcare Ltd, an AI-enabled fertility clinic. H.C.O. holds shares in Hertility Health Ltd, though this entity is unrelated to the work described here and is not judged by the authors to pose a competing interest. C.D. is a shareholder in The Evewell (Harley Street) Ltd. F.V, M.L. and C.O. declare no competing financial or non-financial interests.","formattedTitle":"Embryo selection tools in IVF can favour XY embryos: implications for equitable reproductive AI","fulltext":[{"header":"Main text","content":"\u003cp\u003eGlobal fertility rates are at an all-time low and have declined sharply over recent decades, with many individuals delaying parenthood into ages associated with reduced natural fertility. As a result, reliance on assisted reproductive technologies (ART) such as \u003cem\u003ein vitro\u003c/em\u003e fertilisation (IVF) has increased, accounting for up to 8.5% of all live births in Greece, 7.9% in Spain and around 9% in Denmark, among the highest rates worldwide\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e)\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eA critical stage in IVF is embryo selection: identifying the embryo(s) with the highest chance of implantation and live birth. Traditionally, this has been performed manually by embryologists using morphological criteria, most often the Gardner grading system\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e)\u003c/sup\u003e, assessing the inner cell mass (ICM) and trophectoderm (TE) quality at the blastocyst stage\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e)\u003c/sup\u003e. More recently, time-lapse incubators and AI-based grading algorithms have been introduced to improve selection accuracy, reduce inter-operator variability, and increase efficiency\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e)\u003c/sup\u003e. These systems integrate developmental timings (known as \u003cem\u003ekinetics\u003c/em\u003e) with morphological data, a combined approach referred to as \u003cem\u003emorphokinetics\u003c/em\u003e\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e)\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eHowever, as in other domains, AI in reproductive medicine may encode and propagate biases present in training data or arising from model design. Algorithmic bias in medical AI has been documented across sex, race, and socioeconomic lines in diagnostics and treatment recommendations\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e)\u003c/sup\u003e, but its potential role in embryo selection has not been investigated systematically. This gap is important: if grading tools systematically favour one sex over the other, the cumulative effect in high-IVF-utilisation populations could alter the sex ratio at birth, with biological, social, and ethical implications.\u003c/p\u003e\n\u003cp\u003eIt is well established that natural conception yields a modest male bias at birth (\u0026asymp;\u0026thinsp;51\u0026ndash;52% XY; sex ratio\u0026thinsp;\u0026asymp;\u0026thinsp;1.05)\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e)\u003c/sup\u003e, and IVF may amplify this baseline skew. Some studies have found that XY embryos cleave and reach the blastocyst stage faster than XX embryos\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e)\u003c/sup\u003e, potentially due to the additional developmental demands of X-chromosome inactivation in XX embryos\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e)\u003c/sup\u003e. Extended in vitro culture may also impose metabolic and oxidative stress that disproportionately affects XX embryos\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e)\u003c/sup\u003e. Together, these morphokinetic differences may favour XY embryos and be inadvertently reinforced by embryo selection systems designed under the assumption of sex-neutral development, even in the absence of deliberate sex selection.\u003c/p\u003e\n\u003cp\u003eTo investigate this, we retrospectively evaluated three widely used embryo grading approaches: the Gardner system (a manual grading system based on blastocyst morphology)\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/sup\u003e, KIDScore D3\u0026trade;(Vitrolife) (a classical machine learning model based on annotated morphokinetic events)\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e)\u003c/sup\u003e, and CHLOE EQ\u0026trade; (a deep learning model that analyses time-lapse videos without manual feature selection)\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e)\u003c/sup\u003e. Together, these three systems represent the spectrum of current clinical practice, spanning traditional human-led techniques to modern automated approaches.\u003c/p\u003e\n\u003cp\u003eWe assessed whether each tool assigned different scores to XX vs XY embryos, and modelled the potential demographic impact if such biases were applied at scale (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSex Bias in Embryo Grading\u003c/strong\u003e - In the full cohort of 1,334 embryos, manual morphological grading assigned \u0026lsquo;good\u0026rsquo; grades more often to XY embryos (462/668) than XX embryos (351/614; \u0026chi;\u0026sup2;=19.843, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). KIDScore D3\u0026trade; also scored XY embryos (M\u0026thinsp;\u0026plusmn;\u0026thinsp;SD: 4.182\u0026thinsp;\u0026plusmn;\u0026thinsp;1.353, n\u0026thinsp;=\u0026thinsp;642) higher than XX (4.022\u0026thinsp;\u0026plusmn;\u0026thinsp;1.420, n\u0026thinsp;=\u0026thinsp;692; U\u0026thinsp;=\u0026thinsp;207604, p\u0026thinsp;=\u0026thinsp;0.0182). CHLOE EQ\u0026trade; scores showed no difference between XX (0.787\u0026thinsp;\u0026plusmn;\u0026thinsp;p.276, n\u0026thinsp;=\u0026thinsp;628) and XY embryos (0.802\u0026thinsp;\u0026plusmn;\u0026thinsp;0.254, n\u0026thinsp;=\u0026thinsp;679; U\u0026thinsp;=\u0026thinsp;204621, p\u0026thinsp;=\u0026thinsp;0.208) (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ea).\u003c/p\u003e\n\u003cp\u003eWhen restricted to euploid embryos, manual grading again favoured XY embryos (\u0026chi;\u0026sup2;=15.839, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), but KIDScore D3\u0026trade; and CHLOE EQ\u0026trade; showed no statistically significant difference (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eb).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDrivers of Bias\u003c/strong\u003e - Post-hoc analyses indicated that the disparity in manual grading was driven largely by higher TE grades in XY embryos (2.809\u0026thinsp;\u0026plusmn;\u0026thinsp;0.628) than XX embryos (2.656\u0026thinsp;\u0026plusmn;\u0026thinsp;0.623, n\u0026thinsp;=\u0026thinsp;616; U\u0026thinsp;=\u0026thinsp;179584, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), with no significant differences in ICM scores or other morphokinetic parameters (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ec, Supplementary Table\u0026nbsp;1). Aneuploidy rates between XX (73/642) and XY embryos (75/692) did not differ (\u003cem\u003eX\u003c/em\u003e\u003csup\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sup\u003e (1, N\u0026thinsp;=\u0026thinsp;1334)\u0026thinsp;=\u0026thinsp;0.0958, p\u0026thinsp;=\u0026thinsp;0.757).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredicting Sex from Embryo Grades\u003c/strong\u003e - Machine learning models trained to predict embryo sex from TE grade and KIDScore D3\u0026trade;achieved low predictive performance (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ed), confirming that while sex differences in grades exist, grades cannot be used to reliably infer embryo sex for individual embryos.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModelling of Potential Population Impacts -\u003c/strong\u003e Monte Carlo simulations (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003ee) suggest that in high-IVF-utilisation countries, even small grading biases could modestly skew the sex ratio at birth. In simulations of IVF birth populations that assumed correlation between grades and live birth (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003ef), embryo selection skewed the sex ratio toward XY for TE grade (XX/XY\u0026thinsp;=\u0026thinsp;0.886, 95% CI [0.658,0.904]), KIDScore D3\u0026trade; (XX/XY\u0026thinsp;=\u0026thinsp;0.900 [0.881,0.920]), and CHLOE EQ\u0026trade; (XX/XY\u0026thinsp;=\u0026thinsp;0.940 [0.924, 0.959]). When correlation was not assumed (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eg), the skew persisted: TE grade (XX/XY\u0026thinsp;=\u0026thinsp;0.951, 95% CI [0.930, 0.971]), KIDScore D3\u0026trade; (XX/XY\u0026thinsp;=\u0026thinsp;0.950, 95% CI [0.929, 0.972]), and CHLOE EQ\u0026trade; (XX/XY\u0026thinsp;=\u0026thinsp;0.984, 95% CI [0.963, 1.005]).\u003c/p\u003e\n\u003cp\u003eTo our knowledge, this analysis of over 1,300 human embryos with known sex, graded using both manual and AI-based tools, is the largest study to date assessing sex bias in both manual and AI-based embryo selection methods. We show that manual Gardner grading and the KIDScore D3\u0026trade; algorithm both assign higher scores to XY embryos, whereas the CHLOE EQ\u003csup\u003e\u0026trade;\u003c/sup\u003e deep learning model does not.\u003c/p\u003e\n\u003cp\u003eThese findings highlight two key points. First, bias is not unique to AI: manual grading by experienced embryologists showed stronger sex bias than either AI model. Second, sex bias in embryo selection is not inevitable and can be avoided.\u003c/p\u003e\n\u003cp\u003eAn obvious starting point for interpreting these results is biology, as the differences observed may reflect intrinsic sex-based developmental dynamics. Some works suggest that XY embryos develop more quickly and may be more resilient to extended culture, traits that correlate with implantation and live birth potential\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e)\u003c/sup\u003e. Clinical data support this, with male-skewed birth ratios more pronounced after blastocyst-stage than cleavage-stage transfer (~\u0026thinsp;54\u0026ndash;58% vs\u0026thinsp;~\u0026thinsp;50\u0026ndash;52%)\u003csup\u003e(16)\u003c/sup\u003e, suggesting that sex-related factors become more influential beyond the cleavage stage and are mirrored in real outcomes.\u003c/p\u003e\n\u003cp\u003eHowever, the absence of bias in CHLOE EQ\u0026trade; shows that, although biological factors may contribute, sex-neutral embryo selection is achievable. The reason for this may lie in algorithm design. Conventional morphological grading relies on coarse visual appraisal of blastocyst structures. While blastocyst morphology and grading do correlate with live birth rates\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e)\u003c/sup\u003e, the correlation is imperfect, leaving room for confounding. Following this reasoning, the skew observed in manual grading may have emerged unintentionally over decades of human-guided refinement of grading practices, rather than from deliberate sex selection.\u003c/p\u003e\n\u003cp\u003eIn contrast, CHLOE EQ\u0026trade; employs a convolutional neural network explicitly optimised for the prediction of implantation. This setup enables CHLOE EQ\u0026trade; to discover its own features, focusing on direct predictive value rather than those that are simply accessible to human observers. In doing so, it may have learnt sex-invariant markers of embryo viability or, if sex-linked developmental differences exist, developed parallel strategies for each sex. Interpretability tools could then reveal the features driving these patterns and generate hypotheses on the role of sex differences in embryo development.\u003c/p\u003e\n\u003cp\u003eSimilarly, the KIDScore D3\u0026trade;algorithm relies on human-provided developmental timings, a feature that, as previously discussed, may introduce sex bias. Yet, as it was also optimised directly for implantation prediction, it acts as a middle way between manual grading and CHLOE EQ\u0026trade;, consistent with the intermediate level of bias observed.\u003c/p\u003e\n\u003cp\u003eAs ART becomes more common, particularly in countries with high IVF utilisation rate, even modest biases in selection tools could translate into measurable demographic shifts. Unlike overt sex selection, which is regulated in most jurisdictions, these biases are unintended and therefore largely invisible to clinicians and regulators. This underscores the need for systematic bias evaluation not only of AI systems but also of long-established manual grading practices.\u003c/p\u003e\n\u003cp\u003eOur study has several limitations. It was conducted in a single centre, and most embryos were cryopreserved, leaving us with limited follow-up on live births. While simulations suggest that biases in trophectoderm grading could modestly skew sex ratios at birth, our modelling was necessarily simplistic and did not account for higher-order pregnancies, potential sex differences in embryo viability, or the influence of clinician judgement in transfer decisions. Larger, multi-centre datasets with live birth outcomes will be needed to confirm and extend these findings.\u003c/p\u003e\n\u003cp\u003eA further limitation is that KIDScore D3\u0026trade;only incorporates morphokinetic events up to day 3 and does not capture blastocyst-stage features, unlike Gardner grading or CHLOE EQ\u0026trade;. More advanced algorithms such as KIDScore D5\u0026trade;(Vitrolife) or iDAScore\u0026reg; (Vitrolife) could not be assessed, as they require access to proprietary incubator software and their methodologies are not publicly available. This restricted our ability to benchmark across the full range of current embryo selection systems.\u003c/p\u003e\n\u003cp\u003eDespite these caveats, our study provides the first large-scale evidence that embryo selection tools used in IVF, whether manual or AI-based, can systematically favour XY embryos, and the first attempt to model the possible demographic consequences of such bias. These findings highlight the need to routinely assess risk of bias in embryo grading systems and to report clinical prediction models with greater transparency in their design and training datasets\u003csup\u003e(\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e)\u003c/sup\u003e. Incorporating fairness constraints during model development offers a pathway to more equitable selection. Ultimately, ensuring that embryo grading methods are both effective and equitable will be essential to prevent unintended demographic shifts and to uphold the fairness of ART practice worldwide.\u003c/p\u003e\n\u003cp\u003eFor panels \u003cstrong\u003ea-c\u003c/strong\u003e, slight variations in sample size reflect missing annotations, and distributions are normalised such that the area under the curve for each sex equals 1. Kinetic features include: \u003cem\u003etX\u003c/em\u003e, time to reach the X-cell stage (e.g., t2\u0026thinsp;=\u0026thinsp;time to 2-cell stage); \u003cem\u003etM\u003c/em\u003e, time to morula; \u003cem\u003etSB\u003c/em\u003e, time to start of blastulation; \u003cem\u003etB\u003c/em\u003e, time to blastocyst; \u003cem\u003etEB\u003c/em\u003e, time to expanded blastocyst; and \u003cem\u003ecc1-3\u003c/em\u003e, first to third cell cycles. For panels \u003cstrong\u003ef\u003c/strong\u003e and \u003cstrong\u003eg\u003c/strong\u003e, error bars represent 95% confidence intervals.\u003c/p\u003e"},{"header":"Online Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eEthics statement\u003c/h2\u003e \u003cp\u003eThis study was conducted in accordance with relevant guidelines and regulations. An electronic ethics application was submitted to the UK Integrated Research Application System (IRAS project ID: 328309). The Committee confirmed that formal review was not required, as the project involved secondary analysis of anonymised data collected during routine clinical care, in which individual patients could not be identified.\u003c/p\u003e \u003cp\u003eAll patients undergoing IVF at the clinic provided written informed consent for the use of anonymised data in research, quality control, and clinical process development. All procedures were undertaken under the clinic\u0026rsquo;s HFEA licence, in line with UK regulatory requirements and best practice for assisted reproduction.\u003c/p\u003e \u003cp\u003eEmbryo imaging was performed solely as part of standard clinical care. Data were anonymised such that patient identities could not be ascertained directly or indirectly, and investigators neither contacted nor attempted to re-identify participants. Importantly, the clinical team did not have access to embryo sex information at any stage, ensuring that no intentional bias could have influenced clinical decision-making.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy design and dataset\u003c/h3\u003e\n\u003cp\u003eWe retrospectively analysed 1,411 diploid embryos from 398 IVF/ICSI cycles at a single UK centre (2018\u0026ndash;2021). Embryos with other karyotypes (n\u0026thinsp;=\u0026thinsp;49) and those not derived from normally fertilised zygotes (n\u0026thinsp;=\u0026thinsp;28) were excluded, leaving 1,334 embryos. The mean maternal age at oocyte retrieval was 38.4\u0026thinsp;\u0026plusmn;\u0026thinsp;3.2 years. Most cycles (n\u0026thinsp;=\u0026thinsp;947) used intracytoplasmic sperm injection (ICSI); the remainder used conventional IVF. All followed an antagonist stimulation protocol with oocyte retrieval 36 h after trigger.\u003c/p\u003e \u003cp\u003eEmbryos were cultured in Geri time-lapse incubators (Genea Biomedx), enabling continuous imaging without removal from culture. Time-lapse videos were exported for manual and AI-based grading.\u003c/p\u003e \u003cp\u003eEmbryos were biopsied on day 5, 6, or 7 for preimplantation genetic testing for aneuploidy (PGT-A). Biopsied cells were analysed by next-generation sequencing (PGTai 1.0 and PGTai 2.0; CooperSurgical Inc). In line with standard clinical reporting, results disclosed to the clinical team included only chromosomes 1\u0026ndash;22. Sex chromosome data were not disclosed to clinicians or patients at any point. Instead, anonymised identifiers were generated, and only the research team had restricted access to sex chromosome information for research purposes. The key linking patient identities to research codes was stored separately in a password-protected, access-restricted file on the clinic\u0026rsquo;s secure drive. Access was strictly limited to authorised personnel.\u003c/p\u003e\n\u003ch3\u003eEmbryo grading systems\u003c/h3\u003e\n\u003cp\u003eFor the manual morphological grading, blastocysts were graded for inner cell mass (ICM) and trophectoderm (TE) quality using a modified Gardner system (2, 13). Grades were binarised into \u0026ldquo;good\u0026rdquo; and \u0026ldquo;poor\u0026rdquo; following McCoy et al. (2023) (20). The computation of the KIDScore D3\u0026trade; decision tree algorithm was implemented according to ​Petersen et al., 2016​ (14) making use of kinetic annotations from CHLOE EQ\u0026trade;. Code for our implementation can be found at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/chlohe/embryo-sex\u003c/span\u003e\u003cspan address=\"https://github.com/chlohe/embryo-sex\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The deep learning-based CHLOE EQ score was assessed using CHLOE (Fairtility).\u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eComparisons between XX and XY embryos for each grading method were performed using two-tailed Mann\u0026ndash;Whitney U tests for continuous scores and χ\u0026sup2; tests for binary grade categories. Post-hoc analyses examined morphokinetic variables (timings of pronuclear fading, cleavage stages, morulation, blastulation) and ploidy status. ICM and TE grades were converted to numerical scales (D\u0026thinsp;=\u0026thinsp;1 to A\u0026thinsp;=\u0026thinsp;4) for analysis.\u003c/p\u003e \u003cp\u003eFor each test, embryos missing all required morphokinetic annotations were excluded. For example, we did not consider any embryos that did not have both a TE and ICM grade when analysing the Gardner gradings. All statistical analysis was carried out using Python (v3.8.10) with the Pandas (v1.3.3) and SciPy (v1.7.1) packages.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSex prediction from embryo grades\u003c/h3\u003e\n\u003cp\u003eWe trained and evaluated four widely-used machine learning models on the task of predicting whether an embryo is XX or XY from variables showing significant associations (TE grade and KIDScore D3\u0026trade;). The models included a decision tree, a random forest, multinomial logistic regression and a multilayer perceptron. The models were compared against two simple baseline models based on simple cut-offs for TE grade and KIDScore D3\u0026trade;.\u003c/p\u003e \u003cp\u003eThese baseline models took the form:\u003c/p\u003e \u003cp\u003e \u003cem\u003eH\u003c/em\u003e(\u003cem\u003ex\u003c/em\u003e)\u0026thinsp;=\u0026thinsp;XX (if x \u0026oplus; \u003cem\u003eX\u003c/em\u003e)\u003c/p\u003e \u003cp\u003eXY (otherwise)\u003c/p\u003e \u003cp\u003ewhere x is the embryo score, \u003cem\u003eX\u003c/em\u003e is some cutoff value and \u0026oplus; is a comparison operator. We treat both \u003cem\u003eX\u003c/em\u003e and \u0026oplus; as hyperparameters.\u003c/p\u003e \u003cp\u003eThe dataset was split into training (N\u0026thinsp;=\u0026thinsp;875) and test (N\u0026thinsp;=\u0026thinsp;375) sets. Hyperparameters for each model were computed by grid search and 5-fold cross validation on the training set (a full list of hyperparameters considered can be found in Supplementary Table\u0026nbsp;2). Each model was evaluated on the test set using the accuracy, sensitivity, specificity and area under the receiver operating characteristic curve (AUC) metrics, with XY being the \u0026lsquo;positive\u0026rsquo; outcome. All predictive modelling was carried out using Python (v3.8.10) with the Sci-kit Learn (v1.0) package.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eMonte Carlo Simulations\u003c/h2\u003e \u003cp\u003eWe modeled a population of \u003cem\u003eN\u003c/em\u003e individuals undergoing IVF. We assumed that each individual obtains \u003cem\u003eM\u0026thinsp;~\u0026thinsp;p\u003c/em\u003e(M) usable (euploid) embryos, and that XX and XY embryos are equally likely to develop to be usable. Here, \u003cem\u003ep\u003c/em\u003e represents an arbitrary probability distribution.\u003c/p\u003e \u003cp\u003eWe assigned each embryo its sex s\u003csub\u003ei\u003c/sub\u003e from the set {\u003cem\u003eXX, XY\u003c/em\u003e} with equal probability. Each embryo is also assigned a score xi\u0026thinsp;~\u0026thinsp;p(x\u003csub\u003ei\u003c/sub\u003e | s\u003csub\u003ei\u003c/sub\u003e) conditioned on its sex. In practice, p(x\u003csub\u003ei\u003c/sub\u003e | s\u003csub\u003ei\u003c/sub\u003e = XX) and p(x\u003csub\u003ei\u003c/sub\u003e | s\u003csub\u003ei\u003c/sub\u003e = XY) are derived empirically from the observed distributions of scores for each sex given a specific grading system.\u003c/p\u003e \u003cp\u003eWe then simulate the transfer of embryos sequentially, in descending order of score until a live birth is achieved. This is in keeping with the strategy of elective single embryo transfer (eSET), which is widely used across IVF clinics. When transferred, each embryo has a fate f\u003csub\u003ei\u003c/sub\u003e \u0026isin; {\u003cem\u003eLive Birth, No Live Birth\u003c/em\u003e} sampled from the bimodal distribution p(f\u003csub\u003ei\u003c/sub\u003e | x\u003csub\u003ei\u003c/sub\u003e) conditioned on x\u003csub\u003ei\u003c/sub\u003e. It is important to note that this modelling approach assumes that embryos with a certain grade all have the same probability of live birth, regardless of sex, and that higher-order pregnancies do not happen.\u003c/p\u003e \u003cp\u003eWe repeat these steps \u003cem\u003eR\u003c/em\u003e times, simulating \u003cem\u003eR\u003c/em\u003e populations, and calculate the mean and 95% confidence intervals for the ratio of XX to XY (that is, XX births per XY birth) births using the percentile method.\u003c/p\u003e \u003cp\u003eIn our experiments, \u003cem\u003eN\u003c/em\u003e\u0026thinsp;=\u0026thinsp;50000, M\u0026thinsp;~\u0026thinsp;Binom(8, 0.3) and \u003cem\u003eR\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1000. Live birth probability distributions were obtained from previously published work mapping scores to implantation rates (Hill et al. 2013 for TE grade (17), Petersen at al. 2023 (14) for KIDScore D3, CHLOE EQ score used directly, as per Erlich et al. 2022 (15)). Moreover, an additional batch of experiments were carried out under the assumption that all embryos have the same probability of live birth, regardless of sex or grade.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflicts of Interest\u003c/h2\u003e \u003cp\u003eT.P. and C.H. are employees of Nuevo Healthcare Ltd, an AI-enabled fertility clinic. H.C.O. holds shares in Hertility Health Ltd, though this entity is unrelated to the work described here and is not judged by the authors to pose a competing interest. C.D. is a shareholder in The Evewell (Harley Street) Ltd. F.V, M.L. and C.O. declare no competing financial or non-financial interests.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003cp\u003eT.P. and C.H. are employees of Nuevo Healthcare Ltd, an AI-enabled fertility clinic. H.C.O. holds shares in Hertility Health Ltd, though this entity is unrelated to the work described here and is not judged by the authors to pose a competing interest. C.D. is a shareholder in The Evewell (Harley Street) Ltd. F.V, M.L. and C.O. declare no competing financial or non-financial interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding Statement\u003c/h2\u003e \u003cp\u003eThis work was supported by a PhD studentship awarded to T.P. from the British Heart Foundation, United Kingdom (FS/19/63/34902D).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eT.P., C.H. and H.C.O. conceived the study. C.H. and M.L. conceived the method and designed the algorithmic techniques. C.H. wrote the codes and performed the computational analysis with input from M.L. and F.V. C.D. and C.S.O. provided the Evewell time-lapse and PGT-A datasets. T.P and C.H. drafted the manuscript with input from H.C.O., C.S.O., C.D., M.L. and F.V. All the authors read the paper and suggested edits. H.C.O. supervised the project.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eWe thank the patients who participated in this study. We are grateful to Hertility Ltd. for providing the illustrations in Fig. 1 and to Fairtility Inc. for providing CHLOE EQ scores for research purposes. We thank Alexis Gkantiragas for helpful comments.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe embryo genetics and imaging datasets were not collected as part of this study, and were analysed retrospectively. The embryo-imaging datasets are available under restricted access owing to reasonable privacy and security concerns. Researchers can request access to the data which will be evaluated on a case-by-case basis. Any requests should be sent to the corresponding author ([email protected]).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSmeenk, J. et al. ART in Europe, 2020: results generated from European registries by ESHRE. \u003cem\u003eHum. Reprod.\u003c/em\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGardner, D. K. \u0026amp; Schoolcraft, W. B. Culture and transfer of human blastocysts. \u003cem\u003eCurr. Opin. Obstet. Gynecol.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e (3), 307\u0026ndash;311 (1999).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang, V. S. \u0026amp; Bormann, C. L. Artificial intelligence in the in vitro fertilization laboratory: a review of advancements over the last decade. \u003cem\u003eFertil. Steril.\u003c/em\u003e \u003cb\u003e120\u003c/b\u003e (1), 17\u0026ndash;23 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFernandez, E. I. et al. Artificial intelligence in the IVF laboratory: overview through the application of different types of algorithms for the classification of reproductive data. \u003cem\u003eJ. Assist. Reprod. Genet.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e (10), 2359\u0026ndash;2376 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosenwaks, Z. Artificial intelligence in reproductive medicine: a fleeting concept or the wave of the future? \u003cem\u003eFertil. Steril.\u003c/em\u003e \u003cb\u003e114\u003c/b\u003e (5), 905\u0026ndash;907 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeseguer, M. et al. The use of morphokinetics as a predictor of embryo implantation. \u003cem\u003eHum. Reprod.\u003c/em\u003e \u003cb\u003e26\u003c/b\u003e (10), 2658\u0026ndash;2671 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMittermaier, M., Raza, M. M. \u0026amp; Kvedar, J. C. Bias in AI-based models for medical applications: challenges and mitigation strategies. \u003cem\u003eNPJ Digit. Med.\u003c/em\u003e \u003cb\u003e6\u003c/b\u003e (1), 113 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePergament, E., Toydemir, P. B. \u0026amp; Fiddler, M. Sex ratio: a biological perspective of 'Sex and the City'. \u003cem\u003eReprod. Biomed. Online\u003c/em\u003e. \u003cb\u003e5\u003c/b\u003e (1), 43\u0026ndash;46 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eD'Alfonso, A. et al. Sex ratio at birth: causes of variation and narrative review of literature. \u003cem\u003eMinerva Obstet. Gynecol.\u003c/em\u003e \u003cb\u003e75\u003c/b\u003e (2), 189\u0026ndash;200 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlfarawati, S. et al. The relationship between blastocyst morphology, chromosomal abnormality, and embryo gender. \u003cem\u003eFertil. Steril.\u003c/em\u003e \u003cb\u003e95\u003c/b\u003e (2), 520\u0026ndash;524 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLyon, M. F. Sex chromatin and gene action in the mammalian X-chromosome. \u003cem\u003eAm. J. Hum. Genet.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e (2), 135\u0026ndash;148 (1962).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDallemagne, M. et al. Oxidative stress differentially impacts male and female bovine embryos depending on the culture medium and the stress condition. \u003cem\u003eTheriogenology\u003c/em\u003e \u003cb\u003e117\u003c/b\u003e, 49\u0026ndash;56 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGardner, D. K., Lane, M., Stevens, J., Schlenker, T. \u0026amp; Schoolcraft, W. B. Blastocyst score affects implantation and pregnancy outcome: towards a single blastocyst transfer. \u003cem\u003eFertil. Steril.\u003c/em\u003e \u003cb\u003e73\u003c/b\u003e (6), 1155\u0026ndash;1158 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePetersen, B. M., Boel, M., Montag, M. \u0026amp; Gardner, D. K. Development of a generally applicable morphokinetic algorithm capable of predicting the implantation potential of embryos transferred on Day 3. \u003cem\u003eHum. Reprod.\u003c/em\u003e \u003cb\u003e31\u003c/b\u003e (10), 2231\u0026ndash;2244 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eErlich, I. et al. Pseudo contrastive labeling for predicting IVF embryo developmental potential. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e (1), 2488 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePerlman, B. E. et al. Increased male live-birth rates after blastocyst-stage frozen-thawed embryo transfers compared with cleavage-stage frozen-thawed embryo transfers: a SART registry study. \u003cem\u003eF S Rep.\u003c/em\u003e \u003cb\u003e2\u003c/b\u003e (2), 161\u0026ndash;165 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHill, M. J. et al. Trophectoderm grade predicts outcomes of single-blastocyst transfers. \u003cem\u003eFertil. Steril.\u003c/em\u003e \u003cb\u003e99\u003c/b\u003e (5), 1283\u0026ndash;9e1 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAwadalla, M., Kim, A., Vestal, N., Ho, J. \u0026amp; Bendikson, K. Effect of Age and Embryo Morphology on Live Birth Rate After Transfer of Unbiopsied Blastocysts. \u003cem\u003eJBRA Assist. Reprod.\u003c/em\u003e \u003cb\u003e25\u003c/b\u003e (3), 373\u0026ndash;382 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins, G. S. \u0026amp; Moons, K. G. M. Reporting of artificial intelligence prediction models. \u003cem\u003eLancet\u003c/em\u003e \u003cb\u003e393\u003c/b\u003e (10181), 1577\u0026ndash;1579 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcCoy, R. C. et al. Meiotic and mitotic aneuploidies drive arrest of in vitro fertilized human preimplantation embryos. \u003cem\u003eGenome Med.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (1), 77 (2023).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8230063/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8230063/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAssisted reproductive technologies (ART) are increasingly used worldwide, with up to 8.5% of babies in some countries conceived via \u003cem\u003ein vitro\u003c/em\u003e fertilisation (IVF). Embryo selection, a critical step in IVF, is shifting from manual morphology-based to artificial intelligence (AI)-based assessment. While these systems use outcome data to select embryos with the highest likelihood of pregnancy, their potential to introduce unintended bias has not been evaluated at scale. We analysed 1,334 human embryos with known sex (from preimplantation genetic testing), graded by two widely-used AI models (KIDScore D3\u0026trade;, CHLOE EQ\u0026trade;) and standard manual morphology-based Gardner grading. Manual grading and the traditional AI model (KIDScore D3\u0026trade;) assigned higher scores to XY embryos; the deep-learning model (CHLOE EQ\u0026trade;) did not. Simulations suggest such biases could modestly skew sex ratios at birth in high-IVF-utilisation settings, which is of global concern. These findings highlight the need for systematic bias evaluation in embryo grading and demonstrate that algorithmic selection without sex bias is achievable with appropriate design.\u003c/p\u003e","manuscriptTitle":"Embryo selection tools in IVF can favour XY embryos: implications for equitable reproductive AI","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-22 17:07:03","doi":"10.21203/rs.3.rs-8230063/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-10T12:29:54+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-05T20:43:40+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-09T16:50:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"223690326943741560932106030546861169376","date":"2026-02-08T14:56:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"207665028443493459947332205195559384221","date":"2026-02-05T22:47:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"129855685192394948711114620307874573528","date":"2026-02-01T15:41:16+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-18T15:08:26+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-12-04T15:35:11+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-12-02T14:34:19+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-12-02T14:33:59+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-11-28T11:35:33+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"55c88fd7-0a49-4fed-8b07-9ea9d56eab8c","owner":[],"postedDate":"December 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":60006760,"name":"Biological sciences/Developmental biology"},{"id":60006761,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-05-16T20:23:18+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-22 17:07:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8230063","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8230063","identity":"rs-8230063","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}