P-283 AI-driven aneuploidy prediction: a non-invasive solution for safer and accurate embryo evaluation in fresh transfers

Abstract

Abstract Study question Can an AI genetics screening model, using embryo time-lapse videos, clinical metadata, and morphokinetic profiles, demonstrate strong predictive ability? Summary answer Strong performance validation and effective stratification by AI scores and aneuploidy rates validate the model as a reliable adjunct for patients unfit for invasive biopsy. What is known already Confirmatory genetic assessment of embryos involves preimplantation genetic testing for aneuploidy (PGT-A), performed via a trophectoderm biopsy. However, many patients face challenges. Occasionally, PGT-A is not a viable option due to cost, invasiveness, potential embryo damage, labour limitations, or fresh transfer need. This creates a gap in available genetic screening options. A noninvasive alternative could provide a quantitative risk assessment for aneuploidy, offering valuable insights for prognosis counseling without financial and logistical burdens of invasive testing. Consequently, there is a growing demand for a reliable AI-based solution that delivers real-time, measurable genetic risk assessments before transfer and/or confirmatory testing. Study design, size, duration This single-center, retrospective cohort study at Hygeia IVF Embryogenesis evaluated the ML model on 335 blastocysts with known PGT-A outcomes (aneuploidy/euploidy) biopsied between 2022-2024. The AI model (EMA Genetics) was trained using 5,000 time-lapse videos, along with associated clinical parameters: female age, Day-5 embryo quality, and morphokinetic parameters. Known ploidy and live-birth outcomes were used as ground truth labels during training. Independent performance evaluation quantified the clinical value of the AI model in the clinic. Participants/materials, setting, methods Logistic regression was used to assess the relationship between AI-score and probability of euploidy. The odds ratio (OR) and 95% CI were calculated to quantify the strength of this association. Model performance was evaluated using the area under the curve (AUC), sensitivity and specificity. The statistical significance of differences in AI-scores between euploid and aneuploid embryos was determined with p < 0.05. Data was further stratified by AI-score brackets to confirm linear associations with euploidy rate. Main results and the role of chance Logistic regression analysis revealed a significant association between the AI score and the probability of euploidy, with an odds ratio (OR) of 4.69 (95% confidence interval [CI]: 2.64 - 8.35). The model demonstrated strong discriminative ability, reflected by an AUC-ROC of 0.72, which aligns with values published in other commercially available AI tools. Sensitivity and specificity at the optimal threshold were 67.2% and 71.2%, respectively, showing a relatively balanced capacity to identify both euploid (true positives) and aneuploid (true negatives) embryos. Stratification of embryos into four AI score brackets 1–32 (n = 198), 33–49 (n = 97), 50–66 (n = 37), 67–99 (n = 3) revealed corresponding aneuploidy rates of 91%, 71%, 57%, and 0%, respectively, further substantiating the model’s capacity to stratify embryos according to their likelihood of euploidy. Results indicate a statistically significant inverse relationship between ascending AI score and decreasing aneuploidy rate. Limitations, reasons for caution Genetic status at the time of implantation or birth was not assessed. Additionally, AI results do not offer diagnostic genetic testing at the chromosome/karyotype level. Wider implications of the findings Study findings validate the utility of an AI-based aneuploidy prediction model as a logical adjunct for ranking embryos for transfer, cryopreservation and/or biopsy. Trial registration number No

Ähnliche Arbeiten