P-264 Artificial intelligence (AI)-based deselection of at-risk aneuploid embryos using a validated score threshold and time-lapse imaging in IVF
D Gilboa, Y Tauber, Y Amar, N Lustgarten, M Shapiro, D SeidmanAbstract
Study question
Can EMA Genetics, an AI-based genetic screening tool, confidently deselect embryos at risk of aneuploidy across diverse, unbalanced data using a validated AI score threshold?
Summary answer
The model identified high-risk aneuploid embryos (AI score ≤ 28) with 85.2% NPV, increasing to 88% for ages >40, consistent across real-world and validation datasets.
What is known already
Embryo selection is critical for optimizing IVF outcomes. Preimplantation genetic testing for aneuploidy (PGT-A) identifies euploid embryos but is invasive, costly, and incompatible with fresh transfer protocols. EMA Genetics offers a non-invasive screening method using time-lapse imaging (TLI) and AI to estimate aneuploidy risk prior to confirmatory testing and/or transfer. Previous validation on balanced datasets demonstrated its ability to deselect embryos at risk of aneuploidy with a score threshold ≤ 28, achieving high negative predictive value (NPV). However, real-world data often involve unbalanced patient cohorts, necessitating further validation to confirm the model’s robustness across diverse clinical settings and patient populations.
Study design, size, duration
The study utilized an independently collected multicenter dataset of 1,920 biopsied blastocysts from four clinics across four countries. Data included oocyte age, PGT-A-confirmed aneuploidy status, and TLI videos. An AI score (1–99) was assigned to each embryo. The score increases linearly with likelihood of euploidy. The deselection threshold (AI score ≤ 28) was evaluated for reliability by comparing its negative predictive value (NPV) with results from the balanced validation dataset (N = 700 embryos).
Participants/materials, setting, methods
The final dataset included 1,049 aneuploid (56.4%) and 808 euploid (43.6%) embryos, with maternal age and aneuploidy rates matching reported clinical distributions. AI scores were compared to known PGT-A labels (aneuploid/euploid). Classification performance metrics were assessed in both unbalanced real-world and balanced validation datasets, and the NPV at the validated deselection score threshold was compared across datasets.
Main results and the role of chance
The AI model identified embryos with AI scores ≤ 28 as high-risk aneuploid with a 85.2% negative predictive value (NPV), confirmed by PGT-A in the independently collected dataset. Performance improved with maternal age, with NPV reaching 88% for patients aged 40–42. The deselection threshold of ≤ 28 remained consistent with the threshold identified in the balanced validation dataset (84.8%), demonstrating consistency across diverse patient cohorts.Logistic regression (OR) analysis for evaluating the association between AI scores and euploidy likelihood in the unbalanced real-world and balanced validation datasets: 2.19 [95% CI: 1.749– 2.733] and 2.79 [95% CI = 2.05-3.82], respectively.
Limitations, reasons for caution
Prospective validation in larger, more diverse cohorts is needed. Additionally, potential variability in clinical protocols among participating centers may influence results. Caution is warranted when extrapolating these findings to clinical populations outside the studied demographic.
Wider implications of the findings
This AI-driven tool has the potential to complement existing diagnostic methods, improve embryo selection strategies, and optimize IVF resource allocation, particularly for patients with advanced maternal age or limited embryos available for transfer.
Trial registration number
No