L26/P-158 Limits of prediction: Why AI cannot fully predict embryo euploidy
Human Reproduction · 01.07.2026 · 41 (Supplement_1)
DOI: 10.1093/humrep/deag083.494 ↗Kurz gefasst
Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von R. Del Gallego und 8 weiteren Autor:innen.
Abstract (Originalsprache)
Study question What is limiting the predictive performance of embryo euploidy through artificial intelligence (AI) multimodal systems? Summary answer Predictive performance is primarily limited by aneuploidy subtype–specific phenotypic overlap: trisomies and mosaic abnormalities resemble euploid embryos, constraining probability calibration despite moderate discrimination. What is known already AI models using time-lapse videos (TLV), morphokinetic and clinical data have been extensively explored for non-invasive embryo ploidy prediction. However, reported performance has consistently plateaued at moderate accuracy, with area under the ROC curve (AUROC) rarely exceeding ∼0.70. This limitation may not reflect model failure, but rather biological heterogeneity, as different aneuploidies might impact embryo morphology and developmental dynamics unevenly. Recently published evidence suggests trisomies involving single chromosomes often resemble euploid embryos morphologically, potentially masking genetic abnormalities from image-based AI models. Study design, size, duration Retrospective study analyzing 7,650 TLV and clinical data from embryos cultured between June 2015 and January 2023, originating from 1,624 patients across 2,166 treatment cycles at a tertiary IVF center. Outcomes included embryo ploidy status and detailed PGT-A–based aneuploidy subclassification. Model evaluation was performed using 5-fold cross-validation. Metrics included the AUROC, accuracy, average euploid prediction score (EPS; where 0.5 was selected as cut-off for euploid/aneuploid), and percentage of samples with score>0.5 (%>0.5). Participants/materials, setting, methods Embryos from patients undergoing IVF/ICSI were included. All embryos were monitored using time-lapse incubation systems and blastocysts graded ≥BL3CC (Gardner) followed trophectoderm PGT-A biopsy, analyzed through next generation sequencing (NGS). Aneuploidies were classified according to the number of chromosomes affected, the presence of chromosomal-gains (trisomies or segmental gains) or chromosomal-losses (monosomies or segmental losses), and uniform or mosaic states. The multimodal deep-learning model incorporated data as TLV images and patient- and treatment-specific clinical variables. Main results and the role of chance A deep-learning model was developed using a two-stage prediction pipeline, where an initial blastocyst quality predicted-score was combined with image-derived embeddings and clinical features to generate a final euploidy prediction. Overall, the model demonstrated moderate discriminative ability for distinguishing euploid from aneuploid embryos, achieving AUROC of 0.69 and accuracy of 0.63. The model correctly assigned scores>0.5 to 61.8% of truly euploid embryos (n = 3,647); but the average EPS was moderate (EPS=0.55), indicating conservative probability estimates. Subgroup analyses revealed marked heterogeneity across aneuploidy categories. Embryos with mixed aneuploidies (one chromosomal-gain and one chromosomal-loss, uniform/mosaic; n = 393) and complex aneuploidies (≥3 aneuploidies, uniform/mosaic; n = 835) were the most distinct from euploid embryos, with the lowest EPS values (0.35 and 0.28, respectively), and fewer embryos misclassified as euploid, 28.5% and 19.6%. Trisomic (n = 1,068) and monosomic (n = 1,707) embryos exhibited moderately reduced EPS (0.45 and 0.42), indicating limited phenotypic deviations detectable by the model, with 43.4% and 38.1% exceeding the 0.5-threshold. When we further subcategorized them, trisomic maintained similar EPS, but monosomic exhibited progressively lower values. Single mosaic chromosomal-losses presented the highest values (EPS=0.50; %>0.5=53.3; n = 90), followed by single uniform chromosomal-losses (EPS=0.43; %>0.5=39.0; n = 1,301) and two uniform/mosaic chromosomal-losses (EPS=0.36; %>0.5=30.4; n = 316), indicating cumulative monosomies produced more pronounced phenotypic deviations. Limitations, reasons for caution The less-frequent aneuploidy subtypes were under-represented in the model training. Larger datasets are required to further train models on subtle morphological effects associated with low-impact chromosomal abnormalities. Wider implications of the findings The model discriminated euploid from aneuploid embryos with moderate accuracy but remained underconfident in specific indistinguishable aneuploidy subtypes. This phenomenon might reflect how current AI approaches are fundamentally limited by aneuploidy-specific phenotypic expression, rather than model error. Future models should incorporate probabilistic interpretation to guide appropriate clinical use. Trial registration number No
Zitieren
APA
Gallego, R. D., Sangeeth, A., Zhang, J., Zhylko, D., Selim, S., Lawrenz, B., Elkhatib, I., Shamout, F., & Fatemi, H. (2026). L26/P-158 Limits of prediction: Why AI cannot fully predict embryo euploidy. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.494
Vancouver
Gallego RD, Sangeeth A, Zhang J, Zhylko D, Selim S, Lawrenz B, et al. L26/P-158 Limits of prediction: Why AI cannot fully predict embryo euploidy. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.494
BibTeX
@article{Gallego2026,
author = {Gallego, R. Del and Sangeeth, A. and Zhang, J. and Zhylko, D. and Selim, S. and Lawrenz, B. and Elkhatib, I. and Shamout, F.E. and Fatemi, H.},
title = {L26/P-158 Limits of prediction: Why AI cannot fully predict embryo euploidy},
journal = {Human Reproduction},
year = {2026},
volume = {41},
number = {Supplement_1},
doi = {10.1093/humrep/deag083.494}
}
EUVIMED