Bridging the gap between embryo euploidy, pregnancy potential and morphology using artificial intelligence for ploidy...

Can a non-invasive artificial intelligence (AI) model based on time-lapse imaging (TLI) data accurately predict embryo ploidy and support embryo selection in IVF?

This retrospective multicentre study involved 4822 blastocyst-stage embryos with known ploidy from three IVF clinics for training and internal validation, and 706 embryos from an independent patient cohort used for external testing. The AI model was trained using spatio-temporal features from TLI videos, producing a continuous score (0-100) reflecting the likelihood of euploidy. Associations with morphology, fetal heartbeat and ploidy subtypes were evaluated.

In a blind test set (n = 705), higher AI ploidy scores were significantly associated with euploidy (odds ratio [OR] = 1.39, P < 0.001) and presence of a fetal heartbeat (FH+; OR = 1.26, P = 0.043). Euploid embryos had higher mean scores than aneuploid embryos (51.3 versus 43.44, P < 0.001); FH+ embryos scored higher than FH- embryos (53.63 versus 48.42, P = 0.013). AI ploidy scores decreased with increasing aneuploid complexity (P < 0.001), and within-patient differences were significant (P = 0.0098). External validation (n = 706) confirmed the model's ability to outperform an existing AI pregnancy predictor in ploidy discrimination (area under the curve [AUC] = 0.64 versus 0.61). Combined with morphology, predictive performance improved (AUC = 0.65). Top-quartile AI ploidy scores were associated with euploidy rates of over 90%.

This validated, non-invasive AI model predicts embryo ploidy from TLI data and adds value to morphology in embryo selection. It demonstrates generalizable performance across clinics and may be especially useful when preimplantation genetic testing for aneuploidies is not feasible.