O-239 A model based on artificial intelligence for the non-invasive prediction of embryo aneuploidy

Abstract Study question Can the automated artificial intelligence (AI) model iDAScore predict embryo aneuploidy? Summary answer The AUC for aneuploidy prediction of iDAScore was AUC 0.640, with optimal cut-off score 7.75, sensitivity 61.0% and specificity 60.9%. What is known already Embryo morphokinetics and image analysis using AI have been proposed as alternative approaches for non-invasive embryo aneuploidy prediction. The automated AI model iDAScore, as well as the morphokinetic embryo selection model KIDScoreD5 have been shown to have correlation with embryo ploidy status. The aim was to study the performance of aneuploidy prediction using the iDAScore, KIDScoreD5 and morphology grading (Gardner criteria). Study design, size, duration Retrospective study including 382 blastocysts of known ploidy from 92 patients undergoing PGT-A in a single centre during January-December 2022. For statistical analysis, continuous and categorical data were compared using ANOVA and Fisher’s exact test/Chi-square, respectively. ROC analysis was used to estimate the discrimination performance of variables to predict aneuploidy. Stepwise logistic regression was used to identify confounding variables and construct a model for aneuploidy prediction. Participants/materials, setting, methods Embryos were cultured in Embryoscope, underwent laser assisted hatching on Day3, were biopsied on Day5-Day6 and were vitrified. Whole-genome amplification and next-generation sequencing were used for PGT-A. iDAScore values were calculated using v1.2.0. KIDScoreD5 v.3 was calculated after annotation of the required morphokinetic parameters in Embryoviewer. For Gardner criteria, inner cell mass (ICM) and trophectoderm (TE) were classified into grades A to C. The expansion in all biopsied blastocysts was > =grade 4. Main results and the role of chance Out of 382 embryos analysed, 127 (33.2%) were euploid and 255 (66.8%) were aneuploid. Euploid embryos were associated with significantly higher iDAScore (7.69, 95% CI 7.44-7.95 vs 6.86, 95% CI 6.66-7.09, p < 0.001), higher KIDScoreD5 (6.41, 95% CI 6.07-6.75 vs 5.74, 95% CI 5.51-5.97, p = 0.001) and lower maternal age (37.10 years, 95% 36.39-37.82 vs 39.83 years, 95% CI 39.36-40.30, p < 0.001) compared to aneuploid embryos. Significantly more embryos with grade A ICM (A:39.4%, B:26.0%, C:20.0%, p = 0.013) and grade A TE (A: 45.9%, B:29.6%, C:23.6%, p = 0.002) were euploid compared with B and C grades, respectively. Blastocysts biopsied on Day 5 and Day 6 had similar proportions of euploidy (34.9% vs 27.6%, p = 0.125). The AUCs for aneuploidy prediction were: iDAScore: 0.640 (95% CI 0.582-0.699), KIDScoreD5: 0.604 (95% CI 0.542-0.665), ICM grade: 0.581 (95% CI 0.521-0.640), TE grade: 0.596 (0.536-0.657), maternal age: 0.694 (0.640-0.748). The optimal cut-off (Youdex index) for aneuploidy prediction of iDAScore was 7.75 (sensitivity 61.0%, specificity 60.9%). Following stepwise multivariate logistic regression, the variables that entered the model were iDAScore, maternal age and TE grade. The model iDAScore+age+TE grade increased the ability of aneuploidy prediction (AUC 0.747, 95% CI 0.696-0.796). KIDScoreD5, ICM grade and day of biopsy failed to enter the model. Limitations, reasons for caution This is a retrospective study and therefore the presence of bias cannot be excluded. iDAScore and KIDScore are models originally developed for the prediction of implantation and not for the diagnosis of embryo aneuploidy. Wider implications of the findings iDAScore appears to associate with embryo aneuploidy. Combination of iDAScore+age+TE increased the AUC for aneuploidy prediction. iDAScore can used as a decision-support tool for prioritising embryos for transfer, cryopreservation or biopsy. The analysis of more embryos is needed to confirm the present findings. Trial registration number not applicable