Independent external pre-clinical validation of iDAScore v1.0 in 1232 PGT-A cycles with 3604 biopsied blastocysts and 808 euploid transfers.

Abstract Study question Is iDAScore v1.0 associated with euploidy and live-births (LBs) after euploid transfers? How often would it have affected embryologists’ clinical choices in a blinded analysis? Summary answer iDAScore v1.0 was associated with both euploidy (AUC:0.60) and LBs after euploid transfers (AUC:0.66). iDAScore v1.0 and embryologists would have frequently disagreed on embryo ranking. What is known already Embryo assessment/ranking to improve IVF efficiency (LB/transfer) is still challenging. The widely used static morphological evaluation suffers from subjectivity and intra-/inter-operator variability. Preimplantation genetic testing for aneuploidies (PGT-A) discriminates euploid/aneuploid embryos, improving IVF efficiency. Nevertheless, it requires manipulation and expertise and ∼50% euploid blastocysts still fail to implant. Time-lapse (TL) technology allows continuous undisturbed monitoring of embryo morphokinetics. The software iDAScore v1.0 is a deep learning algorithm trained on TL videos from implanted/non-implanted blastocysts to generate a score that should predict their implantation potential (score:1.0-9.9, from the lowest to the highest, respectively). Study design, size, duration Retrospective independent external pre-clinical validation of iDAScore v1.0 in PGT-A cycles (N = 1232) with fresh own oocytes and ≥1 biopsied blastocyst (N = 3604) (April-2013 to August-2022). The AI-based tool was investigated for associations with embryologists’ assessment, blastocysts’ karyotype, and LBs. Two simulations were then conducted to estimate how often iDAScore v1.0 would have ranked (i) euploid blastocysts first in presence of aneuploid as well, (ii) reproductively-competent blastocysts before incompetent in presence of ≥ 2 euploid and ≥1 LB. Participants/materials, setting, methods Embryos were cultured in EmbryoScope (Vitrolife). Only the first PGT-A cycles were included (maternal age 38.7 ± 3.4 years). The day of biopsy was defined based on the hours-post-insemination (hpi) and blastocyst morphology based on Gardner. ROC curve analyses were conducted to calculate the AUC for euploidy and LB discrimination based on embryologists’ assessment or iDAScore v1.0. Main results and the role of chance iDAScore v1.0 was associated with ICM-quality (A-grade, N = 2107, score: 7.5 ± 1.8; B-grade, N = 833, score: 5.6 ± 1.9; C-grade, N = 664, score: 4.4 ± 1.7; p < 0.01), trophectoderm-quality (A-grade, N = 1988, score: 7.5 ± 1.8; B-grade, N = 951, score: 5.9 ± 1.9; C-grade, N = 664, score: 4.3 ± 1.6; p < 0.01), and day ( ≤ 120hpi, N = 1462, score: 8.2 ± 1.5; 121-144hpi, N = 1874, score: 5.6 ± 1.7; > 144hpi, N = 268, score: 3.9 ± 1.4;p<0.01). Euploid blastocysts showed the highest score (N = 1443, 7.0 ± 2.1) versus both single-aneuploid (N = 1194, score: 6.5 ± 2.2, p < 0.01) and complex-aneuploid (N = 967, score: 5.8 ± 2.1, p < 0.01). AUC of 0.60 (95%CI 0.59−0.62) and 0.66 (95%CI 0.64−0.68) were reported for iDAScore v1.0 and embryologists’ assessment association with euploidy, respectively. Blastocysts resulting in a LB showed higher score (N = 361, 7.6 ± 1.8) than non-implanted/miscarried (N = 447, score: 6.5 ± 2.2; OR: 1.3, 95%CI 1.2-1.4, p < 0.01). AUC of 0.66 (95%CI 0.62-0-69) and 0.64 (95%CI 0.60-0-67) were reported for iDAScore v1.0 and embryologists’ assessment association with LBs, respectively. The simulations revealed that (i) across 587 cycles with both diagnoses, iDAScore v1.0 would have ranked euploid blastocysts as “top-quality” in 63% of the cases; (ii) across 202 cycles with sibling euploid blastocysts and a LB, it would have been equally, less, and more effective than embryologists, in 52%, 3% and 15% of the cases, respectively. Nonetheless, in 29% of the cases, this comparison was not doable because top-ranked euploid blastocysts were not transferred but a LB was achieved with worse-ranked blastocyst. Limitations, reasons for caution iDAScore v1.0 performance might be suboptimal on a dataset of mostly advanced-maternal-age women and on day-7/poor-quality blastocysts, due to their poor representation in the algorithm training. Although euploidy is a proxy of LB, iDAScore v1.0 was trained only to predict the latter. A prospective randomized study is now warranted Wider implications of the findings iDAScore v1.0 was significantly associated with euploidy and LB with results comparable to embryologists’ performance. Nevertheless, iDAScore v1.0 is more objective and reproducible than embryologists’ assessment. Embryo selection workflows, scientific debate, and patient counselling practice might all benefit from AI-based standardization of blastocyst evaluation. Trial registration number not available