JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):55
Poster Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263594

P-43. Comparative Performance of AI-Based Embryo Scoring Systems for Ploidy Prediction

Dóris Spinosa Chéles¹, Mariana Nicolielo¹, Bruna Lourenço¹, Renata F Erberelli¹, Ana Luisa M Campos², Manuela Almeida³, Janaina Maciel⁴, Mauricio Barbour Chehin¹, José Roberto Alegretti¹, Aline Rodrigues Lorenzon¹

¹Huntington Medicina Reprodutiva - Eugin Group - São Paulo - SP - Brasil
²Huntington Pró-Criar Medicina Reprodutiva - Eugin Group - Belo Horizonte - MG -Brasil
³Huntington Medicina Reprodutiva - Eugin Group - Brasília - DF - Brasil
⁴CENAFERT Centro de Medicina Reprodutiva - Eugin Group - SALVADOR - BA - Brasil

Objective: Assisted reproductive technology (ART) has recently embraced non-invasive embryo assessment methods to improve clinical outcomes, such as time-lapse (TL) imaging and artificial intelligence (AI)-based systems. Morphological AI Assistance (MAIA) evaluates static images of blastocysts cultured in a TL system to predict clinical pregnancy. KIDScore Day 5 is an annotationdependent morphokinetic scoring model developed using Known Implantation Data to estimate implantation potential. IDAScore is a fully automated embryo scoring system that integrates morphokinetic and morphological information to predict implantation. This study aimed to assess whether scores generated by MAIA, KIDScore, and IDAScore are associated with chromosomal ploidy status in ART cycles undergoing preimplantation genetic testing for aneuploidy (PGT-A), providing insight into their potential value for non-invasive embryo selection and treatment planning.
Methods: This multicenter retrospective analysis included biopsied embryos from autologous and donor IVF cycles performed between April/2023 and May/2025 across five centers in four Brazilian states. Embryos were cultured in EmbryoScope+® TL incubators (Vitrolife) and evaluated using the MAIA (in-house model, score from 0-10), KIDScore (Vitrolife, score 0-9.9), and IDAScore (Vitrolife, score 0-10) systems. A total of 1,325 cycles from 1,075 patients were included, comprising 4,055 embryos with valid PGT-A results and annotated MAIA and KIDScore values. For IDAScore, the analysis comprised 403 IVF cycles from 316 patients, totaling 1,220 embryos evaluated between September/2019 and May/2025 at a single center. Embryos were classified by ploidy status as euploid, whole aneuploid (including monosomy, trisomy and complex), segmental, or mosaic. Statistical analysis used Mann–Whitney and Kruskal–Wallis tests; p<0.05 was considered significant. Predictive performance was assessed using classification metrics including accuracy and area under the curve (AUC).
Results: In autologous (94.4%) and donor (5.6%) IVF cycles, the proportion of euploid embryos was 37.7% and 65.1%, respectively (p<0.0001). The mean maternal age in autologous cycles was 36.8±3.7 years for euploid embryos and 38.9±3.4 years for whole aneuploid embryos (p<0.0001). The mean values of MAIA, KIDScore, and IDAScore were significantly higher in euploid embryos (5.7±3.0; 6.9±1.9; 5.4±2.4) compared to whole aneuploid embryos (4.7±3.1; 5.9±2.0; 4.2±2.3), with all comparisons reaching p<0.0001. This difference was consistently observed across all three scoring systems for specific aneuploidy types. Mean values for monosomies were 4.8±3.1 (MAIA), 6.0±2.0 (KIDScore), and 4.2±2.3 (IDAScore); for trisomies: 5.0±3.0, 6.6±1.9, and 4.9±2.5; and for complex aneuploidies: 4.6±3.0, 5.6±1.9, and 4.0±2.2, respectively. Notably, complex aneuploidies consistently showed the lowest mean scores across all algorithms when compared to monosomies, trisomies, and euploid embryos (p<0.0001). When comparing euploid and segmental aneuploid embryos, MAIA, KIDScore, and IDAScore values were significantly higher in euploid embryos (all p<0.0001 for MAIA and KIDScore; p0.04 for IDAScore). However, no significant differences were observed between segmental and whole aneuploidy groups across any of the scoring systems. Regarding mosaic embryos, the mean KIDScore (6.5±1.9) differed significantly from both euploid (p0.03) and whole aneuploid embryos (p0.0002). MAIA values in mosaic embryos (5.3±3.1) were significantly different from whole aneuploid embryos (p0.008), while IDAScore values (4.1±1.9) differed from euploid embryos (p0.0005). In terms of predictive performance for euploidy, MAIA, KIDScore, and IDAScore achieved similar accuracies of 0.58, 0.62, and 0.61, with corresponding AUCs of 0.59, 0.65, and 0.64. The optimal cut-off values for predicting euploidy were 6.66 for MAIA, 6.70 for KIDScore, and 4.73 for IDAScore.
Conclusion: These AI-based scoring systems demonstrated significantly higher mean scores for euploid compared to aneuploid embryos, with notable variability in performance for mosaic cases. AUC values above 0.5 across all models suggest modest but consistent predictive potential for ploidy status. These findings highlight the clinical relevance of AI-driven, non-invasive embryo scoring tools as valuable adjuncts to PGT-A, supporting optimization of embryo selection and treatment strategies in ART.