P-117 Retrospectively Comparing Non-Invasive artificial intelligence (AI) Models and Morphological Assessments for Embryo Euploidy Prediction: Insights from iDAScore, KIDScore, and CHLOE-EQ

Abstract

Abstract Study question Which AI model, including iDAScore, KIDScore, CHLOE-EQ, and traditional blastocyst morphology assessment, demonstrates the highest accuracy in non-invasive prediction of embryo euploidy? Summary answer KIDScore shows the highest accuracy in embryo ploidy prediction, outperforming other AI models and blastocyst scoring by embryologists. What is known already Preimplantation Genetic Testing for Aneuploidy (PGT-A) is an invasive procedure, and outcomes are reliant on biopsy protocols and embryologist expertise. Efforts are underway to develop non-invasive alternatives, with AI playing a key role in embryo assessment. Existing algorithms, such as iDAScore, CHLOE-EQ and KIDScore model offer varying levels of accuracy in predicting embryo euploidy. While studies have assessed the ploidy performance of the Gardner and Schoolcraft blastocyst grading system, iDAScore, KIDScore, and CHLOE-EQ individually, there is a lack of studies comparing them together, in order to understand which model performs best in clinical practice. Study design, size, duration A retrospective study included 116 blastocysts of known ploidy from 23 couples (mean age of women: 30, range: 22–46). The study involved intracytoplasmic sperm injection cycles with ejaculated spermatozoa and PGT-A at Private IVF Center, Cyprus, between March and September 2024. Several statistical models were used to compare systems for predicting embryo ploidy, including bootstrapped binary logistic regression (BBLR), Lasso-penalized logistic regression (LPLR), Bayesian logistic regression (BLR), and Classification and Regression Tree (CRT) analysis. Participants/materials, setting, methods Embryos cultured in the Embryoscope underwent biopsy on Day 5. PGT-A was performed using next-generation sequencing. iDAScore values were calculated using version 1.2.0. KIDScoreD5 v.3 was determined after annotating morphokinetic parameters in Embryoviewer. Chloe-EQ Score was assessed using time-lapse videos. Blastocyst quality scores were derived from multiplication of expansion grade (1–6) by ICM and TE grades (A = 3,B=2,C=1). All biopsied blastocysts had an expansion grade of ≥ 3. Main results and the role of chance Comparison of statistical models showed similar performance but limited sensitivity for detecting aneuploidy between the Decision Tree (accuracy=76.7%, specificity=95.1%, sensitivity=34.3%) and the bootstrapped model (accuracy=75.0%, specificity=91.4%, sensitivity=37.1%). When predicting euploidy, the Lasso model emerged as the strongest performer (AUC=.79), while Bayesian analysis showed weaker discrimination (AUC=.544). KID emerged as the primary predictor across all models, showing significant association in the bootstrapped and Lasso model respectively ( B = 0.453, p = .006, 57.3% odds increase; B = 0.3974, 48.8% odds increase), whilst ChloeRank demonstrated consistent secondary importance therein (B = 0.226, p = .009, 25.4% odds increase; B = 0.1662, 18.1% odds increase). Blast score only indicated a small positive effect using the Lasso model (B = 0.0337, 3.4% odds increase), whilst IDA demonstrated minimal and non-significant predictive effects in the bootstrapped model (B = 0.029, p = .842) and negligible contributions in the Lasso analysis (B = 0.0136). Overall, KID and ChloeRank were robust predictors, demonstrating strong predictor-outcome relationships. Limitations, reasons for caution This study is limited by its retrospective nature, single-center design, small cohort with known PGT-A results, and the predominance of young patients with high euploidy rates. While iDAScore, KIDScore, and CHLOE-EQ models predict implantation potential, their use in diagnosing embryo aneuploidy requires further investigation. Wider implications of the findings KIDScore demonstrates strong euploid prediction in younger patients with lower aneuploidy risk, suggesting potential to reduce PGT-A use in select cases. Validation in advanced maternal age and improved aneuploidy detection remain key. This study’s solid oocyte and patient profiles provide a control for future, more granular research. Trial registration number No

Ähnliche Arbeiten