Artificial intelligence unlocks oocyte competence: MAGENTA outperforms clinical predictors and boosts euploidy modeling

Abstract

OBJECTIVE: To study whether an artificial intelligence (AI)-derived oocyte morphology score (MAGENTA) predicts fertilization, embryo development, and euploidy, and whether integrating MAGENTA with clinical and embryological variables improves the performance of euploidy prediction models. DESIGN: Retrospective cohort study. SUBJECTS: A total of 2,196 patients undergoing intracytoplasmic sperm injection (ICSI) between January 2020 and May 2024, with 13,370 oocytes analyzed, in a single university-affiliated IVF center. Preimplantation genetic testing for aneuploidy (PGT-A) was performed on 3,514 embryos from 1,147 cycles. EXPOSURE: All oocytes underwent standardized imaging and MAGENTA scoring before ICSI. Outcomes included fertilization, blastulation, top-quality blastocyst formation, and euploidy after PGT-A. Multivariable models assessed determinants of MAGENTA. Three predictive frameworks, MAGENTA-only, robust (MAGENTA + blastocyst quality), and full (MAGENTA + embryo morphology + 9 clinical variables), were evaluated using receiver operating characteristic analysis, calibration, Brier scores, permutation importance, and SHapley Additive exPlanations (SHAP) values. MAIN OUTCOME MEASURES: Association of MAGENTA scores with fertilization, blastulation, blastocyst quality, embryo euploidy, and clinical pregnancy. Performance of XGBoost machine learning for euploidy prediction and multivariate modeling of factors influencing MAGENTA scores. RESULTS: The MAGENTA scores increased progressively across embryological milestones. Female age was the strongest negative determinant of MAGENTA (B = -0.063/y). The MAGENTA-only model predicted euploidy with modest discrimination (area under the curve [AUC] 0.542; 95% confidence interval [CI] 0.505-0.579). Incorporating blastocyst morphology improved the AUC to 0.607 (95% CI 0.573-0.641). The full model achieved the highest performance (AUC 0.666; 95% CI 0.632-0.700), with significantly better discrimination than the MAGENTA-only and robust models. Calibration metrics improved in parallel, with Brier scores decreasing from 0.237 to 0.227 and calibration slopes increasing from 0.259 to 0.645. Feature importance and SHAP analyses confirmed MAGENTA as a major contributor to predictive performance. CONCLUSION: The MAGENTA was independently associated with fertilization, blastulation, blastocyst quality, and euploidy. Integrating MAGENTA with embryo morphology and pre-ICSI clinical variables significantly improved discrimination and calibration of euploidy prediction compared with MAGENTA or embryo morphology alone.

Ähnliche Arbeiten