JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):167
Poster Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263921

P-155. Machine learning for non-invasive prediction of euploidy in human embryos based on morphokinetic data

Sabrina Caldas Berno¹, Maria Eduarda Santos Abritta Ferreira¹, George Marsicano Correa¹, Ana Beatriz Nogueira Pacheco², Luiz Carlos Pinheiro Junior², Íris Oliveira Rocha², Bruno Ramalho Carvalho³

¹ Faculdade de Ciências e Tecnologias em Engenharia – FCTE/UNB - Brasília - DF - Brasil
² Genesis Centro de Assistência em Reprodução Humana - Brasília - DF - Brasil
³ Bruno Ramalho - Reprodução Humana - Brasília - DF - Brasil

Objective: To develop an artificial intelligence (AI) tool for prediction of embryo euploidy in a non-invasive and possibly more effective manner than the observed for preimplantation genetic testing for aneuploidy (PGT-A).
Methods: We adopted a quantitative and experimental design, focusing on the analysis of numerical data related to the morphokinetic events of 127 human blastocyts addressed to PGT-A; the following morphokinetic parameters were analyzed: first evidence of cytoplasmic movements prior to first cell division (st2); time frame at which an embryo reaches n number of blastomeres (t2, t3, t4, t5, t8); time frame at which an embryo initiates compaction (tSC); time frame at which the blastocoel is firstly visible (tSB); time frame at which the full blastocyst is formed (tB); and on the intervals between some of them (cc3 [t5-t3], t5-t2, s3 [t8-t5], and tB-tSB). From these data, a machine learning-based tool was built to correlate patterns in embryo development to euploidy, training and validating models from metrics such as accuracy, sensitivity, and specificity. Based on a multilayer perceptron (MLP) artificial neural network, the developed model was integrated to the LIME method, making predictions interpretable by identifying which morphokinetic variables most influenced each individual prediction, an essential feature to ensure transparency and reliability in the biomedical context.
Results: The MLP artificial neural network achieved an accuracy of 88.2% and an area under the receiver operating curve (AUC-ROC) of 0.944, indicating a high discriminative ability to predict euploidy. The evaluation through the confusion matrix demonstrated excellent performance in identifying aneuploid embryos, with perfect recall (1.00) for this class and robust recall (0.75) for euploid embryos, without any false positives.
Conclusion: Our AI tool demonstrated high accuracy in euploidy prediction, offering a promising non-invasive alternative to PGT-A, reducing costs and avoiding blastocyst biopsy, with potential for real-world application in the clinical routine of human assisted reproduction.