JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):21
Poster Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263542

P.09. Analysis of human embryo ploidy using Artificial Intelligence, with morphology, morphokinetics, and patient variables

Vinícius Casado Moraes¹, Mariana Nicolielo², Dóris Spinosa Chéles², Bruna Lourenço², José Roberto Alegretti², Aline Rodrigues Lorenzon², Marcelo Fábio Gouveia Nogueira¹, José Celso Rocha¹

¹Universidade Estadual Paulista "Júlio de Mesquita Filho" - UNESP - Assis - SP - Brasil
²Huntington Medicina Reprodutiva - Eugin Group - São Paulo - SP - Brasil

Objective: To develop a non-invasive method for predicting human embryo ploidy using artificial intelligence, based on a combination of morphological, morphokinetic, and patient data.
Methods: The morphology database (Mph) was obtained from images captured using the time-lapse technique with the EmbryoScope^® embryo incubator, Vitrolife^®. These images were submitted to digital image processing software, which extracted 33 mathematical variables representing blastocyst structures such as the inner cell mass, trophectoderm and blastocoel. Morphokinetic variables (Mkt) were represented by cleavage times t2, t3, t4, t5, t8 and correlations between them. In addition, age, body mass index, number of previous pregnancy attempts, diagnosis of endometriosis and/or tubal factor, and diagnosis of male factor were used to compose the patient database (Pt). For the development of the research, a total of 1,128 cases were used to compose the database, which was grouped into the following variable sets: Mph + Mkt, containing 1,128 data; Mkt + Pt, with 706 data; Mph + Pt, with 764 data; and, finally, Mph + Mkt + Pt, containing 706 data. The database was separated for Artificial Intelligence (AI) learning as follows: for the Mkt + Pt and Mph + Mkt + Pt sets, 506 data were allocated for training and 200 for blind test; for the Mph + Pt set, 564 data were used for learning and 200 for blind test; finally, for the Mph + Mkt database, 958 data were allocated for training and 170 for blind test. With the variables defined and the database structured, learning began using AI software on the Matlab^® platform, utilizing the Multilayer Perceptron Artificial Neural Network technique. This approach allows for the processing of variables through pattern identification, enabling the model to adapt to new data sets. After the continuous AI learning process, the results generated by the model were compared with those obtained from embryonic biopsies performed via preimplantation genetic testing for aneuploidy to assess the accuracy of the AI technique employed.
Results: For the Mph + Mkt set of variables, the accuracy obtained was 89.9% in the learning phase and 60% in the blind test. For the Mkt + Pt data, an accuracy of 87.9% was obtained in the learning phase and 67% in the blind test. For the Mph + Pt set, the accuracy obtained by the AI was 98.2% for learning and 64% in the blind test. Finally, for the Mph + Mkt + Pt set, 96% accuracy was obtained in the AI learning and 70% accuracy in the blind test.
Conclusion: The set of variables Mph + Mkt + Pt was the most effective in differentiating between euploid and aneuploid blastocysts, with 70% accuracy in the blind test. Next came the Mkt + Pt variable set, with an average accuracy of 67%, and finally, the Mph + Pt set, with an average accuracy of 64%. The results demonstrated the potential of AI in predicting embryonic ploidy, with accuracy values between 64% and 70% for the blind test data for the sets, enabling greater chances of identifying ploidy in a non-invasive manner to the embryo, in addition to providing a more effective and reliable tool for embryologists and reproductive medicine professionals.