COMUNICACIONES PÓSTERS P-042 MACHINE LEARNING TO PREDICT BLASTOCYST FORMATION: THE DEVELOPMENT OF AN IN-HOUSE ALGORITHM M. Méndez Justo (1), C. González Trigo (2), C. López Gallardo (2), A. Piñol Bonet (2), F. Oliva Cuyas (3), S. Cívico Vallejos (1) (1) Hospital Clínic de Barcelona - Barcelona (Barcelona), (2) Universidad de Barcelona - Barcelona (Barcelona), (3) Departamento de Genética y Microbiología y Estadística, Facultad de Biología, Universidad de Barcelona - Barcelona (Barcelona) INTRODUCTION: Second, the robustness of the algorithm was tested with a new embryo cohort with same clinical features, but in this case with The introduction of time-lapse technology in the IVF laboratory cycles using own oocytes. Third, the algorithm was re-tested in has led to the study of human embryo kinetics as a potential distinct cohorts of embryos from different patient ages and cause non-invasive tool for embryo selection. However, controversy of sterility to evaluate its overall performance. stillexists regarding whether these amounts of data imply further predictive power given the great range of confounding Statistical analyses were performed with R software v4.0.1. factors reported in the literature. Here we present the process of modelling an in-house machine learning model to predict RESULT: blastocyst formation and its proper validation. A total of 55 kinetic variables were first included in the study. OBJECTIVE: When comparing variances, means and medians between the viable (V) and non-viable embryos (NV) groups, a lot Assess whether kinetics from embryos with the competence of the variables showed significant differences, suggesting to properly reach the blastocyst stage (meet morphology to be the complexity of the present data. Cut-off points for the transferred or frozen) differ from those that failed to reach this categorization using conditional inference trees were performed. stage. Moreover, a selection of 20 variables out of 55 was done by PCA. From predictive modelling, the higher accuracy rates correspond Moreover, we present the evolution/steps to develop an in- to the SVM models with specific kernels. Concrete variable sub- house algorithm from a machine learning perspective, adjusting groups presented a robust percentatge of correct classification for the most widely reported confounding factors. for example SVM with radial kernel using 9 quantitative and 5 categorical variables with an AUC of 0.86 (0.81-0.92) in donation MATERIAL AND METHOD: cycles. This algorithm performed well in cycles with own oocytes with <35 years with a percentage of correct classification of 0.745. A total of 181 embryos from donation cycles were included to model the algorithm. Kinetic parameters annotated by a CONCLUSIONS: single observer included in the study were time of extrusion of the second polar body (tPB2), time of appearance and A methodology for the establishment and evaluation of human disappearance of pronuclei (tPNa, tPNf), and time of cell embryo kinetic parameters, reaching proper cut off for variable divisions (t2, t3, t4, t5, t6, t7, t8) and its differences between them.categorization is settled down. The application of predictive Furthermore, categorization using conditional inference trees of machine learning techniques shows promising results in embryo kinetic variables were also applied due to the asymmetry of the selection. variables. A selection among all the possible kinetic variables to include in the model was performed with principal component analysis (PCA). Different machine learning methods (logistic regression, random forest, boosting and support vector machine (SVM) were trained and assessed using leave-one-out cross- validation (LOOCV), 5-fold cross-validation and resubstitution). 167 ASEBIR. Revista de Embriología Clínica y Biología de la Reproducción. Noviembre 2021 Vol. 26 Nº 2