O-025 Factors affecting biochemical pregnancy: Machine learning-assisted identification

Abstract

Abstract Study question What variables in a PGT-A cycle can influence biochemical pregnancy loss? Summary answer There are paternal, maternal, ovarian stimulation, endometrial and embryo biopsy factors that are associated with the rate of biochemical pregnancy after the euploid embryo transfer. What is known already Biochemical miscarriage is an early termination in the development of pregnancy. Embryonic chromosomal alterations have been proposed to cause biochemical pregnancy. However, even though euploid embryos are transferred in PGT-A cycles, biochemical pregnancy rates do not seem to be modified, so there must be other factors associated with this type of miscarriage. The classical-statistical methods used to establish the factors related to biochemical miscarriage have produced contradictory results and there is no unanimity in the literature. As an alternative to traditional methods, different artificial intelligence algorithms are being used for the analysis of biological data. Study design, size, duration The study design is observational and retrospective. A total of 5892 embryos from 1919 PGT-A cycles were considered (January-2017 to October-2021). Only transferred embryos were included in the study (n = 1161). The trophoectoderm biopsies on D5, D6 or D7 blastocysts were analysed by NGS using the Illumina platform (VeriSeq Illumina®, San Diego, CA, USA). The biopsied embryos were vitrified and transferred in a subsequent cycle. Participants/materials, setting, methods Indications for PGT-A were advanced maternal age, altered karyotype or sperm FISH, history of chromosomal abnormalities in the offspring, repeated miscarriages and recurrent implantation failures. Clinical outcomes were recorded in a database including additional potential factors (n = 48) associated with biochemical pregnancy and related to progenitors, embryos and their biopsy, ovarian stimulation and adjuvant treatments. The association between the different variables and biochemical pregnancy was analysed using SPSS (v20.0) and R (v. 4.0.5) statistical software. Main results and the role of chance In order to determine which factors might increase biochemical pregnancy rates in euploid embryos, a multivariate analysis using logistic regression was initially performed. In the best predictive model (AUC=0.659) with a lower AIC (Akaike information criterion) value, only 3 factors showed a statistically significant association: uterine alterations (OR = 4.88, 95% CI [1.65-12.64]), day of embryo biopsy (OR = 2.19, 95% CI [1.46-3.31]) and mosaicism (number of altered chromosomes: OR = 1.59, 95% CI [1.10-2.22]) which significantly increased the risk of biochemical pregnancy. To identify other variables that might modify biochemical pregnancy rates and could be missed by classic statistical methods, different types of machine learning algorithms were used: unsupervised model (cluster analysis) and supervised predictive models (support vector machines (AUC=0.845), k-nearest neighbors (AUC=0.858), random forest (AUC=0.853), neural networks multilayer (AUC=0.719) and gradient boosting (AUC=0.825). The variables that had the greatest predictive power in the different machine learning algorithms were the variables associated with the embryo biopsy (day, number of laser pulses and biopsied cells), endometrial thickness and variables related to the male factor (sperm aneuploidy and DNA fragmentation). These algorithms apply different methodologies, but all agree on the fundamental role of these variables. Limitations, reasons for caution To confirm that the new identified variables are associated with biochemical pregnancy, it would be necessary to carry out prospective studies. Wider implications of the findings Biochemical pregnancy is the least studied clinical outcome in IVF. Knowledge of the variables that could affect biochemical pregnancy may be relevant as it may be a target for new therapies to reduce biochemical pregnancy rates and thus increase success rates. Trial registration number Not Applicable