Factors affecting biochemical pregnancy loss cycles: machine learning-assisted identification

• Published in: Reproductive biology and endocrinology 2024-08, Vol.22 (1)
• Main Authors: O, Lledó, B, Morales, R, Máéez-Grau, A, Cascales, A, Rodríguez-Arnedo, A, Castillo, Juan C, Bernabeu, A, Bernabeu, R
• Format: Article
• Language: English
• Subjects: Algorithms; Analysis; Artificial intelligence; Data mining; Genetic aspects; Genetic screening; Machine learning; Miscarriage; Pregnant women; Risk factors
• ISSN: 1477-7827; 1477-7827
• DOI: 10.1186/s12958-024-01271-1

Purpose To determine the factors influencing the likelihood of biochemical pregnancy loss (BPL) after transfer of a euploid embryo from preimplantation genetic testing for aneuploidy (PGT-A) cycles. Methods The study employed an observational, retrospective cohort design, encompassing 6020 embryos from 2879 PGT-A cycles conducted between February 2013 and September 2021. Trophectoderm biopsies in day 5 (D5) or day 6 (D6) blastocysts were analyzed by next generation sequencing (NGS). Only single embryo transfers (SET) were considered, totaling 1161 transfers. Of these, 49.9% resulted in positive pregnancy tests, with 18.3% experiencing BPL. To establish a predictive model for BPL, both classical statistical methods and five different supervised classification machine learning algorithms were used. A total of forty-seven factors were incorporated as predictor variables in the machine learning models. Results Throughout the optimization process for each model, various performance metrics were computed. Random Forest model emerged as the best model, boasting the highest area under the ROC curve (AUC) value of 0.913, alongside an accuracy of 0.830, positive predictive value of 0.857, and negative predictive value of 0.807. For the selected model, SHAP (SHapley Additive exPlanations) values were determined for each of the variables to establish which had the best predictive ability. Notably, variables pertaining to embryo biopsy demonstrated the greatest predictive capacity, followed by factors associated with ovarian stimulation (COS), maternal age, and paternal age. Conclusions The Random Forest model had a higher predictive power for identifying BPL occurrences in PGT-A cycles. Specifically, variables associated with the embryo biopsy procedure (biopsy day, number of biopsied embryos, and number of biopsied cells) and ovarian stimulation (number of oocytes retrieved and duration of stimulation), exhibited the strongest predictive power. Keywords: Biochemical pregnancy loss (BPL), NGS, PGT-A, Machine learning, Artificial intelligence (AI), SHAP value