Humidity level during culture does not affect mouse embryo development

Abstract

BACKGROUND: Despite decades of experience with in vitro fertilization, there is still significant uncertainty in predicting patient outcomes. Standard operating procedures for clinical decision-making rely on few variables analyzed independently of each other. While the predictive power of a single variable, such as age, is clinically useful, a more sophisticated and personalized approach could improve the quality of patient care.OBJECTIVE: The purpose of this study was to determine the potential of machine learning algorithms to predict successful blastocyst formation. This information will guide the decision to push patients to a day 5 transfer while minimizing the no-transfer rate due to arrested embryo development.MATERIALS AND METHODS: Using our fresh IVF cycle data from 2014-2018 (n ¼ 1777), we trained multiple classification algorithms in R 3.5.1. The training data was a randomly-selected subset of 70% of the total data. Using data that was available on or before day 3 of embryo development, the classifiers were trained using 10-fold cross-validation and tested for performance. Successful blastocyst formation was defined as successful pregnancy or the presence of blastocysts on day 5. The best performer on the pre-specified test data was a gradient boosted machine (GBM). Algorithm performance and variable importance metrics were calculated and the algorithm was simplified to the 10 most predictive variables before being reapplied to the pre-specified test set.RESULTS: The final GBM accurately predicted blastocyst formation with an AUC of 0.922 and an accuracy of 0.858. The variables with the greatest predictive value were the number of good-quality day 3 embryos, number of mature oocytes retrieved, age at retrieval, maximum FSH and estradiol levels, and semen parameters. This algorithm was significantly more accurate at predicting blastocyst formation than our practice's current standard of using >4 good-quality day 3 embryos (95% CI: 0.80-0.90 vs. 0.64-0.71). Of all patients that received day 3 transfers, the algorithm predicted blastocyst formation in 167 patients, of which 148 (89%) formed successful blastocysts, while 19 patients (11%) did not.CONCLUSION: A machine learning approach provides notable discriminatory power for prediction of blastocyst formation in patients undergoing in vitro fertilization. The use of a GBM in our practice will increase the number of patients receiving a day 5 transfer, which in turn should increase both implantation rate and the number of single embryo transfers while decreasing the incidence of multiples. This tool may help providers optimize embryo transfer protocols and can be used to counsel patients to improve pregnancy outcomes.