Predicting and Improving the Probability of Live Birth for Women Undergoing Frozen-thawed Embryo Transfer: a Data-driven Estimation and Simulation Model

BACKGROUND AND OBJECTIVE Frozen-thawed embryo transfer (FET) is now widely used for the treatment of infertility. For many couples and clinicians, concerns over the probability and how to increase the chance of a successful birth are very common. Currently, there is not a single model to predict the live birth outcomes for FET. To estimate the probability of live birth (PLB) in FET and to provide advice on potential treatment options by a data-driven predictive (DDP) model. METHODS 2,189 FET cycles from Jan 2012 to Dec 2015 were recruited in a single center. 815 cycles of FET outcomes were live births and 1,374 cycles of FET outcomes failed. To verify the consistency of the DDP model, we carried out 10-fold cross-validation, and the mean and standard deviation of the accuracy were measured. Moreover, the performance of this model was evaluated by the mean and standard deviation of receiver operating characteristic curve and area under the curve (AUC). RESULTS Nine dominant factors, including age, BMI, HOMA-IR, basal follicle stimulating hormone, basal luteinizing hormone, basal estradiol, endometrial thickness, the number of embryo transfers and the total number of embryos, were automatically extracted from 28 candidate factors. The accuracy of our prediction model is 76.9%±1.6%, and the AUC is 0.83. Then, the PLB is estimated by the random forest algorithm. On this basis, the DDP model can comprehensively traverse and dynamically visualize the impact of several factors on live birth outcomes. Finally, optimal suggestions for the treatment of patients before FET are attempted to be made by the genetic algorithm. CONCLUSION The DDP model can not only provide satisfactory performance for predicting live birth outcomes in FET but also offer a visual estimation and simulation tool for clinicians to make treatment plans.