In our study, data from 33,915 frozen-thawed embryo transfer cycles were analyzed and 25 features were extracted from 46 candidate predictors that could have an impact on live birth events, from which logistic regression models, random forest models, support vector machine models, and XGBoost models were constructed. Comparison of model performance revealed that the prediction model based on the XGBoost algorithm had the largest area under the ROC curve (0.750, 95% CI 0.746–0.755) and outperformed the other models in terms of accuracy, precision, recall, and F1 score performance.
Female age is one of the most important factors influencing the outcome of frozen-thaw embryo transfer. Many studies have shown that older women have lower clinical pregnancy and live birth rates than younger women, which may be associated with a decrease in ovarian function after the age of 35 years in women[29], and the results of this study confirm this. The type of embryo and the number of embryos transferred are also among the factors affecting FET live births. The results of Pan, Y et al.[30] showed that the number of frozen embryos transferred had a significant effect on the live birth rate, with higher live birth rates for cycles in which two embryos were selected for transfer than for cycles in which one embryo was transferred. It is also evident from the shap summary plot in our study that the number of embryos transferred with two has a positive effect on live birth events. The choice of embryo type can likewise have an impact on pregnancy outcome. The SHAP summary plot of this study has indicated that the choice of blastocyst for transfer is more likely to result in live birth than cleavage stage embryos, consistent with most previous studies[31, 32]. This may be due to the fact that the period of cleavage stage to blastocyst development itself eliminates nearly half of the embryos with developmental potential before they reach the blastocyst stage[33, 34]. Endometrial thickness can also have an impact on the outcome of in vitro fertilization-embryo transfer. A large study conducted by Liu, KE et al.[35]showed a significant decrease in clinical pregnancy and live birth rates in frozen-thawed embryo transfer cycles with endometrial thickness below 7 mm. Bu et al.[36] analyzed the relationship between endometrial thickness on the day of embryo transfer and pregnancy outcome in FET cycles and found a significant effect of endometrial thickness on the outcome of the FET cycle. Our study was consistent with the findings of these studies.
To date, several studies have been used to predict outcomes in IVF-embryo transfer, including prediction of outcomes such as biochemical pregnancy, clinical pregnancy, early miscarriage, and embryo development[37, 38]. Live birth events are the outcomes of most concern to clinicians and infertile couples, and there are relatively few predictive models for live birth outcomes[39]. Traditional statistical methods models are usually based on specific assumptions and a priori knowledge, do not take full advantage of the information in the data and in most cases require manual adjustment of the parameters of the prediction model. In contrast, machine learning methods can automatically uncover patterns in the data and are more flexible. In terms of adjusting model parameters, machine learning models are also automated and can easily handle large amounts of data and complex models, reducing the cost and error of human intervention and providing higher prediction accuracy, especially in complex nonlinear data analysis and prediction tasks[40, 41].
Our study had several advantages that could enrich the current body of knowledge on the outcome of frozen-thawed embryo transfer. First, our study employed a large sample, which may be more representative and generalizable compared to smaller studies. Second, interpretability is crucial for clinical decision making, and the black box of machine learning has been difficult to convince clinicians and patients for predictive models related to clinical outcomes. In our study, the SHAP algorithm was used to explain the contribution of each predictor to the final outcome, allowing clinicians to understand the impact of each factor on the likelihood of live birth. Third, we used four different machine learning algorithms to develop predictive models for freeze-thaw embryo transfer outcomes, reducing the risk of bias as well as overfitting that can occur when using a single model and improving the robustness of the study results. Fourth, the use of the developed prediction models has the potential to improve the outcomes of frozen-thaw embryo transfer and to reduce medical costs and the psychological and financial burden on patients to some extent.
Our study also had some limitations. First, all data were collected from one center, which may affect the effectiveness of our prediction model in applying it to populations from other health care institutions. Future studies may consider including populations from multiple centers or different regions, which may provide a more comprehensive picture of the predictive factors affecting FET outcomes. Second, although our study attempted to include as many predictors as possible, some factors not included may still be relevant, and further studies of a broader range of potential predictors may achieve better results.
In summary, we incorporated a total of 46 predictors for patient demographics, laboratory findings, and FET cycle-related parameters, and identified 25 features after feature screening to construct prediction models, and developed four different algorithms for predicting live birth for frozen-thaw embryo transfer, with the best performing XGBoost algorithm providing the possibility of identifying live birth and non-live birth populations, thus aiding clinicians in their decision making.