One of the most overwhelming challenges in contemporary assisted reproductive technology (ART) is how to narrow the gap in perinatal and neonatal outcomes between spontaneous pregnancy and assisted pregnancy, usually caused by multiple-embryo transfer and implantation [20]. In recent decades, SET has been advocated in IVF to prevent multiple pregnancies while led to the concerns of decreased IVF success [3, 21, 22]. Therefore, precise embryo selection for SET and twin risk warning for DET are particularly necessary. In this paper, a novel hierarchical model was constructed and validated to optimize embryo selection strategies and successfully predict pregnancy for both SET and DET, as well as predict the twin risk of DET for each individual. Subsequently, we validated our model on 2961 patients, and the results showed that our model had achieved an acceptable performance on embryo selection and twin risk prediction for each individual, as shown in Fig. 4 and Fig. 5.
It is generally accepted that embryos with similar morphology present variable implantation probability, depending on other assessed features such as patient characteristics and cycle demographics. Many features, including age, FSH level, anti-Mullerian hormone (AMH) and embryo quality, have been reported as independent impact factors on oocyte variability and embryo implantation potential [5, 23–26]. Machine learning is considered a powerful mathematical tool for correlation analysis when huge data is involved; therefore, AI has been introduced in embryo variability prediction in many articles [27, 28]. By increasing the information of the input features, the predictive power of the proposed model may be improved. Similarly, 19 features from the initial data sets were considered in the current study to construct the predictive power in our model. It is noteworthy that although previous models may have helped predict the implantation potential of a specific embryo, they did not offer a reasonable and optimal guidance to the embryo selection determination in clinical practice.
Minaretzis et al. [29] once presented an embryo selection strategy of transferring one additional good-quality embryo for each five years of incremental increase in maternal age to improve IVF outcome by multivariate analysis of factors predictive of successful live birth. Because maternal age was the only influencing factor to be considered, the recommended strategy was not so accurate or comprehensive. Kaufmann et al. [30] reported a neural networks predicting model in IVF, where four input parameters were included and the overall accuracy was 59%. Uyar et al. [31] proposed a Support Vector Machine (SVM) method in embryo implantation prediction in terms of Area Under ROC curve (0.712 ± 0.032), when 12 features were included.
Vaegter et al. [16] also constructed a predictive model for an embryo transfer strategy, in which only one embryo was transferred if the predictive risk of twin implantation was above 15%, using the two highest scored embryos. While validating this predictive model, the actual twin rate was 3.8%, which was far lower than the expected 15% setting, accompanied by a decreased live birth rate. Unlike previous studies, a novel strategy in embryo selection based on an accurate embryo potential prediction and twin risk assessment was developed in our model. The model not only guides a determination of the number of embryos for transfer and the specific embryo selection, but also presents a relatively accurate prediction of the pregnancy rate and twin risk of the corresponding selection scheme.
Because SET was not arbitrarily implemented in all IVF patients, twin pregnancy was inevitable. However, the acceptable twin rate threshold varied among different countries, even among different IVF centers. The greatest strength of our model was that the embryo selection strategy we developed varied correspondingly to patients’ characteristics and embryo morphology parameters as long as different twin rate thresholds were set. Therefore, the selection strategy model is applicable to any IVF center in any country. In such a situation, embryologists can input their accepted twin rate setting on the model to generate their guided embryo selection strategies and present a predictable pregnancy possibility and twin rate assessment.
More important, the predictive twin risk indicated that the transferred two embryos were neither a simplified embryo implantation nor an implantation of two independent samples. This principle was previously discussed in an embryo-uterus modeling framework [32–34]. A higher twin rate than expected would be achieved if the implantation chances of the embryos transferred together were completely independent of one another [16]. Considering the variability of confounding factors, a two-level algorithm was introduced in our model for DET prediction to achieve a satisfying predictive model. Furthermore, unlike previous researches, which only considered significant influencing factors in model construction [35–37], concomitant investigation in our DET model provided a novel insight: that even if one variable were not statistically significant, it might still be important for machine learning models to predict final outcomes.
To minimize twin risk, some researchers also constructed a predictable model and validated it in subsequent IVF cycles, and the results showed a significantly reduced twin rate (from 25.2–3.8%) [16]. However, the proportion of SETs was greatly increased (11.3–75.5%), and surplus embryos would be preserved for future use, resulting in a reduced live birth rate in fresh cycle (29.0–25.1%). Even though the cumulative live birth rate (CLBR) was not affected, patients had to pay more fees and spend more time for a successful pregnancy in subsequent frozen-thawed cycles. Luke et al. [38] also reported a similar conclusion of comparative CLBR and decreased multiple births in SETs over two cycles, compared to DET in one cycle. In our algorithm, conversely, the pregnancy rate and twin risk were assessed and predicted simultaneously in the fresh cycle, and further validation showed that it effectively reduced twin risk without compromising clinical pregnancy in the fresh cycle.
The recommended embryo selection strategy in our model provides decision support to embryologists with higher accuracy and efficiency. For a large proportion of IVF patients with suboptimal prognosis, which plan to choose was a dilemma to embryologists, because prediction judgment was based on their clinical experience instead of on the analysis of thousands of embryos and patient records prior to each embryo transfer [39, 40]. Besides, it may also act as a counseling tool for clinicians to evaluate the chance of pregnancy before the transfer procedure.
Previous research to establish similar predictive models were problematic due to the limited number of involved features or dramatic distinction between the predicted value and the actual situation [16, 29]. The model in the current study guides a determination of the number of embryos for transfer and the specific embryo selection, based on the analysis of a large number of features. In addition, the predicted pregnancy rate and twin risk of the corresponding selection scheme were relatively more accurate. More significantly, our model provides a flexible strategy, with individualized embryo selection for any given patient and varied operations corresponding with any setting acceptable twin rate threshold.
One shortcoming of our predictive model was that the predicted mean value was slightly elevated compared to the actual observed rate for both pregnancy and twin rates in validation, although a modest discrimination was also reported in the previous prediction model in ART [41–43]. The model should be adjusted and improved to make the predictive value more closely agree with the actual virtue in the future verification. Besides, the model was developed and validated using data from a single center. As future work, we will need to be much more rigorous with additional data sets demonstrating the predictive value, using the same parameters, as well as applying them in different centers.