Construction of the machine learning-based live birth prediction models for the first in vitro fertilization pregnant women

Infertility is defined as the failure of husband and wife to get pregnant after 12 months of regular sexual intercourse and without using any contraceptive method [1]. It is estimated that 8–12% of couples of reproductive ages worldwide face this problem, affecting family well-being and social stability worldwide [2]. Thereby, the demand for in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI) is increasing, with over 8 million babies born through IVF or other assisted reproductive technology treatments since the world’s first in 1978 [3]. Although the success rate of assisted reproductive technology increases with newer techniques, IVF does not guarantee success in fertilization [4]. Considering the high financial burden and inconclusive health risk of IVF [5], it is necessary to explore the factors related to the live birth of women with IVF and construct prediction models to predict the live birth rate of IVF pregnant women for patients counseling and shaping expectations.

Previous studies have reported that age, duration of infertility, previous live birth, previous miscarriage, previous abortion, and type of infertility were associated with the IVF outcome [6, 7]. Using these parameters that can be obtained before the IVF cycle to evaluate live births may improve the treatment efficiency of assisted reproductive technology and reduce the waste of resources. Moreover, assisted reproductive process-related factors, such as stimulation protocols, and number of embryos transferred were also found to be related to the outcome of assisted reproductive technology [8, 9]. The outcome prediction model based on the above post-cycle factors is conducive to doctor-patient communication and can help clinicians adjust assisted reproductive planning. The traditional prediction relies on the comprehensive evaluation of the doctor based on the age of the patient and the live birth rate of the reproductive center, which is highly subjective [4]. Up till now, reliable and accurate prediction of IVF outcomes has always been an outstanding issue. Machine-learning algorithms have recently been used to predict IVF outcomes on the basis of multiple clinical variables [7]. A recent study reported that the predicted probability of live birth in IVF varies during important stages throughout the treatment [10]. However, few studies have established prediction models respectively before the first cycle and after the first cycle. Assessing outcomes before and after the first cycle is critical to provide individualized treatment for infertile patients. Moreover, previous prediction models for live birth were mostly based on fresh embryo transfer [11, 12], but did not fully consider frozen embryo transfer. In view of the development of frozen embryo transfer, it is necessary to conduct prediction models to predict live births of based on the fresh and frozen embryo transfer.

Herein, the objective of the present study is to use machine learning methods to conduct prediction based on parameters before and after the first cycle, respectively, to predict live births in women who received IVF for the first time.

In this study, among the 1,857 women, 851 women (45.83%) ended up with live birth. We respectively constructed 8 machine learning models to predict the live birth for the first complete IVF attempt or ICSI using predictive factors before and after the first cycle. The result showed that the LGBM model achieved a robust performance before and after the first cycle compared with LR, XGB, and RF. LGBM showed better predictive performance on the live birth chance after the first cycle compared with the prediction models before the first cycle. The predictive factors before the first cycle included maternal age, AMH, basal FSH, basal P, basal E2, infertility duration, and the number of left sinus follicles. Sixteen predictive factors were identified that affect the live birth before and after the first cycle of IVF/ICSI including maternal age, AMH, basal E2, basal PRL, infertility duration, the number of pregnancies, the number of left sinus follicles, serum HCG level, EMT on HCG day, LH level on HCG day, E2 level on HCG day, Gn, method of fertilizations, types of transferred embryos, stimulation protocols, and endometrial preparation protocol.

Several predictive models have been developed to assess the outcome of IVF treatment. One of the early and most accepted prediction models is the McLernon model [16, 17], which utilizes only discrete LR to predict the chance of live birth for a couple. In addition, the McLernon model did not take several important factors such as BMI and AMH into account. Unlike classical statistics, machine learning algorithms can take into account complex associations between different parameters, and thus can better exploit the synergy between these correlated parameters [12]. ZoharBarnett-Itzhaki et al. [12] used age, BMI, and clinical characteristics to predict IVF outcomes in 136 women undergoing fresh IVF cycles. The author found that compared with LR, the accuracy of artificial neural network (NN) and support vector machine (SVM) was 0.69 to 0.9 and 0.45 to 0.77, respectively, while the accuracy was 0.34 to 0.74 using LR model, indicating that machine learning algorithms based on age, BMI, and clinical data are superior to LR in predicting IVF outcomes. CelineBlank et al. used the parameters before and after the cycle to develop a RF model to predict the implantation rate after fresh blastocyst transfer and compared it with the LR model [18]. The results showed that the AUC of the RF model was 0.74, and the AUC of the traditional LR model was 0.66. Qiu et al. used a machine learning method to establish a personalized prediction of live birth prior to the first IVF and found that the XGBoost model achieved an area under the ROC curve of 0.73. However, this model was only established based on the factors before the first cycle [7]. In this study, we used machine learning techniques to predict live birth in women undergoing IVF treatment or ICSI based on the predictive factors before and after the first cycle. Our results showed that the LGBM model had a robust performance before and after the first cycle of IVF treatment in comparison with LR. Our study also indicated that the prediction model indicated a better predictive performance on the live birth chance after the first cycle compared with the prediction models before the first cycle. We speculate that this possible reason may be that the prediction model based on the predictive factors after the first cycle is closer to the live birth outcome, so it is more suitable to predict the live births. This study might be a promising step to provide personalized estimates of live birth chance of the IVF or ICSI before and after the first cycle based on the machine learning algorithms.

In previous prediction models, female age was the most established predictive factor associated with live birth after IVF/ICSI [7, 12, 19]. A study by Hassan et al. proposed a hill-climbing feature selection algorithm with five different machine learning models to analyze and predict IVF pregnancies with higher accuracy and has found that age was the most important factor affecting IVF pregnancy outcomes [6]. Our study demonstrated that maternal age was related to live birth for women before and after the first cycle of IVF/ICSI. Female ovarian reserve gradually declines with age, and the quantity and quality of oocytes decrease significantly, which has a significant negative impact on the success of pregnancy in patients undergoing IVF/ICSI treatment [20]. We observed that the infertility duration, the number of pregnancies, and the number of left sinus follicles were associated with the birth rate in women with IVF treatment. Mo et al. found that the number of infertility years, and sinus follicles were related to the success of treatment of IVF-embryo transfer [21]. A prediction model has reported that the duration of infertility was the predictive factor of live birth prior to the first IVF treatment [7].

In this study, hormone levels could be the predictive factors for live birth in women with IVF or ICSI. AMH, basal FSH, basal P, and basal E2 were associated with live birth before the first cycle of IVF/ICSI and AMH, basal E2, basal PRL, serum HCG level, LH level on HCG day, E2 level on HCG day, and Gn were found to affect the live birth after the first cycle of IVF/ICSI. Gn has been used to induce multiple follicle development and increase ovarian stimulation efficiency in infertile women since the 1960s [22]. AMH can inhibit follicular proliferation and growth by limiting the functions of growth factors and Gn, and it directly reflects the condition of the primordial follicles and effectively predicts the patient's ovarian reserve [23]. A previous systematic review suggested that AMH has some value in the prediction of live birth, and can be a predictor of live birth in women undergoing assisted conception [24]. Estrogen levels in women would be reduced as ovarian reserve decreases, resulting in an increase in FSH secretion. The elevated basal FSH level is associated with the decline of ovarian reserve [25]. A study evaluating the impact of different etiologies of diminished ovarian reserve on pregnancy outcome in IVF-ET cycles indicated that elevated basal FSH affects the number of oocytes retrieved and the implantation rates after IVF/ICSI [26]. Farhi et al. reported that elevated E2 levels during IVF cycles were linked to a higher incidence of adverse pregnancy outcomes [27]. Lower basal E2 was associated with better pregnancy rates and ovarian reserve during IVF utilizing a gonadotropin releasing hormone (GnRH) antagonist [28]. Morales et al. reported serum E2 on the day of the trigger as a predictor of metaphase II oocytes in antagonist cycles encourage greater oocyte maturity and fertilization [29]. Chen et al. found that high serum E2 level on HCG day was associated with decreased live-birth rates in patients with frozen embryo transfer [30]. Previous studies showed that the alterations of these hormones such as basal LH, basal PRL, and basal P levels were associated with pregnancy in embryo transfer [31‐33]. In addition to the hormone levels, a study [34] evaluating the probability of live birth after a freeze-all based on an IVF treatment strategy found that EMT was one of the most important predictive factors. A meta-analysis indicated that women with thinner EMT are inversely associated with IVF/ICSI success, whereas higher endometrium tends to confer the best outcome [35]. Wen et al. [19] have identified EMT as one of the independent predictors affecting IVF/ICSI success. EMT on HCG day was highlighted as a predictive factor for live birth in our model after the first cycle of IVF/ICSI.

In this study, the method of fertilization and types of transferred embryos were associated with live births in women who received IVF or ICSI. Previous studies indicated that the quality and implantation potential of frozen embryos is similar to or even better than that of fresh embryos [36‐38]. In practice, the freeze-all strategy can reduce the risk of ovarian hyperstimulation syndrome (OHSS) in the ovarian stimulation cycle by avoiding pregnancy and obtaining better results [39]. However, the results of a previous meta-analysis showed no differences in fertilization rate, total fertilization failure rate, good embryo quality rate, fresh embryo implantation clinical pregnancy rate, fresh embryo transfer live birth rate, miscarriage rate, neonatal preterm birth rate and neonatal low birth weight rate in those treated with ICSI compared with IVF [40]. The relationship between the fertilization method and live birth needs further clarification. We found that stimulation protocols and endometrial preparation protocols were related to the live births in women who received IVF or ICSI. A previous study indicated that live birth rates per transferred embryo were significantly increased by GnRH agonist administration in women stimulated either with the long protocol or with a GnRH antagonist [41]. In women undergoing single euploid frozen blastocyst transfer, the natural cycles group had a lower pregnancy loss rate and ultimately a higher live birth rate compared with the hormone replacement treatment group [42]. A retrospective study of 12,950 frozen embryo transfer cycles published by Li et al. concluded a comparable clinical pregnancy rate and a lower live birth rate when hormone replacement treatment cycles were compared with natural cycles [42].

Our prediction models not only estimate the chances of success in couples before commencing IVF but also are able to revise these chances on the basis of the couple’s response to a first treatment cycle. The prediction model before the first cycle can be used to interrogate couples before treatment, while the predictive model after the first cycle predicts the chances of future success for couples considering further treatment. Our study has a number of strengths. First of all, we introduced the parameters before and after the first cycle when constructing the models for live birth, which provided the summary information of the success opportunities before and after the treatment. Secondly, machine learning methods were used to predict live birth in pregnant women undergoing the first IVF in the Chinese population, including fresh and frozen embryo transfer. Thirdly, the factors included are comprehensive, which improves the prediction ability of the prediction models. However, several limitations deserve to be mentioned. Firstly, the study was retrospective, single-center, and encompassed a limited number of patients. Secondly, the lack of external validation and the study population was pregnant women undergoing IVF for the first time, which may limit the popularization of our research results. Thirdly, without an impact analysis, it is impossible to verify whether the clinical decisions made by clinicians combined with the prediction performance of the models can improve the live birth rate or reduce the treatment cost. Finally, the prediction models were constructed based on limited predictive factors obtained before and after IVF treatment, however, pregnancy is a dynamic and ongoing process, and there are many other confounders that have an effect at different time points.

This study indicated that machine learning-based predictive model could be used as a tool to predict live birth in women with IVF cycles, This study may help predict the chances of live birth before and after the first cycles of IVF or ICSI using personalized information, helping shape couples’ expectations of their IVF or ICSI outcome, allowing them to plan their treatments more efficiently and prepare emotionally and financially.