Prediction of risk factors for first trimester pregnancy loss in frozen-thawed good-quality embryo transfer cycles using machine learning algorithms

Purpose Can the risk factors that cause first trimester pregnancy loss in good-quality frozen-thawed embryo transfer (FET) cycles be predicted using machine learning algorithms?

This is a retrospective cohort study conducted at Sisli Memorial Hospital, ART and Reproductive Genetics Center, between January 2011 and May 2021. A total of 3805 good-quality FET cycles were included in the study. First trimester pregnancy loss rates were evaluated according to female age, paternal age, body mass index (BMI), diagnosis of infertility, endometrial preparation protocols (natural/artificial), embryo quality (top/good), presence of polycystic ovarian syndrome (PCOS), history of recurrent pregnancy loss (RPL), recurrent implantation failure (RIF), severe male infertility, adenomyosis and endometriosis.

The first trimester pregnancy loss rate was 18.2% (693/ 3805). The presence of RPL increased first trimester pregnancy loss (OR = 7.729, 95%CI = 5.908–10.142, P = 0.000). BMI, which is > 30, increased first trimester pregnancy loss compared to < 25 (OR = 1.418, 95%CI = 1.025–1.950, P = 0.033). Endometrial preparation with artificial cycle increased first trimester pregnancy loss compared to natural cycle (OR = 2.101, 95%CI = 1.630–2.723, P = 0.000). Female age, which is 35–37, increased first trimester pregnancy loss compared to 37, increased first trimester pregnancy loss compared to < 30 (OR = 2.286, 95%CI = 1.146–4,38, P = 0.016). The presence of PCOS increased first trimester pregnancy loss (OR = 1.693, 95%CI = 1.198–2.390, P = 0.002). The number of previous IVF cycles, which is > 3, increased first trimester pregnancy loss compared to < 3 (OR = 2.182, 95%CI = 1.708–2.790, P = 0.000).

History of RPL, RIF, advanced female age, presence of PCOS, and high BMI (> 30 kg/m2) were the factors that increased first trimester pregnancy loss.

First trimester pregnancy loss, Frozen-thawed embryo transfer (FET) cycles, In vitro fertilization (IVF), Machine learning algorithms, Infertility

The success rates of ART applications have increased significantly over the last 44 years. Developments in laboratories, including embryo culture systems, embryo cryopreservation techniques, and embryo selection methods, have contributed to this success. In addition, advances in clinical applications such as the controlled ovarian hyperstimulation (COH) protocols have also contributed [1]. However, despite all these developments, the rates of early pregnancy loss are still 15–25% (Society for Assisted Reproductive Technology:SART,2020). Many studies have shown that early pregnancy loss causes serious physical and psychological problems for couples. Psychological problems include the risk of anxiety, depression, and stress disorders. The physical consequences depend mainly on surgical complications, such as bleeding, infection, and intrauterine synechiae [2–4]. It is well-known that good embryo quality, endometrial receptivity, and the synchronization of embryo and endometrial development are required for the success of assisted reproductive technology (ART). Endometrial thickness is considered as an indicator receptivity of the endometrium. Studies have shown that early pregnancy loss is more common in cases with damaged and thin endometrium [5–7]. Nevertheless, pregnancy sometimes ends in early pregnancy loss in ART, even though these conditions are met. Although there are some studies investigating the risk factors for early pregnancy loss in FET, as far as we know, this is the first study to exclude endometrial and uterine problems and to discuss only good-quality embryo transfer cycles. The main objective of this study was to construct a model with well-known risk factors that cause first trimester pregnancy loss in frozen-thawed embryo transfer (FET) cycles using machine learning (ML) and statistical models. Using this model, we analyzed the relative predictive power of each feature. Machine learning, a subfield of artificial intelligence, is a powerful and flexible tool for analyzing and predicting outcomes from clinical data. It is used extensively in various fields. Since it can automate the process while analyzing large amounts of data, the use of ML is inevitable in medicine as well, as recent studies have shown [8, 9]. Islam et al. recently published a systematic review of ML for predicting pregnancy outcomes. According to this study, a total of 26 articles published in the last two decades were selected from the original set of 241 articles. While the study focuses on predicting the optimal mode of childbirth and detecting various complications during childbirth through ML, only three studies out of them, namely, by Liu et al. [10], Hassan et al. [11], and Qui et al. [12], are related to IVF treatment [13]. In this context, we believe that our study will fill the gap in the relevant literature.

Ethical approval This study was approved by the Institutional Review Board of Istanbul Memorial Sisli Hospital, Istanbul, Turkey (approval number 24.12.2021/008). Patient selection This study is a retrospective study. We investigated pregnancy outcomes in frozen-thawed top-quality or good-quality embryo transfer cycles at the IVF and Genetics Center of Istanbul Memorial Sisli Hospital between January 2011 and May 2021. A total of 3805 transfer cycles were included in the study. The inclusion criteria were cases with frozen-thawed top-/good-quality blastocyst transfer cycles with normal endometrial thickness. The exclusion criteria were women over 42 years of age, cases with congenital and acquired uterine abnormalities, patients with endometrial thickness less than 7 mm, cases with untreated endocrinological diseases, and cycles with preimplantation genetic diagnosis/screening. The rates of first trimester pregnancy loss were evaluated according to various parameters such as female age, paternal age, body mass index (BMI), diagnosis of infertility, endometrial preparation protocols (natural vs artificial), embryo grade top-quality (TQ)/good-quality(GQ), the presence of polycystic ovarian syndrome (PCOS), the history of recurrent pregnancy loss (RPL), recurrent implantation failure (RIF), severe male infertility, adenomyosis and endometriosis. The subjects were divided into four groups by age (< 30, 30–34, 35–37, 37–42) and three groups by BMI ( 30 kg/m2). First trimester pregnancy loss was defined according to the definition of Clinical Practice and Coding policies by The American College of Obstetricians and Gynecologists (ACOG): “Non-viable intrauterine pregnancy with no fetal heart activity, empty gestational sac or gestational sac containing embryo or fetus within the first 12 weeks.” Ongoing pregnancy was defined when the pregnancy had completed > 12 weeks. RPL was defined by the spontaneous loss of two or more pregnancies [14]. RIF was defined as the absence of a positive pregnancy test after three consecutive transfers of good-quality embryos [15]. Modified natural cycle This protocol was applied to cases with regular menstrual cycles between 24 and 35 days. The uterus and ovaries were examined with transvaginal ultrasound on the second day of the spontaneous menstrual cycle. There were no hormone-producing cysts in the ovaries and/or problems with the endometrium such as polyps or fibroids, and the endometrial preparation cycle was started. The following examination was performed on the tenth or twelfth day of the menstrual cycle, depending on the cycle length. Once the dominant follicle reached a mean diameter of 16–20 mm and serum level LH > 15 IU/L, 6500 IU recombinant human chorionic gonadotropin (r-hCG) (Ovitrelle, Merck Serono, Switzerland) was administered to induce ovulation. We only included in our study those cases that did not have problems with the endometrium and where the endometrium was larger than 7 mm and had a triple-line pattern. To support the luteal phase, micronized vaginal progesterone (Crinone® 8%; Merck Serono, Switzerland) once daily or natural vaginal progesterone tablets (Lutinus 100 mg; Ferring, Germany) twice daily were started 36 h after the ovulation trigger. The top-quality or good-quality blastocyst was thawed and transferred 4 days after progesterone support. The pregnancy test was performed 9 days after transfer. If the pregnancy test was positive, luteal phase support was continued until the 10th gestational week. Artificial cycle The transdermal estradiol patch 3.9 mg (Climara®, Bayer Turk, Turkey) or the estradiol tablet 2 mg (Estrofem®, Novo Nordisk, Denmark) was used on the second day of menstruation after a transvaginal ultrasound examination. The estradiol tablet was taken three times a day. The ultrasound examination was repeated on the 11th and 15th day. Luteal phase support was initiated when the thickness of the endometrium was more than 7 mm and after estradiol had been taken for at least 12 days. Micronized progesterone vaginal gel (Crinone® 8%; Merck Serono, Switzerland) twice daily or natural vaginal tablets (Lutinus 100 mg; Ferring, Germany) 200 mg twice daily were administered to support the luteal phase. Embryo was transferred 5 days after the start of progesterone administration. The use of estradiol was continued until the 8th week of pregnancy and the use of progesterone until the 10th week of pregnancy. Embryo scoring All blastocysts were scored by an experienced embryologist using the scoring system described by Gardner and Schoolcraft. TQ embryo comprises 3AA-4AA-5AA-6AA, and GQ embryo comprises 3AB-4AB-5AB-6AB-3BA-4BA-5BA,6BA. Inferior quality embryos were excluded from this study. Good- or top-quality blastocysts were frozen on day 5 and 6 in the morning using Kitazato Vitrification Media (Kitazato, Japan) according to the manufacturer’s instructions. Blastocysts were thawed according to Kitazato Warming Media. Embryos with grade loss after thawing were not included in the study. Machine learning Machine learning (ML) is a set of statistics-based approaches that make predictions about new and future observations by using past experiences and/or observations [16]. It is known as “learning and improving from experience without being explicitly programmed” and includes all algorithms from linear regression to deep learning [17]. In general, it is a subfield of artificial intelligence (AI) and is classified into two basic groups depending on whether or not they are trained under human supervision or not, supervised or unsupervised [18]. In addition to these, reinforcement learning, which is based on estimated errors as rewards or punishments and semi-supervised learning which uses both labelled and unlabelled data for training, also belongs to categories of ML. In this study, random forest (RF), one of the tree-based machine learning methods, was preferred to select the most important and relevant features for determining the risk factors leading to pregnancy loss in the first trimester in the FET cycle. A stepwise multivariate logistic regression model (LR) was then built using the features identified by RF. Before briefly explaining RF and LR, some terms should be clarified. In the equation, Yi = f(Xi), f(.) expresses the model, while Xi = [x1, x2, …, xn] vector i represents the samples, Y is the dependent/target variable that needs to be predicted, and xj's are the independent/explanatory variables, J:1,…,n. Each sample represents which vector corresponds to n explanatory variables. In this study, the explanatory variables are defined as the characteristics that are patient-specific factors including demographic and clinical factors; the dependent/target variable is labeled as the first trimester pregnancy loss (positive) and ongoing pregnancy (negative). Random forest Random forest is an ensemble of decision trees proposed by Leo Breiman in 2001 [19]. In addition to classification and regression tasks, it also calculates the importance of a feature to measure its contribution. RF generates randomness from a standard data set using the bagging method. To create a forest, the dataset is resampled for each tree based on the bootstrap approach. Finally, the prediction results of the tree-based models are combined using either the mean value for regression or majority voting for classification to obtain the committee’s decision. Generally, decision trees for classification are built based on the degree of purity. The decision tree calculates the importance level according to how much each feature reduces impurity. Therefore, the final importance level of each feature in RF is determined by the average of the impurity reduction collected from each feature in all trees. To avoid overfitting, RF tends to create multiple trees that have been trained slightly differently [20]. In addition to this property, the parameters should be properly selected. Therefore, RF can be considered as a robust classifier that is less prone to overfitting. In this study, the tuning of the hyper-parameters was performed using the grid search by fivefold cross-validation. Logistic regression LR is basically a discriminative classification algorithm where the dependent/target variable (Y) is binomial, i.e., each case belongs to one of two categories [21]. It is worth noting that there are other types of LR besides binomial; it can be multinomial, i.e., multiclass, if the dependent variable has more than two possible categories, and it can be ordinal if the dependent variable has ordered categories. In addition to classification task, LR also tries to extract the best fitting model to define the relationship between the dependent and independent variables using a logistic function. Statistical analysis Since the demographic and clinical characteristics of patients were categorically used, all the descriptive analyses were conducted with Chi-square, and results were given in frequencies and percentages. According to the importance level of the patient-specific factors obtained with the RF, those with an importance level greater than a certain threshold were included in the model to be used in the analysis. It is important to note that the threshold varies according to datasets. Independent risk factors affecting first trimester pregnancy loss were determined by establishing a stepwise multivariable LR model with the characteristics determined by the RF. The dataset with selected characteristics based on the RF method was randomly split into two distinct subsets: a training set (70%) to construct the model and a test set (30%) to test the model. The LR model was established using the training set. After obtaining the LR model, it was tested with a test set to see if the model was valid for future patients. All analyses (descriptive, inferential, classification, and performance evaluations) were performed using Python [22] and R [23] with several libraries and packages. Pandas [24] and NumPy [25] for data loading and preprocessing, Matplotlib [26] for visualization, and Sklearn [27] for calculating importance levels of the variables and model evaluation were used in Python. stats package [28] in R was used for stepwise multivariable logistic regression.

In this study, 3805 frozen embryo transfer cycles with 16 patient-specific factors were analyzed. Six hundred ninety-three cycles (18%) were labeled as first trimester pregnancy loss, and the rest were labeled as ongoing pregnancy. All patient-specific factors were examined in terms of missing values. For determining risk factors that cause first trimester pregnancy loss in FET cycles, patients’ clinical and cycle characteristics are described in Table 1. Before applying RF model, hyper-parameter tuning were conducted using grid search through fivefold cross-validation. The optimal values for the hyper-parameters were obtained as follows: the number of trees to grow is 100, the number of features randomly sampled as candidates at each split is 7, the maximum number of depth in each decision tree is 5, and the minimum number of data points placed in a node before the node split is 5. The importance level that yielded by RF for each patient-specific factor was calculated, and the descending order of the importances was presented in Fig. 1. According to the order given in Fig. 1, 11 of the patient-specific factors were chosen for further analysis. Those with an importance level greater than the threshold which was chosen as 0.05 based on the results were included in the model to be used in the classification. These factors are recurrent pregnancy loss, number of previous IVF cycles, endometrial preparation, maternal age, diagnosis of infertility, paternal age, duration of infertility, number of transferred embryos, number of total oocytes, BMI, and PCOS. The logistic regression model was established using training data with these 11 variables with the highest importance level, and the results obtained by the established model are given in Table 2. According to this, the presence of RPL, high BMI, previous IVF number >3, advanced female age, the presence of PCOS, and endometrial preparation with artificial cycle were the most important risk factors for first trimester pregnancy loss. The presence of RPL increased first trimester pregnancy loss (OR=7.729, 95%CI=5.908–10.142, P=0.000). BMI, which is >30, increased first trimester pregnancy loss compared to <25 (OR=1.418, 95%CI=1.025–1.950, P=0.033). Endometrial preparation with artificial cycle increased first trimester pregnancy loss compared to natural cycle (OR=2.101, 95%CI=1.630–2.723, P=0.000). Female age, which is 35–37, increased first trimester pregnancy loss compared to 37, increased first trimester pregnancy loss compared to 3, which increased first trimester pregnancy loss compared to <3 (OR=2.182, 95%CI=1.708–2.790, P=0.000). Table 1. | Characteristic | No. of transfer cycle | No. of first pregnancy loss (%of all pregnancy) | P value | |---|---|---|---| | Total | 3805 | 693 (18.2) | | | Maternal age | ||| | < 30 30–34 35–37 37–42 | 1715 1581 401 108 | 290 (16.9) 284 (18) 87 (21.7) 32 (29.6) | 0.002 | | BMI | ||| | 30 | 2263 1018 509 | 374 (16.5) 188 (18.5) 131 (25.7) | 0.000 | | Endometrial preparation | ||| | Artificial cycles Natural cycles | 2285 1520 | 526 (23) 167 (11) | 0.000 | | Presence of absolute tubal factor | ||| | Yes No | 291 3514 | 53 (18.2) 640 (18.2) | 1.000 | | Diagnosis of infertility | ||| | Unexplained infertility Male factor Female factor Combined factor | 793 1377 1069 565 | 117(14.8) 224(16.3) 214(20) 138 (24.4) | 0.000 | | Presence of severe male infertility | ||| | Yes No | 1834 1971 | 344 (18.8) 349 (17.7) | 0.402 | | Presence of Hashimoto thyroiditis | ||| | Yes No | 130 3675 | 30 (23.1) 663 (18) | 0.144 | | Presence of RPL | ||| | < 2 ≥ 2 | 3332 473 | 417 (12.5) 276 (58.4) | 0.000 | | Presence of PCOS | ||| | Yes No | 1034 2771 | 250 (24.2) 443 (16) | 0.000 | | Presence of adenomyosis | ||| | Yes No | 36 3769 | 7(19.4) 686 (18.2) | 0.874 | | Presence of endometriosis | ||| | Yes No | 98 3707 | 15 (15.3) 678 (18.3) | 0.450 | | Duration of infertility(year) | ||| | 10 | 2184 1091 426 | 396 (18.1) 198 (18.1) 81 (19) | 0.907 | | Paternal age | ||| | 39 | 758 1608 989 442 | 138 (18.2) 274 (17) 191 (19.3) 87 (19.7) | 0.402 | | Number of previous IVF cycles | ||| | ≤ 3 > 3 | 2039 1766 | 239 (11.7) 454 (25.7) | 0.000 | | Total oocytes | ||| | 25 | 190 1615 1217 675 | 28 (14.7) 296 (18.3) 216 (17.7) 144 (21.3) | 0.118 | | Number of transferred embryos | ||| | 1 2 | 2693 1111 | 478 (17.7) 214 (19.3) | 0.272 | BMI, body mass index; RPL, recurrent pregnancy loss; PCOS, polycystic ovarian syndrome Table 2. | Characteristics | OR | 95% CI | P -value | |---|---|---|---| | Presence of RPL | ||| | < 2 ≥ 2 | 1.00* 7.729 | 5.908–10.142 | 0.000 | | BMI | ||| | 30 | 1.00* 1.030 1.418 | 0.796–1.329 1.025–1.950 | 0.820 0.033 | | Endometrial preparation | ||| | Natural cycles | 1.00* | || | Artificial cycles | 2.101 | 1.630–2.723 | 0.000 | | Female age | ||| | < 30 | 1.00* | || | 30–34 | 1.058 | 0.830–1.348 | 0.482 | | 35–37 | 1.617 | 1.120–2.316 | 0.018 | | 38–42 | 2.286 | 1.146–4.38 | 0.016 | | Presence of PCOS | ||| | No | 1.00* | || | Yes | 1.693 | 1.198–2.390 | 0.002 | | Number of previous IVF cycles | ||| | ≤ 3 | 1.00* | || | > 3 | 2.182 | 1.708–2.790 | 0.000 | As mentioned above, the dataset used in this study represents an imbalanced distribution of first trimester pregnancy loss (18%) and ongoing pregnancy (82%). It is essential to adjust the output threshold for imbalanced data set while applying classification algorithms. Therefore, ROC curve was employed to find the optimum probability threshold considering the trade-off in sensitivity and false alarm rate. As a result of the analysis, the threshold probability was estimated as 0.15 for LR. Once obtaining the model by LR using a training set, a test set was used to validate the model. In order for fair evaluation, a confusion matrix was calculated based on the results obtained using the test set. According to the confusion matrix, given in Table 3, the correctly classified first trimester pregnancy loss rate, i.e., sensitivity, was calculated as 69%; the correctly classified ongoing pregnancy rate, i.e. specificity, was calculated as 68%; and the overall success rate was calculated as 68%. The area under curve (AUC) which is drawn by true positive rate (sensitivity) vs false positive rate (1. specificity) can be seen in Fig. 2. As seen in the figure, the model has a 76.6% AUC score which can be an acceptable value [29]. Table 3.

The main objective of this study is to determine the risk factors that cause first trimester pregnancy loss in good-quality frozen-thawed embryo transfer (FET) cycles using machine learning and statistical models. Recently, studies similar to our study on using ML in IVF have been published. Liu et al. used machine learning approach to predict early pregnancy loss after the appearance of embryonic cardiac activity undergoing IVF-ET [10]. After comparing several classification algorithms, they concluded that random forest model can assist to make clinical decisions. In order to reach this conclusion, they used 31,030 cases with 64 clinical characteristics. They compared their studies in terms of AUC values with similar studies published by Uyar et al. and Yi et al. [8, 9]. Uyar et al. aimed to construct a decision support system based on ML on the number of embryos transferred to predict the implantation outcome of individual embryos in an IVF cycle. In their retrospective cohort study, 2453 embryos, with 18 clinical features, transferred on day 2 or day 3 after intracytoplasmic sperm injection were included in the dataset. The best AUC score (0.754) was obtained in the study using the naive Bayes algorithm [8]. It should be noted that the data set used in Uyar’s study represents an imbalanced distribution of positive ( non-implanted embryos) and negative (implanted embryos) implantation classes of the embryos dataset. In 2016, Yi et al. published a study to predict early pregnancy loss (EPL) after the occurrence of embryonic cardiac activity in patients who had undergone in vitro fertilization embryo transfer (IVF-ET) treatment [9]. In the study, a logistic model was constructed to assess the differences in ultrasound parameters between group with miscarriage and a group with ongoing pregnancy during the 1st trimester. They analyzed 2400 patients who had ongoing pregnancy and an additional 201 patients who experienced miscarriage after fetal cardiac activity in the first trimester. The results show that maternal age, mean diameter of gestational sac, crown–rump length, yolk sac diameter, fetal heart rate, duration of infertility, and fluid accumulation around the gestational sac each correlated with EPL after IVF in infertile patients. Hassan et al. [11] proposed an automated tool based on ML algorithms to predict IVF pregnancy success using a dataset of 1048 cases. They obtained promising results by combining the hill-climbing feature selection method with several classifiers used in the study. In 2019, Qui et al. suggested a model based on ML to provide personalized estimates of the cumulative chance live birth in the first full IVF cycle using pre-treatment variables such as BMI and AMH [12]. In order to address this, clinical data of 7188 women who underwent their first IVF was used. The best AUC score (0.73) was obtained by the XGBoost model, one of the ML algorithms, and is considered a promising step to provide personalized estimates. In this study, we found that first trimester pregnancy loss was 7 times higher in women with a history of RPL than in women without a history of RPL. In cases with a history of RPL, good-quality embryo transfer and normal endometrium do not reduce pregnancy losses because many factors cause RPL. According to the studies, the most important causes of RPL are congenital or acquired problems of the uterus [30]; genetic factors of the oocyte, sperm, and embryo [31]; thrombophilia; and immunological factors [32]. These are the most important causes of RPL. Since we did not include the causes of congenital and acquired problems of a uterine anomaly in our study, we have excluded these problems among the factors causing first trimester pregnancy loss in our group. Chromosomal abnormalities in the oocyte or sperm are among the most important causes of RPL. According to Morin et al., exome or genome sequencing studies have demonstrated that in 12% of couples with a history of RPL, one partner is a carrier of a chromosomal rearrangement, including inversions and balanced translocations. Unfortunately, traditional karyotypes might not detect 40% of these rearrangements. A quantitative high-density assessment of the parent’s genome is necessary to detect chromosomal rearrangements [33]. Besides, studies have identified many mutations related to RPL. From these mutations, the mutations in the genes of REC114 (required for double-strand breaks in the unsynapsed regions during recombination) and MEI1 (also involved in double-strand break formation) lead to meiotic division errors and excessively aneuploid embryo, which causes abortion [34, 35]. In their recent review, Alecsandru et al. stated that abortion might result from disruption of maternal–fetal immune homeostasis. They stated that since acquired thrombophilia was the leading cause of RPL, it is necessary to focus on antiphospholipid antibody syndrome for screening tests and treatment [32]. For all these reasons, abortion rates are higher in cases with RPL, even if a good-quality embryo transfer is performed. Several studies have been published on the pregnancy outcomes of different endometrial preparations, but the results have been inconsistent. There are studies indicating no difference between the artificial cycle and the natural cycle in terms of pregnancy outcomes [36–40]. However, some studies have shown that first trimester pregnancy losses are higher in artificial cycles than in natural FET cycles [41–44]. Our study found that the first trimester pregnancy loss rate was higher in artificial cycles than in natural FET cycles. Morozov et al. stated in their study that pregnancy rates are higher in the natural cycle, and the reason for this is that the cases in the natural cycle had lower estradiol levels and greater endometrial thicknesses than did the artificial cycles [42]. Some studies showed that increased estrogen concentrations during early pregnancy caused apoptosis in trophoblast cells, preventing the normal development of the placenta and leading to first trimester pregnancy loss [45, 46]. Our study found that the rate of first trimester pregnancy loss was 1.5 times higher in women with PCOS than in women without it. PCOS is the most important endocrinological problem affecting women of reproductive age [47]. Although many studies show that PCOS increases abortion rates, the mechanisms causing abortion are controversial. There are many endocrinological problems underlying the abortion rate in PCOS. Studies have shown that hyperandrogenism and high BMI increase the rate of abortion in PCOS, and insulin resistance adversely affects oocyte development, leading to implantation failure and abortions [48, 49]. In addition, studies showed that LH elevation in the follicular phase might cause premature luteinization and inhibit inhibitors and oocyte maturation, leading to premature oocyte maturation, which may impair oocyte and embryo quality [50–52]. Elevated androgen levels, especially the increase in testosterone in the oocyte, may be associated with decreased calcium oscillations, inhibiting cytoplasmic maturation and preventing meiotic maturation. These causes may increase the rate of abortion in PCOS patients [51, 53]. Although studies show that the rate of aneuploidy increases in the embryo due to impaired endocrine control of meiosis due to high LH, Weghover et al. showed that PCOS does not increase the rate of aneuploidy in the embryo [54]. There are also studies stating that elevated testosterone damages the endometrium through reduced proliferation of endometrial stromal cells and leads to abortion [55–58]. Similar to our study, many studies are showing higher abortion rates in cases with high BMI [59–68]. Although many studies show that obesity increases abortion rates, some studies have argued that the reason for this increase is due to poor oocyte quality and embryo development. In contrast, some studies have argued that endometrial receptivity is completely impaired. In a comprehensive study examining 22,043 FET high-quality embryo transfer cycles, it was shown that abortion rates in the obese group increased due to the impairment of endometrial receptivity [65]. Bellver et al. examined GQ embryo transfer cycles obtained from donor oocytes of young, healthy women and transferred to recipients with different BMI groups. They found low implantation, clinical pregnancy, and live birth rates in the recipient group with high BMI. They stated that the cause of abortion in the group with high BMI was not related to embryo quality but to the alteration that obesity caused in the uterine environment. In a recent study, endometrial biopsies were taken during the implantation period, and endometrial gene expressions were examined by microarray analysis. It was determined that impaired gene expression was detected in obese patients compared to non-obese patients. It was stated that higher abortion rates in obese patients might be related to impaired endometrial gene expression [69]. Rittenberg et al. investigated the effect of BMI on IVF outcomes in their systematic review and meta-analysis of 33 studies and 47,000 cycles. They stated that endocrine and paracrine changes due to obesity, hyperandrogenism, abnormal leptin concentrations, and insulin resistance impair oocyte maturation and embryo competence. They also expressed that increased inflammatory markers such as IL6 and tumor necrosis factor in the follicles and endometrium of obese cases might increase abortion rates by impairing oocyte quality and implantation. Studies showing that obesity also increases the abortion rate in euploid embryo transfer cycles are significant evidence that this situation is not related to oocyte or embryo quality [70, 71]. Our study shows that the high abortion rate in cases with TQ/GQ embryo transfers indicates that obesity increases abortion rates regardless of embryo quality. Studies have demonstrated today that abortion rates increase with the advancement of maternal age [72–76]. The advanced maternal age decreases the quality and number of oocytes, increasing the possibility of chromosomal aneuploidy, which leads to high abortion rates. Studies show that maternal age-related structural and functional changes in mitochondria increase abortion rates [77]. Capalbo et al. estimated in their study that a baseline of 20% of human oocytes is aneuploid, and this increases exponentially from 30 to 35 years, reaching on average 80% by 42 years. As a result, they started that the genetic quality of oocytes rapidly declines from 30 years of age [78]. Women should be informed that their fertility will decrease due to these factors that are related to advanced age: decreased oocyte reserve and quality, the increase in the number of oocytes with chromosomal abnormalities, and age-related structural and functional changes in the mitochondria. Our study showed that abortion rates in cases with more than 3 IVF cycles (recurrent implantation failure) were higher than those with less than 3 IVF cycles. There are many causes of recurrent implantation failure. Genetics of the embryo or partners, hematological diseases, endocrinological disorders, obesity, insulin resistance, endometriosis, adenomyosis, and diseases related to the maternal immune system are among the known causes of recurrent implantation failure [79, 80]. Cases with a history of RIF should be evaluated and treated in detail in terms of these diseases. As a result, top- or good-quality embryo transfer and the absence of endometrial problems are not sufficient to reduce the first trimester pregnancy loss rates. Our study provides the model based on machine learning for predicting pregnancy loss in the first trimester. Using this model, clinicians can provide more accurate recommendations for follow-up and subsequent management of patients with first trimester pregnancy loss. We found that the rate of first trimester pregnancy loss was very high in cases with a history of RPL. Therefore, we recommend that these cases are investigated in detail and that the condition causing RPL should be identified and treated before embryo transfer. In addition, the presence of PCOS, high body mass index, advanced maternal age, and recurrent implantation failure are risk factors for first trimester pregnancy loss. Consequently, physicians should individually assess couples who experienced first trimester pregnancy loss. In addition, physicians and patients should know the risk factors to prevent another pregnancy loss. Declarations Conflict of interest The authors declare no competing interests. Footnotes Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Gonul Ozer, Email: [email protected]. Aysu Akca, Email: [email protected]. Beril Yuksel, Email: [email protected]. Ipek Duzguner, Email: [email protected]. Semra Kahraman, Email: [email protected].

- 1.Niederberger C, et al. Forty years of IVF. Fertility and Sterility. 2018;110(2):185–324 e5. doi: 10.1016/j.fertnstert.2018.06.005. [DOI] [PubMed] [Google Scholar] - 2.Cumming G, et al. The emotional burden of miscarriage for women and their partners: trajectories of anxiety and depression over 13 months BJOG: Int J Obstet Gynaecol. 2007;114(9):1138–1145. doi: 10.1111/j.1471-0528.2007.01452.x. [DOI] [PubMed] [Google Scholar] - 3.Jurkovic D, Overton C. , Bender-Atik R. Diagnosis and management of first trimester miscarriage. BMJ. 2013: 346 [DOI] [PubMed] - 4.Quenby S, et al. Miscarriage matters: the epidemiological, physical, psychological, and economic costs of early pregnancy loss. The Lancet. 2021;397(10285):1658–1667. [DOI] [PubMed] - 5.Turner K, et al. Stress and anxiety scores in first and repeat IVF cycles: a pilot study. Plos One. 2013;8(5):e63743. doi: 10.1371/journal.pone.0063743. [DOI] [PMC free article] [PubMed] [Google Scholar] - 6.Winter E, et al. Early pregnancy loss following assisted reproductive technology treatment. Hum Reprod. 2002;17(12):3220–3223. doi: 10.1093/humrep/17.12.3220. [DOI] [PubMed] [Google Scholar] - 7.Hipp H, et al. First trimester pregnancy loss after fresh and frozen in vitro fertilization cycles. Fertil Steril. 2016;105(3):722–728. doi: 10.1016/j.fertnstert.2015.11.012. [DOI] [PMC free article] [PubMed] [Google Scholar] - 8.Uyar A, Bener A, Ciray HN. Predictive modeling of implantation outcome in an in vitro fertilization setting: an application of machine learning methods. Med Decis Making. 2015;35(6):714–725. doi: 10.1177/0272989X14535984. [DOI] [PubMed] [Google Scholar] - 9.Yi Y, et al. A logistic model to predict early pregnancy loss following in vitro fertilization based on 2601 infertility patients. Reprod Biol Endocrinol. 2016;14(1):1–7. doi: 10.1186/s12958-016-0147-z. [DOI] [PMC free article] [PubMed] [Google Scholar] - 10.Liu L, et al. Machine learning algorithms to predict early pregnancy loss after in vitro fertilization-embryo transfer with fetal heart rate as a strong predictor. Comput Methods Programs Biomed. 2020;196:105624. doi: 10.1016/j.cmpb.2020.105624. [DOI] [PubMed] [Google Scholar] - 11.Hassan MR, et al. A machine learning approach for prediction of pregnancy outcome following IVF treatment. Neural Comput Appl. 2020;32(7):2283–2297. doi: 10.1007/s00521-018-3693-9. [DOI] [Google Scholar] - 12.Qiu J, et al. Personalized prediction of live birth prior to the first in vitro fertilization treatment: a machine learning method. J Transl Med. 2019;17(1):1–8. doi: 10.1186/s12967-019-2062-5. [DOI] [PMC free article] [PubMed] [Google Scholar] - 13.Islam MN, et al. Machine learning to predict pregnancy outcomes: a systematic review, synthesizing framework and future research agenda. BMC Pregnancy Childbirth. 2022;22(1):1–19. doi: 10.1186/s12884-022-04594-2. [DOI] [PMC free article] [PubMed] [Google Scholar] - 14.Practice Committee of the American Society for Reproductive Medicine. Definitions of infertility and recurrent pregnancy loss: a committee opinion. Fertil Steril. 2013; 99(1):63. [DOI] [PubMed] - 15.Mascarenhas M et al. Management of recurrent implantation failure: British Fertility Society policy and practice guideline. Human Fertil 2021:1–25. [DOI] [PubMed] - 16.Géron A. Hands-on machine learning with scikit-learn, Keras, and TensorFlow: concepts, tools, and techniques to build intelligent systems. 2019: " O'Reilly Media, Inc.". - 17.Mitchell, T.M, Mitchell TM, Machine learning. Vol. 1. 1997: McGraw-hill New York. - 18.Tan P-N, Steinbach M, Kumar V. Data mining introduction. Beijing: People’s Posts and Telecommunications Publishing House; 2006. [Google Scholar] - 19.Breiman L. Random forests. Mach Learn. 2001;45(1):5–32. doi: 10.1023/A:1010933404324. [DOI] [Google Scholar] - 20.Grekousis G, et al. Ranking the importance of demographic, socioeconomic, and underlying health factors on US COVID-19 deaths: a geographical random forest approach. Health Place. 2022;74:102744. doi: 10.1016/j.healthplace.2022.102744. [DOI] [PMC free article] [PubMed] [Google Scholar] - 21.Hosmer DW Jr, Lemeshow S, RX. Sturdivant, Applied logistic regression. Vol. 398. 2013: John Wiley & Sons. - 22.Van Rossum G, Drake F. Python/C Api Manual-Python. 2009;3: CreateSpace. - 23.Team RC. R: A language and environment for statistical computing. 2013 - 24.McKinney W. Data structures for statistical computing in python. in Proceedings of the 9th Python in Science Conference. 2010. Austin - 25.Harris CR, et al. Array programming with NumPy. Nature. 2020;585(7825):357–362. doi: 10.1038/s41586-020-2649-2. [DOI] [PMC free article] [PubMed] [Google Scholar] - 26.Hunter JD. Matplotlib: a 2D graphics environment. Compt Sci Eng. 2007;9(03):90–95. doi: 10.1109/MCSE.2007.55. [DOI] [Google Scholar] - 27.Pedregosa F, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825–2830. [Google Scholar] - 28.Team RC, R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. - 29.Hosmer DW, Lemeshow S, Cook E. Applied logistic regression. 2. New York: Jhon Wiley and Sons Inc; 2000. [Google Scholar] - 30.Carbonnel M, et al. Uterine factors in recurrent pregnancy losses. Fertil Steril. 2021;115(3):538–545. doi: 10.1016/j.fertnstert.2020.12.003. [DOI] [PubMed] [Google Scholar] - 31.Klimczak AM, et al. Role of the sperm, oocyte, and embryo in recurrent pregnancy loss. Fertil Steril. 2021;115(3):533–537. doi: 10.1016/j.fertnstert.2020.12.005. [DOI] [PubMed] [Google Scholar] - 32.Alecsandru D, et al. Immunologic causes and thrombophilia in recurrent pregnancy loss. Fertil Steril. 2021;115(3):561–566. doi: 10.1016/j.fertnstert.2021.01.017. [DOI] [PubMed] [Google Scholar] - 33.Morin SJ, et al. Translocations, inversions and other chromosome rearrangements. Fertil Steril. 2017;107(1):19–26. doi: 10.1016/j.fertnstert.2016.10.013. [DOI] [PubMed] [Google Scholar] - 34.Nguyen NMP, et al. Causative mutations and mechanism of androgenetic hydatidiform moles. Am J Human Genet. 2018;103(5):740–751. doi: 10.1016/j.ajhg.2018.10.007. [DOI] [PMC free article] [PubMed] [Google Scholar] - 35.Kuliev A, et al. Chromosomal abnormalities in a series of 6733 human oocytes in preimplantation diagnosis for age-related aneuploidies. Reprod Biomed Online. 2003;6(1):54–59. doi: 10.1016/S1472-6483(10)62055-X. [DOI] [PubMed] [Google Scholar] - 36.Kim YJ, et al. Does a vitrified blastocyst stage embryo transfer program need hormonal priming for endometrial preparation? Journal of Obstetrics and Gynaecology Research. 2010;36(4):783–788. doi: 10.1111/j.1447-0756.2010.01243.x. [DOI] [PubMed] [Google Scholar] - 37.Kawamura T, et al. Clinical outcomes of two different endometrial preparation methods for cryopreserved-thawed embryo transfer in patients with a normal menstrual cycle. Reprod Med Biol. 2007;6(1):53–57. doi: 10.1111/j.1447-0578.2007.00165.x. [DOI] [PMC free article] [PubMed] [Google Scholar] - 38.Alur-Gupta S, et al. Impact of method of endometrial preparation for frozen blastocyst transfer on pregnancy outcome: a retrospective cohort study. Fertil Steril. 2018;110(4):680–686. doi: 10.1016/j.fertnstert.2018.05.013. [DOI] [PMC free article] [PubMed] [Google Scholar] - 39.Groenewoud ER, Cohlen BJ, Macklon NS. Programming the endometrium for deferred transfer of cryopreserved embryos: hormone replacement versus modified natural cycles. Fertil Steril. 2018;109(5):768–774. doi: 10.1016/j.fertnstert.2018.02.135. [DOI] [PubMed] [Google Scholar] - 40.Lathi RB, et al. Frozen blastocyst embryo transfer using a supplemented natural cycle protocol has a similar live birth rate compared to a programmed cycle protocol. J Assist Reprod Genet. 2015;32(7):1057–1062. doi: 10.1007/s10815-015-0499-x. [DOI] [PMC free article] [PubMed] [Google Scholar] - 41.Armas DFC et al. Frozen-thawed blastocyst transfer in natural cycle increase implantation rates compared artificial cycle. Gynecol Endocrinol. 2019. [DOI] [PubMed] - 42.Morozov V, et al. Natural cycle cryo-thaw transfer may improve pregnancy outcome. J Assist Reprod Genet. 2007;24(4):119–123. doi: 10.1007/s10815-006-9100-y. [DOI] [PMC free article] [PubMed] [Google Scholar] - 43.Givens CR, et al. Outcomes of natural cycles versus programmed cycles for 1677 frozen–thawed embryo transfers. Reprod Biomed Online. 2009;19(3):380–384. doi: 10.1016/S1472-6483(10)60172-1. [DOI] [PubMed] [Google Scholar] - 44.Tomás C, et al. Pregnancy loss after frozen-embryo transfer—a comparison of three protocols. Fertil Steril. 2012;98(5):1165–1169. doi: 10.1016/j.fertnstert.2012.07.1058. [DOI] [PubMed] [Google Scholar] - 45.Hancke K, et al. Patients undergoing frozen-thawed embryo transfer have similar live birth rates in spontaneous and artificial cycles. J Assist Reprod Genet. 2012;29(5):403–407. doi: 10.1007/s10815-012-9724-z. [DOI] [PMC free article] [PubMed] [Google Scholar] - 46.Patel S, et al. Estradiol elicits proapoptotic and antiproliferative effects in human trophoblast cells. Biology of reproduction. 2015;93(3):7–1-10. doi: 10.1095/biolreprod.115.129114. [DOI] [PMC free article] [PubMed] [Google Scholar] - 47.Teede HJ, et al. Recommendations from the international evidence-based guideline for the assessment and management of polycystic ovary syndrome. Fertil Steril. 2018;110(3):364–379. doi: 10.1016/j.fertnstert.2018.05.004. [DOI] [PMC free article] [PubMed] [Google Scholar] - 48.Craig LB, Ke RW, Kutteh WH. Increased prevalence of insulin resistance in women with a history of recurrent pregnancy loss. Fertil Steril. 2002;78(3):487–490. doi: 10.1016/S0015-0282(02)03247-8. [DOI] [PubMed] [Google Scholar] - 49.Tian L, et al. Insulin resistance increases the risk of spontaneous abortion after assisted reproduction technology treatment. J Clin Endocrinol Metab. 2007;92(4):1430–1433. doi: 10.1210/jc.2006-1123. [DOI] [PubMed] [Google Scholar] - 50.Tesarik J. Effects of LH on oocyte yield and developmental competence. Hum Reprod. 2003;18(6):1358–1360. doi: 10.1093/humrep/deg267. [DOI] [PubMed] [Google Scholar] - 51.van der Spuy ZM, Dyer SJ. The pathogenesis of infertility and early pregnancy loss in polycystic ovary syndrome. Best Pract Res Clin Obstet Gynaecol. 2004;18(5):755–771. doi: 10.1016/j.bpobgyn.2004.06.001. [DOI] [PubMed] [Google Scholar] - 52.Kurzawa R, et al. Comparison of embryological and clinical outcome in GnRH antagonist vs GnRH agonist protocols for in vitro fertilization in PCOS non-obese patients A prospective randomized study. J Assist Reprod Genet. 2008;25(8):365–374. doi: 10.1007/s10815-008-9249-7. [DOI] [PMC free article] [PubMed] [Google Scholar] - 53.Tesarik J, Mendoza C. Nongenomic effects of 17 beta-estradiol on maturing human oocytes: relationship to oocyte developmental potential. J Clin Endocrinol Metab. 1995;80(4):1438–1443. doi: 10.1210/jcem.80.4.7714121. [DOI] [PubMed] [Google Scholar] - 54.Weghofer A, et al. Lack of association between polycystic ovary syndrome and embryonic aneuploidy. Fertil Steril. 2007;88(4):900–905. doi: 10.1016/j.fertnstert.2006.12.018. [DOI] [PubMed] [Google Scholar] - 55.Freis A, et al. Effects of a hyperandrogenaemic state on the proliferation and decidualization potential in human endometrial stromal cells. Arch Gynecol Obstet. 2017;295(4):1005–1013. doi: 10.1007/s00404-017-4295-y. [DOI] [PubMed] [Google Scholar] - 56.Tuckerman EM, et al. Do androgens have a direct effect on endometrial function? An in vitro study. Fertil Steril. 2000;74(4):771–779. doi: 10.1016/S0015-0282(00)00711-1. [DOI] [PubMed] [Google Scholar] - 57.Liu L, et al. A comparison of the miscarriage rate between women with and without polycystic ovarian syndrome undergoing IVF treatment. Eur J Obstet Gynecol Reprod Biol. 2014;176:178–182. doi: 10.1016/j.ejogrb.2014.02.041. [DOI] [PubMed] [Google Scholar] - 58.Shang K, et al. Endometrial abnormality in women with polycystic ovary syndrome. Reprod Sci. 2012;19(7):674–683. doi: 10.1177/1933719111430993. [DOI] [PubMed] [Google Scholar] - 59.Rittenberg V, et al. Effect of body mass index on IVF treatment outcome: an updated systematic review and meta-analysis. Reprod Biomed Online. 2011;23(4):421–439. doi: 10.1016/j.rbmo.2011.06.018. [DOI] [PubMed] [Google Scholar] - 60.Sermondade N, et al. Female obesity is negatively associated with live birth rate following IVF: a systematic review and meta-analysis. Hum Reprod Update. 2019;25(4):439–451. doi: 10.1093/humupd/dmz011. [DOI] [PubMed] [Google Scholar] - 61.Bellver J, et al. Obesity and poor reproductive outcome: the potential role of the endometrium. Fertil Steril. 2007;88(2):446–451. doi: 10.1016/j.fertnstert.2006.11.162. [DOI] [PubMed] [Google Scholar] - 62.Gosman GG, Katcher HI, Legro RS. Obesity and the role of gut and adipose hormones in female reproduction. Hum Reprod Update. 2006;12(5):585–601. doi: 10.1093/humupd/dml024. [DOI] [PubMed] [Google Scholar] - 63.Wang JX, Davies MJ, Norman RJ. Obesity increases the risk of spontaneous abortion during infertility treatment. Obes Res. 2002;10(6):551–554. doi: 10.1038/oby.2002.74. [DOI] [PubMed] [Google Scholar] - 64.Yang X, Zheng B, Wang Y. Effect of pre-pregnancy body mass index on neonatal outcomes in women undergoing autologous frozen-thawed embryo transfer. Fertil Steril. 2021;4(116):1010–19. [DOI] [PubMed] - 65.Zhang J, et al. Effect of body mass index on pregnancy outcomes in a freeze-all policy: an analysis of 22,043 first autologous frozen-thawed embryo transfer cycles in China. BMC Med. 2019;17(1):1–9. doi: 10.1186/s12916-019-1354-1. [DOI] [PMC free article] [PubMed] [Google Scholar] - 66.Bellver J, et al. Female obesity impairs in vitro fertilization outcome without affecting embryo quality. Fertil Steril. 2010;93(2):447–454. doi: 10.1016/j.fertnstert.2008.12.032. [DOI] [PubMed] [Google Scholar] - 67.Luke B, et al. Female obesity adversely affects assisted reproductive technology (ART) pregnancy and live birth rates†. Hum Reprod. 2010;26(1):245–252. doi: 10.1093/humrep/deq306. [DOI] [PubMed] [Google Scholar] - 68.Bellver J et al. Obesity reduces uterine receptivity: clinical experience from 9,587 first cycles of ovum donation with normal weight donors. Fertil Steril. 2013;100(4):1050–1058. e2. [DOI] [PubMed] - 69.Comstock IA, et al. Does an increased body mass index affect endometrial gene expression patterns in infertile patients? A functional genomics analysis. Fertil Steril. 2017;107(3):740–748.e2. doi: 10.1016/j.fertnstert.2016.11.009. [DOI] [PubMed] [Google Scholar] - 70.Boots CE, Bernardi LA, Stephenson MD. Frequency of euploid miscarriage is increased in obese women with recurrent early pregnancy loss. Fertil Steril. 2014;102(2):455–459. doi: 10.1016/j.fertnstert.2014.05.005. [DOI] [PubMed] [Google Scholar] - 71.Tremellen K, Pearce K, Zander-Fox D. Increased miscarriage of euploid pregnancies in obese women undergoing cryopreserved embryo transfer. Reprod Biomed Online. 2017;34(1):90–97. doi: 10.1016/j.rbmo.2016.09.011. [DOI] [PubMed] [Google Scholar] - 72.Schoolcraft WB, et al. Preimplantation aneuploidy testing for infertile patients of advanced maternal age: a randomized prospective trial. Fertil Steril. 2009;92(1):157–162. doi: 10.1016/j.fertnstert.2008.05.029. [DOI] [PubMed] [Google Scholar] - 73.Harton GL, et al. Diminished effect of maternal age on implantation after preimplantation genetic diagnosis with array comparative genomic hybridization. Fertil Steril. 2013;100(6):1695–1703. doi: 10.1016/j.fertnstert.2013.07.2002. [DOI] [PubMed] [Google Scholar] - 74.Seshadri S, et al. Assisted conception in women of advanced maternal age. Best Pract Res Clin Obstet Gynaecol. 2021;70:10–20. doi: 10.1016/j.bpobgyn.2020.06.012. [DOI] [PubMed] [Google Scholar] - 75.Franasiak JM et al. The nature of aneuploidy with increasing age of the female partner: a review of 15,169 consecutive trophectoderm biopsies evaluated with comprehensive chromosomal screening. Fertil steril. 2014;101(3):656–663. e1. [DOI] [PubMed] - 76.Ubaldi FM, et al. Advanced maternal age in IVF: still a challenge? The present and the future of its treatment. Front Endocrinol. 2019;10:94. doi: 10.3389/fendo.2019.00094. [DOI] [PMC free article] [PubMed] [Google Scholar] - 77.Van Blerkom J. Mitochondrial function in the human oocyte and embryo and their role in developmental competence. Mitochondrion. 2011;11(5):797–813. doi: 10.1016/j.mito.2010.09.012. [DOI] [PubMed] [Google Scholar] - 78.Capalbo A, et al. Human female meiosis revised: new insights into the mechanisms of chromosome segregation and aneuploidies from advanced genomics and time-lapse imaging. Hum Reprod Update. 2017;23(6):706–722. doi: 10.1093/humupd/dmx026. [DOI] [PubMed] [Google Scholar] - 79.Hill MJ. Recurrent implantation failure: Sapere aude. Fertil Steril. 2021;116(6):1430–1431. doi: 10.1016/j.fertnstert.2021.09.030. [DOI] [PubMed] [Google Scholar] - 80.Franasiak JM, et al. A review of the pathophysiology of recurrent implantation failure. Fertil Steril. 2021;116(6):1436–1448. doi: 10.1016/j.fertnstert.2021.09.014. [DOI] [PubMed] [Google Scholar]