Prediction model for missed abortion of patients treated with IVF-ET based on XGBoost: a retrospective study

Background

In missed abortion, the embryo stops developing for various reasons, but the dead embryo remains in the uterine cavity (Segawa et al., 2017; Wu et al., 2017). Missed abortion is diagnosed by ultrasound and has an incidence of 8–20% in clinically confirmed intrauterine pregnancy (Fang et al., 2018). Various factors affect embryonic development, including the age of women, genetic factors, endocrine diseases, immune factors, pregnancy infection, pregnancy history, behavioral factors, etc. Due to the complexity of the etiology, the pathogenesis of missed abortion is not known (Fang et al., 2018; Guo et al., 2016; Zhao et al., 2017; Zhao et al., 2019; Zhou et al., 2019).

Missed abortion seriously affects the health of the affected women. If the dead embryos are not found on time and remain in the uterine cavity for long, they can cause abnormal maternal blood coagulation, uterine adhesion, disseminated or diffuse intravascular coagulation (DIC), and even threaten the life of the individual (Fang et al., 2018). For women who require assisted reproductive technology, missed abortion causes greater physical and mental harm. Additionally, they and their family have to bear economic and psychological pressure (Hu et al., 2018). IVF-ET is an effective reproductive technology for the clinical treatment of infertility. However, the incidence of missed abortion among individuals treated with IVF-ET can be as high as 18–30% (Qiao & Li, 2013). For couples wanting children, a diagnosis of missed abortion affects individuals emotionally. To reduce such events, high-risk factors need to be identified, and targeted preventive measures need to be provided. Although several studies have screened the high-risk factors of missed abortion (Feng et al., 2021; Jiang, Yang & Luo, 2022), studies targeting infertile people are limited. The method to predict high-risk groups is not well-defined, and therefore, a method for predicting missed abortion for patients treated with IVF-ET needs to be developed. Such a theoretical basis for targeted personalized diagnosis and treatment might enhance the success rate of pregnancy.

Predictive models are widely used in the medical field, and the common modeling methods include traditional logical regression (Lian et al., 2022; Xu et al., 2022) and machine learning (Krenz et al., 2022; Yan et al., 2021). A traditional prediction model is a well-known modeling and analysis method in the field of ART, but its application is limited. First, the accuracy of the model is low. A study constructed a traditional logistic regression model and machine learning model to predict the potential of embryo implantation, and the AUC was 0.66 and 0.74, respectively (Blank et al., 2019). Barnett-Itzhaki evaluated whether the machine learning method is better than the traditional statistical modeling in predicting the outcome of IVF and found that the accuracy is 0.69∼0.90 and 0.34∼0.74, respectively (Barnett-Itzhaki et al., 2020). Second, the method to collect and store medical data has improved considerably, and the data can be integrated and shared in large-capacity information systems. Unlike machine learning, traditional logical regression cannot deal very well with a large number of multi-class features or variables (Altman, 2017). Additionally, the traditional model relies on predetermined equations and does not have the ability of autonomous learning, and thus, it is impossible to use traditional models for building an automated clinical decision-making system to help doctors make decisions (Krittanawong et al., 2017).

Machine learning is the core of artificial intelligence, and its data processing, induction, and synthesis abilities are better than those of other statistical methods. A lifting algorithm based on the regression tree called extreme gradient boosting (XGBoost) is a machine learning hot spot. Because of its short training time and high precision, it is widely used in the medical field (Chen & Guestrin, 2016). The XGBoost algorithm can make up for the limitations of traditional logical regression. It can simulate nonlinear effects and has high efficiency and accuracy. The XGBoost algorithm can perform parallel operations and run large-scale data quickly; it can automatically optimize split nodes and can effectively deal with irregular data with many outliers and missing values. It can also learn independently, and the model constructed is interpretable and flexible (Chen & Guestrin, 2016; Gao et al., 2018). In this study, we integrated many clinical characteristics of IVF-ET patients to construct an XGBoost prediction model for predicting the risk of missed abortion. We then compared this model with the traditional logical regression prediction model to evaluate the prediction performance of the model.

Methods

Results

Construction and evaluation of the prediction model

We constructed the prediction model based on the influencing factors presented in Table 3. The hyperparameters used in the XGBoost model included the following: objective-binary: logistic, learning rate = 0.01, max depth = 4, min child weight = 2, and reg lambda = 1. The hyperparameters used in the logical regression model included the following: Regularization factor (C) = 1, max iter = 100, penalty = l2, and tol = 0.0001. The AUC score and the FI score were used to evaluate the performance of the XGBoost model and the logistic regression model (Table 4). The AUC score of the XGBoost model (0.877 ±0.014) was significantly higher than that of the logistic model (0.713 ±0.013). The ROC curves of the two prediction models are shown in Fig. 1A. The F1 score of the XGBoost model (0.730 ±0.019) was also significantly higher than that of the logistic model (0.568 ±0.026). The XGBoost model performed better than the logistic model. The predictability of the two models was evaluated by the resampling method, and the XGBoost model performed better (Table 5 and Fig. 1B). Therefore, the ranking of importance was also based on the XGBoost model (Fig. 2).

Table 4:

The results of the training set.

	Model	AUC (95% CI)	Accuracy	Sensitivity	Specificity	F1 Score
Mean	logistic	0.713 (0.673∼0.753)	0.668	0.657	0.676	0.568
SD	logistic	0.013 (0.014∼0.013)	0.021	0.093	0.073	0.026
Mean	XGBoost	0.877 (0.851∼0.904)	0.802	0.795	0.807	0.730
SD	XGBoost	0.014 (0.015∼0.012)	0.021	0.038	0.045	0.019

DOI: 10.7717/peerj.14762/table-4

Notes:

AUC

area under the curve

CI

confidence interval

F1 Score: harmonic average of precision and recall.

Table 5:

The results of the test set.

	Model	AUC (95% CI)	Accuracy	Sensitivity	Specificity	F1 Score
Mean	logistic	0.695 (0.633∼0.758)	0.635	0.659	0.657	0.550
SD	logistic	0.030 (0.032∼0.028)	0.042	0.074	0.065	0.049
Mean	XGBoost	0.759 (0.698∼0.820)	0.705	0.587	0.861	0.566
SD	XGBoost	0.023 (0.026∼0.021)	0.021	0.071	0.045	0.042

DOI: 10.7717/peerj.14762/table-5

Notes:

AUC

area under the curve

CI

confidence interval

F1 Score: harmonic average of precision and recall.

Discussion

Assisted reproductive technology has been used in the clinic for many years. IVF-ET technology is also the last hope of many infertile couples. Although the clinical pregnancy rate of patients treated with IVF-ET can increase by up to 46.9%, the live birth rate is very low, around 38.1% only (Sunderam et al., 2017). Missed abortion decreases the success rate of pregnancy and causes deep emotional disturbance among patients (Zhang et al., 2021). Identifying the clinical influencing factors of missed abortion after IVF-ET can facilitate etiological treatment. In this study, we determined eight independent influencing factors.

We found an abnormal coagulation function, such as the shortening of APTT and PT, in women with missed abortions, suggesting the hypercoagulable state of the blood. This might be because the hypercoagulable state of the blood can selectively affect blood circulation in the uterus and placenta, form microthrombus in the placenta, cause a local placental infarction, decrease placental blood supply, and cause ischemia and hypoxia in the embryo and fetus, thus arresting the growth of the embryo (Dong & Du, 2013). Therefore, if blood hypercoagulability is found before or during early pregnancy, appropriate exercise can be recommended to promote blood circulation along with drug treatment, such as heparin.

The hypothalamic-pituitary-ovarian axis regulates the complex endocrine system. Abnormalities in any part of the axis can lead to adverse pregnancy outcomes. In this study, we found that the level of PRL also increased significantly in the observation group, and hyperprolactinemia was common in pituitary dysfunction or space-occupying lesions in the pituitary. An increase in prolactin levels can inhibit the synthesis and release of gonadotropin, adversely affect the development of follicles and embryos, and cause ovulation disorders, resulting in infertility or missed abortion (Grigg et al., 2017). Additionally, the level of AMH in such cases decreases in the placenta and is often accompanied by structural lesions in the ovary. AMH is negatively correlated with the risk of early spontaneous abortion (Tarasconi et al., 2017), and the structural changes in the ovary can affect its function. Progesterone is required for the successful implantation of fertilized eggs and pregnancy, and proper functioning of the ovary is essential for normal pregnancy (Patel et al., 2017). Hence, when the PRL level increases due to pituitary adenomas, or the ovarian function decreases due to substantial ovarian lesions, patients need surgical resection of the lesions and postoperative adjuvant therapy with drugs.

The age of the individual strongly affects the success of pregnancy. As the age of a woman increases, the quality of oocytes decreases, leading to errors during oocyte meiosis, the formation of aneuploidy, chromosome translocation, inversion, etc. Embryos formed by such gametes are at a greater risk of spontaneously stopping development (Hansen, 1986; Qiao & Yang, 2017). The results of our study not only confirmed the above-mentioned findings but also showed that male age is an important influencing factor. Some studies have shown that the male sperm quality decreases with an increase in male age, and the probability of gene mutation increases (Brahem et al., 2011; Cooper et al., 2010). In our study, sperm abnormality did not affect the results. This might be because the standard of abnormal sperm was defined based on a clinical diagnosis, such as oligospermia, asthenospermia, and sperm deformity, rather than on the quality of sperm used for in vitro fertilization. Although China has implemented the three-child policy, the reproductive desire of young women has decreased, and the proportion of conception among relatively older women in China has increased. Therefore, improving the early pregnancy monitoring of elderly patients receiving ART is important.

We also found that the level of FT4 in women with embryo termination was lower than that in normal people, and the positive rate of TPO-Ab was significantly higher, suggesting that thyroid dysfunction might affect embryos adversely. Hypothyroidism might cause adverse effects, such as embryo termination and fetal malformations (Alexander et al., 2017). TPO-Ab and TG-Ab are specific indicators of thyroid autoimmunity. Abnormal levels of TPO-Ab and TG-Ab can cause autoimmune hypothyroidism. Some studies (Grossmann et al., 2013; Ji, Wang & Song, 2019) have shown that positive TPO-Ab in women during early pregnancy increases the risk of abortion. Therefore, pre-pregnancy thyroid function should be examined, and appropriate treatment should be administered to women diagnosed with thyroid diseases and abnormal laboratory indices; if necessary, multi-disciplinary treatment (MDT) should be provided.

The last influencing factor we discuss in this study is the immune factor. ACA and ANA are human autoimmune antibodies. We found that the level of these two antibodies was higher in women with missed abortions. ACA might act on the membrane phospholipids of placental vascular endothelial cells and platelets in the early stage of pregnancy, block prostacyclin synthesis, and lead to placental embolism. ANA might influence DNA replication and immune abnormalities (Yu et al., 2020). Only a positive laboratory test for ACA or ANA may not cause missed abortion. Usually, when patients have symptoms, they should be administered timely treatment. Some researchers have found that for ACA-positive patients, prednisone and low-dose aspirin therapy can improve pregnancy outcomes (Zhu et al., 2013).

Based on the above factors, we constructed an XGBoost-based missed abortion risk prediction model for patients treated with IVF-ET. The results showed that the prediction performance of the model was better than the prediction performance of the traditional logical regression model. Although many researchers have studied the risk factors for missed abortion, Yi et al. (2016) analyzed the relationship between ultrasound factors and EPL based on the data collected from initial transvaginal ultrasonography scans of 2,601 females with viable singleton pregnancies who underwent IVF-ET. Using these data, they designed a logistic model to predict EPL. Kapfhamer et al. (2018) predicted the loss of pregnancy by measuring the average sac diameter and head-hip length after IVF. Puget et al. (2018) predicted early pregnancy outcomes based on serial human chorionic gonadotropin (hCG) and progesterone levels. The input variables of the prediction models mentioned above were few and focused on one aspect. However, our study integrated 48 possible high-risk factors and analyzed eight independent risk factors through binary logistic regression. We established a comprehensive prediction model and considered the influence of multiple factors to improve the prediction performance of the model. Additionally, the influencing factors included by some researchers were not commonly detected. Yang et al. (2021) detected serum fibroblast growth factor 21 (FGF21) and fatty acid binding protein 4 (FABP4) levels of missed abortion by performing an enzyme-linked immunosorbent assay. Gao et al. (2021) collected fecal specimens from patients and extracted bacterial DNA, and performed bacterial chip assays. They found that a high abundance of Actinobacteria was one of the high-risk factors for missed abortion (Gao et al., 2021). The variables included in our study are routine testing items in the process of IVF-ET treatment. They are easy to extract from the clinical database, either for the next step of external verification or multicenter applications. Most applications of XGBoost prediction models are focused on the automation technology field and are uncommon in the medical field. Qiu et al. (2019) selected age, AMH, BMI, duration of infertility, previous live birth, previous miscarriage, previous abortion, and the type of infertility as predictors to develop four machine learning models to predict live birth. They found that XGBoost provided the most accurate prediction (AUC of the training dataset = 0.74 ± 0.02, AUC of the test dataset = 0.73) on the cumulative live birth chance for IVF cycles (Qiu et al., 2019). We also used XGBoost to construct a prediction model for missed abortions. Besides age and AMH, the coagulation level, thyroid dysfunction, and immune factors caused an adverse pregnancy outcome. Our XGBoost model had high accuracy (AUC of the training dataset = 0.877 ±0.014, AUC of the test dataset = 0.713 ±0.013) and might be used as a theoretical basis for the prevention of missed abortion.

Along with our prediction model, when predicting the risk of missed abortion, clinicians should identify high-risk factors in advance, perform personalized prevention and treatment of these high-risk factors before IVF-ET treatment and during early pregnancy, and monitor patients after confirming clinical pregnancy to avoid a missed diagnosis, which might cause embryos to remain in the uterus for a long time.

With the development of big data, similar prediction models based on machine learning might be constructed by more researchers. These studies might effectively deal with massive clinical databases and use an unbiased method to identify new information variables that are not easy for clinicians to find, to increase the convenience for personalized diagnosis and treatment platforms. However, its application still has some limitations. First, machine learning lacks human emotion and cannot completely replace the human brain in making decisions. Clinicians need to use it as a theoretical basis to make decisions according to the actual situation of the patients. Second, there is a problem of information leakage. Constructing a model requires patient information. Our health system needs to be improved and fortified to ensure the privacy and security of patients.

Limitations

One limitation of this study was that the inclusion index did not reach an ideal state. Embryonic chromosomes, bad living habits, a history of close contact with pets, and environmental factors were not included because this was a retrospective study. Some of the clinical data used had certain limitations, and standardizing the subjective indicators was challenging. A more accurate forecasting model needs to be constructed with all the indicators. Another limitation was that the study population included individuals from a small area and lacked geographical and ethnic variation. Future studies should consider including data from multiple medical centers for the external verification of the model.

Conclusions

To summarize, we found that the independent influencing factors of missed abortion included female age, male age, abnormal ovarian structures, PRL, AMH, APTT, ACA, and TPO-Ab. We constructed a prediction model based on the XGBoost algorithm, which could be used to accurately predict the risk of missed abortion in patients with IVF-ET. This model performed better than the traditional logical regression model. The findings of this study might provide a theoretical basis for preventing abortion among women who are at a high risk of missed abortion.

Supplemental Information