Recurrent pregnancy loss (RPL) is one of the most common pregnancy complications in obstetrics and gynecology, and the incidence rate of RPL is about 2% (1–3). Patients with RPL usually have experienced multiple miscarriages before diagnosis, which is not only a serious threat to the physical and mental health of patients, but also a heavy economic burden to patients and their families (1). Due to the complex and diverse etiology of RPL, as well as the lack of specific clinical manifestations in patients before pregnancy loss occurs, risk prediction and early intervention of RPL have become important issues that urgently need to be addressed in the field of reproductive health (1, 4).

Epidemiological investigation showed the number of expected abortions was the main risk factors for pregnancy loss (2, 5). However, a systematic meta-analysis showed that there were no differences in the incidence of uterine abnormalities, chromosomal abnormalities, thrombotic diseases, and thyroid diseases between patients with 2 and ≥ 3 failed pregnancies, indicating statistical differences between RPL patients and the normal pregnancy women (6). These differences may already exist before the first pregnancy loss. If these differences can be detected earlier, it is of great significance for identifying high-risk populations for RPL and improving prognosis (3, 5).

The causes of RPL mainly include chromosomal or genetic abnormalities, anatomical abnormalities (including congenital and acquired), autoimmune diseases, pre-thrombotic states, endocrine factors, infectious factors, male factors, and environmental psychological factors (2, 6). It should be pointed out that the specific causes and pathogenesis of about 40% of RPL are still unknown (7). In clinical diagnosis and treatment, the screening indicators of RPL involve multiple personalized examination items such as reproductive tract anatomy, chromosome karyotype, autoimmune antibodies and thromboela-stogram, which are often used for etiological analysis after the diagnosis of RPL, but not suitable for evaluating the risk of RPL before the miscarriage occurs (7). It is very difficult to conduct effective early intervention before the diagnosis of RPL.

Thyroid function indicators, 25-(OH) D, ferritin, homocysteine, vitamin B12 and folic acid are relatively routine inspections for women in the first trimester of pregnancy (7–11). Previous studies have shown that some of these factors may be related to RPL, but the clinical reliability and accuracy of predicting the risk of RPL based on a single indicator are difficult to guarantee (7, 10). In recent years, artificial intelligence and machine learning have been successfully applied in disease diagnosis, disease development prediction, disease risk factor identification, and new drug development in the medical field (11–13). Logistic regression is a classic machine learning method, which can statistically analyze the impact of multiple variables on diseases and is more helpful for clinical practical applications (13–15).

The main purpose of this study is to construct a risk prediction model for RPL based on routine inspections in the first trimester of pregnancy, in order to provide clinical physicians with indications of whether the patients are at high risk before the miscarriage occurs. This study has important value for early intervention and improving prognosis in patients with RPL. To our knowledge, this study is the first time to assess the risks of RPL based on routine inspections in the first trimester of pregnancy using logistic regression.

This was a retrospective study between January 2019 and December 2023. Characteristics of 805 participants from Henan Provincial People’s Hospital was used for model training and internal testing, and characteristics of 229 participants from Shangqiu Maternal and Child Health Hospital was used for external validation. In this study, we investigated the relationship between routine inspections in the first trimester of pregnancy and RPL, and tried to predict the risk of RPL before the first miscarriage. According to the results of chi-square test, the logistic regression model was established using the training set with TPO-Ab, TG-Ab, 25-(OH) D, Hcy and FA. The confusion matrix, ROC curve, and calibration curve of internal testing set and external validation set showed that the model proposed had acceptable accuracy and universality in the risk prediction of RPL. The DCA decision curves indicated that the proposed model had clinical significance. Considering that the inspection items were collected before the first pregnancy loss and some RPL causes such as chromosomal abnormalities may not be clearly reflect in the detection results, we believe that the accuracy of the proposed model is acceptable and has clinical guidance significance. There is no significant deterioration in the results of external verification, whicDh confirmed the generalizability of the study results.

Endocrine function is crucial to the establishment and maintenance of pregnancy (5, 11, 16, 17). Thyroid abnormalities associated with RPL include hypothyroidism, hyperthyroidism, and thyroid autoantibody abnormalities (5, 18). TGAb and TPOAb are both thyroid specific antibodies. Clinical studies have found that abnormally high level of TPOAb can lead to the occurrence of autoimmune diseases, which in turn produce autoimmune effects on fetal cells and cause miscarriage (16, 18). Recent studies have shown that women with positive TPOAb are at higher risk of recurrent miscarriage (16, 18). Abnormal elevation of TGAb level indicates abnormal thyroid follicular structure in patients, leading to increased responsiveness to the placenta and affecting the quality of the placenta and embryo, thereby increasing the risk of miscarriage (5). In this study the TG-Ab level and TPOAb level in the RPL group were both higher than those in the control group and the difference was statistically significant, consistent with relevant research conclusions.

Vitamin D is a kind of vitamin synthesized by ultraviolet light or a small amount of food, 25 hydroxyvitamin D is the main circulating form of vitamin D in the blood, good stability, is recognized as a reliable indicator to evaluate the nutritional status of human vitamin D (10, 19–22). Its role in immune function, healthy cell division, and bone health has been supported by extensive research, and many studies have found that low serum vitamin D levels are not only associated with certain types of cancer, autoimmune diseases, but also with the occurrence of spontaneous abortion (23). Some studies have found that vitamin D may be involved in the genetic polymorphism of CYP2R1, leading to the occurrence of recurrent abortion. The AG and GG genotypes of rsl2794714 of CYP2R1 increase the risk of RPL, while the AG genotype of rsl2794714 of CYP2R1 is closely related to vitamin D levels (24). In a cross-sectional study, plasma levels of 25 hydroxyvitamin D and 25 hydroxyvitamin D-1α hydroxylase were measured in early pregnancy (7 to 9 weeks) of 120 women of childbearing age who were not pregnant, active pregnancy, and spontaneous abortion (8). The results suggested that decreased serum vitamin D levels may increase the risk of spontaneous abortion. Studies have shown that in mouse models that do not express vitamin D receptors, platelet aggregation is increased, antithrombin gene and thrombomodulin are decreased, and tissue factors are increased, while supplementation with 25 (OH) D analogs can regulate the expression of thrombomodulin and tissue factors in monocytes to improve coagulation status which can reduce the risk of RPL (25). Some studies have found that the expression of FOXP3 gene in peripheral blood of patients with recurrent abortion is significantly decreased, and the expression of FOXP3 is increased after supplementing vitamin D (21, 26). Studies have shown that vitamin D is involved in follicle development and steroid hormone production, promotes follicle maturation and ovulation by inhibiting the expression of AMH receptor and follicle-stimulating hormone receptor, affects embryo quality, promotes granule cell luteinization by inducing the reaction of key steroid enzymes, such as 3β hydroxysteroid dehydrogenase, and maintains progesterone levels to prevent miscarriage (10, 20). In summary, the etiological mechanism of vitamin D and recurrent abortion is directly or indirectly related to genetics, pre-thrombotic state, immunity and other aspects, which is also consistent with the conclusion of this study.

Serum homocysteine is a sulfur-containing amino acid present in plasma and an intermediate product of the metabolism of methionine and cysteine (27, 28). As a metabolite to maintain the normal operation of the human body, the serum homocysteine concentration tends to decline during normal pregnancy (27). In early pregnancy, high levels of homocysteine can significantly inhibit the formation of villous vessels, thus reducing the amount of blood supply to the embryo and leading to embryo death (27). Folic acid is an essential substance in Hcy metabolic pathway. Folic acid deficiency is one of the important reasons leading to abnormal Hcy metabolism and hyperhomocysteinemia. Relevant studies have shown that the serum Hcy level of non-pregnant RPL patients are higher than that of healthy women with normal reproductive history, and the folic acid level is lower than that of healthy women with normal reproductive history (11). A retrospective cohort study in Japan found that folic acid supplementation and reduction of homocysteine levels in women with assisted reproductive technology were effective in improving pregnancy outcomes (9). In this study, the FA level of the RPL group was lower than that of the control group and the difference was statistically significant, consistent with the conclusions of relevant studies.

Machine learning methods have been widely used in the fields of obstetrics and gynecology and epidemiology, including exploring biomarkers for disease diagnosis, analyzing the correlation between test indicators and diseases, predicting the prognosis of patients, and so on (12–15, 29). A study tried to investigate the biomarkers and molecular mechanisms associated with RPL using ANN model and identified three hub genes, WBP11, ACTR2, and NCSTN (15). This study contributed to understanding the molecular mechanisms and treatment strategies of RPL, but was not helpful for predicting RPL before it occurs. A study used XGBoost algorithm to establish a predictive model for natural miscarriage after in vitro fertilization and embryo transfer, which featured female age, ovarian structural abnormalities, lactation, anti-Mullerian hormone, activated partial thromboplastin time, anticardiolipin antibodies, and thyroid peroxidase antibodies. The features included in this model are not routine test items, and its clinical application value is limited for women of childbearing age who naturally conceive (15). A study used clinical information, vitamin D and thyroid function measurements to establish a framework for conducting an effective analysis for RPL (29). Although the model adopted innovative machine learning algorithms, the clinical data involved in this study was limited and there was no clear indicator collection time, so the predictive value of the constructed model for RPL is limited.

Current research either delves into the molecular mechanisms underlying the pathogenesis of recurrent pregnancy loss (RPL), such as signaling pathways associated with lncRNA PART1, circRNA, and CYR61, or establishes models to analyze the prognosis of RPL patients (30–33). Unlike other studies, this study aims to provide early intervention for potential patients before RPL occurs. The primary contribution of this paper lies in the establishment of a multi-parameter combined prediction model, which can predict the risk of RPL using only routine first-trimester pregnancy examinations, making it convenient for clinical application and implementation. This model allows for the screening of high-risk groups for RPL prior to the occurrence of the first miscarriage, facilitating early clinical intervention. By applying this model, clinical physicians can assess the risk of RPL in women undergoing pre-pregnancy examinations, providing a reference for further personalized examinations and early interventions.

Due to factors such as the data collection cycle and sample size, this study still has some limitations in data analysis and model construction. Retrospective studies inherently carry the risk of potential inaccuracies in data collection due to the reliance on existing records. To mitigate this concern, we took several measures to ensure the quality and accuracy of the information utilized in our analysis. These included rigorous data validation procedures, cross-verification with multiple data sources when available, and the exclusion of any cases with incomplete or ambiguous data. Despite these efforts, we recognize that the retrospective design may still introduce some degree of uncertainty. In future studies, we plan to address this limitation by conducting prospective research, which would allow for more controlled data collection and potentially yield more definitive results.