A method for predicting spontaneous abortion by integrating learning combining pre-pregnancy serum metabolome

Abstract

This invention relates to a method for predicting spontaneous abortion using ensemble learning combined with pre-pregnancy serum metabolomics. The method includes the following steps: S1: Collection of peripheral blood samples from women planning pregnancy, input of clinical information variables, and follow-up of pregnancy outcomes; S2: Construction of a pre-pregnancy serum metabolomics dataset; S3: Artificial intelligence modeling; S4: Incorporation of variable values ​​from the dataset in S2 into the artificial intelligence modeling in S3; S5: Model feedback validation. This invention combines patient peripheral blood serum metabolomics, clinical information variables, and follow-up patient pregnancy outcomes, using machine learning methods such as XGBoost, RF, GBDT, and DT to construct a pregnancy outcome prediction model. The optimized model achieves a prediction accuracy of 87%, demonstrating significant predictive value. The established artificial intelligence model utilizes ensemble feature engineering to evaluate the contribution of feature values, further revealing that the levels of several amino acids are variables with significant influence after the number of miscarriages.

Description

Technical Field

[0001] This invention relates to the fields of reproductive medicine, metabolomics, and artificial intelligence. Specifically, it is a method for predicting spontaneous abortion by integrating learning with pre-pregnancy serum metabolomics. Background Technology

[0002] Maternal and infant health has a profound and wide-ranging social impact on a nation and country, while pregnancy-related diseases are increasingly harming women of childbearing age, offspring, families, and society. About 15% of couples of childbearing age worldwide suffer from spontaneous abortion [1]. Epidemiological studies have shown that for each additional miscarriage, the risk of recurrent miscarriage after subsequent pregnancies increases by about 10%, and the risk of subsequent miscarriage increases to 42% in women with a history of three or more miscarriages [2]. Spontaneous abortion not only causes serious physical and mental harm to women of childbearing age, but also imposes a huge economic burden on families and even society. At present, the causes of spontaneous abortion are complex and diverse, and most of them are unknown, which has greatly limited and challenged clinical diagnosis and treatment. Therefore, whether the risk of spontaneous abortion can be identified as early as possible and timely intervention can be carried out has become a clinical focus of concern for scholars at home and abroad.

[0003] With the maturity of 5G technology, big data, the Internet of Things, cloud computing and other technologies are also developing rapidly. Society is gradually moving from the information age to the intelligent age, and all industries are vying to promote the pace of intelligentization [3]. In the field of medical and health care, the deep application of artificial intelligence is a hot topic in current clinical medical research. It can quickly and accurately process big data and mine the potential information behind it, and can play an important role in lesion identification, auxiliary diagnosis and efficacy evaluation [4]. In recent years, artificial intelligence has also been increasingly used in the field of obstetrics and gynecology.

[0004] Current research reports that machine learning can be used to predict the risk of postpartum hemorrhage in women after delivery, and it has good discriminative ability [5]. In addition, predictions of preterm birth, preeclampsia, emergency cesarean section, early gestational diabetes, fetal growth, and the probability of vaginal delivery in subsequent pregnancies after cesarean section have all been reported to be accurately predicted by building machine learning models based on clinical data variables [6-10].

[0005] In recent years, artificial intelligence modeling combined with omics data for disease prediction has become a hot topic in clinical medical research. Several pioneering works published in CNS journals have highlighted the good predictive value of post-conception blood biomarkers (RNA or metabolites) for gestational age, preterm birth, and preeclampsia [11-13]. However, for women planning to become pregnant, post-conception warnings do not provide sufficiently early risk assessment and are more likely to cause anxiety and panic. Therefore, pre-conception data is more important for predicting pregnancy outcomes and can provide important risk warnings before pregnancy. Unfortunately, to date, no existing research has confirmed that the risk of miscarriage can be predicted based on pre-conception blood data.

[0006] Therefore, our team has developed a pregnancy outcome prediction model and software for women preparing for pregnancy, especially those with a history of recurrent miscarriages, to provide pre-pregnancy outcome prediction and assessment. By analyzing the pre-pregnancy blood metabolome of women preparing for pregnancy and combining it with our team's established ensemble learning-based pre-pregnancy early warning model for spontaneous abortion risk, we can provide rapid, convenient, accurate, and efficient pregnancy outcome prediction for the target population. This prediction model and software development meet current societal medical needs, offering a unique, scarce, and highly novel and innovative service that can significantly reduce social medical costs and greatly promote reproductive health and eugenics. Summary of the Invention

[0007] The first objective of this invention is to provide a method for predicting spontaneous abortion by integrating learning with pre-pregnancy serum metabolomics, addressing the shortcomings of existing technologies.

[0008] A second objective of this invention is a composition of amino acids for predicting spontaneous abortion.

[0009] A third objective of this invention is a detection product for predicting amino acid compositions associated with spontaneous abortion.

[0010] A fourth object of the present invention is the first use of a detection product for predicting amino acid compositions associated with spontaneous abortion and for predicting amino acid compositions associated with spontaneous abortion.

[0011] A fifth object of the present invention is the second use of a detection product for predicting amino acid compositions associated with spontaneous abortion and for predicting amino acid compositions associated with spontaneous abortion.

[0012] To achieve the first objective mentioned above, the technical solution adopted by this invention is: a method for predicting spontaneous abortion by integrating learning with pre-pregnancy serum metabolomics, the method comprising the following steps:

[0013] S1: Clinical information variable entry and pregnancy outcome follow-up for peripheral blood sample collection from women preparing for pregnancy;

[0014] S2: Construct a preconception serum metabolomics dataset;

[0015] S3: Artificial Intelligence Modeling;

[0016] S4: Incorporate the variable values ​​from the dataset in S2 into the artificial intelligence modeling in S3;

[0017] S5: Model feedback verification;

[0018] In step S3, the artificial intelligence modeling uses the extreme gradient boosting algorithm of machine learning to build a prediction model, supplemented by random forest, gradient boosting decision tree, and decision tree to verify the model's predictive ability.

[0019] The S3 further includes data processing, which is as follows:

[0020] S31: Establish a model with all variables;

[0021] S32: Extract the important variables that affect the target variable;

[0022] S33: Screening important influencing factors affecting the target variable;

[0023] S34: Analyze the important influencing factors that have a significant impact on the target variable;

[0024] In S32, the importance scores of each variable in the model are calculated and sorted in descending order. The variables are: number of miscarriages, histidine, glutamine, lysine, arginine, glutamic acid, tryptophan, phenylalanine, threonine, age, body mass index, leucine, serine, tyrosine, methionine, proline, asparagine, isoleucine, and valine.

[0025] As a preferred example, in the method for predicting spontaneous abortion, the extreme gradient boosting algorithm in step S3 is an ensemble learning method consisting of a set of classification and regression trees (CART), each CART (fk(x)) can be defined by an independent tree structure q and leaf weights w, f k (x)=w q(x) ,

[0026] Here, for a given sample xi, it is classified into a leaf using the decision rule in the tree (given by q), and the final prediction is calculated by summing the scores in the corresponding leaves (given by w).

[0027]

[0028] The minimization regularization objective is defined as:

[0029]

[0030] Where n is the number of training samples (determined by the data split from 10x cross-validation), y_i^real is the actual value of the training sample, and the second part of the regularization term defines the complexity of XGBoost to limit the number of leaf nodes and smooth the scores of leaf nodes:

[0031]

[0032] Where T represents the number of leaf nodes in the decision tree, and ψ and λ are the default coefficients of the regularization term.

[0033] As a preferred example, in the method for predicting spontaneous abortion, step S32 calculates the importance score of each variable in the model. The importance scores are as follows: number of miscarriages (10.19), histidine (8.14), glutamine (7.87), lysine (7.19), arginine (6.82), glutamic acid (5.29), tryptophan (5.24), phenylalanine (4.99), threonine (4.74), age (4.58), body mass index (4.53), leucine (4.49), serine (4.27), tyrosine (4.06), methionine (3.83), proline (3.62), asparagine (3.44), isoleucine (3.4), and valine (3.31).

[0034] To achieve the second objective mentioned above, the technical solution adopted by the present invention is as follows:

[0035] An amino acid composition for predicting spontaneous abortion, the amino acid composition comprising the following amino acids: histidine, glutamine, lysine, arginine, glutamic acid, tryptophan, phenylalanine, threonine, leucine, serine, tyrosine, methionine, proline, asparagine, isoleucine, and valine.

[0036] To achieve the third objective mentioned above, the technical solution adopted by the present invention is as follows:

[0037] A detection product for predicting spontaneous abortion-related amino acid compositions includes a product for detecting the expression level of the predicted spontaneous abortion-related amino acid compositions in the peripheral blood of a test individual.

[0038] To achieve the fourth objective mentioned above, the technical solution adopted by the present invention is as follows:

[0039] Application of amino acid compositions for predicting spontaneous abortion and detection products for amino acid compositions for predicting spontaneous abortion in the establishment of a system for predicting spontaneous abortion.

[0040] To achieve the fifth objective mentioned above, the technical solution adopted by the present invention is as follows:

[0041] Application of amino acid compositions for predicting spontaneous abortion and detection products of amino acid compositions for predicting spontaneous abortion in the preparation of diagnostic products for predicting spontaneous abortion.

[0042] The advantages of this invention are:

[0043] This invention provides a method for predicting spontaneous abortion by integrating learning with pre-pregnancy serum metabolomics. (1) It is the first large-scale integration of resources from the fields of reproductive medicine, metabolism, and artificial intelligence to comprehensively manage patients and observe the comprehensive predictive ability of pre-pregnancy serum metabolite indicators and clinical indicators. This provides important clinical diagnostic and intervention basis for the field of reproductive medicine, helps individualized and precise intervention for patients with spontaneous abortion, and achieves a major breakthrough in improving eugenics. (2) Compared with biomarkers or clinical symptoms during pregnancy, this invention can accurately predict pregnancy outcomes before pregnancy, which is expected to prevent and / or intervene in the occurrence of miscarriage at a very early stage, greatly reduce social medical costs, and promote the development of maternal and child health and reproductive health.

[0044] In this embodiment of the invention, (1) by combining the patient's peripheral blood serum metabolomics, clinical information variables, and follow-up patient pregnancy outcomes, a pregnancy outcome prediction model was constructed using machine learning methods such as XGBoost, RF, GBDT, and DT. The optimized model achieved a prediction accuracy of 87%, demonstrating significant predictive value. (2) In the established artificial intelligence model, ensemble feature engineering was used to evaluate the contribution of feature values, further revealing that the levels of various amino acids are variables with a significant impact after the number of miscarriages. Histidine, the most important metabolite, was selected for mechanistic research, and animal and cell models were used to demonstrate that high histidine levels lead to increased embryonic resorption. This expands the understanding of the pathological mechanisms of spontaneous abortion, especially recurrent spontaneous abortion, and provides more strategies for subsequent diagnosis and treatment.

[0045] This invention develops an integrated learning system combining pre-pregnancy serum metabolomics to predict spontaneous abortion, namely, an interactive platform for predicting spontaneous abortion risk using a pre-pregnancy early warning model (AI-MP). This allows clinicians to easily and quickly input data and predict the risk of spontaneous abortion using the established early warning model in a very short time (no more than 5 milliseconds). Furthermore, the AI-MP platform will allow researchers to further optimize the early warning model using continuously accumulating clinical and metabolomics data and custom parameters to achieve better predictive performance. Attached Figure Description

[0046] Figure 1 Technology roadmap.

[0047] Figure 2 Machine learning model building.

[0048] Figure 3Non-targeted LC-MS detection of peripheral blood metabolome in NOR / RSM patients before pregnancy.

[0049] Figure 4 Feeding mice with high histidine diets can lead to embryo loss.

[0050] Figure 5 Miscarriage prediction software based on a machine-trained pre-pregnancy early warning model for the risk of natural miscarriage.

[0051] (AI-MP). Detailed Implementation

[0052] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the description of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0053] Example 1

[0054] Non-targeted / targeted metabolomics was used to detect the levels of small molecule metabolites in the pre-pregnancy serum of women of childbearing age who were trying to conceive. Simultaneously, clinical data from patients (including basic information, history of miscarriages, and relevant peripheral blood test indicators) were collected to establish a database of candidate predictive variables for the predictive model. Based on ensemble algorithms, the XGBoost algorithm from machine learning was used, supplemented by decision trees, gradient boosting decision trees, random trees, and other algorithms to establish and validate a spontaneous abortion prediction model. Simultaneously, AI-MP, an artificial intelligence-based pregnancy outcome prediction software, was developed and made available to the public to assist in clinical diagnosis and treatment. The technology roadmap is available in [link to technology roadmap]. Figure 1 .

[0055] 1. Collect clinical samples

[0056] Peripheral blood serum samples were collected from patients who underwent initial consultation at the Department of Reproductive Immunology, Shanghai First Maternity and Infant Hospital between September 2020 and March 2021 and were in the preconception period. Comprehensive clinical information was collected, including: basic information (age, body mass index, reproductive history, occupation, education level, smoking and drinking history), genetic factors (presence of chromosomal abnormalities in either parent or in the embryo), immune factors (presence of antiphospholipid syndrome, Sjögren's syndrome, etc.), anatomical factors (presence of congenital or acquired uterine anatomical abnormalities), thrombotic tendency (MTHFR mutation, protein C / S deficiency, coagulation factor V Leiden mutation, etc.), endocrine factors (presence of insulin resistance, polycystic ovary syndrome, hypothyroidism, hyperprolactinemia, luteal insufficiency, etc.), and infectious factors (presence of reproductive tract pathogen infections, such as Ureaplasma urealyticum, Mycoplasma hominis, Chlamydia trachomatis, Listeria monocytogenes, Toxoplasma gondii, rubella, cytomegalovirus, herpesvirus, etc.).

[0057] 1.1 Selection and Enrollment of Subjects: Patients who were newly diagnosed at the Department of Reproductive Immunology of Shanghai First Maternity and Infant Hospital from September 2020 to December 2020 were selected as subjects for model training and internal testing, and patients who were newly diagnosed at the Department of Reproductive Immunology of Shanghai First Maternity and Infant Hospital from January 2021 to March 2021 were selected as subjects for external validation.

[0058] 1.2 Inclusion criteria: Females aged 18 to 40 years, and the subject (or their legal representative) has signed an informed consent form.

[0059] 1.3 Exclusion criteria: history of major cardiovascular or cerebrovascular disease; liver or kidney dysfunction; clear history of mental illness; participation in other clinical studies within 2 months prior to the start of the study; patient or their authorized representative is unwilling to sign a written informed consent form or comply with the study protocol.

[0060] 1.4 Recruitment Methods: Patients newly diagnosed at the Department of Reproductive Immunology, Shanghai First Maternity and Infant Hospital were recruited.

[0061] 1.5 Data Acquisition

[0062] (1) Medical history was collected from all included subjects.

[0063] Basic information: age, body mass index, pregnancy history, occupation, education level, smoking and drinking history. Genetic factors: presence of chromosomal abnormalities in both parents and in the embryo.

[0064] Immune factors: whether there is antiphospholipid syndrome, Sjögren's syndrome, etc.

[0065] Anatomical factors: presence or absence of congenital uterine anatomical abnormalities: arcuate, septate, unicornuate, and bicornuate uterus, etc.; presence or absence of acquired uterine anatomical abnormalities: uterine fibroids, endometrial polyps, intrauterine adhesions, etc.

[0066] Thrombosis tendency: MTHFR mutation, protein C / S deficiency, coagulation factor V Leiden mutation, etc. Endocrine factors: presence of insulin resistance, polycystic ovary syndrome, hypothyroidism, hyperprolactinemia, luteal insufficiency, etc.

[0067] Infectious factors: Whether there is infection with reproductive tract pathogens, such as Ureaplasma urealyticum, Mycoplasma hominis, Chlamydia trachomatis, Listeria monocytogenes, Toxoplasma gondii, Rubella, Cytomegalovirus, Herpes simplex virus, etc.

[0068] (2) Fasting venous blood was collected from all subjects within 24 hours, and small molecule metabolites in the serum were detected by full scan using liquid chromatography-LC-MS.

[0069] 1.6 Follow-up

[0070] All patients were followed up by members of the research team, and their conditions at different stages of pregnancy were recorded, including biochemical events, embryonic arrest, miscarriage, and full-term delivery.

[0071] 2. Detection and analysis of metabolites

[0072] Serum samples were pretreated by mixing with methanol before being processed. Metabolites were then separated using an Agilent 1290 ultra-high performance liquid chromatography (UHPLC) system, coupled with an Agilent 6530B quadrupole time-of-flight mass spectrometer (qTOF-MS) for mass-charge ratio scanning. Agilent software was used for substance identification and abundance quantification of the mass spectrometric peaks.

[0073] 3. Establish an early warning model

[0074] An Extreme Gradient Boosting (XGBoost) ensemble learning algorithm was employed to build a miscarriage risk prediction model. This method consists of a set of classification and regression trees, each defined by an independent tree structure and leaf weights. The training process optimizes the tree structure and leaf scores, employing an additive strategy: with a fixed tree structure and learned leaf scores, only one new tree is added at a time, and gradient optimization is performed. We found that for building a miscarriage risk warning model, XGBoost is more efficient than other machine learning methods such as Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Decision Tree (DT), and the resulting model exhibits higher accuracy and precision. The model was validated using various methods, including random split validation, cross-validation, and Bootstrap repeated sampling validation.

[0075] 3.1 Data Processing

[0076] (1) Establish a model with all variables

[0077] All data is entered into the system by designated personnel and then "normalized": continuous variables are transformed into a single value between 0 and 1 using a linear function; binary variables are represented as either 0 or 1. This data processing method can prevent data divergence to some extent.

[0078] (2) Extract important variables that affect the target variable

[0079] Using XGBoost supplemented by random trees, gradient boosting decision trees, and decision tree models, the importance scores of each variable in the model are calculated and sorted in descending order. At the same time, a 5-fold cross-validation method is used to select all important variables with an importance score > 0 in the 5-fold cross-validation, and the intersection of the two is the important variable that affects the target variable.

[0080] (3) Screening important influencing factors affecting the target variable

[0081] The important variables obtained are segmented and one-hot encoded according to the normal value reference range. The remaining continuous variables are segmented and one-hot encoded according to the upper and lower quartiles based on the data distribution.

[0082] (4) Analyze the important influencing factors that have a significant impact on the target variable.

[0083] By using propensity score matching (control variables being pregnancy outcome: clinical pregnancy or not; delivery outcome: live birth or not), the baseline levels of the experimental group and the control group were made to be basically consistent. Then, based on the re-screened samples of each group, the remaining important influencing factors were subjected to chi-square tests to obtain the factors that had a significant impact on the target variable.

[0084] 3.2 Model Construction and Validation

[0085] The prediction model is built using the Extreme Gradient Boosting (XGBoost) algorithm of machine learning, and is further supported by Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Decision Tree (DT) to verify the predictive ability of the model.

[0086] Extreme gradient boosting is an ensemble learning method consisting of a set of classification and regression trees (CARTs). Each CART (fk(x)) can be defined by an independent tree structure q and leaf weights w, where f... k (x)=w q(x) Here, for a given sample xi, we classify it into a leaf using the decision rule in the tree (given by q), and compute the final prediction by summing the scores in the corresponding leaves (given by w).

[0087]

[0088] Our minimized regularization objective is defined as:

[0089]

[0090] Where n is the number of training samples (determined by the data split from 10x cross-validation), and y_i^real is the actual value of the training sample. The second part of the regularization term defines the complexity of XGBoost to limit the number of leaf nodes and smooth the scores of the leaf nodes:

[0091]

[0092] Where T represents the number of leaf nodes in the decision tree, and ψ and λ are the default coefficients of the regularization term.

[0093] The training process of this model involves optimizing the tree structure and leaf scores, consistent with traditional decision tree training. In XGBoost, an additive strategy is used during training: the tree structure and the already learned leaf scores are fixed, and a new tree is added one at a time, with gradient optimization performed.

[0094] The verification methods include: internal verification within the model development queue and external verification independent of the development queue.

[0095] Internal validation: Available methods include random split validation, cross-validation, and Bootstrap repeated sampling validation.

[0096] External validation: Available methods include time-based validation, spatial validation, and domain validation.

[0097] This study employed time-period validation, externally validating the predictive model by prospectively and continuously including women trying to conceive from January 2023 to December 2023.

[0098] 4. Develop early warning model software

[0099] An interactive platform for miscarriage prediction (AI-MP) based on a trained pre-pregnancy early warning model for spontaneous abortion risk was developed. This platform allows experimenters to easily and quickly input data and predict spontaneous abortion risk using the established early warning model within a very short time (no more than 5 milliseconds). Furthermore, the AI-MP platform will allow researchers to further optimize the early warning model using continuously accumulating clinical and metabolomics data and custom parameters to achieve better predictive performance.

[0100] Research findings:

[0101] We first collected pre-pregnancy peripheral blood samples from 481 patients with RSM and normal individuals, and performed non-targeted metabolomics LC-MS analysis. Based on the pre-pregnancy serum metabolomics data, combined with clinical information, and following up on pregnancy outcomes, we constructed optimal models for predicting pregnancy outcomes using machine learning methods such as XGBoost, RF, GBDT, and DT. The optimized machine learning model, XGBoost, achieved a prediction accuracy of 87%, demonstrating good predictive value. Furthermore, the model based on XGBoost revealed that histidine is the second most important influencing factor on pregnancy outcome after the number of miscarriages (see...). Figure 2 ), Figure 2 In the following data: PNM: Number of miscarriages (10.19); L-Histidine: Histidine (8.14); L-Glutamine: Glutamine (7.87); L-Lysine: Lysine (7.19); L-Arginine: Arginine (6.82); L-Glutamate: Glutamate (5.29); L-Tryptophan: Tryptophan (5.24); L-Phenylalanine: Phenylalanine (4.99); L-Threonine: Threonine (4.74); Age: Age (4.5) 8); BMI: Body Mass Index (4.53); L-leucine: Leucine (4.49); L-Serine: Serine (4.27); L-Tyrosine: Tyrosine (4.06); L-Methionine: Methionine (3.83); L-Proline: Proline (3.62); L-Asparagine: Asparagine (3.44); L-Isoleucine: Isoleucine (3.4); L-Valine: Valine (3.31). Left graph: Horizontal axis represents iterations; vertical axis represents accuracy. This is to illustrate that the XGBoost algorithm establishes a stable model within 100 iterations, as can be seen from the prediction accuracy of each iteration on the training dataset and the 10x CV on the test set. Right graph: Horizontal axis represents the included indicator data (input features of machine learning); vertical axis represents the importance of the input features. The figure on the right illustrates the model based on XGBoost, revealing the importance ranking of input features to obtain deeper analysis and insights into pregnancy outcomes.

[0102] 1. Compared with the peripheral blood of normal patients, the histidine content in the peripheral blood of RSM patients was significantly increased (see...). Figure 3 ).

[0103] To investigate the effect of histidine on spontaneous abortion, we conducted an animal experiment using 8-week-old male and female C57BL / 6 mice fed a special diet. Eight-week-old female C57BL / 6 mice (weighing 20-23g, n=24) were randomly divided into two groups: a control group (fed a normal diet) and a high-histidine group (fed a diet containing 10% histidine). The different diets for each group were started two weeks in advance. Two weeks later, peripheral blood and uterine horns of four mice from each group were collected, and mass spectrometry was used to detect changes in histidine levels. The remaining 16 female mice were then combined with male mice in a 1:2 ratio. The vaginal plugs were examined at 6 PM the previous evening; the presence of a vaginal plug was defined as 0.5 days of pregnancy. On day 12.5, the mice were euthanized by cervical dislocation, and uterine horns were collected to observe embryo resorption and calculate the embryo resorption rate. We found that feeding mice a high-histidine diet two weeks in advance increased histidine levels in peripheral blood before pregnancy and increased embryo resorption rate after pregnancy (see...). Figure 4 ).

[0104] 2. Based on an ensemble learning-based natural model for predicting pre-pregnancy serum metabolomics, code was written to develop pregnancy outcome prediction software. This software achieves pregnancy outcome prediction speeds within 5 milliseconds, significantly reducing medical testing cycles and aiding clinical diagnosis and treatment. Furthermore, this software will allow researchers to further optimize the early warning model using continuously accumulating clinical and metabolomics data and custom parameters to obtain even better predictive performance (see...). Figure 5 ).

[0105] in conclusion:

[0106] Preconception serum metabolomics combined with artificial intelligence to predict pregnancy outcomes

[0107] Therefore, constructing a miscarriage prediction model based on the pre-pregnancy serum metabolome and all potential clinical predictors was established as the research content of our group, and we conducted preliminary experimental studies. We first collected peripheral blood samples from 481 patients with RSM and normal patients before pregnancy and performed non-targeted metabolomics LC-MS detection. Principal component analysis and orthogonal partial least squares discriminant analysis revealed that compared with the peripheral blood of normal patients, the lysine and histidine content in the peripheral blood of RSM patients was significantly increased. Many studies have reported that the reduced activity of histamine-degrading enzyme DAO in the placenta can lead to excessive accumulation of histamine. Excessive histamine can stimulate vasoconstriction, uterine contraction, and an overactive immune response, thereby triggering the occurrence of RSM (24,25). The activity of histidine decarboxylase HDC and histamine-degrading enzyme DAO is activated by hormone regulation after pregnancy (26). In animal experiments, this invention found that feeding mice with a high concentration of histidine diet in advance can lead to an increase in histidine content in the peripheral blood of mice before pregnancy and an increase in embryonic absorption after pregnancy. These results further suggest that excessive histidine may be a potential high-risk factor for miscarriage in RSM patients, and that preconception serum metabolomics based on artificial intelligence is expected to become a very good miscarriage prediction model.

[0108] Based on the above findings, we combined pre-pregnancy serum metabolomics data, clinical information, and pregnancy outcome follow-up to construct an optimal model for predicting pregnancy outcomes using machine learning methods such as XGBoost, RF, GBDT, and DT. The optimized model achieved a prediction accuracy of 87%, demonstrating significant predictive value. Furthermore, the machine learning model revealed that histidine levels are a variable with a significant impact after the number of miscarriages. In addition, our research team developed a miscarriage prediction software (AI-MP) based on a machine-trained pre-pregnancy early warning model for spontaneous abortion risk. This allows experimenters in miscarriage early warning platforms to easily and quickly input data and predict spontaneous abortion risk using the established early warning model within a very short time (no more than 5 milliseconds).

[0109] This study is the first to explore the clinical predictive value of preconception serum metabolomics data based on artificial intelligence, providing important evidence-based medicine for the very early treatment of RSM. It pioneered the accurate prediction of miscarriage before pregnancy, and is expected to prevent and / or intervene in miscarriage at an early stage, which will be a major breakthrough in improving the health and disease of pregnant women.

[0110] References

[0111] [1]S.Quenby,IDGallos,RKDhillon-Smith,M Podesek,MDStephenson,JFisher,JJBrosens,JBrewin,R Ramhorst,ESLucas,RCMcCoy,R Anderson,S Daher,LReg an, Al-Memar M, Bourne T, DMacIntyre, R Rai, OBChristiansen, Sugiura-Ogasawara M, Odendaal J, Devall AJ, PRBennett, Petrou S, Coomarasamy A, Miscarriage matters:the epidemiological,physical,psychological,and economiccosts of early pregnancy loss,Lancet(London,England)397(10285)(2021)1658-1667.

[0112] [2]A.Coomarasamy,RKDhillon-Smith,A.Papadopoulou,M.Al-Memar,J.Brewin,VMAbrahams,A.Maheshwari,OBChristiansen,MDStephenson,MGoddijn,OTOladap o,CNWijeyaratne,D.Bick,H.Shehata,R.Small,PRBennett,L.Regan,R.Rai,T.Bourne,R.Kaur,O.Pickering,JJBrosens,AJDevall,IDGallos,S.Quenby,Appellant miscarriage:evidence to accelerate action,Lancet(London,England)397(10285)(2021)1675-1682.

[0113] [3]Q.Yue,A.Yuan,X.Che,M.Huynh,C.Zhou,Batch Model for BatchedTimestamps Data Analysis with Application to the SSADisability Program,KDD:proceedings.International Conference on Knowledge Discovery&DataMining2016(2016)343-352.

[0114] [4]N.M.Cohen,O.Schwartzman,R.Jaschek,A.Lifshitz,M.Hoichman,R.Balicer,L.I.Shlush,G.Barbash,A.Tanay,Personalizedlabtestmodelstoquantifydiseasepotentials inhealthyindividuals,Naturemedicine27(9)(2021)1582-1591.

[0115] [5]K.K.Venkatesh,R.A.Strauss,C.A.Grotegut,R.P.Heine,N.C.Chescheir,J.S.A.Stringer,D.M.Stamilio,K.M.Menard,J.E.Jelovsek,Machine Learning andStatistical Models to Predict Postpartum Hemorrhage,Obstetrics and gynecology135(4)(2020)935-944.

[0116] [6]M.Lipschuetz,J.Guedalia,A.Rottenstreich,M.Novoselsky Persky,S.M.Cohen,D.Kabiri,G.Levin,S.Yagel,R.Unger,Y.Sompolinsky,Prediction ofvaginalbirth after cesarean deliveries using machine learning,American journal ofobstetricsandgynecology222(6)(2020)613.e1-613.e12.

[0117] [7]H.Liu,J.Li,J.Leng,H.Wang,J.Liu,W.Li,H.Liu,S.Wang,J.Ma,J.C.Chan,Z.Yu,G.Hu,C.Li,X.Yang,Machine learning risk score for prediction ofgestational diabetes in early pregnancy in Tianjin,China,Diabetes / metabolismresearchandreviews 37(5)(2021)e3397.

[0118] [8]I. A.Tsur,N.Aghaeepour,A.Montanari,D.K.Stevenson,G.M.Shaw,V.D.Winn,Early prediction of preeclampsiaviamachinelearning,Americanjournalofobstetrics&gynecologyMFM2(2)(2020)100100.

[0119] [9]A.I.Naimi,R.W.Platt,J.C.Larkin,Machine Learning for Fetal GrowthPrediction,Epidemiology(Cambridge,Mass.)29(2)(2018)290-298.

[0120]

[10] R.Raja,I.Mukherjee,B.K.Sarkar,A Machine Learning-Based PredictionModel for Preterm Birth in Rural India,Journalofhealthcareengineering2021(2021)6665573.

[0121]

[11] L.Liang,M.H.Rasmussen,B.Piening,X.Shen,S.Chen,H. J.K.Snyder,R.Tibshirani,L.Skotte,N.C.Lee,K.Contrepois,B.Feenstra,H.Zackriah,M.Snyder,M.Melbye,Metabolic Dynamics and Prediction of GestationalAgeandTimetoDeliveryinPregnantWomen,Cell 181(7)(2020)1680-1692.e15.

[0122]

[12] T.T.M.Ngo,M.N.Moufarrej,M.H.Rasmussen,J.Camunas-Soler,W.Pan,J.Okamoto,N.F.Neff,K.Liu,R.J.Wong,K.Downes,R.Tibshirani,G.M.Shaw,L.Skotte,D.K.Stevenson,J.R.Biggio,M.A.Elovitz,M.Melbye,S.R.Quake,Noninvasive bloodtests for fetal development predict gestational age and preterm delivery,Science(New York,N.Y.)360(6393)(2018)1133-1136.

[0123]

[13] M.Rasmussen,M.Reddy,R.Nolan,J.Camunas-Soler,A.Khodursky,NMScheller,DECantonwine,L.Engelbrechtsen,JDMi,A.Dutta,T.Brundage,F.Siddiqui,M.Thao,EPSGee,J.La,C.Baru ch-Gravett,MKSantillan,S.Deb,SMAme,SMAli,M.Adkins,MADePristo,M.Lee,E.Namsaraev,DJGybel-Brask,L.Skibsted,JALitch,DASantillan,S.Sazawal,RMTribe,JMRoberts,M.Jain,E. C. Holzman, SR Quake, MA Elovitz, TF McElrath, RNA profiles reveal signatures of future health and disease inpregnancy, Nature 601 (7893) (2022) 422-427.

[0124] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the method of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A method of integrating learning combining pre-pregnancy serum metabolome to predict spontaneous abortion, characterized in that: The method includes the following steps: S1: Collect peripheral blood samples from women preparing for pregnancy, enter clinical information variables, and follow up on pregnancy outcomes; S2: Construct a preconception serum metabolomics dataset; S3: Artificial Intelligence Modeling; S4: Incorporate the variable values ​​from the dataset in S2 into the artificial intelligence modeling in S3; S5: Model feedback verification; In step S3, the artificial intelligence modeling uses the extreme gradient boosting algorithm of machine learning to build a prediction model, supplemented by random forest, gradient boosting decision tree, and decision tree to verify the model's predictive ability. The S3 further includes data processing, which is as follows: S31: Establish a model with all variables; S32: Extract the important variables that affect the target variable; S33: Screening important influencing factors affecting the target variable; S34: Analyze the important influencing factors that have a significant impact on the target variable; In S32, the importance scores of each variable in the model are calculated and sorted in descending order. The variables are: number of miscarriages, histidine, glutamine, lysine, arginine, glutamic acid, tryptophan, phenylalanine, threonine, age, body mass index, leucine, serine, tyrosine, methionine, proline, asparagine, isoleucine, and valine.

2. The method for predicting spontaneous abortion according to claim 1, characterized in that: The extreme gradient boosting algorithm in step S3 is an ensemble learning method composed of a set of classification and regression trees (CART), each of which (fk(x)) can be defined by an independent tree structure q and leaf weights w, f k (x) = w q(x) , Here, for a given sample xi, the decision rules in the tree are used to classify it into a leaf, and the final prediction is calculated by summing the scores in the corresponding leaves. : The minimization regularization objective is defined as: Where n is the number of training samples, determined by the data split from 10x cross-validation, y_i^real is the actual value of the training sample, and the second part of the regularization term defines the complexity of XGBoost to limit the number of leaf nodes and smooth the scores of leaf nodes: Where T represents the number of leaf nodes in the decision tree, and ψ and λ are the default coefficients of the regularization term.

3. The method for predicting spontaneous abortion according to claim 1 or 2, characterized in that: In step S32, the importance scores of each variable in the model are calculated as follows: number of miscarriages 10.19, histidine 8.14, glutamine 7.87, lysine 7.19, arginine 6.82, glutamic acid 5.29, tryptophan 5.24, phenylalanine 4.99, threonine 4.74, age 4.58, body mass index 4.53, leucine 4.49, serine 4.27, tyrosine 4.06, methionine 3.83, proline 3.62, asparagine 3.44, isoleucine 3.4, valine 3.31.

Images