Method and system for predicting severe brain injury probability of newborn

Abstract

The method for predicting the severe brain injury probability of the newborn is used for predicting the severe brain injury probability of a selective fetal growth limited double-fetal newborn, and the selective fetal growth limited double-fetal newborn comprises a to-be-detected large fetus and a to-be-detected small fetus. The method comprises the following steps: acquiring a meconium sample of a large fetus to be detected and a meconium sample of a small fetus to be detected; the method comprises the following steps: pretreating and detecting a meconium sample of a large fetus to be detected and a meconium sample of a small fetus to be detected to obtain the concentration of marker metabolites in the meconium sample to be detected; according to the concentration of the marker metabolite of the big fetus and the concentration of the marker metabolite of the small fetus, the probability of severe brain injury of the big fetus and the small fetus is predicted; wherein the big fetus marker metabolite is different from the small fetus marker metabolite; according to the embodiment of the invention, the characteristic of recording the intrauterine metabolism state of the fetus in the whole gestation period by virtue of meconium is utilized, and the severe brain injury risk can be predicted in the early stage after birth. Different specific markers are screened for fetuses of different sizes, a prediction model is established, and the prediction result is accurate.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology, specifically to a method and system for predicting the probability of severe brain injury in twin newborns with selective fetal growth restriction. Background Technology

[0002] Selective fetal growth restriction (sFGR) is a serious complication of monochorionic diamniotic (MCDA) twin pregnancies, significantly increasing the risk of adverse perinatal outcomes. According to the developmental origin of health and disease (DOHaD) theory, the unfavorable intrauterine environment in sFGR may have profound and lasting effects on the neurological development of offspring. Therefore, affected infants are particularly susceptible to brain injuries such as periventricular leukomalacia (PVL), diffuse white matter lesions, and intraventricular hemorrhage (IVH), which in turn increases the risk of neurodevelopmental disorders and functional deficits such as cerebral palsy later in life. Thus, early identification and risk stratification of neonatal brain injury in sFGR remains a critical need in the clinical management of sFGR.

[0003] Currently, the diagnosis of perinatal brain injury can only rely on the results of cranial ultrasound or MRI scans during the mid-to-late stages of pregnancy and after birth. However, since the development of brain injury is a dynamic process, many fetuses with potentially serious abnormalities are difficult to identify early through imaging examinations, which to some extent hinders the early prediction and intervention process for severe neonatal brain injury in clinical practice. Summary of the Invention

[0004] This application provides a method for predicting the probability of severe brain injury in newborns, specifically for predicting the probability of severe brain injury in selectively fetal growth restriction twins, where the selectively fetal growth restriction twins include a larger fetus and a smaller fetus to be examined. The method includes: Obtain meconium samples from the large fetus to be examined and the small fetus to be examined; Preprocessing of meconium samples from the large fetus to be examined and meconium samples from the small fetus to be examined; The concentration of fetal marker metabolites in the meconium samples of the pretreated fetuses was obtained by detecting the meconium samples of the fetuses to be tested. The concentration of fetal marker metabolites in the pretreated meconium samples of the fetuses to be tested was obtained. The probability of severe brain injury in a large fetus can be predicted by the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined. The probability of severe brain injury in a fetus can be predicted by the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined. Among them, the marker metabolites of large fetuses and small fetuses are different.

[0005] This application also provides a system for predicting the probability of severe brain injury in newborns, used to predict the probability of severe brain injury in selectively fetal growth restriction twin newborns, the selectively fetal growth restriction twin newborns including the larger fetus to be examined and the smaller fetus to be examined, the system includes: The sample acquisition module is used to acquire meconium samples from the large fetus to be examined and the small fetus to be examined. The preprocessing module is used to preprocess meconium samples from large fetuses and small fetuses to be examined. The metabolite concentration detection module is used to detect the meconium sample of the pretreated large fetus to obtain the concentration of large fetal marker metabolites in the meconium sample of the large fetus to be tested; and to detect the meconium sample of the pretreated small fetus to be tested to obtain the concentration of small fetal marker metabolites in the meconium sample of the small fetus to be tested. The probability prediction module is used to predict the probability of severe brain injury in the large fetus based on the concentration of fetal marker metabolites in the meconium sample of the large fetus to be tested, and to predict the probability of severe brain injury in the small fetus based on the concentration of fetal marker metabolites in the meconium sample of the small fetus to be tested. Among them, the marker metabolites of large fetuses and small fetuses are different. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating a method for predicting the probability of severe brain injury in newborns according to an embodiment of this application.

[0007] Figure 2 This is a schematic diagram of a method for predicting the probability of severe brain injury in newborns according to an embodiment of this application.

[0008] Figure 3 This is a schematic diagram of a marker metabolite screening method according to an embodiment of this application.

[0009] Figure 4 The ROC curve of the first candidate prediction model in control case 1 is shown.

[0010] Figure 5 The ROC curve is shown for the second candidate prediction model in control case 1. Detailed Implementation

[0011] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0012] This invention provides a method 10 for predicting the probability of severe brain injury in newborns. Method 10 is used to predict the probability of severe brain injury in twin newborns with selective fetal growth restriction.

[0013] According to the standards of the International Society of Ultrasound in Obstetrics and Gynecology, a monochorionic diamniotic (MCDA) twin pregnancy with an estimated fetal weight difference exceeding 25% (calculated as 100 × (larger fetal weight - smaller fetal weight) / larger fetal weight) and where the estimated fetal weight of the smaller fetus is below the 10th percentile is defined as selective fetal growth restriction (sFGR). Larger and smaller fetuses in sFGR neonates are designated as sFGR-L and sFGR-S, respectively.

[0014] Selective fetal growth restriction (sFGR) twins to be examined include the larger twin and the smaller twin. The larger twin is the one with the higher birth weight in the selected sFGR twins to be examined, and the smaller twin is the one with the lower birth weight in the selected sFGR twins to be examined.

[0015] When assessing brain injury, focus on periventricular leukomalacia (PVL) and intraventricular hemorrhage (IVH).

[0016] The PVL classification refers to the de Vries classification: - Grade I: Periventricular echogenicity persists for more than 7 days, with no cystic cavities present; - Grade II: Localized increased echogenicity around the ventricles, which later becomes a localized small cystic cavity; - Grade III: Extensive periventricular echogenicity, which later transforms into extensive cystic cavities; - Grade IV: Extensive periventricular echogenicity involving subcortical white matter, later progressing to diffuse periventricular and subcortical cystic lesions.

[0017] IVH classification is based on the Papile standard: - Grade I: Mild IVH or subependymal hemorrhage; - Grade II: IVH occupies 10%-50% of the ventricular area; - Grade III: IVH occupies >50% of the ventricular area (accompanied by ventricular dilatation); - Grade IV: Cerebral parenchymal hemorrhage (most likely caused by hemorrhagic venous infarction).

[0018] In this application, mild brain injury is defined as grade I-II IVH or grade I PVL, and severe brain injury is defined as grade III-IV IVH or grade II-IV PVL.

[0019] like Figure 1 As shown, method 10 includes: Step 110: Obtain meconium samples from the large fetus to be examined and the small fetus to be examined.

[0020] Meconium forms in the fetal intestine during the second month of pregnancy and is usually expelled within 24-48 days after birth. Its components include shed mucosal epithelial cells, water, bile, bile acids, epithelial cells, and other biomolecules ingested by the fetus when swallowing amniotic fluid. Meconium is an important matrix for analyzing fetal metabolic characteristics, directly reflecting the fetus's long-term exposure to the intrauterine environment throughout pregnancy.

[0021] For example, when collecting meconium samples, use a sterile plastic spoon to collect approximately 1-1.5 grams of meconium sample and quickly transfer it to a sterile centrifuge tube to avoid sample contamination. Immediately after collection, label the sample information, such as fetal identity and collection time, and temporarily freeze the sample at -20°C within 1 hour. Transport the sample to the laboratory using dry ice within 24 hours and finally store it long-term in an ultra-low temperature freezer at -80°C. Moreover, the sample is only allowed to undergo one freeze-thaw cycle before subsequent analysis.

[0022] Meconium was chosen as a sample due to its unique advantages. Its components include fetal intestinal mucosal epithelial cells shed, bile acids, biomolecules swallowed from amniotic fluid, and accumulated products of maternal-fetal metabolic activity, providing a complete reflection of the fetus's intrauterine metabolic environment and exposure throughout pregnancy. Compared to maternal plasma, which is limited by the placental barrier, and umbilical cord blood, which only reflects the state at the time of delivery, meconium is more suitable for tracing the link between long-term fetal metabolic disorders and brain damage. Furthermore, the first meconium sample is unaffected by the postnatal external environment, preserving the fetus's intrauterine metabolic characteristics to the greatest extent. Strict aseptic techniques and low-temperature storage prevent metabolite degradation, ensuring accurate subsequent test results. At the same time, clearly distinguishing between samples from large and small fetuses is fundamental for subsequent screening of specific marker metabolites for their different pathological mechanisms, avoiding confounding between groups.

[0023] Step 120: Preprocess the meconium samples of the large fetus to be examined and the small fetus to be examined.

[0024] In step 120, metabolites are extracted from the meconium sample through pretreatment.

[0025] For example, step 120 includes steps 1201-1204.

[0026] Step 1201: Sample drying and weighing. For example, the sample is removed from a -80°C freezer, thawed at room temperature, dried at a low temperature below 40°C using a Labconco SpeedVac vacuum concentrator, and then 10 mg ± 1 mg of the dried meconium powder is accurately weighed.

[0027] Step 1202: Metabolite Extraction. For example, 600 μL of 100% methanol containing an internal standard mixture is added to the weighed meconium powder. Methanol, as a polar solvent, can efficiently extract small-molecule metabolites from meconium while precipitating large-molecule impurities such as proteins. The internal standard mixture, such as n-hexane-d14, can be used to correct for detection errors.

[0028] Step 1203: Homogenization and centrifugation. For example, using a Qiagen Tissuelyser-II tissue homogenizer (Germany), homogenization was performed at 30 Hz for 30 seconds to break up the meconium particle structure and allow the metabolites to fully dissolve in methanol. Subsequently, the mixture was centrifuged at 17,000 × g for 15 minutes at 4°C. The supernatant was the metabolite extract, and the bottom precipitate was discarded to reduce detection interference.

[0029] Step 1204: Derivatization. For example, the supernatant is dried again using a SpeedVac concentrator until no liquid residue remains, and 50 μL of methyl chloroformate (MCF) derivatization reagent is added. The mixture is then incubated at 37°C for 30 minutes. This step converts highly polar, non-volatile metabolites into volatile derivatives, meeting the requirements for subsequent gas chromatography-mass spectrometry (GC-MS) detection.

[0030] Step 130: Detect the meconium sample of the pretreated fetus to obtain the concentration of fetal marker metabolites in the meconium sample of the fetus to be tested. In step 130, gas chromatography and mass spectrometry can be used to detect the meconium sample of the pretreated fetus to be tested, and the concentration of fetal marker metabolites in the meconium sample of the fetus to be tested can be obtained.

[0031] Large fetal marker metabolites are metabolites that are pre-screened from meconium samples of large fetuses and can be used to predict the probability of severe brain injury in large fetuses.

[0032] Specifically, the instrument used was an Agilent 7890 gas chromatography system (Agilent 7890 GC), coupled with a 5975 mass spectrometer detector (Agilent 5975 MSD, Agilent Technologies, Inc.) with an electron impact ionization source (70 eV).

[0033] Step 130 includes steps 1301-1304.

[0034] Step 1301: Perform chromatographic separation on the pretreated meconium sample of the large fetus to be tested, i.e., the meconium sample filtrate of the large fetus.

[0035] The purpose of chromatographic separation is to separate the various metabolites mixed together in the filtrate according to their characteristics, so that each metabolite can "queue up" separately for mass spectrometry detection.

[0036] For example, the chromatographic separation conditions in step 1301 are as follows: The chromatographic column was an HSS T3 C18 column (2.1 × 100 mm, 1.8 μm; Waters), and the mobile phase was solvent A: 0.1% formic acid aqueous solution; solvent B: 0.1% formic acid acetonitrile solution. Gradient elution was performed using different concentrations of solvent B, with a total run time of 25 min. The specific process is shown in Table 1.

[0037] Table 1 Step 1302: Perform mass spectrometry detection on the chromatographically separated products to obtain raw metabolite data.

[0038] The purpose of mass spectrometry is to detect individual metabolites at the molecular level after chromatographic separation in order to determine the concentration of the metabolites.

[0039] For example, the mass spectrometry conditions were as follows: data acquisition was performed using an Agilent 7890 GC-5975 MSD system, configured with an electron impact ionization source (EI, 70 eV) and a ZB-1701 gas chromatographic column (30 m × 250 μm id × 0.15 μm, Phenomenex); the carrier gas was helium (flow rate 1 mL / min), the injection port temperature was 290°C, and the mass spectrometry parameters were set as follows: ion source temperature 150°C, quadrupole temperature 230°C, transfer line temperature 250°C, mass scan range m / z 38-550, and scan rate 1.562 μs. - ¹, solvent delay 5.5 min.

[0040] Step 1303: Clean and annotate the raw metabolite data to obtain the concentration of fetal marker metabolites in the meconium sample of the fetus to be tested.

[0041] In step 1303, the raw metabolite data detected by mass spectrometry is processed to transform the messy signal into clear metabolite information.

[0042] For example, in step 1303, peak extraction (signal-to-noise ratio S / N>3) is first performed to screen out the real metabolite signals and eliminate noise. Secondly, retention time alignment (offset tolerance <0.2 min) is used to unify the identification criteria for the same metabolite in different samples.

[0043] When different samples (such as meconium from different fetuses) are tested, the retention time (i.e., the time at which it is detected) of the same metabolite may vary slightly due to minor operational differences (such as injection time and column temperature fluctuations). Retention time alignment corrects for these minor differences: by setting an offset tolerance of <0.2 min, as long as the retention time difference of the same metabolite is within 0.2 min, it is considered to be the same metabolite, ensuring that the same metabolite is not misclassified as different metabolites when comparing samples in the future.

[0044] Next, background subtraction of blank samples is performed to eliminate interference signals not originating from meconium samples.

[0045] Next, metabolite annotation was performed.

[0046] After three cleaning steps—peak extraction, retention time alignment, and blank sample background subtraction—real metabolite signal peaks from meconium samples have been obtained. However, it is still unknown which specific metabolite these peaks correspond to. Annotation is the solution to this problem, essentially labeling unknown metabolite signals with metabolite names.

[0047] For example, the metabolite data obtained after the three-step cleaning process can be compared with the chromatograms of substances in the NIST 2020 mass spectrometry library. If the similarity is greater than a threshold (e.g., >80%), it is preliminarily determined that a peak in the metabolite data corresponds to that substance. Then, retention index verification is performed. The retention index (a standardized value reflecting retention time) of each metabolite on a specific chromatographic column is fixed. By verifying whether the detected retention index is consistent with the retention index of substances in the mass spectrometry library, the identity of the metabolite is further confirmed, avoiding misjudgments of similar chromatograms that are actually different substances.

[0048] Finally, a data matrix containing the peak area of ​​the detected metabolites is obtained, which characterizes the concentration of the metabolites. Thus, the concentration of fetal marker metabolites in the meconium sample of the fetus to be tested can be obtained.

[0049] For example, in the embodiments of this application, 211 metabolites were identified in a neonatal meconium sample. Thus, the data matrix of metabolite peak areas can contain 211 columns, each column representing a metabolite. The metabolite peak area is the area of ​​the mass spectrometry signal peak, and its value can characterize the concentration of the metabolite.

[0050] The peak area of ​​the metabolite corresponding to the fetal marker metabolite in the meconium sample of the fetus to be tested can be found from the data matrix containing the peak areas of all detected metabolites, and this peak area can be used as the concentration of the fetal marker metabolite in the meconium sample of the fetus to be tested.

[0051] Step 140: Detect the meconium sample of the pretreated fetus to obtain the concentration of fetal marker metabolites in the meconium sample of the fetus to be tested.

[0052] Step 140 is similar to step 130, except that the object of the test is a pre-processed meconium sample of the fetus to be tested, which will not be described in detail here.

[0053] It should be noted that in the embodiments of this application, the marker metabolites of the larger fetus and the marker metabolites of the smaller fetus are different. The inventors of this application have discovered that the different pathological mechanisms of brain injury in sFGR (smalformed fetal growth restriction) fetuses are: the smaller fetus is mainly affected by "citric acid cycle disorder leading to energy deficiency," while the larger fetus is mainly affected by "abnormal amino acid metabolism leading to cellular stress." It is difficult to find a unified meconium marker metabolite to accurately predict brain injury; only by detecting specific metabolites separately can the individual brain injury risks be accurately reflected. Using differentiated meconium metabolite indicators to predict brain injury in fetuses of different sizes can accurately match the different pathological mechanisms of brain injury in both, thereby accurately predicting the probability of severe brain injury in both fetuses.

[0054] Step 150: Based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined, predict the probability of severe brain injury in the fetus.

[0055] For example, the concentration of fetal biomarker metabolites can be substituted into a preset mapping relationship between fetal biomarker metabolite concentration and fetal severe brain injury probability to directly obtain the corresponding fetal severe brain injury probability; alternatively, the concentration of fetal biomarker metabolites can be input into a pre-trained fetal-specific prediction model (hereinafter referred to as the first prediction model), and the model can calculate and output the fetal severe brain injury probability.

[0056] For example, in step 150, the concentrations of fetal marker metabolites in the meconium sample of the fetus to be tested can first be transformed using Log10 and standardized using Z-score, and then substituted into the mapping relationship or input into the prediction model. Log10 transformation can narrow the data span, for example, converting the concentration range of 1-1000 μmol / L into a logarithmic range of 0-3. Z-score standardization can make the mean of each metabolite 0 and the standard deviation 1, thereby eliminating the model weight bias caused by the difference in concentration magnitude.

[0057] Step 160: Based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined, predict the probability of severe brain injury in the fetus.

[0058] For example, the concentration of fetal marker metabolites can be used to establish a pre-defined mapping relationship between the concentration of fetal marker metabolites and the probability of severe brain injury in fetuses, directly obtaining the probability of severe brain injury in fetuses. Alternatively, the concentration of fetal marker metabolites can be input into a pre-trained fetal-specific prediction model (hereinafter referred to as the second prediction model), which will calculate and output the probability of severe brain injury in fetuses.

[0059] For example, after performing steps 150 and 160, thresholds for severe brain injury in the larger fetus and the smaller fetus can be set. If the probability of severe brain injury in the larger fetus predicted in step 150 is greater than the threshold for severe brain injury in the larger fetus, it is determined that severe brain injury exists; otherwise, it is determined that severe brain injury does not exist. Similarly, if the probability of severe brain injury in the smaller fetus predicted in step 160 is greater than the threshold for severe brain injury in the smaller fetus, it is determined that severe brain injury exists; otherwise, it is determined that severe brain injury does not exist.

[0060] This application uses meconium as a sample, leveraging its ability to record the intrauterine metabolic state of the fetus throughout pregnancy. It enables immediate early warning of severe brain injury risk in newborns via metabolic biomarkers, overcoming the limitations of traditional transcranial ultrasound and MRI, which require continuous dynamic monitoring to determine outcomes (transcranial ultrasound has a detection rate of <30% within 72 hours of birth) and cannot provide immediate prediction. This provides a critical time window for clinical intervention. Furthermore, considering the differences in brain injury mechanisms between fetuses of different sizes in sFGR twin pregnancies, different specific biomarkers are screened for both, and a classification and prediction model is established, significantly improving the prediction accuracy and success rate of each. This provides reliable decision support for early clinical management.

[0061] In some embodiments, the large fetal marker metabolites include at least one of histidine, trans-4-hydroxyproline, pyruvate, and 1-phenylethanol; the small fetal marker metabolites include at least one of hippuric acid, nicotinamide, succinic acid, and citrate.

[0062] For example, large fetal marker metabolites include histidine, trans-4-hydroxyproline, pyruvate, and 1-phenylethanol; or, large fetal marker metabolites include histidine and trans-4-hydroxyproline; for example, small fetal marker metabolites include hippuric acid, nicotinamide, succinic acid, and citrate, or, small fetal marker metabolites include hippuric acid and nicotinamide.

[0063] Through extensive experiments and calculations, the inventors of this application screened numerous non-targeted metabolites contained in meconium samples to identify marker metabolites that can predict the probability of severe brain injury. Using one or more of these marker metabolites in combination can accurately predict the probability of severe brain injury in twins with selective fetal growth restriction. The specific screening process will be described later.

[0064] In some embodiments, step 150 includes: predicting the probability of severe brain injury in the large fetus based on the concentration of fetal marker metabolites in the meconium sample of the large fetus to be examined using a first prediction model; step 160 includes: predicting the probability of severe brain injury in the small fetus based on the concentration of fetal marker metabolites in the meconium sample of the small fetus to be examined using a second prediction model. The model parameters of the first prediction model and the second prediction model are different.

[0065] For example, the first and second prediction models can be machine learning models such as logistic regression, random forest, or extreme gradient boosting (XG Boost). They can also be the same model with different parameters.

[0066] The first and second prediction models were designed separately for the large and small fetuses to achieve accurate prediction. The core of this approach is based on the differences in the pathological mechanisms of brain injury in sFGR twins: brain injury in sFGR-L (large fetus) is mostly related to abnormal amino acid metabolism. The parameters of the first prediction model (such as the weight coefficients of each amino acid metabolism marker) are trained around this mechanism and can accurately capture the risk signals related to amino acid metabolism and cellular stress. On the other hand, brain injury in sFGR-S (small fetus) is centered on energy metabolism disorders (such as the citric acid cycle disorder mediated by hippuric acid and succinic acid). The parameters of the second prediction model (such as the weights of energy metabolism markers and constant terms) are matched to this pathological logic.

[0067] This design of using different parameters to predict fetal size avoids the limitations of traditional single models that use uniform parameters to predict fetal size. On the other hand, different model parameters can maximize the correlation between their respective markers and brain injury, ultimately making the prediction of the probability of severe brain injury in fetuses of different sizes more accurate and providing a reliable basis for developing differentiated intervention plans for fetuses of different sizes in clinical practice.

[0068] Figure 2 This is a schematic diagram illustrating a method for predicting the probability of severe brain injury in newborns according to an embodiment of this application. Meconium samples are collected from newborns with severe fetal brain injury (sFGR) upon their first expulsion after birth. After preprocessing, the samples are analyzed using an Agilent 7890 GC-5975 system to obtain the concentrations of histidine and trans-4-hydroxyproline, marker metabolites of the larger fetus, and nicotinamide and hippuric acid, marker metabolites of the smaller fetus. The concentrations of histidine and trans-4-hydroxyproline are input into a first prediction model, which is used to predict the probability of severe brain injury in the larger fetus of sFGR. The concentrations of nicotinamide and hippuric acid are input into a second prediction model, which is used to predict the probability of severe brain injury in the smaller fetus of sFGR.

[0069] The following describes how biomarker metabolites for large and small fetuses are screened.

[0070] In one example, the filtering process includes steps 210-320.

[0071] Step 210: Obtain meconium samples from G-group MCDA twin newborns as modeling samples. The modeling samples have brain injury labels, including severe brain injury labels and non-severe brain injury labels. G-group MCDA twin newborns include twin newborns with selective fetal growth restriction. Each group of modeling samples includes one meconium sample from the larger fetus and one meconium sample from the smaller fetus.

[0072] G can be an integer greater than or equal to 20, and the number of twin newborns in sFGR is greater than or equal to 20.

[0073] For example, group G MCDA twin newborns can be all sFGR twin newborns, or partly sFGR twin newborns and partly MCDA-C twin newborns (where the MCDA-C larger fetus is denoted as MCDA-L, and the MCDA-C smaller fetus is denoted as MCDA-S). Each modeling sample is labeled with a severe brain injury tag (1) and a non-severe brain injury tag (0), where non-severe brain injury includes mild brain injury and no brain injury. Brain injury tags can be determined by cranial ultrasound or cranial MRI. Each modeling sample group includes one meconium sample from the modeling larger fetus and one meconium sample from the modeling smaller fetus.

[0074] Step 220: For each modeling sample in the G group of modeling samples, Preprocess the modeling sample; The concentration of non-target metabolites in the preprocessed modeling sample was obtained by detecting the sample.

[0075] This step is similar to steps 120, 130, and 140, and will not be described again.

[0076] Step 230: Perform partial least squares discriminant analysis on non-target metabolites in the meconium samples of the fetuses in the G group modeling sample. Non-target metabolites with variable importance projection (VIP) greater than or equal to the first threshold are used as primary screening metabolites for fetuses. The variable importance projection of non-target metabolites greater than or equal to the first threshold indicates that they play a role in distinguishing between fetuses with severe brain injury and fetuses in the control group.

[0077] Understandably, when all twins in group G with MCDA are sFGR twins, the control group for the larger fetus is the sFGR-L non-severe brain injury group (including sFGR-L with mild brain injury and sFGR-L without brain injury). When some twins in group G with MCDA are sFGR twins and some are MCDA-C, the control group for the larger fetus is the sFGR-L non-severe brain injury group and all MCDA-C twins. That is, twins in the MCDA-C group, regardless of size, are considered as the control group for the larger fetus. Similarly, when all twins in group G with MCDA are sFGR twins, the control group for the smaller fetus is the sFGR-S non-severe brain injury group (including sFGR-S with mild brain injury and sFGR-S without brain injury). When some twins in group G with MCDA are sFGR twins and some are MCDA-C, the control group for the smaller fetus is the sFGR-S non-severe brain injury group and all MCDA-C twins. In other words, twins in the MCDA-C group, regardless of their size, were used as the control group for the smaller fetus.

[0078] For example, all non-targeted metabolites can be extracted from the meconium of the modeling fetus in the G group modeling sample, and PLSDA analysis can be performed using the MetaboAnalyst 6.0 platform. For example, the metabolite concentration can be first transformed by Log10 and standardized by Z-score, and then a PLSDA model can be constructed to calculate the variable importance projection (VIP) value of each metabolite. The first threshold is set to 1, and metabolites with VIP≥1 are defined as the initial screening metabolites of the fetus.

[0079] For example, the G group modeling sample includes 20 pairs of sFGR twins and 13 pairs of MCDA-C twins. In step 230, based on brain injury labels, the large fetal severe brain injury group (including sFGR-L moderate to severe brain injury samples) and the large fetal control group (including sFGR-L moderate to non-severe brain injury samples + MCDA-C samples) are identified from the G group modeling sample. The PLSDA analysis in step 230 uses the large fetal severe brain injury group and the large fetal control group as input as the overall dataset and calculates the VIP value of all non-targeted metabolites. Its core is to extract the metabolites that can best distinguish the large fetal severe brain injury group vs. the large fetal control group. The VIP value reflects the contribution of the metabolite to the group separation. A VIP ≥ 1 means that the metabolite has a significant effect on distinguishing severe brain injury from non-severe brain injury / normal, and the metabolite can be used as a primary screening metabolite.

[0080] This application does not construct a method for predicting the probability of mild brain injury because the inventors, through extensive research and experiments, discovered that PLSDA results show no significant separation between the metabolic profiles of the mild brain injury group and the normal group (no injury), while only the metabolic profiles of the severe brain injury group and the other two groups (mild / normal) show significant separation. The inventors believe this phenomenon is due to the extremely mild metabolic disturbances in mild brain injury, which do not form distinguishable metabolic characteristics. Mild brain injury (such as grade I-II IVH, grade I PVL) is mostly temporary and reversible, with minimal impact on fetal intrauterine metabolism, causing only minor fluctuations in a small amount of metabolites; while the metabolism of uninjured fetuses is in a normal homeostasis state, and the difference in metabolite concentration between the two does not reach the separation threshold that can be captured by PLSDA, therefore, PLSDA cannot effectively distinguish them.

[0081] Furthermore, from a clinical perspective, prioritizing the prediction of severe brain injury is more meaningful. Severe brain injury (Grades III-IV IVH, Grade II-IV PVL) is the core cause of long-term neurodevelopmental disorders (such as cerebral palsy and cognitive impairment) in newborns with sFGR, and the damage is irreversible, requiring urgent clinical intervention; while mild brain injury can often recover spontaneously, with a low risk of adverse long-term outcomes. Focusing on severe brain injury can provide more accurate early warning for high-risk cases requiring urgent intervention, meeting the clinical need to prioritize the treatment of severe risks. Step 240: Perform univariate analysis on the primary screening metabolites of large fetuses to screen out the secondary screening metabolites of large fetuses that showed significant differences between the large fetuses with severe brain injury group and the large fetuses control group.

[0082] For example, for the initial screening metabolites obtained in step 230, the Shapiro-Wilk test can be used to determine the distribution type of metabolite concentration. If it is normally distributed, the concentration difference between the large fetus with severe brain injury group and the large fetus control group can be compared using an independent samples t-test; if it is not normally distributed, the Mann-Whitney U test can be used; finally, metabolites that simultaneously meet the criteria of fold change (FC) ≥1.5 or ≤0.67 and statistical significance P<0.05 are selected as the second screening metabolites for large fetuses.

[0083] Understandably, PLSDA is a multivariate analysis, which may result in multiple metabolites contributing synergistically but individual metabolite differences not being significant. Univariate analysis can further verify the independent differences of individual metabolites between the two groups. Through rigorous statistical tests, metabolites with high contribution in PLSDA but no significant differences between the actual groups are eliminated, ensuring that the metabolites screened in the second screening are truly differentiating metabolites between the two groups.

[0084] Step 250: Based on the metabolites from the second screening of large fetuses and brain injury labels, construct at least one first candidate prediction model to predict the probability of severe brain injury in large fetuses; each of the at least one first candidate prediction models corresponds to a subset of the metabolites from the second screening of large fetuses.

[0085] In step 250, one or more component subsets can be selected from the large fetal second screening metabolites screened in step 240, and different candidate prediction models can be obtained using metabolites in different subsets.

[0086] For example, there are 15 distinct subsets of the four major fetal screening metabolites. Therefore, when there are four major fetal screening metabolites, a maximum of 15 candidate models can be obtained, with each candidate model corresponding to a subset. For instance, histidine and trans-4-hydroxyproline are selected as subsets from the four major fetal screening metabolites to obtain candidate prediction model A; histidine, trans-4-hydroxyproline, pyruvate, and 1-phenylethanol are selected as subsets from the four major fetal screening metabolites to obtain candidate prediction model B.

[0087] Different subsets of metabolites have varying predictive capabilities. Constructing multiple candidate models allows for experimentation with various combinations, such as single biomarker and multi-biomarker combinations, providing a foundation for selecting the optimal prediction scheme and avoiding insufficient predictive performance due to using only a single subset.

[0088] Step 260: Evaluate the predictive power of at least one first candidate prediction model.

[0089] For example, for each first candidate prediction model, the metabolite concentration of a subset of the modeling samples corresponding to that first candidate prediction model can be input into the model to obtain the predicted probability value of severe brain injury for each sample; based on the predicted value, a receiver operating characteristic (ROC) curve is plotted, and the area under the curve (AUC) is calculated. The closer the AUC is to 1, the stronger the model's predictive ability; at the same time, it can be used to assist in the evaluation of indicators such as sensitivity, true positive rate, and specificity.

[0090] In this way, the performance of different candidate models can be compared using objective quantitative indicators.

[0091] Step 270: The first candidate prediction model that meets the prediction ability requirement is taken as the first prediction model, and the non-targeted metabolites in the subset of large fetal second screening metabolites corresponding to the first candidate prediction model that meets the prediction ability requirement are taken as large fetal marker metabolites.

[0092] Understandably, the threshold for severe brain injury in large fetuses that yields the best predictive index for the first predictive model can also be determined based on the ROC curve.

[0093] For example, when the modeling sample consists of 20 pairs of sFGR twins and 13 pairs of MCDA-C twins, in step 230, from the meconium non-targeted metabolites of the modeling large fetal samples (a total of 46, including 20 sFGR-L large fetuses and 26 MCDA-C control group fetuses of the same weight), partial least squares discriminant analysis (PLSDA) is used to screen out 73 metabolites with a VIP≥1 (first threshold). In step 240, univariate analysis is performed on the 73 initial screening metabolites to finally screen out the large fetal secondary screening metabolites, the core of which include histidine, trans-4-hydroxyproline, pyruvate, 1-phenylethanol, etc. In step 250, different subsets are selected from the secondary screening metabolites to construct the first candidate prediction model. In steps 260 and 270, after assessing the predictive ability through receiver operating characteristic (ROC) curves, the subset of metabolites corresponding to the model with the largest AUC (predictive ability meets the requirements) is defined as the large fetal marker metabolites, namely histidine, trans-4-hydroxyproline, pyruvate and 1-phenylethanol.

[0094] For example, when the modeling sample is 20 pairs of sFGR twins, the large fetal marker metabolites are identified as histidine and trans-4-hydroxyproline through steps 230 to 270.

[0095] Step 280: Perform partial least squares discriminant analysis on non-target metabolites in the meconium samples of the fetuses in the G group modeling samples. Non-target metabolites with variable importance projection greater than or equal to the second threshold are used as the initial screening metabolites of the fetuses. The variable importance projection of non-target metabolites greater than or equal to the second threshold indicates that they play a role in distinguishing between the fetuses with severe brain injury group and the fetuses control group. Step 290: Perform univariate analysis on the primary screening metabolites of small fetuses to screen out the secondary screening metabolites of small fetuses that showed significant differences between the small fetuses with severe brain injury group and the small fetuses control group. Step 300: Based on the metabolites from the second screening of small fetuses and brain injury labels, construct at least one second candidate prediction model to predict the probability of severe brain injury in small fetuses; each of the at least one second candidate prediction model corresponds to a subset of the metabolites from the second screening of small fetuses; Step 310: Evaluate the predictive power of at least one second candidate prediction model; Step 320: Select the second candidate prediction model that meets the prediction ability requirement as the second prediction model, and select the non-targeted metabolites in the subset of small fetal second screening metabolites corresponding to the second candidate prediction model that meets the prediction ability requirement as small fetal marker metabolites.

[0096] For an explanation of steps 280-320, please refer to the description of steps 230-270, which will not be repeated here.

[0097] For example, when the modeling sample consists of 20 pairs of sFGR twins and 13 pairs of MCDA-C twins, in step 280, after performing PLSDA analysis on the meconium non-targeted metabolites of the modeling fetuses (a total of 46, including 20 sFGR-S and 26 MCDA-C control fetuses), 66 non-targeted metabolites with VIP≥1 were screened out. In step 290, univariate analysis was performed on the 66 primary screening metabolites, and the secondary screening metabolites of the fetuses included hippuric acid, nicotinamide, succinic acid, citrate, and pyruvic acid. In step 300, different subsets were selected from the secondary screening metabolites to construct a second candidate predictive model, and after evaluating the predictive ability through ROC curves in step 310, in step 320, the subset of metabolites corresponding to the model with the largest AUC (predictive ability meets the requirements) was designated as the fetal marker metabolites. The fetal marker metabolites are hippuric acid, nicotinamide, succinic acid, and citrate.

[0098] For example, when the modeling sample is 20 pairs of sFGR twins, the fetal marker metabolites are identified as hippuric acid and nicotinamide through steps 280 to 320.

[0099] It should be noted that different control groups were used in different embodiments of this application: Control Group Scenario 1: The control group consisted of sFGR non-severe brain injury samples; Control Group Scenario 2: The control group included both sFGR non-severe brain injury samples and MCDA non-severe brain injury samples. It is understood that MCDA-C samples are generally healthy, full-term newborns with a low complication rate. These newborns are not routinely subjected to cranial ultrasound according to clinical guidelines; cranial ultrasound is only performed after evaluation by a pediatrician if signs of brain injury are suspected. The MCDA-C samples used in the embodiments of this application are all without brain injury. For Scenario 1, the model can focus on the metabolic differences within sFGR twins. The baseline is a metabolic state unique to sFGR, and even if there is mild disturbance, it is consistent with the pathological background of the same population. The trained model is more targeted in the sFGR twin population and can more accurately capture the metabolic signals unique to severe brain injury in this population, avoiding interference from healthy metabolism in normal MCDA samples. This is suitable for clinical scenarios where only the risk level of sFGR twins needs to be assessed. For scenario 2, the model's metabolic baseline is broader (covering the metabolic status of undamaged sFGR and normal MCDA twins). The trained model can not only distinguish the risk of damage within sFGR, but also compare the degree of metabolic deviation between sFGR and normal twins. It has stronger generalization ability, a wider range of clinical applications, and can more intuitively judge the severity of sFGR metabolic abnormalities by comparing with the normal baseline, providing a more comprehensive reference for intervention decisions.

[0100] Figure 3 This illustrates a specific marker metabolite screening process in control group scenario 2. The modeling samples consisted of 20 pairs of sFGR twins and 13 pairs of MCDA-C twins. After birth, meconium samples were collected for metabolomics analysis to obtain the concentrations of different metabolites in each modeling sample. Simultaneously, cranial ultrasound and the ASQ neurodevelopmental questionnaire were performed within one month of birth to obtain brain injury labels for the modeling samples. Subsequently, correlation analysis was conducted between the concentrations of different metabolites in each modeling sample and the brain injury labels. Specifically, the correlation analysis included: screening for secondary screening metabolites using PLSDA and univariate analysis; constructing candidate models based on the concentrations of secondary screening metabolites and brain injury labels to predict the probability of severe fetal brain injury; evaluating the predictive ability of at least one candidate model; and selecting the candidate model whose predictive ability met the requirements as the final prediction model.

[0101] This application embodiment preprocesses the modeling samples and detects non-targeted metabolites. Differential metabolites are screened by PLSDA primary screening and univariate analysis secondary screening for fetuses of different sizes. Then, the optimal prediction model and specific marker metabolites are determined through candidate model construction and evaluation. This ensures that the modeling samples are relevant to the clinical high-risk group and the data is reliable. Furthermore, by subgrouping, metabolites and models that match the pathological differences in brain injury between fetuses of different sizes are accurately screened. This ensures that the first and second prediction models are accurate and reproducible, providing a practical tool for predicting severe brain injury in sFGR twin newborns in the clinic.

[0102] In some embodiments, step 240 includes: for fetal primary screening metabolites with normally distributed concentrations, verifying whether there is a significant difference between fetal primary screening metabolites and fetal control groups using an independent samples t-test, and using fetal primary screening metabolites with significant differences as fetal secondary screening metabolites; for fetal primary screening metabolites with non-normally distributed concentrations, verifying whether there is a significant difference between fetal primary screening metabolites and fetal control groups using a Mann-Whitney U test, and using fetal primary screening metabolites with significant differences as fetal secondary screening metabolites.

[0103] Step 290 includes: for fetal primary screening metabolites with normally distributed concentrations, an independent samples t-test is used to verify whether there is a significant difference between the fetal primary screening metabolites in the fetal severe brain injury group and the fetal control group, and the fetal primary screening metabolites with significant differences are used as fetal secondary screening metabolites; for fetal primary screening metabolites with non-normally distributed concentrations, a Mann-Whitney U test is used to verify whether there is a significant difference between the fetal primary screening metabolites in the fetal severe brain injury group and the fetal control group, and the fetal primary screening metabolites with significant differences are used as fetal secondary screening metabolites.

[0104] This application's embodiments employ a univariate analysis design that first determines the type of metabolite concentration distribution and then matches it with corresponding statistical tests, providing precise and rigorous technical support for screening metabolites in the second screening of large and small fetuses. Different metabolites exhibit fundamentally different concentration distributions (e.g., histidine concentration in large fetuses shows a normal distribution while 1-phenylethanol shows a non-normal distribution, and succinic acid in small fetuses shows a non-normal distribution while nicotinamide shows a normal distribution). If a single test method is used uniformly (e.g., t-tests for all), it can lead to biases in judging inter-group differences of non-normally distributed metabolites (e.g., misjudging significant differences or missing true differences). However, this application's embodiments employ independent samples t-tests (adapted to normal distributions, accurately calculating the statistical significance of mean differences) and Mann-Whitney U tests (adapted to non-normal distributions, avoiding extreme value interference based on rank analysis) according to the distribution type, accurately reflecting the concentration differences of metabolites in the severe brain injury group versus the control group, providing a reliable basis for the construction of subsequent candidate models.

[0105] In some embodiments, step 250 includes steps 2501-2503.

[0106] Step 2501: Select at least a portion of the fetal secondary screening metabolites from the large fetal screening metabolites as the first target subset. There are M types of large fetal secondary screening metabolites, and the first target subset contains m types of large fetal secondary screening metabolites; m M.

[0107] For example, if M is 4, then m can be 1, 2, 3, or 4, resulting in a maximum of 15 different first target subsets.

[0108] Step 2502: Define the logistic regression function corresponding to the first candidate prediction model as: P(Y=1|X)= z = WX + w0; Where X = [x1, ..., x] m ],x1, …, x mThe concentrations of m types of fetal secondary screening metabolites in the first target subset; W=[w1, …, w m [] represents the weight coefficient matrix corresponding to X; w0 is a constant term; Step 2503: Based on the concentrations of m types of fetal secondary screening metabolites in the G-group modeling samples and the brain injury state labels of the G-group modeling samples, solve for W and w0 to obtain the first candidate prediction model corresponding to the first target subset. For example, W and w0 can be solved using the maximum likelihood estimation method.

[0109] Step 300 includes steps 3001-3003.

[0110] Step 3001: Select at least a portion of the fetal second screening metabolites from the fetal second screening metabolites as the second target subset. There are N types of fetal second screening metabolites in total, and the second target subset contains n types of fetal second screening metabolites; n N; Step 3002: Define the logistic regression function corresponding to the second candidate prediction model as: P(Y=1|X)= z = WX + w0; Where X=[x1, ..., x n ],x1, …, x n The concentrations of n different fetal screening metabolites in the second target subset; W=[w1, …w n [] represents the weight coefficient matrix corresponding to X; w0 is a constant term; Step 3003: Based on the concentrations of n fetal second screening metabolites in the G-group modeling samples and the brain injury state labels of the G-group modeling samples, solve for W and w0 to obtain the second candidate prediction model corresponding to the second target subset. For example, W and w0 can be solved using the maximum likelihood estimation method.

[0111] For example, in control group scenario 1, in step 240, the large fetal second screening metabolites histidine and trans-4-hydroxyproline are obtained. In step 2501, a subset containing histidine and trans-4-hydroxyproline is selected as the first target subset. In step 2502, X=[x1, x2] is set as the independent variable matrix, and W=[w1, w2] is set as the weight coefficient matrix, then the linear function z=WX+w0 is obtained. Further, the probability function P(Y=1|X)= In step 2503, the weighting coefficient matrix W and constant term w0 are determined based on the histidine and trans-4-hydroxyproline concentrations in the 20 sFGR-L meconium samples from the modeling sample. The weighting coefficient matrix W for the larger fetus is [2.135, -3.216]; w0 is -1.430. Thus, the first candidate prediction model P(Y=1|X) corresponding to the first target subset is obtained. z = WX + w0; where W = [2.135, -3.216] and w0 = -1.430.

[0112] Similarly, for control group scenario 1, in step 290, the fetal second screening metabolites hippuric acid and nicotinamide are obtained. In step 3001, a subset containing hippuric acid and nicotinamide is selected as the second target subset. Through steps 3003 and 3004, the weight coefficient matrix W = [8.070, -2.690]; w0 = -78.602 for the fetus is obtained. Thus, the second candidate prediction model P(Y=1|X) corresponding to the second target subset is obtained. z = WX + w0; where W = [8.070, -2.690]; w0 = -78.602.

[0113] In some embodiments, step 260 includes: for each first candidate prediction model, plotting a receiver operating characteristic (ROC) curve based on the probability prediction values ​​generated by the first candidate prediction model, and determining at least one of the following indicators based on the ROC curve: area under the ROC curve, specificity, and sensitivity; step 310 includes: for each second candidate prediction model, plotting a ROC curve based on the probability prediction values ​​generated by the second candidate prediction model, and determining at least one of the following indicators based on the ROC curve: area under the ROC curve, specificity, and sensitivity.

[0114] For example, after plotting the receiver operating characteristic (ROC) curve, the area under the curve, as well as indicators such as specificity and sensitivity, can be calculated to evaluate the predictive ability of the candidate model.

[0115] Continuing the previous example. In step 260, the concentrations of histidine and trans-4-hydroxyproline in the 20 sFGR-L modeling samples were substituted into the first candidate prediction model to obtain 20 predicted probabilities of severe brain injury. A ROC object was constructed using the "roc" and "ggroc" functions of the R package "pROC", the area under the ROC curve was calculated, and the ROC curve was plotted. The area under the ROC curve (AUC) for large fetuses was 0.922, with a 95% CI of 0.786⁻¹ (DeLong). In step 310, the concentrations of hippuric acid and nicotinamide in the 20 sFGR-S modeling samples were substituted into the first candidate prediction model to obtain 20 predicted probabilities of severe brain injury. Receiver operating characteristic (ROC) curves were plotted in a similar manner, and the AUC, specificity, and sensitivity of the ROC curve were determined. The AUC for small fetuses was 0.952, with a 95% CI of 0.866⁻¹ (DeLong).

[0116] For details of the ROC curves for large and small fetuses, see [link to ROC curves]. Figure 4 and Figure 5 . Figure 4 and Figure 5 The dots in the text represent the optimal cutoff value, and the parentheses after the cutoff value represent the true positive rate (TPR), which is also known as sensitivity, and the false positive rate (FPR), which is also known as specificity.

[0117] The 95% CI (Confidence Interval) represents the 95% confidence interval, which is the range of uncertainty in the AUC estimate. It means that based on the sample data, there is a 95% chance that the estimated true value of the AUC will fall within this interval. Generally, the narrower the confidence interval, the higher the accuracy of the AUC estimate. AUC represents the area under the ROC curve and the coordinate axis.

[0118] The DeLong method is a commonly used nonparametric approach for estimating the 95% confidence interval of the AUC. This method does not rely on assumptions about the data distribution; it generates multiple AUC values ​​through bootstrap sampling and then calculates the confidence interval of the AUC based on the distribution of these values.

[0119] Based on the Youden index (the sum of sensitivity and specificity minus 1), the optimal cutoff value for the first candidate prediction model was determined to be 0.154. Prediction and classification were performed using this optimal cutoff value, resulting in a sensitivity of 100% and a specificity of 82.4% for the correct classification of the dependent variable. Conversely, the optimal cutoff value for the second candidate prediction model was 0.192. Prediction and classification were performed using this optimal cutoff value, resulting in a sensitivity of 100% and a specificity of 85.7% for the correct classification of the dependent variable.

[0120] There can be one or more first-candidate prediction models and second-candidate prediction models. For example, when there are four metabolites in the second screening of a large fetus, there can be no more than 15 first-candidate prediction models and no more than 1 first-candidate prediction model, with each first-candidate prediction model corresponding to a first target subset. For example, when there are four metabolites A, B, C, and D in the second screening of a large fetus, the first target subset can be {A}, {B}, {C}, {D}, {A,B}, {A,C}, {A,D}, {B,C}, {B,D}, {C,D}, {A,B,C,D}, {A,B,C,D}. When there are two metabolites in the second screening of a large fetus, there can be no more than 3 second-candidate prediction models and no more than 1 second-candidate prediction model, with each second-candidate prediction model corresponding to a second target subset.

[0121] In step 270, the first candidate prediction model that meets the prediction capability requirement is selected as the first prediction model, and the non-targeted metabolites in the subset of large fetal secondary screening metabolites corresponding to the first candidate prediction model that meets the prediction capability requirement are selected as large fetal marker metabolites. In step 320, the second candidate prediction model that meets the prediction capability requirement is selected as the second prediction model, and the non-targeted metabolites in the subset of small fetal secondary screening metabolites corresponding to the second candidate prediction model that meets the prediction capability requirement are selected as small fetal marker metabolites.

[0122] For example, the predictive ability meeting the requirements can be that the indicator meets the threshold requirement, the indicator is the highest, the indicator is the highest that meets the threshold requirement, one indicator meets the threshold requirement and another indicator is the highest, etc.

[0123] For example, the first candidate prediction model with the largest area under the curve (AUC) can be selected as the first prediction model, or the first candidate prediction model with an AUC greater than a threshold and the best specificity and sensitivity can be selected as the first prediction model. Similarly, the second candidate prediction model with the largest AUC can be selected as the second prediction model, or the second candidate prediction model with an AUC greater than a threshold and the best specificity and sensitivity can be selected as the second prediction model. For instance, if the candidate prediction model has an AUC ≥ 0.85, and a sensitivity ≥ 85% and a specificity ≥ 80%, then its predictive ability is considered to meet the requirements.

[0124] Continuing with the previous example, there is only one first candidate predictive model and one second candidate predictive model. The first candidate predictive model has an AUC of 0.922, with a sensitivity of 100% and a specificity of 82.4% for correctly classifying the dependent variable. The second candidate predictive model has an AUC of 0.952, with a sensitivity of 100% and a specificity of 85.7% for correctly classifying the dependent variable. The predictive ability requirement is an AUC greater than 0.8. Therefore, the first candidate predictive model meets the requirement and can be used as the first predictive model to identify histidine and trans-4-hydroxyproline as the first marker metabolites. The second candidate predictive model meets the requirement and can be used as the second predictive model to identify nicotinamide and hippuric acid as the second marker metabolites.

[0125] For example, in control group scenario 2 (20 sFGR modeling samples plus 26 MCDA-C modeling samples), there are 5 first candidate prediction models, corresponding to the following first target subsets: {histidine}, {trans-4-hydroxyproline}, {pyruvate}, {1-phenylethanol}, and {histidine, trans-4-hydroxyproline, pyruvate, 1-phenylethanol}. The metrics of the 5 first candidate models are as follows: Histidine: AUC = 0.922 (95% CI: 0.765–1) trans-4-hydroxyproline: AUC = 0.902 (95% CI: 0.745–1) Pyruvate: AUC = 0.882 (95% CI: 0.706–1) 1-Phenylephrine: AUC = 0.882 (95% CI: 0.706–1) Histidine, trans-4-hydroxyproline, pyruvate, and 1-phenylethanol: AUC = 0.947 (95% CI: 0.748–1).

[0126] The first candidate prediction model for {histidine, trans-4-hydroxyproline, pyruvate, 1-phenylethanol} has the best AUC and is therefore selected as the first prediction model. Histidine, trans-4-hydroxyproline, pyruvate, and 1-phenylethanol are identified as the first marker metabolites.

[0127] For control group scenario 2 (20 sFGR modeling samples plus 26 MCDA-C modeling samples), there are 5 second candidate prediction models, corresponding to the following second target subsets: {nicotinamide}, {hippuric acid}, {citric acid}, {succinic acid}, and {nicotinamide, hippuric acid, citric acid, succinic acid}. The metrics of the 5 second candidate models are as follows: Nicotinamide: AUC = 0.833 (95% CI: 0.571–1) Hippuric acid: AUC = 0.821 (95% CI: 0.607–0.976) Citric acid: AUC=0.81 (95% CI: 0.548–0.976) Succinic acid: AUC=0.81 (95% CI: 0.571–0.952) Four types of nicotinamide, hippuric acid, citric acid, and succinic acid: AUC=0.888 (95% CI: 0.611–1).

[0128] The second candidate prediction model for {nicotinamide, hippuric acid, citric acid, and succinic acid} had the best AUC and was therefore selected as the second prediction model. Nicotinamide, hippuric acid, citric acid, and succinic acid were then identified as the second marker metabolites.

[0129] In some embodiments, method 10 further includes: Step 330: Import the metabolites from the second screening of large fetuses into the KEGG database for pathway enrichment analysis, determine the first enriched pathway corresponding to the metabolites from the second screening of large fetuses, and use the metabolites from the second screening of large fetuses related to the first enriched pathway as the metabolites from the second screening of large fetuses used to construct the first prediction model. Step 340: Import the fetal second screening metabolites into the KEGG database for pathway enrichment analysis, determine the second enriched pathway corresponding to the fetal second screening metabolites, and use the fetal second screening metabolites related to the second enriched pathway as the fetal second screening metabolites used to construct the second prediction model.

[0130] Step 330 can be executed after step 250 and before step 260; step 340 can be executed after step 300 and before step 310.

[0131] In this embodiment, although the second-screening metabolites met the requirement of significant differences between groups, they may include some indirectly differentially expressed metabolites that are not directly related to brain injury. Through KEGG (Kyoto Encyclopedia of Genes and Genomes) pathway enrichment analysis, pathways related to the pathological mechanisms of severe brain injury can be identified, such as the amino acid metabolism pathway in larger fetuses and the citric acid cycle pathway in smaller fetuses. Only the second-screening metabolites within these pathways are retained for modeling. This eliminates metabolites with significant differences but no pathological association, ensuring that the model relies on core metabolic signals directly related to the brain injury mechanism and avoiding statistically significant but biologically insignificant biases in model predictions caused by irrelevant metabolites.

[0132] This application discloses several embodiments for determining a first prediction model and a second prediction model. According to these embodiments, a first prediction model and a second prediction model can be constructed by modeling samples, thereby using the first prediction model and the second prediction model to predict the probability of severe brain injury in the fetus to be examined.

[0133] In another embodiment of this application, a system for predicting the probability of severe brain injury in newborns is provided. The system is used to predict the probability of severe brain injury in selectively fetal growth-restricted twin newborns, which include a larger fetus and a smaller fetus to be examined. The system includes: a sample acquisition module for acquiring meconium samples from the larger fetus and the smaller fetus; a preprocessing module for preprocessing the meconium samples from both the larger and smaller fetuses; and a metabolite concentration detection module for detecting the preprocessed meconium sample from the larger fetus to obtain the concentration of a large fetal marker metabolite in the meconium sample; and detecting the preprocessed meconium sample from the smaller fetus to obtain the concentration of a small fetal marker metabolite in the meconium sample. The probability prediction module is used to predict the probability of severe brain injury in the large fetus based on the concentration of fetal marker metabolites in the meconium sample of the large fetus to be examined, and to predict the probability of severe brain injury in the small fetus based on the concentration of fetal marker metabolites in the meconium sample of the small fetus to be examined; wherein the fetal marker metabolites of the large fetus and the small fetus are different.

[0134] In some embodiments, predicting the probability of severe brain injury in a large fetus based on the concentration of fetal marker metabolites in the meconium sample of the large fetus to be examined includes: predicting the probability of severe brain injury in a large fetus based on the concentration of fetal marker metabolites in the meconium sample of the large fetus to be examined using a first prediction model; predicting the probability of severe brain injury in a small fetus based on the concentration of fetal marker metabolites in the meconium sample of the small fetus to be examined includes: predicting the probability of severe brain injury in a small fetus based on the concentration of fetal marker metabolites in the meconium sample of the small fetus to be examined using a second prediction model; wherein the model parameters of the first prediction model and the second prediction model are different.

[0135] In some embodiments, the system further includes a marker metabolite screening module for obtaining meconium samples from G-group MCDA twin newborns as modeling samples. The modeling samples have brain injury labels, including severe brain injury labels and non-severe brain injury labels. G-group MCDA twin newborns include twin newborns with selective fetal growth restriction. Each group of modeling samples includes one meconium sample from the larger modeling fetus and one meconium sample from the smaller modeling fetus. For each modeling sample in the G-group modeling samples, the modeling sample is preprocessed. The preprocessed modeling sample is detected to obtain the concentration of non-target metabolites in the modeling sample. The meconium from the larger modeling fetus in the G-group modeling samples is then preprocessed. Partial least squares discriminant analysis was performed on non-targeted metabolites in the sample. Metabolites with a variable importance projection greater than or equal to a first threshold were used as primary screening metabolites for large fetuses. A variable importance projection greater than or equal to the first threshold indicates that the non-targeted metabolites are effective in distinguishing between the large fetuses with severe brain injury and the large fetus control group. Univariate analysis was performed on the primary screening metabolites to identify secondary screening metabolites that showed significant differences between the large fetuses with severe brain injury and the large fetus control group. Based on the secondary screening metabolites and brain injury labels, at least one first-candidate prediction model was constructed to predict the probability of severe brain injury in large fetuses. Each corresponding secondary screening metabolite in the at least one first-candidate prediction model... A subset of the metabolites; evaluate the predictive power of at least one first candidate predictive model; select the first candidate predictive model that meets the predictive power requirement as the first predictive model, and select the non-targeted metabolites in the subset of large fetal secondary screening metabolites corresponding to the first candidate predictive model that meets the predictive power requirement as large fetal marker metabolites; perform partial least squares discriminant analysis on the non-targeted metabolites in the meconium samples of small fetuses in the G group modeling sample, and select the non-targeted metabolites whose variable importance projection is greater than or equal to the second threshold as small fetal primary screening metabolites; the variable importance projection of non-targeted metabolites greater than or equal to the second threshold indicates that they play a role in distinguishing between the small fetal severe brain injury group and the small fetal control group; for small Univariate analysis was performed on fetal primary screening metabolites to identify fetal secondary screening metabolites that showed significant differences between the fetal severe brain injury group and the fetal control group. Based on fetal secondary screening metabolites and brain injury labels, at least one second candidate predictive model was constructed to predict the probability of fetal severe brain injury. Each of the at least one second candidate predictive model corresponds to a subset of fetal secondary screening metabolites. The predictive ability of the at least one second candidate predictive model was evaluated. The second candidate predictive model that meets the predictive ability requirement was selected as the second predictive model, and the non-targeted metabolites in the subset of fetal secondary screening metabolites corresponding to the second candidate predictive model that meets the predictive ability requirement were selected as fetal marker metabolites.

[0136] In some embodiments, the large fetal control group is the large fetal non-severe brain injury group, and the small fetal control group is the small fetal non-severe brain injury group; or, the G group MCDA twin newborns also include twin newborns of the same weight monochorionic diamniotic group; the large fetal control group consists of the large fetal non-severe brain injury group and the monochorionic diamniotic group of the same weight; the small fetal control group consists of the small fetal non-severe brain injury group and the monochorionic diamniotic group of the same weight.

[0137] In some embodiments, univariate analysis was performed on the primary screening metabolites of large fetuses to screen for secondary screening metabolites that showed significant differences between the large fetuses with severe brain injury and the large fetuses control group. This included: for primary screening metabolites with normally distributed concentrations, an independent samples t-test was used to verify whether there were significant differences between the primary screening metabolites of large fetuses with severe brain injury and the large fetuses control group, and the primary screening metabolites showing significant differences were identified as secondary screening metabolites of large fetuses; for primary screening metabolites with non-normally distributed concentrations, a Mann-Whitney U test was used. The U-test was used to verify whether there were significant differences in the primary screening metabolites of large fetuses between the large fetuses with severe brain injury and the large fetuses control group. Metabolites with significant differences in primary screening metabolites of large fetuses were used as secondary screening metabolites of large fetuses. Univariate analysis was performed on the primary screening metabolites of small fetuses to screen for secondary screening metabolites of small fetuses with significant differences between the small fetuses with severe brain injury and the small fetuses control group. This included: for primary screening metabolites of small fetuses with normally distributed concentrations, the independent samples t-test was used to verify whether there were significant differences in primary screening metabolites of small fetuses between the small fetuses with severe brain injury and the small fetuses control group. Metabolites with significant differences in primary screening metabolites of small fetuses with non-normally distributed concentrations were used to verify whether there were significant differences in primary screening metabolites of small fetuses between the small fetuses with severe brain injury and the small fetuses control group. Metabolites with significant differences in primary screening metabolites of small fetuses were used as secondary screening metabolites of small fetuses.

[0138] In some embodiments, based on fetal secondary screening metabolites and brain injury labels, at least one first candidate prediction model is constructed to predict the probability of severe brain injury in fetuses, including: selecting at least a portion of fetal secondary screening metabolites as a first target subset, wherein there are M types of fetal secondary screening metabolites, and the first target subset contains m types of fetal secondary screening metabolites; m M; Define the logistic regression function corresponding to the first candidate prediction model as: P(Y=1|X)= z = WX + w0; where X = [x1, ..., x] m ],x1, …, x m The concentrations of m types of fetal secondary screening metabolites in the first target subset; W = [w1, …, wm [x] represents the weight coefficient matrix corresponding to X; w0 is a constant term; based on the concentrations of m types of fetal second screening metabolites in the G-group modeling samples and the brain injury status labels of the G-group modeling samples, solve for W and w0 to obtain the first candidate prediction model corresponding to the first target subset; based on fetal second screening metabolites and brain injury labels, construct at least one second candidate prediction model to predict the probability of severe brain injury in fetuses, including: selecting at least a portion of fetal second screening metabolites as the second target subset, where there are N types of fetal second screening metabolites and the second target subset contains n types of fetal second screening metabolites; n N; Define the logistic regression function corresponding to the second candidate prediction model as: P(Y=1|X)= z = WX + w0; where X = [x1, ..., x n ],x1, …, x n The concentrations of n different fetal screening metabolites in the second target subset; W = [w1, …w n ] is the weight coefficient matrix corresponding to X; w0 is a constant term; based on the concentrations of n kinds of fetal second screening metabolites in the G group modeling samples and the brain injury status labels of the G group modeling samples, solve for W and w0 to obtain the second candidate prediction model corresponding to the second target subset.

[0139] In some embodiments, evaluating the predictive ability of at least one first candidate prediction model includes: for each first candidate prediction model, plotting a receiver operating characteristic (ROC) curve based on the probability prediction values ​​generated by the first candidate prediction model; and determining at least one of the following metrics based on the ROC curve: area under the ROC curve, specificity, and sensitivity. Evaluating the predictive ability of at least one second candidate prediction model includes: for each second candidate prediction model, plotting a ROC curve based on the probability prediction values ​​generated by the second candidate prediction model; and determining at least one of the following metrics based on the ROC curve: area under the ROC curve, specificity, and sensitivity.

[0140] In some embodiments, evaluating the predictive ability of at least one first candidate predictive model further includes: importing large fetal second screening metabolites into a KEGG database for pathway enrichment analysis to determine a first enriched pathway corresponding to the large fetal second screening metabolites, and using large fetal second screening metabolites related to the first enriched pathway as large fetal second screening metabolites for constructing the first predictive model; importing small fetal second screening metabolites into a KEGG database for pathway enrichment analysis to determine a second enriched pathway corresponding to the small fetal second screening metabolites, and using small fetal second screening metabolites related to the second enriched pathway as small fetal second screening metabolites for constructing the second predictive model.

[0141] Principles and steps not explicitly described in this invention are all obtainable by those skilled in the art through conventional technical means, and therefore will not be elaborated upon. Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the probability of severe brain injury in newborns, characterized in that, The method is used to predict the probability of severe brain injury in selectively fetal growth-restricted twin newborns, which includes a larger fetus and a smaller fetus to be examined. The method includes: Obtain meconium samples from the large fetus to be examined and meconium samples from the small fetus to be examined; The meconium samples from the large fetus to be examined and the meconium samples from the small fetus to be examined are pre-processed; The concentration of fetal marker metabolites in the meconium sample of the pretreated fetus was obtained by detecting the meconium sample of the fetus. The concentration of fetal marker metabolites in the pretreated meconium sample of the fetus to be tested was obtained. The probability of severe brain injury in the fetus can be predicted based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined. The probability of severe brain injury in the fetus can be predicted based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined. The marker metabolites of the large fetus and the marker metabolites of the small fetus are different.

2. The method according to claim 1, wherein, The large fetal marker metabolites include at least one of histidine, trans-4-hydroxyproline, pyruvate, and 1-phenylethanol; the small fetal marker metabolites include at least one of hippuric acid, nicotinamide, succinic acid, and citric acid.

3. The method according to claim 2, wherein, The method of predicting the probability of severe brain injury in a fetus based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined includes: using a first prediction model to predict the probability of severe brain injury in a fetus based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined. The method of predicting the probability of severe brain injury in a fetus based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined includes: using a second prediction model to predict the probability of severe brain injury in a fetus based on the concentration of fetal marker metabolites in the meconium sample of the fetus to be examined. The model parameters of the first prediction model and the second prediction model are different.

4. The method according to claim 3, wherein, The method further includes: Meconium samples from twin newborns in group G of MCDA were obtained as modeling samples. The modeling samples had brain injury labels, which included severe brain injury labels and non-severe brain injury labels. The twin newborns in group G of MCDA included twin newborns with selective fetal growth restriction. Each group of modeling samples included one meconium sample from the larger fetus and one meconium sample from the smaller fetus. For each modeling sample in the G group of modeling samples Preprocess the modeling sample; The concentration of non-target metabolites in the preprocessed modeling sample was obtained by detecting the sample. Partial least squares discriminant analysis was performed on non-targeted metabolites in meconium samples from large fetuses in the G group modeling sample. Non-targeted metabolites with variable importance projection greater than or equal to the first threshold were used as primary screening metabolites for large fetuses. The variable importance projection of non-targeted metabolites greater than or equal to the first threshold indicates that they play a role in distinguishing between large fetuses with severe brain injury and large fetuses in the control group. Univariate analysis was performed on the metabolites of the large fetuses in the initial screening to identify the metabolites of the large fetuses in the second screening that showed significant differences between the large fetuses with severe brain injury group and the large fetuses in the control group. Based on the metabolites from the second screening of large fetuses and the brain injury label, at least one first candidate prediction model is constructed to predict the probability of severe brain injury in large fetuses; each of the at least one first candidate prediction models corresponds to a subset of the metabolites from the second screening of large fetuses. Evaluate the predictive power of the at least one first candidate prediction model; The first candidate prediction model that meets the prediction capability requirements is used as the first prediction model, and the non-targeted metabolites in the subset of metabolites in the second screening of large fetuses corresponding to the first candidate prediction model that meets the prediction capability requirements are used as the marker metabolites in the large fetuses. Partial least squares discriminant analysis was performed on non-target metabolites in meconium samples of fetuses in the G group modeling sample. Non-target metabolites with variable importance projection greater than or equal to the second threshold were used as primary screening metabolites for fetuses. The variable importance projection of non-target metabolites greater than or equal to the second threshold indicates that they play a role in distinguishing between fetuses with severe brain injury and fetuses in the control group. Univariate analysis was performed on the primary screening metabolites of the small fetuses to identify secondary screening metabolites of the small fetuses that showed significant differences between the small fetuses with severe brain injury group and the small fetuses control group. Based on the fetal second screening metabolites and the brain injury label, at least one second candidate prediction model is constructed to predict the probability of severe brain injury in the fetus; each of the at least one second candidate prediction models corresponds to a subset of the fetal second screening metabolites. Evaluate the predictive power of the at least one second candidate prediction model; The second candidate prediction model that meets the prediction requirements is used as the second prediction model, and the non-targeted metabolites in the subset of small fetal second screening metabolites corresponding to the second candidate prediction model that meets the prediction requirements are used as small fetal marker metabolites.

5. The method according to claim 4, wherein, The large fetal control group was the large fetal non-severe brain injury group, and the small fetal control group was the small fetal non-severe brain injury group. Alternatively, the G group of MCDA twin newborns also includes twin newborns in the monochorionic diamniotic group with the same weight; the large fetal control group consists of the large fetal non-severe brain injury group and the monochorionic diamniotic group with the same weight; the small fetal control group consists of the small fetal non-severe brain injury group and the monochorionic diamniotic group with the same weight.

6. The method according to claim 4, wherein, Univariate analysis was performed on the primary screening metabolites of the large fetuses to identify secondary screening metabolites that showed significant differences between the large fetuses with severe brain injury and the large fetuses control group. These metabolites included: For the primary screening metabolites of large fetuses with normally distributed concentrations, the independent samples t-test was used to verify whether there was a significant difference between the primary screening metabolites of large fetuses with severe brain injury and the control group of large fetuses. The primary screening metabolites of large fetuses with significant differences were used as the secondary screening metabolites of large fetuses. For the primary screening metabolites of large fetuses with nonnormally distributed concentrations, the Mann-Whitney U test was used to verify whether there was a significant difference between the primary screening metabolites of large fetuses with severe brain injury and the control group of large fetuses. The primary screening metabolites of large fetuses with significant differences were used as the secondary screening metabolites of large fetuses. Univariate analysis was performed on the primary screening metabolites of the small fetuses to identify secondary screening metabolites that showed significant differences between the small fetuses with severe brain injury group and the small fetuses control group, including: For the primary screening metabolites of small fetuses with a normal concentration distribution, the independent samples t-test was used to verify whether there was a significant difference between the primary screening metabolites of small fetuses with severe brain injury and the small fetus control group. The primary screening metabolites of small fetuses with significant differences were used as the secondary screening metabolites of small fetuses. For the primary screening metabolites of small fetuses with non-normally distributed concentrations, the Mann-Whitney U test was used to verify whether there was a significant difference between the primary screening metabolites of small fetuses in the group with severe brain injury and the control group. The primary screening metabolites of small fetuses with significant differences were used as the secondary screening metabolites of small fetuses.

7. The method according to claim 4, wherein, Based on the metabolites from the second screening of the large fetus and the brain injury label, at least one first candidate prediction model is constructed to predict the probability of severe brain injury in the large fetus, including: At least a portion of the fetal secondary screening metabolites are selected as a first target subset from the aforementioned fetal secondary screening metabolites. There are M types of fetal secondary screening metabolites in total, and the first target subset contains m types of fetal secondary screening metabolites. M; Define the logistic regression function corresponding to the first candidate prediction model as: P(Y=1|X)= z=WX+w0; Where X = [x1, ..., x] m ],x1, …, x m The concentration of m types of fetal secondary screening metabolites in the first target subset; W=[w1, …, w m [] represents the weight coefficient matrix corresponding to X; w0 is a constant term; Based on the concentrations of the m types of fetal second screening metabolites in the modeling samples of Group G and the brain injury status labels of the modeling samples of Group G, W and w0 are solved to obtain the first candidate prediction model corresponding to the first target subset. Based on the metabolites from the second screening of the small fetus and the brain injury label, at least one second candidate prediction model is constructed to predict the probability of severe brain injury in the small fetus, including: At least a portion of the fetal second screening metabolites are selected as a second target subset from the fetal second screening metabolites. There are N types of fetal second screening metabolites, and the second target subset contains n types of fetal second screening metabolites. N; Define the logistic regression function corresponding to the second candidate prediction model as: P(Y=1|X)= z=WX+w0; Where X=[x1, ..., x n ],x1, …, x n The concentration of n kinds of fetal screening metabolites in the second target subset; W=[w1, …w n [] represents the weight coefficient matrix corresponding to X; w0 is a constant term; Based on the concentrations of the n fetal secondary screening metabolites in the modeling samples of Group G and the brain injury status labels of the modeling samples of Group G, W and w0 are solved to obtain the second candidate prediction model corresponding to the second target subset.

8. The method according to claim 4, wherein, The evaluation of the predictive ability of the at least one first candidate prediction model includes: For each first candidate prediction model, a receiver operating characteristic (ROC) curve is plotted based on the probability prediction values ​​generated by the first candidate prediction model; at least one of the following indicators is determined based on the ROC curve: area under the ROC curve, specificity, and sensitivity. The evaluation of the predictive ability of the at least one second candidate prediction model includes: For each second candidate prediction model, a receiver operating characteristic (ROC) curve is plotted based on the probability prediction values ​​generated by the second candidate prediction model. At least one of the following indicators is determined based on the ROC curve: area under the ROC curve, specificity, and sensitivity.

9. The method according to claim 4, wherein, The method further includes: The metabolites of the second screening of large fetuses were imported into the KEGG database for pathway enrichment analysis to determine the first enriched pathway corresponding to the metabolites of the second screening of large fetuses. The metabolites of the second screening of large fetuses associated with the first enriched pathway were used as the metabolites of the second screening of large fetuses for constructing the first prediction model. The metabolites from the second screening of the fetus were imported into the KEGG database for pathway enrichment analysis to determine the second enriched pathways corresponding to the metabolites from the second screening of the fetus. The metabolites from the second screening of the fetus associated with the second enriched pathway were used as the metabolites from the second screening of the fetus for constructing the second prediction model.

10. A system for predicting the probability of severe brain injury in newborns, characterized in that, The system is used to predict the probability of severe brain injury in selectively fetal growth-restricted twin newborns, which includes a larger fetus and a smaller fetus to be examined. The system includes: The sample acquisition module is used to acquire meconium samples from the large fetus to be examined and meconium samples from the small fetus to be examined. A preprocessing module is used to preprocess the meconium samples of the large fetus to be examined and the meconium samples of the small fetus to be examined. The metabolite concentration detection module is used to detect the meconium sample of the pretreated large fetus to be tested, and to obtain the concentration of large fetal marker metabolites in the meconium sample of the large fetus to be tested; and to detect the meconium sample of the pretreated small fetus to be tested, and to obtain the concentration of small fetal marker metabolites in the meconium sample of the small fetus to be tested. The probability prediction module is used to predict the probability of severe brain injury in the large fetus based on the concentration of fetal marker metabolites in the meconium sample of the large fetus to be examined, and to predict the probability of severe brain injury in the small fetus based on the concentration of fetal marker metabolites in the meconium sample of the small fetus to be examined. The marker metabolites of the large fetus and the marker metabolites of the small fetus are different.

Images