An early pregnancy four-factor combined clinical index early warning method and system for predicting preeclampsia of twins

Abstract

The application provides an early warning method and system for predicting twin preeclampsia in early pregnancy by four-factor joint clinical indicators, and is applied to the field of data processing. The XGBoost model is used to construct the early warning system. The concentrations of four proteins ANXA2, ANXA3, CHIT1 and FTL in the plasma of pregnant women at 11-14 weeks of pregnancy are detected by ELISA, and the protein characteristics are obtained by logarithmic conversion and Z-score standardization. At the same time, the clinical indicators such as age, pre-pregnancy BMI and mean arterial pressure are standardized, and the binary classification indicators are directly coded to form the standardized clinical characteristics. The two types of characteristics are fused into a 9-dimensional input vector, and the model is trained with fixed hyperparameters. The 0-1 risk probability is obtained by integrated tree calculation and Sigmoid mapping. The optimal threshold is determined by relying on the ROC curve and the Youden index, the model performance is verified by combining information gain, Delong method and internal and external double queue, and finally the risk classification output is completed according to the threshold.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method and system for predicting preeclampsia in twin pregnancies using a combination of four clinical indicators in early pregnancy. Background Technology

[0002] Twin pregnancies are considered a high-risk pregnancy, with a significantly increased risk of complications compared to singleton pregnancies. This is particularly true in the area of ​​early warning and risk screening for preeclampsia, where there is a clear need for clinical data processing and predictive modeling. Current early risk assessments for preeclampsia in twin pregnancies largely rely on single clinical indicators or traditional statistical analysis methods, lacking specific early warning methods for twins. Furthermore, standardized processing mechanisms for multi-source heterogeneous data are insufficient, making it difficult to effectively integrate protein biomarkers with maternal clinical indicators. This results in unstable and poorly reproducible risk assessment results.

[0003] In the combined application of clinical data and biomarkers, existing technologies have not established a unified preprocessing system adapted to machine learning models: continuous indicators and categorical indicators have not been aligned with different scales, the skewed distribution of protein concentration has not been effectively corrected, and features of different dimensions and numerical ranges are directly input into the model, which can easily lead to problems such as model training bias, slow convergence speed, and insufficient generalization ability, and cannot meet the requirements of stable output and objective judgment in clinical screening.

[0004] In scenarios where machine learning models are used to calculate risks in early pregnancy, existing technologies generally suffer from defects such as inconsistent model parameters, uncontrollable stochastic processes, and inconsistent threshold determination methods, and lack a complete model performance verification system and standardized output process.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore includes information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] According to one aspect of this application, a method for predicting preeclampsia in twin pregnancies using a four-factor combined clinical index in early pregnancy is provided, comprising: obtaining the application requirements for predicting preeclampsia risk in twin pregnancies; selecting XGBoost as the model architecture; setting a fixed random seed to ensure reproducible model training; performing ELISA quantitative detection of four proteins (ANXA2, ANXA3, Chitinase 1, and FTL) on plasma samples from pregnant women at 11 to 14 weeks of gestation; first eliminating heteroscedasticity through natural logarithmic transformation, and then performing Z-score standardization to generate standardized protein characteristic data; and directly performing continuous indexes such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy. Z-score standardization directly encodes the binary indicators of obstetric history and conception method without conversion, generating standardized maternal clinical feature data. A 9-dimensional input vector is constructed by fusing 4 protein features and 5 clinical features. XGBoost model training is completed according to preset fixed hyperparameters. After outputting the original prediction score via an ensemble tree, risk probability data in the 0-1 range is output through Sigmoid mapping. A risk probability cutoff threshold is determined based on ROC curves and Youden's index. Feature importance is calculated with information gain as the core, and the DeLong method is used for evaluation. Model performance is validated through an internal modeling queue and an independent external validation queue. Based on the risk probability and cutoff threshold, risk level classification is completed, generating a prediction result for preeclampsia risk in twin pregnancies.

[0007] Another aspect of this application discloses an early warning system for predicting preeclampsia in twin pregnancies using a four-factor combined clinical index in early pregnancy, comprising: a risk prediction requirement and model configuration module, used to obtain the application requirements for risk prediction of preeclampsia in twin pregnancies, select XGBoost as the model architecture, and set a fixed random seed to ensure reproducible model training; a targeted protein detection and standardization module, used to perform ELISA quantitative detection of ANXA2, ANXA3, CHIT1, and FTL proteins in plasma samples from pregnant women at 11 to 14 weeks of gestation, and generate standardized protein feature data through natural logarithmic transformation and Z-score standardization; and a maternal clinical feature preprocessing module, used to perform Z-score standardization on continuous indicators such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy, and to preprocess maternal clinical features such as obstetric history and mode of conception. The system employs several modules: a classification index direct encoding to generate standardized maternal clinical feature data; a multi-source feature fusion and model training module to fuse 4 protein features and 5 clinical features to construct a 9-dimensional input vector, train the XGBoost model according to preset fixed hyperparameters, output the original prediction score through an ensemble tree, and output risk probability data in the 0-1 interval through Sigmoid mapping; a model evaluation and threshold determination module to determine the risk probability cutoff threshold based on ROC curves and Youden's index, calculate feature importance using information gain, and verify model performance using the DeLong method through an internal modeling queue and an independent external validation queue; and a risk level determination and result output module to classify risk levels based on risk probability and cutoff threshold, generate and uniformly output the risk prediction results for preeclampsia in twin pregnancies.

[0008] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described method for predicting preeclampsia in twin pregnancies using a combination of four clinical indicators in early pregnancy.

[0009] The beneficial effects of this application are as follows: Using XGBoost ensemble learning as the core architecture, a predictive system is constructed by detecting the concentrations of four proteins (ANXA2, ANXA3, CHIT1, and FTL) via ELISA and combining them with five clinical indicators: age, pre-pregnancy BMI, mean arterial pressure in early pregnancy, obstetric history, and method of conception. Protein data are first transformed by natural logarithm and then standardized using Z-scores. Continuous clinical indicators are directly standardized, and binary clinical indicators are directly encoded, fusing the two types of features into a nine-dimensional input vector. The model is trained using fixed hyperparameters and fixed random seeds, and a risk probability of 0-1 is obtained through Sigmoid mapping. The optimal threshold is determined based on the ROC curve and Youden's index. Model validation is completed by combining information gain, the DeLong method, and internal and external cohorts. Finally, the risk level is output according to the threshold.

[0010] This application eliminates dimensional and distributional differences through differentiated preprocessing, enabling unified input of protein and clinical features and improving model learning balance. Fixed hyperparameters and random seeds ensure reproducible training and prediction results, making it suitable for standardized clinical deployment. Validated by internal and external cohorts, the model exhibits high sensitivity, high specificity, and strong generalization ability, demonstrating excellent early risk differentiation performance. The process is simple and computationally efficient, requiring no complex parameter tuning, and can be quickly adapted to clinical screening scenarios, providing objective data support for early warning of twin pregnancies.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0012] Figure 1 This document illustrates a flowchart of an early warning method for predicting preeclampsia in twin pregnancies using a combination of four clinical indicators in early pregnancy, provided by an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the structure of an early warning system for predicting preeclampsia in twin pregnancies using four combined clinical indicators in early pregnancy, provided in an embodiment of this application. Detailed Implementation

[0013] 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.

[0014] The following is combined Figure 1 This application describes an early warning method for predicting preeclampsia in twin pregnancies using a combination of four clinical indicators in early pregnancy, according to an exemplary embodiment of the present application: S101, Obtain the application requirements for predicting the risk of preeclampsia in twin pregnancies, select XGBoost as the model architecture, and set a fixed random seed to ensure that the model training is repeatable.

[0015] In one implementation, combining the engineering requirements for reproducible deployment of machine learning models for early risk prediction of preeclampsia in twin pregnancies, and leveraging the technical characteristics of XGBoost ensemble learning, feature standardization, and probabilistic output, a model construction system for early pregnancy risk prediction adapted to clinical data is established by introducing three core constraints: fixed random seeds, data distribution correction, and fixed hyperparameters. The model system consists of four layers from top to bottom: a demand access layer, a feature processing layer, a model training layer, and a probability output layer, with sequential connections between layers. The demand access layer receives the task of predicting the risk of preeclampsia in twin pregnancies, specifying four protein indicators and five clinical indicators for twin pregnant women at 11 to 14 weeks of gestation, and outputting a risk probability in the zero-to-one range and a binary risk level. The feature processing layer is used to complete the unified transformation between protein data and clinical data. It performs natural logarithmic transformation on the concentrations of four proteins, ANXA2, ANXA3, CHIT1, and FTL, and then performs Z-score standardization on the transformed data to unify the features into a standard distribution with a mean of zero and a standard deviation of one. It directly performs Z-score standardization on three continuous clinical indicators, namely age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy, while keeping the original codes unchanged for two binary indicators, namely obstetric history and method of conception.

[0016] The model training layer employs the XGBoost ensemble learning architecture for feature learning and fitting, with fixed hyperparameters for training. The total number of decision trees is set to 200, the maximum tree depth to 8, the learning rate to 0.3, the minimum split loss to 1, the feature column sampling rate to 0.5, the minimum child node weight to 1, and the sample row sampling rate to 0.5. The probability output layer converts the model output into standardized risk values, using a Sigmoid mapping to transform the original predicted scores into risk probabilities in the zero-to-one range. A fixed random seed (20251001) is used to lock in random sampling during training, ensuring consistency in the random sequences of data partitioning, feature sampling, and sample sampling, resulting in identical predictions across multiple training iterations. Data distribution correction unifies the distribution of multi-source clinical data, eliminating heteroscedasticity of protein concentration through logarithmic transformation and eliminating prediction bias caused by dimensional differences and data shifts through standardization. Fixed hyperparameters are used to lock the model structure and training rules. The number of decision trees, maximum depth, learning rate, feature sampling rate, and sample sampling rate are all set to fixed values ​​and are not dynamically adjusted during training to avoid parameter fluctuations affecting the stability of the results.

[0017] Following the prediction standards of high sensitivity, high specificity, and strong generalization, the three key processes of logarithmic transformation of protein biomarkers, standardization of clinical features, and nine-dimensional input vectors are jointly matched to align the model input with the training process. The prediction standards aim for early clinical screening, requiring the model to have a high detection rate and a low false positive rate. The logarithmic transformation of protein biomarkers uses the natural logarithm transformation formula to perform logarithmic calculations on the original plasma absolute concentrations of four proteins—ANXA2, ANXA3, CHIT1, and FTL—with the natural constant as the base, converting the skewed protein concentration data into an approximately normal distribution and eliminating heteroscedasticity.

[0018] Clinical feature standardization employs the Z-score standardization formula to transform three continuous clinical indicators: age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy. Using the corresponding indicators from all modeling samples as the statistical population, the population mean and standard deviation are calculated. Each value is subtracted from its corresponding mean and divided by the standard deviation to convert it to a uniform scale with a mean of zero and a standard deviation of one. The nine-dimensional input vector is constructed by concatenating four standardized protein features and five preprocessed clinical features in a fixed order, directly serving as input data for the XGBoost model. The three processing steps are sequentially linked: protein biomarkers undergo logarithmic transformation before Z-score standardization; continuous clinical indicators are directly standardized using Z-score standardization; and binary clinical indicators retain their original encoding. The results of the preceding processing steps serve as input data for the subsequent processing steps, ultimately forming a uniform nine-dimensional feature vector, achieving complete alignment between the data flow and the model training process.

[0019] Based on the balance requirements of prediction accuracy, fitting stability, and overfit suppression, fixed hyperparameters, the number of decision trees, and tree depth of XGBoost are configured to complete the training constraints of the prediction model. The training constraints are directly configured in the XGBoost model training module, which includes a decision tree generation unit, a feature sampling unit, a sample sampling unit, and a loss calculation unit. The total number of decision trees is set to 200 to control the ensemble size; the tree depth is set to 8 to limit the maximum splitting level of a single tree; the learning rate is set to 0.3 to control the weight update magnitude of each tree; the feature sampling rate is set to 0.5 to control the proportion of features used by each tree; the sample sampling rate is set to 0.5 to control the proportion of samples used by each tree; the minimum splitting loss is set to 1; and the minimum child node weight is set to 1. All parameters are configured with fixed values ​​and are not dynamically adjusted. The total number of decision trees ensures that the model fully learns feature information. The tree depth prevents overfitting caused by excessive complexity of a single tree. The learning rate ensures smooth model convergence. Sampling settings improve the model's generalization ability on independent data.

[0020] The model system inputs four protein concentration detection data, maternal clinical characteristic data, and preprocessing parameters, and outputs the original predicted scores and risk probabilities according to the training rules. The model system receives three types of data at its data input: concentration detection data for ANXA2, ANXA3, CHIT1, and FTL; clinical indicators such as age, pre-pregnancy body mass index, mean arterial pressure in early pregnancy, obstetric history, and method of conception; and the mean and standard deviation parameters used for standardization. The data enters the feature processing layer for transformation and concatenation. First, a natural logarithmic transformation is performed on the four protein concentrations. Then, Z-score standardization is performed on the transformed data. Z-score standardization is directly performed on the three continuous clinical indicators: age, pre-pregnancy body mass index, and mean arterial pressure in early pregnancy. The two binary indicators, obstetric history and method of conception, retain their original encoding, forming a unified scale feature with a mean of zero and a standard deviation of one. These features are then concatenated in a fixed order to form a nine-dimensional feature vector, which is then input into the model training layer. The training layer consists of a fixed number of regression decision trees, with a total of 200 trees and a maximum depth of 8 for each tree. Each tree performs node splitting and feature weighting based on a nine-dimensional feature vector. The outputs of all trees are summed to obtain the original prediction score. The original prediction score is converted into a value between zero and one using the Sigmoid mapping formula. This value represents the probability of preeclampsia in twin pregnancies.

[0021] By using a fixed random seed, Sigmoid probability mapping, and ROC target threshold, the model training and prediction process is optimized, generating a reproducible XGBoost risk prediction model construction strategy suitable for clinical screening of twin pregnancies. The fixed random seed, with a value of 20251001, is used throughout the entire training and prediction process. This consistent random sequence is maintained in data partitioning, feature sampling, and sample sampling to ensure identical results across multiple runs and guarantee process reproducibility. The Sigmoid probability mapping converts the model's raw scores into intuitive risk probabilities. Using a Sigmoid mapping formula with a base of the natural constant, raw scores of any range are converted into values ​​between zero and one, facilitating clinical understanding and threshold determination. The ROC target threshold is determined based on the principle of maximizing the Youden exponent. By traversing all candidate risk probabilities and calculating the Youden exponent, the value with the largest exponent is selected as the optimal cutoff threshold, balancing sensitivity and specificity to achieve accurate segmentation between high-risk and low-risk populations. These three processes respectively affect training stability, output readability, and judgment accuracy, forming a closed-loop optimization process. The strategy includes data access standards, feature processing procedures, model structure configuration, training parameter settings, probability transformation rules, and threshold determination methods. The model training parameters are set as follows: total number of decision trees 200, maximum tree depth 8, learning rate 0.3, feature column sampling rate 0.5, sample row sampling rate 0.5, minimum split loss 1, and minimum child node weight 1. It can be directly used for the early pregnancy risk prediction of preeclampsia in twin pregnancies, realizing a deep integration of machine learning algorithms and clinical screening services.

[0022] S102 performs ELISA quantitative detection of ANXA2, ANXA3, CHIT1, and FTL proteins in plasma samples from pregnant women at 11 to 14 weeks of gestation. The results are first processed by natural logarithmic transformation to eliminate heteroscedasticity, and then Z-score standardization is performed to generate standardized protein characteristic data.

[0023] In one embodiment, plasma samples from pregnant women at 11 to 14 weeks of gestation are quantitatively analyzed using enzyme-linked immunosorbent assay (ELISA) to obtain the absolute concentrations of four target proteins: annexin A2 (ANXA2), annexin A3 (ANXA3), chitinase 1 (CHIT1), and ferritin light chain (FTL). The concentration units are nanograms per milliliter (ng / mL). The detection process is performed using a dedicated kit. Precision recovery and linearity validation are conducted for each protein. The intra-assay coefficient of variation (COP) for ANXA2, ANXA3, CHIT1, and FTL is less than 8%, and the inter-assay COP is less than 10%, ensuring accurate and stable concentration values. The test samples are limited to peripheral blood plasma from pregnant women in twin pregnancies at 11 to 14 weeks of gestation. The output result is the concentration value of each protein, which serves as the sole raw input for subsequent data processing.

[0024] The original concentration values ​​of the four proteins were substituted into the natural logarithmic transformation formula, x' = ln(x), where x is the original concentration value and x' is the logarithmically transformed value. This transformed the data distribution and eliminated heteroscedasticity caused by the right-skewed distribution. The original protein concentration data showed a significant right-skewed distribution, which could affect training stability if directly input into the model. The natural logarithmic transformation was applied to each concentration value, resulting in a data distribution closer to a normal distribution with more stable fluctuations, providing a stable foundation for subsequent standardization. The transformation process acted independently on each protein, without altering the relative differences between the data. The logarithmically transformed protein values ​​were then substituted into the Z-score standardization formula, which is: ,in, The standardized value of the j-th protein in the i-th sample is ultimately used as the model input. Let be the overall mean concentration of the j-th protein after logarithmic transformation; Let be the overall standard deviation of the concentration of the j-th protein after logarithmic transformation. Based on the overall mean and overall standard deviation, scale unification is completed to generate standardized protein characteristic data.

[0025] Standardization uses the logarithmically transformed values ​​of all samples as the statistical population. First, the mean and standard deviation of each protein are calculated. Then, each value is subtracted from the corresponding mean and divided by the standard deviation, transforming the feature data of each protein into a standard distribution with a mean of zero and a standard deviation of one. After processing, the four proteins have a uniform scale and can be directly integrated with clinical features into the model. The logarithmically transformed and Z-score standardized values ​​of the four proteins are combined to form standardized protein feature data with fixed dimensions. Each sample corresponds to a set of four standardized values, representing the processed features of annexin A2 (ANXA2), annexin A3 (ANXA3), chitinase 1 (CHIT1), and ferritin light chain, respectively. This set of data is unitless and bias-free, and can be stably input into the XGBoost model, forming a nine-dimensional input vector together with clinical features, achieving alignment and fusion of multi-source data.

[0026] S103 directly performs Z-score standardization on continuous indicators such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy, and directly encodes binary indicators such as obstetric history and conception method without conversion, generating standardized maternal clinical characteristic data.

[0027] In one implementation, combining the clinical data adaptation requirements for predicting preeclampsia risk in twin pregnancies with the modeling logic of unified input of multiple types of features, a differentiated processing rule of continuous feature standardization and direct input of binary feature classification is introduced to build a standardized preprocessing system for maternal clinical features. The preprocessing system is divided into a continuous feature processing unit and a binary feature processing unit. These two units receive raw clinical data in parallel, and the output results are fused by a feature concatenation unit before being uniformly output. The system is designed for clinical screening scenarios in twin pregnancies, converting different types of maternal indicators into a unified format that can be directly input into the XGBoost model, ensuring balanced weights and stable computation of various features during subsequent model training. The continuous feature processing unit performs Z-score standardization on three continuous indicators: age, pre-pregnancy body mass index, and mean arterial pressure in early pregnancy. This standardizes indicators with different dimensions and numerical ranges into a standard distribution with a mean of zero and a standard deviation of one, eliminating dimensional differences in age, pre-pregnancy body mass index, and mean arterial pressure.

[0028] Standardization is achieved by subtracting the overall mean age of all modeling samples from the original age value and then dividing by the overall standard deviation of all modeling samples' ages. The same calculation method is used to standardize the pre-pregnancy body mass index (BMI) and early pregnancy mean arterial pressure (MAP), ensuring that all three continuous indicators fall within the same numerical range. The binary classification feature processing unit directly retains the original codes for the two binary classification indicators, obstetric history and conception method, without logarithmic transformation or scaling, preserving their original classification meaning. Obstetric history is directly coded as a fixed value for primiparous and multiparous women, and conception method is directly coded as a fixed value for natural conception and assisted reproduction, maintaining consistency throughout the encoding process without any numerical distortion. After processing different types of features separately, the feature concatenation unit combines them in a fixed order to form a five-dimensional clinical feature vector. This vector sequentially includes standardized age, standardized pre-pregnancy BMI, standardized early pregnancy MAP, coded obstetric history, and coded conception method. This vector can be directly concatenated with four standardized protein features to form a nine-dimensional model input vector, ensuring the stability and interpretability of the model input.

[0029] Following the XGBoost model's preprocessing guidelines of being on the same scale and dimensionless, the three continuous indicators—age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy—were directly standardized using Z-scores. The two binary indicators—obstetric history and method of conception—remained in their original encoding, thus aligning the scale of clinical features with the model input. The XGBoost model requires all input features to have similar numerical ranges; therefore, continuous indicators are standardized to ensure that clinical indicators from different sources and of different magnitudes participate in model training under a uniform distribution. Age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy are continuous numerical indicators, and were directly converted to a uniform scale using the Z-score standardization formula. Specifically, using the corresponding indicators of all modeling samples as the statistical population, the population mean and standard deviation were first calculated. Then, each original value was subtracted from its corresponding mean and divided by the standard deviation, ensuring that the processed values ​​strictly conformed to a standard distribution with a mean of zero and a standard deviation of one. The Z-score standardization formula is: ,in, This represents the original value of the k-th continuous clinical feature of the i-th sample; These are the standardized values ​​ultimately used as model inputs; is the overall mean of the k-th continuous clinical feature; denoted as the overall standard deviation of the k-th continuous clinical feature.

[0030] The raw pre-pregnancy BMI value was subtracted from the overall mean of the pre-pregnancy BMI of all modeling samples, and then divided by the overall standard deviation of the pre-pregnancy BMI of all modeling samples to complete the standardization transformation. The transformation process strictly followed the Z-score standardization formula mentioned above. Obstetric history and method of conception were used as classification labels and were not numerically transformed; their original encoding states were directly retained. Obstetric history was assigned a fixed numerical code of zero for primiparous women and one for multiparous women. Method of conception was assigned a fixed numerical code of zero for natural conception and one for assisted reproduction. No logarithmic transformation or scaling was performed. After processing the two types of indicators separately, both continuous and binary indicators could be input into the model in the same dimension. The numerical ranges of each feature remained consistent to avoid model training bias due to differences in numerical ranges and to ensure that the model's learning weights for each feature remained balanced.

[0031] Based on the core requirements of stable feature distribution and efficient model training convergence, continuous clinical features are standardized with a mean of zero and a standard deviation of one. Binary clinical features are not subjected to logarithmic transformation or scaling, thus completing the feature constraint setting for the preprocessing system. The standardization constraints for continuous clinical features are calculated using global statistics. Taking the corresponding indicators of all modeled samples as the population, the mean and standard deviation are calculated. Then, each value is subtracted from the mean and divided by the standard deviation, ensuring that the processed features strictly conform to a standard distribution with a mean of zero and a standard deviation of one. Continuous clinical features include three indicators: age, pre-pregnancy body mass index, and mean arterial pressure in early pregnancy. All are transformed using the Z-score standardization formula. When processing the mean arterial pressure in early pregnancy, the population mean and population standard deviation of the mean arterial pressure in early pregnancy for all modeled samples are first calculated. Then, the original mean arterial pressure value for each pregnant woman is subtracted from the population mean, and the result is divided by the population standard deviation, finally yielding a standardized value with a mean of zero and a standard deviation of one. Binary clinical features are not subject to any mathematical transformations, logarithmic operations, or scaling, maintaining the original classification meaning. The binary clinical features include two indicators: obstetric history and mode of conception. Obstetric history is coded using a fixed code for primiparous and multiparous women, while the mode of conception is coded using a fixed code for natural conception and assisted reproduction. This coding remains constant throughout the process. This constraint ensures a stable feature distribution, avoids model training bias caused by differences in units and numerical shifts, reduces the risk of overfitting, accelerates iterative convergence, and improves the predictive stability of the model in both the internal modeling cohort and the independent external validation cohort.

[0032] Raw data on maternal age, BMI, mean arterial pressure, obstetric history, and method of conception are input into the preprocessing system. After continuous feature standardization and binary feature direct encoding, a uniformly scaled clinical feature vector is output. The preprocessing system receives five raw clinical data items: age, pre-pregnancy body mass index, mean arterial pressure in early pregnancy, obstetric history, and method of conception. Among them, age, pre-pregnancy body mass index, and mean arterial pressure in early pregnancy are three continuous data items, which are uniformly encoded using Z-scores. The raw values ​​are processed using the score standardization formula to convert them into standardized values ​​with a mean of zero and a standard deviation of one. The two binary categories of obstetric history and conception method are entered into direct coding units, retaining pre-set fixed codes without any transformation or scaling. The raw age data is standardized by subtracting the overall mean age of all modeling samples and then dividing by the overall standard deviation. Pre-pregnancy body mass index and mean arterial pressure in early pregnancy are standardized using the same rules. Obstetric history is coded as zero for primiparous women and one for multiparous women; conception method is coded as zero for natural conception and one for assisted reproduction, and these values ​​remain unchanged after processing.

[0033] The five processed values ​​are combined in a fixed order: standardized age, standardized pre-pregnancy body mass index, standardized early pregnancy mean arterial pressure, encoded obstetric history, and encoded mode of conception. This forms a fixed-dimensional five-dimensional clinical feature vector. All elements in this vector have been freed from dimensional bias and have similar numerical distribution ranges. They can be directly concatenated with the four standardized protein features (ANXA2, ANXA3, CHIT1, and FTL) in a fixed order to form the nine-dimensional input vector required for model training. This provides unified, stable, and standardized input data for subsequent XGBoost model calculations.

[0034] Differential preprocessing preserves the original semantics of features and eliminates dimensional differences, generating standardized parent clinical feature data adapted to the input of the XGBoost model, thus achieving numerical alignment between clinical indicators and model training. Differential preprocessing removes the influence of units for continuous indicators while preserving diagnostically relevant semantics for categorical indicators, maintaining consistency in their numerical ranges. Continuous indicators are processed using Z-multiplexing... The score standardization process converts age, pre-pregnancy body mass index (BMI), and early pregnancy mean arterial pressure into a standardized distribution with a mean of zero and a standard deviation of one, completely eliminating the differences caused by the original units and numerical magnitudes. Age is expressed in years, pre-pregnancy BMI in kilograms per square meter, and early pregnancy mean arterial pressure in millimeters of mercury; after standardization, these are all converted into unitless standardized values, no longer constrained by the original dimensions. Categorical indicators undergo no mathematical transformation, directly retaining their original coded meaning. The classification status of obstetric history and conception method directly reflects clinical risk-related information without altering the diagnostic semantics of the features themselves. The processed data can be directly input into the feature input layer of the XGBoost model, sharing the same numerical distribution range as the four standardized protein features. This allows the model to learn the contribution of each type of feature equally during training, without bias towards any particular feature due to differences in numerical range. Continuous clinical features, binary clinical features, and protein features are all on a unified scale, resulting in a more balanced feature weight allocation, more stable model training, and more efficient convergence. This enables the model to fully learn the association between multi-source features and the risk of preeclampsia in twin pregnancies, achieving seamless integration of clinical business data and machine learning algorithms.

[0035] S104 integrates 4 protein features and 5 clinical features to construct a 9-dimensional input vector. The XGBoost model is trained according to preset fixed hyperparameters. After the raw prediction score is output through the ensemble tree, the risk probability data in the 0 to 1 interval is output through the Sigmoid mapping.

[0036] In one implementation, considering the need for multi-source feature fusion modeling for predicting preeclampsia risk in twin pregnancies, protein biomarker standardization, clinical feature differentiation preprocessing, and XGBoost ensemble learning rules are introduced to determine a unified modeling input scheme combining four proteins and five clinical indicators. For the early risk prediction scenario of twin pregnancies between 11 and 14 weeks of gestation, the concentration data of four proteins (ANXA2, ANXA3, CHIT1, and FTL) and five clinical data (age, pre-pregnancy body mass index, early pregnancy mean arterial pressure, obstetric history, and method of conception) are incorporated into the same modeling framework. The protein biomarkers are processed using a rule of first performing a natural logarithmic transformation followed by Z-score standardization. The original concentrations of the four proteins are first subjected to a logarithmic transformation with a base of the natural constant. The logarithmic transformation formula is as follows: ,in, Let J be the absolute concentration of the j-th protein in the original plasma of the i-th sample. The concentration values ​​are logarithmically transformed. Then, the mean and standard deviation are calculated using the logarithmically transformed results of all modeled samples as the statistical population, converting the values ​​to a standard distribution with a mean of zero and a standard deviation of one. After performing logarithmic transformation on the original ANXA2 concentration, the population mean after logarithmic transformation is subtracted and divided by the population standard deviation to complete the standardization process. Clinical characteristics employ continuous indicator standardization and direct differentiation rules for binary indicators. Z-score standardization is directly applied to the three continuous indicators: age, pre-pregnancy body mass index, and mean arterial pressure in early pregnancy. The standardization formula is z''=(z μ) / σ, where z is the original value of the continuous indicator, μ is the overall mean of the corresponding indicator, σ is the overall standard deviation of the corresponding indicator, and z'' is the standardized value. For the two binary indicators of obstetric history and conception method, the fixed codes are directly retained: zero for primiparous women, one for multiparous women, zero for natural conception, and one for assisted reproduction, without any mathematical transformation. The two types of rules are fully compatible with the input requirements of the XGBoost model, forming a fixed multi-source data access and transformation scheme. All features are processed into a dimensionless, identically distributed standard format, providing a unified execution standard for subsequent feature splicing and fusion.

[0037] Preprocessing was performed on four protein features (ANXA2, ANXA3, CHIT1, and FTL) and five maternal clinical characteristics (age, BMI, mean arterial pressure, obstetric history, and mode of conception) to convert both types of features into a standardized, dimensionless model input format. The four protein features were first transformed using the natural logarithm to eliminate heteroscedasticity caused by skewed distributions, and then standardized using Z-scores to a standard distribution with a mean of zero and a standard deviation of one. The formula for the natural logarithm transformation is as follows: This is used to transform the original plasma concentrations of four proteins—ANXA2, ANXA3, CHIT1, and FTL—from a right-skewed distribution to an approximately normal distribution, eliminating model training bias caused by data fluctuations; the Z-score standardization formula is... The mean μ and standard deviation σ of the statistical population are calculated using the logarithmic transformation results of all modeled samples, and the values ​​are standardized to the standard distribution. The original concentration of ANXA3 protein is first taken as the natural logarithm, then the population mean after logarithmic transformation is subtracted and divided by the population standard deviation to complete the standardization process.

[0038] Of the five clinical characteristics, age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy (three continuous indicators) were directly standardized using Z-scores. The standardization formula is as follows: The overall statistics are calculated using the indicators corresponding to all modeled samples, eliminating the influence of different units such as age, kg / m², and mmHg. For example, the original mean arterial pressure value in early pregnancy is subtracted from the overall mean of all modeled samples, and then divided by the overall standard deviation to achieve scale uniformity. The two binary indicators, obstetric history and method of conception, retain their original codes without modification. Obstetric history is fixedly coded as 0 for primiparous women and 1 for multiparous women, and method of conception is fixedly coded as 0 for natural conception and 1 for assisted reproduction. No logarithmic transformation or scaling is performed, fully preserving the clinical classification meaning.

[0039] After processing, all features are free from dimensional differences, and their numerical ranges converge to similar intervals with consistent distribution patterns. This prevents the model training from being biased towards a particular type of feature due to differences in the original numerical values. The features can be directly used for feature fusion and XGBoost model input, ensuring a stable model learning process and balanced weight distribution.

[0040] An XGBoost ensemble learning network driven by fixed hyperparameters was constructed, with fixed network structure and training constraints such as the number of decision trees, maximum depth, learning rate, and sampling rate. The network employs an additive ensemble architecture, consisting of a feature input layer, a regression tree ensemble layer, and a score output layer from top to bottom. The feature input layer receives a pre-processed nine-dimensional standardized feature vector, directly inputting four protein features and five clinical features simultaneously into the regression tree ensemble layer. The regression tree ensemble layer comprises multiple regression decision trees, trained sequentially using a forward stepwise algorithm, with each new tree learning from the prediction residuals of its predecessor. The score output layer sums the outputs of all regression trees to obtain the final raw prediction score, which serves as the sole basis for subsequent probability transformations.

[0041] Key parameters are configured with fixed settings, all pre-defined based on experimental verification results. Specific parameters and their functions are as follows: The total number of decision trees is set to 200 to control the ensemble size and ensure the model can fully learn the correlation information between features and risks; the maximum depth is set to 8 to limit the maximum splitting level of a single tree, preventing overfitting due to excessive complexity of individual trees; the learning rate is set to 0.3 to control the weight update magnitude of each tree, ensuring smooth convergence during model iteration; the feature column sampling rate is set to 0.5, randomly using only half of the features during training for each tree, improving the model's generalization ability; the sample row sampling rate is set to 0.5, randomly using only half of the samples during training for each tree, further reducing the risk of overfitting; the minimum splitting loss is set to 1, and the minimum child node weight is set to 1, constraining the tree's growth conditions and improving model stability.

[0042] During model training, a fixed random seed value of 20251001 is used to ensure that the feature sampling, sample sampling, and random sequence generated by the tree structure are completely consistent in each run. All parameters are configured once before training begins. No grid search or adaptive adjustment is performed during training, and the parameters remain unchanged throughout the process. This ensures that the model structure is stable, the training process is reproducible, and the risk probability results obtained from multiple training runs are completely consistent with the predicted results.

[0043] The fused 9-dimensional standardized feature vector is input into the XGBoost network. After iterative training with multiple regression trees, node splitting, and feature weighting, the integrated raw prediction score is output. The nine-dimensional vector is composed of four standardized protein features and five preprocessed clinical features concatenated in a fixed order and directly input to the feature input layer. This fixed order is as follows: four standardized protein features: ANXA2, ANXA3, CHIT1, and FTL, followed by five clinical features: standardized age, standardized pre-pregnancy BMI, standardized early pregnancy mean arterial pressure, encoded obstetric history, and encoded conception method. Each sample strictly follows this order to form a nine-dimensional feature vector with fixed dimensions, without arbitrarily changing the order, ensuring consistent and reproducible model input structure.

[0044] The regression tree ensemble layer performs node splitting and feature contribution weighting based on feature values. During the training of each regression decision tree, node splitting is performed using a fixed random seed 20251001 and fixed hyperparameters: the optimal splitting feature and splitting threshold are selected based on the numerical values ​​of the nine-dimensional features, the samples are divided into different leaf nodes, and a corresponding contribution weight is assigned to each feature. For example, the model can perform node splitting based on ANXA3 standardized values ​​or early pregnancy mean arterial pressure standardized values, assigning higher weights to features with stronger risk discrimination capabilities, allowing the model to focus more on key indicators.

[0045] Each tree independently outputs a mapping result for the current sample. The outputs of all regression trees are successively added according to the forward addition rule to obtain the original prediction score representing the overall judgment result of the model. In this model, the total number of regression trees is fixed at 200, that is, the total score is obtained by adding the outputs of the 200 trees in sequence. This score is not affected by the unit of measurement and only reflects the comprehensive correlation between features and risks. It serves as the sole basis for subsequent Sigmoid probability transformation and does not introduce other calculation factors.

[0046] The original predicted scores are mapped using the Sigmoid activation function to generate the risk probability of preeclampsia in twins, taking values ​​between 0 and 1. Substituting the original predicted scores into the Sigmoid transformation formula, exponential and fractional calculations are performed using the natural constant as the base, compressing scores from any numerical range into a closed interval between zero and one. The Sigmoid transformation formula is as follows: ,in, This indicates the occurrence of preeclampsia. Let P be the final predicted probability of preeclampsia in twin pregnancies; e is the natural constant, and P is the converted risk probability of preeclampsia in twin pregnancies, strictly limited to the range of 0 to 1. If the original predicted score output by the model is 2.0, substituting it into the formula yields P≈0.88; if the original predicted score is -1.0, substituting it into the formula yields P=1 / (1+e^(1.0))≈0.27. Regardless of whether the original score is positive or negative, it can be stably mapped to the 0~1 interval through this formula.

[0047] The transformed value directly represents the likelihood of preeclampsia in twin pregnancies; a higher value indicates a higher risk, aligning with the needs of intuitive clinical understanding and threshold determination. This probability value can be used for risk stratification without additional conversion. Combined with the optimal cutoff threshold determined by the Youden index, it can directly distinguish between high-risk and low-risk individuals. It is understandable without requiring a professional background, meeting the practical needs of rapid clinical screening and clear, readable results.

[0048] By integrating preprocessed protein features, clinical features, and model output probabilities, multi-source feature fusion modeling and risk probability calculation information suitable for clinical early warning is generated. The information set includes a complete data processing workflow, feature fusion method, model structure configuration, training parameter settings, score calculation logic, and probability conversion rules. The entire process is centered on data calculation and does not involve disease diagnosis or medical intervention. The data processing workflow clearly covers all operational steps, including the natural logarithmic transformation and Z-score standardization of four proteins, the direct standardization of three continuous clinical indicators, and the fixed encoding of two binary clinical indicators. The feature fusion method strictly follows a fixed concatenation order of "four protein features first, five clinical features last," forming a unified nine-dimensional input vector. The model structure configuration clearly defines the XGBoost additive ensemble architecture and the three-layer network structure. The training parameter settings fully specify all fixed hyperparameters, including a total of 200 decision trees, a maximum depth of 8, a learning rate of 0.3, a feature sampling rate of 0.5, a sample sampling rate of 0.5, a minimum split loss of 1, and a minimum child node weight of 1. The score calculation logic clearly states that the outputs of 200 regression trees are successively accumulated to obtain the original predicted score. The probability transformation rules fully specify the Sigmoid mapping formula and the 0-1 interval probability output method. All steps use reproducible numerical calculations and formula operations, and do not include any physician subjective judgment, disease diagnosis, or treatment suggestions.

[0049] When a pregnant woman carrying twins at 11-14 weeks of gestation enters clinical data including ANXA2, ANXA3, CHIT1, FTL concentrations, age, pre-pregnancy BMI, mean arterial pressure, obstetric history, and method of conception, the system simply performs standardization, splicing, model calculation, probability mapping, and threshold comparison according to a predetermined formula. The final output is only the risk probability value and the high-risk classification result, without providing a diagnostic conclusion, medication guidance, or replacing the professional judgment of a physician.

[0050] This information can be directly used for early risk screening in twin pregnancies, providing standardized and reproducible computational basis for clinical early warning. Using the exact same processing flow and parameter configuration in both the internal modeling cohort and the independent external validation cohort, highly consistent prediction results can be obtained. When running at different times and on different devices, the output results are completely reproducible due to the fixed random seed 20251001 and all hyperparameters being locked. This facilitates unified use, quality control, and result comparison by clinical institutions, and can stably support the standardized implementation of early screening procedures for preeclampsia in twin pregnancies.

[0051] S105 determines the risk probability cutoff threshold based on ROC curves and Youden's index, calculates feature importance with information gain as the core, evaluates using the DeLong method, and completes model performance verification through an internal modeling queue and an independent external verification queue.

[0052] In one implementation, the continuous risk probabilities output by the model are used as input to construct an ROC curve. All probability values ​​are used as candidate cut points, and the sensitivity and specificity corresponding to each cut point are calculated iteratively. The Youden exponent formula is used to calculate the exponent value at each cut point. The Youden exponent formula is as follows: ,in, (Sensitivity) refers to the proportion of pregnant women who actually develop preeclampsia and are correctly identified as positive by the model. Specificity refers to the proportion of pregnant women who do not actually have preeclampsia but are correctly identified as having it by the model. The cutoff point with the highest Youden index is selected as the optimal cutoff threshold to classify high-risk and low-risk groups, achieving the best balance between missed diagnosis and false diagnosis rates. The optimal cutoff threshold calculated by this model is 0.5, at which the sensitivity reaches 0.92 and the specificity reaches 0.94.

[0053] Based on the trained XGBoost model, information gain was used as the core metric to calculate the contribution of each input feature to the prediction result. A higher information gain value indicates a stronger role for the feature in risk differentiation. The contribution values ​​of four protein features and five clinical features were calculated and ranked sequentially to form a feature importance sequence, which was used to explain the model's decision-making basis. Among them, ANXA3 (0.218), early pregnancy mean arterial pressure (0.215), ANXA2 (0.200), and CHIT1 (0.171) were the four features with the highest contribution in the model.

[0054] Substituting the ROC curve data into the DeLong method calculation process, the model's AUC value and corresponding 95% confidence interval were obtained to quantify the model's overall discriminative ability. The AUC of the internal cohort was 0.97, with a 95% confidence interval of 0.93–1.00; the AUC of the external validation cohort was 0.95, with a 95% confidence interval of 0.88–1.00. This method is a non-parametric statistical test, which can stably evaluate the model's predictive performance and provide statistical support for clinical applications. Using all 28 samples from which the model was constructed as the internal cohort, the optimal cutoff threshold was applied to the risk probability, and the model's sensitivity and specificity in this cohort were statistically analyzed. The validation results reflect the model's ability to fit known data, ensuring effective model training and accurate discrimination.

[0055] Twenty-three twin pregnancies with no overlap with the modeling cohort and consistent inclusion and exclusion criteria were selected as an external validation cohort. Preprocessed features were input into the trained model, and the risk probabilities were output, with the same optimal threshold used to determine the risk level. The AUC, sensitivity, and specificity of the external cohort were statistically analyzed to validate the model's predictive ability on novel data, ensuring it possesses the generalization ability required for clinical application. A complete model performance validation report was compiled, summarizing the optimal cutoff threshold, feature importance ranking, AUC and confidence intervals, and internal and external validation metrics.

[0056] S106, based on the risk probability and cutoff threshold, completes the risk level classification output and generates the risk prediction result of preeclampsia in twin pregnancies.

[0057] In one implementation, combining the requirements of high discrimination, low misjudgment, and strong generalization in predicting preeclampsia risk in twin pregnancies with ROC curve evaluation standards, a target cutoff value determination mechanism and risk level classification logic are introduced to determine the core calculation rules for the prediction output stage. An output calculation framework is established for early screening scenarios in twin pregnancies, relying solely on data calculation and not involving disease diagnosis. This framework only standardizes, fuses features, performs model calculations, and performs probability classification on the input protein detection data and clinical indicator data. It does not generate disease diagnosis conclusions or provide medical intervention suggestions throughout the process; all outputs are objective calculation results based on fixed algorithms and parameters. The core calculation rules revolve around risk probability values, including three stages: threshold determination, level classification, and sample-by-sample judgment. The entire process follows the ROC curve evaluation and Youden's index optimization principles to ensure objective and stable output results. In the threshold determination stage, based on the 0-1 risk probability output by the model, all candidate probability values ​​are iterated, using the Youden's index formula: Youden's Index = Sensitivity + Specificity. 1. Calculate the exponent value at each cut point and select the value with the largest exponent as the optimal cutoff threshold. In this model, the optimal threshold is 0.5.

[0058] In the risk classification stage, a fixed binary classification rule is adopted: a risk probability > 0.5 is classified as high risk; a risk probability ≤ 0.5 is classified as low risk. No intermediate levels are set, and no subjective adjustments are made; the classification standard is consistent throughout the process. In the sample-by-sample assessment stage, the risk probability of each sample is directly compared with a fixed threshold, and the level result is output according to preset rules, without manual correction or experience-based adjustments. For example, a twin pregnancy calculated by the model has a risk probability of 0.82, which is greater than 0.5, and the system directly classifies it as high risk; the other twin pregnancy has a probability of 0.37, which is less than or equal to 0.5, and the system directly classifies it as low risk. The entire assessment process relies solely on numerical comparison; the procedure is fixed, the results are reproducible, and it meets the standardized and regulated needs of clinical screening.

[0059] Extract the preeclampsia risk probability values ​​from the model output, determine the target cutoff threshold according to the Youden index maximization principle, and define the judgment boundary for risk level classification. Use the risk probability in the zero-to-one interval of the model output as input, iterate through all candidate probability values, and calculate the corresponding sensitivity and specificity for each. Sensitivity is the proportion of samples that actually have preeclampsia correctly classified as high-risk, and specificity is the proportion of samples that actually do not have preeclampsia correctly classified as low-risk. Calculate the index for each candidate value using the Youden index formula: Youden Index = Sensitivity + Specificity. 1. The value with the largest exponent is selected as the optimal cutoff threshold. Calculations show that the optimal cutoff threshold for this model is 0.5. This threshold is the unique boundary for risk level classification, which can maximize the balance between model sensitivity and specificity. In the internal modeling queue, the sensitivity can reach 0.92 and the specificity can reach 0.94, effectively reducing missed diagnoses and false diagnoses.

[0060] When the candidate threshold is 0.4, sensitivity increases but specificity decreases; when the candidate threshold is 0.6, specificity increases but sensitivity decreases; only when the threshold is 0.5 does the Youden index reach its maximum, at which point the false negative and false positive rates are in optimal balance. This threshold determination rule is based entirely on statistical calculations, without incorporating manual adjustments or subjective experience. It is applicable to both internal modeling cohorts and independent external validation cohorts, ensuring stable, consistent, and reproducible risk grading results.

[0061] Based on the numerical comparison between risk probability and target cutoff threshold, a binary risk classification result is defined, clarifying the corresponding meanings of high risk, low risk, and probability intervals. The zero-to-one risk probability output by the model is used as input, and all candidate probability values ​​are iterated to calculate the corresponding sensitivity and specificity. Sensitivity reflects the model's ability to correctly identify positive preeclampsia cases, while specificity reflects its ability to correctly identify negative healthy cases; both are core evaluation indicators for clinical screening. The Youden index formula is used to calculate the index for each candidate value: Youden index = sensitivity + specificity - 1. The value with the largest index is selected as the optimal cutoff threshold. Calculations show that the optimal cutoff threshold for this model is 0.5. This threshold serves as the unique boundary for risk level classification, maximizing the balance between model sensitivity and specificity. In the internal modeling cohort, the sensitivity reaches 0.92 and the specificity reaches 0.94, effectively reducing missed diagnoses and misdiagnoses in clinical use.

[0062] During the threshold iteration process, if a threshold of 0.4 is used, the sensitivity increases to 0.96 but the specificity decreases to 0.86, which can easily lead to false positives; if a threshold of 0.6 is used, the specificity increases to 0.97 but the sensitivity decreases to 0.85, which can easily lead to false negatives; only when the threshold is 0.5 does the Youden index reach its maximum value, at which point the sensitivity and specificity achieve the optimal balance. This threshold determination method is entirely based on statistical calculations and does not rely on human experience. It can be used in both internal modeling queues and independent external validation queues, ensuring that the risk classification results are stable, reproducible, and comparable.

[0063] By balancing model sensitivity and specificity, two classification rules are established: a probability greater than a threshold indicates high risk, and a probability less than or equal to the threshold indicates low risk. These rules define the criteria for risk classification. The classification rules strictly adhere to numerical comparison logic, without introducing additional conditions, relying on clinical experience for adjustment, or changing the criteria based on the sample source. Each sample is classified solely based on the relationship between its own risk probability and a fixed threshold. The rules remain consistent throughout the process: a risk probability > 0.5 is classified as high risk; a risk probability ≤ 0.5 is classified as low risk.

[0064] One sample, calculated by the model, has a risk probability of 0.73, which is greater than 0.5, and the system directly classifies it as high-risk. Another sample has a probability of 0.41, which is less than or equal to 0.5, and the system directly classifies it as low-risk. Regardless of whether the samples come from the internal modeling cohort or the external validation cohort, the same set of rules and the same threshold are used for judgment, without any differentiation. This judgment standard is applicable to both the internal modeling cohort and the independent external validation cohort, and the threshold or rules are not changed due to different datasets, ensuring that the model maintains stable classification performance on different datasets. In the internal modeling cohort of 28 cases, this rule achieves a sensitivity of 0.92 and a specificity of 0.94; in the independent external validation cohort of 23 cases, it still maintains a sensitivity of 0.90 and a specificity of 0.92, demonstrating stable classification performance and reproducible results, meeting the practical needs of unified, standardized, and comparable clinical screening.

[0065] Based on a numerical comparison classification algorithm, the risk probability and target cutoff threshold are substituted into the grading rules for sample-by-sample calculation to dynamically determine the risk level. This generates risk level and prediction results for preeclampsia in twin pregnancies, eliminating subjective interference and relying solely on the data model output. The system reads the risk probability value for each sample, compares it with the threshold, and outputs the corresponding level according to preset rules. In actual calculations, the system sequentially reads the risk probability in the 0-1 range calculated by the model for each twin pregnancy, directly compares this probability value with the optimal cutoff threshold of 0.5, and strictly outputs a high-risk or low-risk level according to preset rules without any additional corrections. Specifically, if the risk probability > 0.5, it is classified as high-risk; if the risk probability ≤ 0.5, it is classified as low-risk.

[0066] A sample being tested had a calculated risk probability of 0.83. The system compared this to a probability of 0.5; 0.83 > 0.5, therefore it was directly classified as high risk. Another sample had a calculated probability of 0.36; compared to 0.5, 0.36 ≤ 0.5, therefore it was directly classified as low risk. The entire judgment process was completed numerically, without relying on clinical experience or human judgment, without subjective adjustments based on medical history, and without introducing additional parameters. The entire process was executed automatically by a fixed algorithm. The final risk level and prediction results only reflect the risk probability classification results calculated by the model based on four protein features and five clinical features. They do not constitute a disease diagnosis or provide treatment recommendations, but can be directly used for clinical early warning reference, providing medical staff with objective, standardized, and reproducible early risk alerts.

[0067] In one implementation, such as Figure 2 As shown, this application also provides an early warning system for predicting preeclampsia in twin pregnancies using a combination of four factors and clinical indicators in early pregnancy, comprising: The Risk Prediction Requirements and Model Configuration Module 201 is used to obtain the application requirements for predicting the risk of preeclampsia in twin pregnancies, select XGBoost as the model architecture, and set a fixed random seed to ensure that the model training is repeatable. The targeted protein detection and standardization module 202 is used to perform ELISA quantitative detection of ANXA2, ANXA3, CHIT1 and FTL proteins in plasma samples from pregnant women at 11 to 14 weeks of gestation. After natural logarithmic transformation and Z-score standardization, standardized protein characteristic data are generated. The maternal clinical feature preprocessing module 203 is used to perform Z-score standardization on continuous indicators such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy, and to directly encode binary indicators such as obstetric history and conception method to generate standardized maternal clinical feature data. The multi-source feature fusion and model training module 204 is used to fuse 4 protein features and 5 clinical features to construct a 9-dimensional input vector, complete the XGBoost model training according to preset fixed hyperparameters, output the original prediction score through the ensemble tree, and output the risk probability data in the interval of 0 to 1 through the Sigmoid mapping. The model evaluation and threshold determination module 205 is used to determine the risk probability cutoff threshold based on the ROC curve and Youden index, calculate the feature importance with information gain, and complete the model performance verification by adopting the DeLong method and through an internal modeling queue and an independent external verification queue. The risk level determination and result output module 206 is used to classify risk levels based on risk probability and cutoff threshold, generate risk prediction results for preeclampsia in twin pregnancies, and output them uniformly.

[0068] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of the early warning method, electronic device, electronic device, and readable storage medium for assessing the four-factor combined clinical indicators for predicting preeclampsia in twin pregnancies are basically similar to the above-described embodiment of the early warning method for predicting preeclampsia in twin pregnancies, so the description is relatively simple. Relevant parts can be referred to in the description of the above-described embodiment of the early warning method for predicting preeclampsia in twin pregnancies.

Claims

1. A method for predicting early warning of preeclampsia in twin pregnancies using a combination of four clinical indicators in early pregnancy, characterized in that, include: To identify the application requirements for predicting the risk of preeclampsia in twin pregnancies, XGBoost was selected as the model architecture, and a fixed random seed was set to ensure that the model training was repeatable. Plasma samples from pregnant women at 11 to 14 weeks of gestation were quantitatively detected by ELISA for ANXA2, ANXA3, CHIT1, and FTL proteins. The heteroscedasticity was first eliminated by natural logarithmic transformation, and then Z-score standardization was performed to generate standardized protein characteristic data. Z-score standardization was performed directly on continuous indicators such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy. Dichotomous indicators such as obstetric history and method of conception were directly coded without conversion to generate standardized maternal clinical characteristic data. A 9-dimensional input vector is constructed by fusing 4 protein features and 5 clinical features. The XGBoost model is trained according to preset fixed hyperparameters. After the raw prediction score is output through the ensemble tree, the risk probability data in the interval of 0 to 1 is output through the Sigmoid mapping. The risk probability cutoff threshold is determined based on ROC curve and Youden index. The feature importance is calculated with information gain as the core. The DeLong method is used for evaluation, and the model performance is verified through internal modeling queue and independent external verification queue. Based on the risk probability and cutoff threshold, the risk level is classified and output, generating the risk prediction results for preeclampsia in twin pregnancies.

2. The method as described in claim 1, characterized in that, To determine the application requirements for predicting preeclampsia risk in twin pregnancies, XGBoost was selected as the model architecture. A fixed random seed was set to ensure repeatable model training, including: Combining the engineering requirements of early risk prediction of preeclampsia in twin pregnancies with the reproducible deployment of machine learning models, and the technical characteristics of XGBoost ensemble learning, feature standardization, and probability output, this paper introduces three core constraints: fixed random seed, data distribution correction, and fixed hyperparameters, to build a model construction system for early pregnancy risk prediction that is adapted to clinical data. Following the prediction criteria of high sensitivity, high specificity, and strong generalization, the three key processes of logarithmic transformation of protein biomarkers, standardization of clinical features, and 9-dimensional input vector are jointly matched to align the model input with the training process. Based on the balance requirements of prediction accuracy, fitting stability, and overfit suppression, the fixed hyperparameters, number of decision trees, and tree depth of XGBoost are set to complete the training constraints of the prediction model. Input the four protein detection data, maternal clinical characteristic data, and preprocessing parameters into the model system, and output the original prediction score and risk probability according to the training rules; By using a fixed random seed, Sigmoid probability mapping, and ROC target threshold determination, the model training and prediction process is optimized, and a reproducible XGBoost risk prediction model construction strategy suitable for clinical screening of twin pregnancies is generated.

3. The method as described in claim 1, characterized in that, Z-score standardization was directly applied to continuous indicators such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy. Binary indicators such as obstetric history and mode of conception were directly coded without conversion, generating standardized maternal clinical characteristic data, including: Combining the clinical data adaptation needs for predicting the risk of preeclampsia in twin pregnancies with the modeling logic of unified input of multiple types of features, we introduce differentiated processing rules for the standardization of continuous features and the direct input of binary features to build a standardized preprocessing system for maternal clinical features. Following the XGBoost model input preprocessing specifications of same scale and dimensionless interference, the three continuous indicators of age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy were directly standardized by Z-score, while the two binary indicators of obstetric history and conception method were kept in their original encoding, thus completing the scale alignment between clinical features and model input. Based on the core requirements of stable feature distribution and efficient model training convergence, a standardization transformation with a mean of zero and a standard deviation of one is performed on continuous clinical features, while no logarithmic transformation or scaling is performed on binary clinical features, thus completing the feature constraint setting of the preprocessing system. The raw data of maternal age, BMI, mean arterial pressure, obstetric history, and method of conception are input into the preprocessing system. After continuous feature standardization and binary feature direct encoding, a uniform scale clinical feature vector is output. By preprocessing the features to preserve their original semantics and eliminating dimensional differences, standardized parent clinical feature data adapted to the input of the XGBoost model is generated, thus achieving numerical alignment between clinical indicators and model training.

4. The method as described in claim 1, characterized in that, A 9-dimensional input vector is constructed by fusing 4 protein features and 5 clinical features. The XGBoost model is trained using preset fixed hyperparameters. After outputting the raw prediction score via an ensemble tree, risk probability data in the 0-1 interval is output through a Sigmoid mapping, including: In response to the need for multi-source feature fusion modeling for predicting the risk of preeclampsia in twin pregnancies, this study introduces protein biomarker standardization, clinical feature differentiation preprocessing, and XGBoost ensemble learning rules to determine a unified modeling input scheme that combines four proteins with five clinical indicators. Preprocessing was performed on four protein features (ANXA2, ANXA3, CHIT1, and FTL) and five maternal clinical features (age, BMI, mean arterial pressure, obstetric history, and mode of conception) to convert both types of features into a standardized, dimensionless model input format. An XGBoost ensemble learning network driven by fixed hyperparameters was constructed, with fixed network structure and training constraints such as the number of decision trees, maximum depth, learning rate, and sampling rate. The fused 9-dimensional standardized feature vector is input into the XGBoost network. After iterative training with multiple regression trees, node splitting, and feature weighting, the integrated original prediction score is output. The original predicted score is mapped and transformed by the Sigmoid activation function to generate the risk probability of preeclampsia in twins with values ​​in the range of 0 to 1. By integrating preprocessed protein features, clinical features, and model output probabilities, multi-source feature fusion modeling and risk probability calculation information adapted for clinical early warning is generated.

5. The method as described in claim 4, characterized in that, Based on risk probability and cutoff threshold, risk level classification is completed and output, generating a risk prediction result for preeclampsia in twin pregnancies, including: Combining the need for high discrimination, low misjudgment, and strong generalization in predicting the risk of preeclampsia in twin pregnancies with the ROC curve assessment standard, a target cutoff value determination mechanism and risk level classification logic are introduced to determine the core calculation rules for the prediction output stage. Extract the preeclampsia risk probability value from the model output, determine the target cutoff threshold according to the Youden index maximization principle, and define the judgment boundary for risk level classification. Based on the numerical comparison between risk probability and target cutoff threshold, a binary risk level result is defined, clarifying the corresponding meanings of high risk, low risk and probability interval; By combining the model's sensitivity and specificity balance constraints, two classification rules are set: a probability greater than the threshold is considered high risk, and a probability less than or equal to the threshold is considered low risk, thus defining the criteria for risk classification. Based on the numerical comparison classification algorithm, the risk probability and target cutoff threshold are substituted into the grading rules for sample-by-sample calculation to complete the dynamic determination of risk level and generate information on the risk level and prediction results of preeclampsia in twin pregnancies that eliminates subjective interference and is based solely on the data model output.

6. A pre-pregnancy early warning system for twin preeclampsia using a combination of four clinical indicators in early pregnancy, characterized in that, The system includes: The risk prediction requirement and model configuration module is used to obtain the application requirements for predicting the risk of preeclampsia in twin pregnancies, select XGBoost as the model architecture, and set a fixed random seed to ensure that the model training is repeatable. The targeted protein detection and standardization module is used to perform ELISA quantitative detection of ANXA2, ANXA3, CHIT1, and FTL proteins in plasma samples from pregnant women at 11 to 14 weeks of gestation. After natural logarithmic transformation and Z-score standardization, standardized protein characteristic data are generated. The maternal clinical feature preprocessing module is used to perform Z-score standardization on continuous indicators such as age, pre-pregnancy BMI, and mean arterial pressure in early pregnancy, and to directly encode binary indicators such as obstetric history and conception method to generate standardized maternal clinical feature data. The multi-source feature fusion and model training module is used to fuse 4 protein features and 5 clinical features to construct a 9-dimensional input vector, complete the XGBoost model training according to preset fixed hyperparameters, output the original prediction score through the ensemble tree, and output the risk probability data in the interval of 0 to 1 through the Sigmoid mapping. The model evaluation and threshold determination module is used to determine the risk probability cutoff threshold based on the ROC curve and Youden index, calculate the feature importance using information gain, and complete the model performance verification through the DeLong method and an internal modeling queue and an independent external verification queue. The risk level assessment and result output module is used to classify risk levels based on risk probability and cutoff threshold, generate risk prediction results for preeclampsia in twin pregnancies, and output them uniformly.

7. An electronic device, characterized in that, include: First processor; And a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the early warning method for predicting preeclampsia in twin pregnancies using four-factor combined clinical indicators in early pregnancy as described in any one of claims 1 to 5 by executing the executable instructions.

Images