Newborn birth weight prediction method and system based on XGBoost model
Through the XGBoost model, the birth weight prediction method for newborns is predicted by combining ultrasound measurement and parental attribute data, multiple CART trees are constructed and the model is optimized, which solves the accuracy and cost of newborns' weight prediction, and improves robustness and generalization , adapt to multidimensional causal relationship, reduce human dependence, and improve diagnostic accuracy.
Patent Information
- Application Number
- CN202411018259.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-26
AI Technical Summary
The existing neonatal weight prediction methods have problems such as low prediction accuracy, high cost and time-consuming, poor model robustness and generalization, and no parental factors are considered.
A newborn birth weight prediction method based on the XGBoost model was adopted. By obtaining ultrasound measurement data and parent attribute data for preprocessing and splicing, multiple CART trees were constructed and the model was optimized using gradient enhancement algorithm to consider parental factors to make predictions.
It improves the accuracy of neonatal weight prediction, reduces cost and time-consuming, improves the robustness and generalization of the model, adapts to the causal relationship of multidimensional data, reduces human dependence, and improves the diagnostic accuracy.
Smart Images

Figure CN119028599A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of auxiliary measurement and diagnosis of neonatal prenatal examination, and more specifically, relates to a method and system for predicting the birth weight of neonates based on an XGBoost model. Background Art
[0002] The perinatal neonatal mortality rate is closely related to the occurrence of premature babies and intrauterine growth restriction. The weight of the newborn reflects the growth and development of the newborn in the mother's body to a certain extent. Therefore, accurate estimation of the weight of the newborn before delivery can improve the diagnosis rate of macrosomia and low birth weight babies, thereby reducing the risk of maternal delivery, neonatal disability and mortality.
[0003] Among newborns in my country, low birth weight babies account for about 3.2% to 6.3%, and macrosomia accounts for about 7%. The mortality rate of low birth weight babies in the perinatal period is 2 to 5 times that of normal birth weight babies, and they are more likely to suffer from diseases such as neonatal infection, pulmonary dysplasia, hypoglycemia and polycythemia. Macrosomia are more likely to suffer from birth injuries, shoulder dystocia and other conditions during delivery, and there is a certain chance of suffering from diabetes and cardiovascular related diseases in adulthood. Accurately predicting the birth weight of newborns can assist obstetricians in diagnosing low birth weight babies and macrosomia. It is of great significance for decisions such as maternal nutrition management, timing and mode of delivery of newborns. Therefore, the weight of the newborn is crucial in the entire obstetric management and is one of the key indicators for assessing the health status of newborns.
[0004] There are three main methods for predicting newborn weight. The first is a regression model based on ultrasound measurement data proposed by Hadlock et al. in 1985. The second is a weight prediction method based on three-dimensional ultrasound and magnetic resonance imaging. The third is a newborn weight prediction method based on a machine learning model.
[0005] However, the above three methods for predicting newborn weight all have some defects that cannot be ignored:
[0006] First, the existing prediction method of neonatal weight based on regression model has a low prediction accuracy (only 67% to 86%), and most normal newborns will be mistakenly diagnosed as small for gestational age.
[0007] Second, the existing newborn weight prediction methods based on three-dimensional ultrasound and magnetic resonance imaging are not conventional measurement methods, and they are expensive and time-consuming, so they are difficult to be widely promoted and applied;
[0008] Third, although the existing newborn weight prediction methods based on machine learning models have improved the prediction accuracy to a certain extent, most of these methods have only been verified on small-scale data, resulting in poor model robustness and generalization;
[0009] Fourth, the existing newborn weight prediction method based on machine learning model does not take into account the impact of parental factors (such as parental height, weight, etc.), which will reduce the accuracy of weight prediction. Summary of the invention
[0010] In response to the above defects or improvement needs of the prior art, the present invention provides a method and system for predicting the birth weight of a newborn based on an XGBoost model, which aims to solve the technical problems that the existing method for predicting the weight of a newborn based on a regression model has a low prediction accuracy and most normal newborns are mistakenly diagnosed as small for gestational age, and the existing method for predicting the weight of a newborn based on three-dimensional ultrasound and magnetic resonance is not a conventional measurement method, which is expensive and time-consuming, and therefore difficult to be widely promoted and applied, and the existing method for predicting the weight of a newborn based on a machine learning model has a technical problem that most of the methods have only been verified on small-scale data, resulting in poor model robustness and generalization, and the existing method for predicting the weight of a newborn based on a machine learning model does not take into account the influence of parental factors, thereby reducing the accuracy of weight prediction.
[0011] To achieve the above object, according to one aspect of the present invention, a method for predicting the birth weight of a newborn based on an XGBoost model is provided, comprising the following steps:
[0012] (1) Obtain the ultrasound measurement data and parent attribute data of the newborn to be predicted.
[0013] (2) Preprocessing the ultrasound measurement data and parent attribute data of the newborn to be predicted obtained in step (1) respectively, and concatenating the preprocessed ultrasound measurement data and the preprocessed parent attribute data to obtain a comprehensive feature vector of the newborn to be predicted.
[0014] (3) The comprehensive feature vector of the newborn to be predicted obtained in step (2) is input into the pre-trained XGBoost model to obtain the birth weight prediction result of the newborn to be predicted.
[0015] Preferably, the process of preprocessing the ultrasound measurement data of the newborn to be predicted and the parent attribute data obtained in step (1) in step (2) comprises the following sub-steps:
[0016] (2-1) interpolating and smoothing the ultrasonic measurement data obtained in step (1) to obtain preprocessed ultrasonic measurement data (the purpose of which is to fill in missing values and remove outliers);
[0017] (2-2) performing feature extraction on the preprocessed ultrasonic measurement data obtained in step (2-1) to obtain a set of feature vectors;
[0018] (2-3) normalizing the parent attribute data obtained in step (1) to obtain normalized parent attribute data;
[0019] (2-4) Concatenate the set of feature vectors obtained in step (2-2) with the normalized parent attribute data obtained in step (2-3) to obtain a comprehensive feature vector of the newborn.
[0020] Preferably, the XGBoost model is trained by the following steps:
[0021] (3-1) obtaining a data set, wherein the data set includes data of multiple newborns, including ultrasound measurement data of each newborn, parent attribute data of each newborn, and birth weight data of each newborn, wherein each data exists in the form of a feature vector;
[0022] (3-2) preprocessing the data set obtained in step (3-1) to obtain a preprocessed data set, and dividing the preprocessed data set into a training set and a test set in a ratio of 8:2;
[0023] (3-3) Using all the data in the training set obtained in step (3-2), construct K classification and regression trees CART, and all CART trees constitute a CART tree set; where K is the number of CART trees;
[0024] (3-4) For each newborn baby corresponding to the data set obtained in step (3-1), the data of the newborn baby is input into the CART tree corresponding to the feature vector obtained in step (3-3) to obtain a prediction score for the newborn baby corresponding to the CART tree;
[0025] (3-5) Set iteration counter t = 0;
[0026] (3-6) Determine whether the iteration counter t is equal to the preset maximum number of iterations, if so, proceed to step (3-13), otherwise proceed to step (3-7);
[0027] (3-7) For each newborn corresponding to the data set obtained in step (3-1), the predicted value of the i-th data of the newborn at the t-th iteration is obtained according to the prediction scores corresponding to all CART trees obtained in step (3-4);
[0028] (3-8) For each newborn baby corresponding to the data set obtained in step (3-1), the objective function at the t-th iteration is obtained according to the predicted value of the i-th data of the newborn baby at the t-th iteration obtained in step (3-5):
[0029] (3-9) According to the objective function of the t-th iteration obtained in step (3-8), the t-th CART tree f corresponding to the i-th data of the newborn at the t-th iteration is obtained. t (x i (t) );
[0030] (3-10) The tth CART tree f corresponding to the i-th data of the newborn at the t-th iteration obtained in step (3-9) t (x i (t) ) Update the objective function at the t-th iteration obtained in step (3-8) to obtain an updated objective function at the t-th iteration;
[0031] (3-11) performing a second-order Taylor expansion on the objective function at the t-th iteration after the update in step (3-10) to obtain the expanded objective function at the t-th iteration;
[0032] (3-12) Set t = t + 1 and return to step (3-7);
[0033] (3-13) Based on the objective function at the tth iteration after the expansion of step (3-11), the CART tree set obtained in step (3-3) is optimized and solved using a gradient boosting algorithm to obtain an optimal CART tree set;
[0034] (3-14) Combine the optimal CART tree set obtained in step (3-13) to generate a trained XGBoost prediction model.
[0035] Preferably, the ultrasound measurement data in step (3-1) are four routine ultrasound examination data of the mother of the newborn, namely, NT ultrasound examination at 11 to 14 weeks, level III ultrasound examination at 18 to 26 weeks, level II ultrasound examination at 29 to 31 weeks, level I ultrasound examination at 36 weeks, and ultrasound examination within 7 days before birth, involving indicators such as the newborn's head-to-hip length, biparietal diameter, head circumference, abdominal circumference, femur length, and newborn's heart rate;
[0036] The parental attribute data of the newborn include the age, ethnicity, place of origin, gestational age at delivery, number of pregnancies and parities, pre-pregnancy weight, weight and height at each ultrasound examination of the newborn's mother, and weight and height of the newborn's father.
[0037] Preferably, in the process of using data to construct K CART trees in step (3-3), first, the data of all newborns are extracted from the preprocessed training set, and then, according to the dimension of the data of each newborn, a CART tree corresponding to the data is constructed, and all CART trees constitute a CART tree set.
[0038] Preferably, the predicted value of the i-th data of the newborn at the t-th iteration in step (3-7) is:
[0039] where f k ∈τ
[0040] Among them, y i (t) represents the predicted value of the i-th data of the newborn at the t-th iteration, which corresponds to the i-th leaf node in each CART tree, x i (t) represents the i-th data of the newborn at the t-th iteration, τ represents the CART tree set obtained in step (3-3), and f k (x i (t) ) represents the kth CART tree corresponding to the i-th data of the newborn at the t-th iteration, f k represents the kth CART tree, and has i∈[1, the total number of data of the newborn obtained in step (3-1) n], k∈[1, K].
[0041] Preferably, the objective function for obtaining the tth iteration in step (3-8) is obtained by using the following formula:
[0042]
[0043] where y i represents the true value of the i-th data of the newborn at the t-th iteration, represents the mean square error (MSE) function, Ω(f t ) represents the regularization term.
[0044] The mean square error function is specifically:
[0045]
[0046] The regularization term is:
[0047]
[0048] Where γ and λ are regularization parameters, ω i (t) Represents the prediction score of the i-th leaf node in the CART tree at the t-th iteration.
[0049] Preferably, step (3-9) is performed using the following formula:
[0050]
[0051] Steps (3-10) are performed using the following formula:
[0052]
[0053] Where C is a constant term.
[0054] Preferably, step (3-11) is performed using the following formula:
[0055]
[0056] Among them, g i and h i are the first and second order derivatives. Here, g i is the loss function right The first derivative of i is the loss function for The second derivative of .
[0057] According to another aspect of the present invention, a newborn birth weight prediction system based on an XGBoost model is provided, comprising:
[0058] The first module is used to obtain ultrasound measurement data and parent attribute data of the newborn to be predicted.
[0059] The second module is used to preprocess the ultrasound measurement data and the parent attribute data of the newborn to be predicted obtained by the first module, respectively, and splice the preprocessed ultrasound measurement data and the preprocessed parent attribute data to obtain a comprehensive feature vector of the newborn to be predicted;
[0060] The third module is used to input the comprehensive feature vector of the newborn to be predicted obtained in the second module into the pre-trained XGBoost model to obtain the birth weight prediction result of the newborn to be predicted.
[0061] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:
[0062] (1) The present invention adopts step (3-3) and step (3-8), and by constructing and optimizing multiple CART trees, it is possible to fully explore the complex feature relationships in the data and improve the accuracy of prediction. Therefore, it is possible to solve the technical problem that the prediction accuracy of the existing neonatal weight prediction method based on the regression model is low and most normal newborns are mistakenly diagnosed as small for gestational age.
[0063] (2) The present invention adopts step (3-1) and step (3-2), obtains and pre-processes a large amount of newborn data, and uses conventional ultrasound examination data and parent attribute data for prediction, thereby avoiding the high cost and time-consuming problems of three-dimensional ultrasound and magnetic resonance examinations. Therefore, it can solve the technical problem that the existing newborn weight prediction method based on three-dimensional ultrasound and magnetic resonance is not a conventional measurement method, which is expensive and time-consuming, and therefore difficult to be widely promoted and applied;
[0064] (3) The present invention adopts step (3-1) and step (3-2), and improves the robustness and generalization of the model by training the XGBoost model on a large-scale data set. Therefore, it can solve the technical problem that the existing newborn weight prediction method based on machine learning model is mostly verified only on small-scale data, resulting in poor model robustness and generalization.
[0065] (4) The present invention adopts step (3-1) and step (2-3), and improves the accuracy of weight prediction by incorporating parental attribute data into model training. Therefore, it can solve the technical problem that the existing newborn weight prediction method based on machine learning model does not consider the influence of parental factors, thereby reducing the accuracy of weight prediction;
[0066] (5) The XGBoost model established by the present invention performs regression prediction on multidimensional data, finds the causal relationship between the data of newborns and pregnant women at different stages, adapts to the theoretical basis that the weight of newborns is associated with multiple factors, and provides a strong guarantee for the innovation of making more accurate predictions of the weight of newborns;
[0067] (6) The present invention uses a close combination of prenatal ultrasound and artificial intelligence, which can greatly reduce heavy and time-consuming repetitive manual labor and reduce the dependence on the experience of ultrasound doctors, making the diagnosis more accurate, the work process smoother, and the time cost lower. It can significantly enhance medical safety and have huge social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 It is a flow chart of the method for predicting the birth weight of newborns based on the XGBoost model of the present invention. DETAILED DESCRIPTION
[0069] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0070] The basic idea of the present invention is to provide a method for predicting the birth weight of newborns based on the XGBoost model, which realizes accurate prediction of the birth weight of newborns before birth by learning a large amount of newborn birth weight information, prenatal ultrasound examination biometric data from NT to one week before birth, and information such as the height and weight of their parents through the model. After the doctor obtains the prenatal ultrasound measurement indicators of the newborn and the parents' information, the method can accurately predict the birth weight of the newborn. The present invention can solve the technical problems of inaccurate prediction of the birth weight of newborns caused by the existing regression model established based on ultrasound measurement data, such as low prediction accuracy, high misdiagnosis rate, poor robustness, failure to consider parental factors, and insufficient personalization.
[0071] like Figure 1 As shown, the present invention provides a method for predicting the birth weight of a newborn based on an XGBoost model, comprising the following steps:
[0072] (1) Obtaining ultrasound measurement data of the newborn and parental attribute data;
[0073] (2) preprocessing the ultrasound measurement data and the parent attribute data of the newborn to be predicted obtained in step (1) respectively, and concatenating the preprocessed ultrasound measurement data and the preprocessed parent attribute data to obtain a comprehensive feature vector of the newborn to be predicted;
[0074] The advantage of this step (1) is that, through data preprocessing and feature concatenation, outliers can be effectively removed and missing values can be filled, thereby improving the quality and consistency of the data.
[0075] Specifically, in this step, the ultrasound measurement data of the newborn to be predicted and the parent attribute data obtained in step (1) are preprocessed separately. This process includes the following sub-steps:
[0076] (2-1) interpolating and smoothing the ultrasonic measurement data obtained in step (1) to obtain preprocessed ultrasonic measurement data (the purpose of which is to fill in missing values and remove outliers);
[0077] (2-2) performing feature extraction on the preprocessed ultrasonic measurement data obtained in step (2-1) to obtain a set of feature vectors;
[0078] (2-3) normalizing the parent attribute data obtained in step (1) to obtain normalized parent attribute data;
[0079] (2-4) concatenating the set of feature vectors obtained in step (2-2) with the normalized parent attribute data obtained in step (2-3) to obtain a comprehensive feature vector of the newborn;
[0080] The advantage of the above sub-steps (2-1) to (2-4) is that, through data preprocessing and feature concatenation, the quality and consistency of the data can be effectively improved, providing high-quality input for subsequent machine learning models.
[0081] (3) The comprehensive feature vector of the newborn to be predicted obtained in step (2) is input into the pre-trained XGBoost model to obtain the birth weight prediction result of the newborn to be predicted.
[0082] Specifically, the XGBoost model of the present invention is trained by the following steps:
[0083] (3-1) obtaining a data set, wherein the data set includes data of multiple newborns, including ultrasound measurement data of each newborn, parent attribute data of each newborn, and birth weight data of each newborn, wherein each data exists in the form of a feature vector;
[0084] Specifically, the ultrasound measurement data are the data of four routine ultrasound examinations of the newborn's mother (NT ultrasound examination from 11 to 14 weeks, level III ultrasound examination from 18 to 26 weeks, level II ultrasound examination from 29 to 31 weeks, level I ultrasound examination at 36 weeks and ultrasound examination within 7 days before birth), involving indicators such as the newborn's head-to-hip length, biparietal diameter, head circumference, abdominal circumference, femur length, and newborn heart rate; the newborn's parental attribute data include the newborn's mother's age, ethnicity, native place, gestational age at delivery, pregnancy and parity, pre-pregnancy weight, weight and height at each ultrasound examination, as well as the weight and height of the newborn's father.
[0085] (3-2) preprocessing the data set obtained in step (3-1) to obtain a preprocessed data set, and dividing the preprocessed data set into a training set and a test set in a ratio of 8:2;
[0086] The process of preprocessing the training data set in this step is exactly the same as step (2) above, and will not be repeated here.
[0087] The advantage of the above steps (3-1) to (3-2) is that by acquiring and preprocessing large-scale newborn and parent data, rich and diverse data can be provided for model training, thereby improving the robustness and generalization of the model.
[0088] (3-3) Using all the data in the training set obtained in step (3-2), construct K classification and regression trees (CART for short), and all CART trees constitute a CART tree set; wherein K is the number of CART trees, and its value range is 1 to 1000, and preferably 200;
[0089] Specifically, the process of using data to construct K CART trees is to first extract the data of all newborns from the preprocessed training set (which includes ultrasound measurement data of newborns, parent attribute data of all newborns, and birth weight data of all newborns), and then, based on the dimension of each newborn's data, construct a CART tree corresponding to the data. All CART trees constitute a CART tree set.
[0090] The advantage of this step (3-3) is that by constructing multiple CART trees, the complex feature relationships in the data can be fully explored to improve the accuracy of the prediction.
[0091] (3-4) For each newborn baby corresponding to the data set obtained in step (3-1), the data of the newborn baby is input into the CART tree corresponding to the feature vector obtained in step (3-3) to obtain a prediction score for the newborn baby corresponding to the CART tree;
[0092] (3-5) Set iteration counter t = 0;
[0093] (3-6) Determine whether the iteration counter t is equal to the preset maximum number of iterations, if so, proceed to step (3-13), otherwise proceed to step (3-7);
[0094] Specifically, the maximum number of iterations preset in the present invention is equal to 500;
[0095] (3-7) For each newborn corresponding to the data set obtained in step (3-1), the predicted value of the i-th data of the newborn at the t-th iteration is obtained according to the prediction scores corresponding to all CART trees obtained in step (3-4);
[0096] Specifically, the predicted value of the i-th data of the newborn at the t-th iteration is:
[0097] where f k ∈τ
[0098] Among them, y i (t) represents the predicted value of the i-th data of the newborn at the t-th iteration, which corresponds to the i-th leaf node in each CART tree, x i (t) represents the i-th data of the newborn at the t-th iteration, τ represents the CART tree set obtained in step (3-3), and f k (x i (t) ) represents the kth CART tree corresponding to the i-th data of the newborn at the t-th iteration, f krepresents the kth CART tree, with i∈[1, the total number of newborn data obtained in step (3-1) n], k∈[1, K];
[0099] (3-8) For each newborn corresponding to the data set obtained in step (3-1), the objective function at the t-th iteration is obtained according to the predicted value of the i-th data of the newborn at the t-th iteration obtained in step (3-5) (the objective function is used to evaluate and optimize the performance of the CART tree set):
[0100]
[0101] where y i represents the true value of the i-th data of the newborn at the t-th iteration, represents the mean square error (MSE) function, Ω(f t ) represents the regularization term;
[0102] Specifically, the mean square error function is:
[0103]
[0104] The regularization term is defined as:
[0105]
[0106] Where γ and λ are regularization parameters, ω i (t) Represents the prediction score of the i-th leaf node in the CART tree at the t-th iteration;
[0107] (3-9) According to the objective function of the t-th iteration obtained in step (3-8), the t-th CART tree f corresponding to the i-th data of the newborn at the t-th iteration is obtained. t (x i (t) );
[0108] Specifically, this step uses the following formula:
[0109]
[0110] (3-10) The tth CART tree f corresponding to the i-th data of the newborn at the t-th iteration obtained in step (3-9) t (x i (t) ) Update the objective function at the t-th iteration obtained in step (3-8) to obtain an updated objective function at the t-th iteration;
[0111] Specifically, this step uses the following formula:
[0112]
[0113] Where C is a constant term;
[0114] (3-11) performing a second-order Taylor expansion on the objective function at the t-th iteration after the update in step (3-10) to obtain the expanded objective function at the t-th iteration;
[0115] Specifically, this step uses the following formula:
[0116]
[0117] Among them, g i and h i are the first and second order derivatives. Here, g i is the loss function right The first derivative of i is the loss function for The second derivative of
[0118] (3-12) Set t = t + 1 and return to step (3-7);
[0119] (3-13) Based on the objective function at the tth iteration after the expansion of step (3-11), the CART tree set obtained in step (3-3) is optimized and solved using a gradient boosting algorithm to obtain an optimal CART tree set;
[0120] (3-14) Combine the optimal CART tree set obtained in step (3-13) to generate a trained XGBoost prediction model.
[0121] The advantage of the above steps (3-1) to (3-14) is that, through systematic data processing and model optimization, an efficient, robust and accurate newborn weight prediction model can be constructed.
[0122] Specifically, the process of constructing the CART tree of the present invention first obtains a large number of comprehensive data sets of newborns and their parents, pre-processes the data sets, and divides them into training sets and test sets. Then, K CART trees are constructed based on features of different dimensions, and each tree is trained. Subsequently, the feature vectors in the training data set are input into each CART tree, and the prediction score of each tree is calculated. Thereafter, the prediction scores of all CART trees are added, the optimization loss function is defined, and a new CART tree is introduced in each iteration, the Taylor expansion approximation is performed on the objective function, and a regularization term is added to optimize the objective function. Finally, the gradient boosting algorithm is used for optimization and solution to obtain the optimal CART tree set, thereby forming a complete XGBoost prediction model.
[0123] Through the above steps, the XGBoost model for predicting the birth weight of newborns can be trained. This model can fully explore the complex feature relationships in the data and improve the accuracy of prediction by constructing multiple CART trees and optimizing the tree set using the gradient boosting algorithm. In actual application, the feature vector of the newborn to be predicted is input into the trained XGBoost model to quickly obtain the predicted value of birth weight.
[0124] Test Results
[0125] The present invention obtained data of 1158 qualified newborns and divided them into a training set and a test set in a ratio of 8:2, and then input these data into the XGBoost model for training and performance verification.
[0126] To evaluate the accuracy of the model, the present invention used the following criteria: prediction error less than 10%, prediction error less than 20%, prediction error within 250 grams, and prediction error within 500 grams. The results showed that the model used in the present invention has a high accuracy rate in prenatal prediction of newborn birth weight.
[0127] Table 1 Accuracy indicators
[0128]
[0129] In general, the beneficial effects of the present invention are as follows:
[0130] (1) By training the XGBoost model to learn multidimensional data such as prenatal ultrasound and parental attributes, the intrinsic correlation between different dimensions was mined, significantly improving the accuracy of birth weight prediction;
[0131] (2) The XGBoost algorithm was used for regression prediction, which made full use of the dynamic data of newborns at different gestational weeks and mothers at different stages, established the causal relationship between relevant factors and birth weight, and laid the foundation for personalized prediction;
[0132] (3) Applying the trained model to prediction greatly reduces the workload of obstetricians, improves diagnostic efficiency, and reduces dependence on physician experience, which has significant social and economic benefits.
[0133] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting birth weight of newborns based on XGBoost model, characterized in that: The following steps are involved: (1) Obtain the ultrasound measurement data and parent attribute data of the newborn to be predicted. (2) Preprocessing the ultrasound measurement data and parent attribute data of the newborn to be predicted obtained in step (1) respectively, and concatenating the preprocessed ultrasound measurement data and the preprocessed parent attribute data to obtain a comprehensive feature vector of the newborn to be predicted. (3) The comprehensive feature vector of the newborn to be predicted obtained in step (2) is input into the pre-trained XGBoost model to obtain the birth weight prediction result of the newborn to be predicted.
2. The method for predicting birth weight of newborns based on the XGBoost model according to claim 1, characterized in that: In step (2), the ultrasound measurement data of the newborn to be predicted and the attribute data of the parents obtained in step (1) are preprocessed respectively. This process includes the following sub-steps: (2-1) interpolating and smoothing the ultrasonic measurement data obtained in step (1) to obtain preprocessed ultrasonic measurement data; (2-2) performing feature extraction on the preprocessed ultrasonic measurement data obtained in step (2-1) to obtain a set of feature vectors; (2-3) normalizing the parent attribute data obtained in step (1) to obtain normalized parent attribute data; (2-4) Concatenate the set of feature vectors obtained in step (2-2) with the normalized parent attribute data obtained in step (2-3) to obtain a comprehensive feature vector of the newborn.
3. The method for predicting the birth weight of newborns based on the XGBoost model according to claim 1 or 2, characterized in that: The XGBoost model is trained by following the steps below: (3-1) obtaining a data set, wherein the data set includes data of multiple newborns, including ultrasound measurement data of each newborn, parent attribute data of each newborn, and birth weight data of each newborn, wherein each data exists in the form of a feature vector; (3-2) preprocessing the data set obtained in step (3-1) to obtain a preprocessed data set, and dividing the preprocessed data set into a training set and a test set in a ratio of 8:2; (3-3) Using all the data in the training set obtained in step (3-2), construct K classification and regression trees CART, and all CART trees constitute a CART tree set; where K is the number of CART trees; (3-4) For each newborn baby corresponding to the data set obtained in step (3-1), the data of the newborn baby is input into the CART tree corresponding to the feature vector obtained in step (3-3) to obtain a prediction score for the newborn baby corresponding to the CART tree; (3-5) Set iteration counter t = 0; (3-6) Determine whether the iteration counter t is equal to the preset maximum number of iterations, if so, proceed to step (3-13), otherwise proceed to step (3-7); (3-7) For each newborn corresponding to the data set obtained in step (3-1), the predicted value of the i-th data of the newborn at the t-th iteration is obtained according to the prediction scores corresponding to all CART trees obtained in step (3-4); (3-8) For each newborn baby corresponding to the data set obtained in step (3-1), the objective function at the t-th iteration is obtained according to the predicted value of the i-th data of the newborn baby at the t-th iteration obtained in step (3-5): (3-9) According to the objective function of the t-th iteration obtained in step (3-8), the t-th CART tree f corresponding to the i-th data of the newborn at the t-th iteration is obtained. t (x i (t) ); (3-10) The tth CART tree f corresponding to the i-th data of the newborn at the t-th iteration obtained in step (3-9) t (x i (t) ) Update the objective function at the t-th iteration obtained in step (3-8) to obtain an updated objective function at the t-th iteration; (3-11) performing a second-order Taylor expansion on the objective function at the t-th iteration after the update in step (3-10) to obtain the expanded objective function at the t-th iteration; (3-12) Set t = t + 1 and return to step (3-7); (3-13) Based on the objective function at the tth iteration after the expansion of step (3-11), the CART tree set obtained in step (3-3) is optimized and solved using a gradient boosting algorithm to obtain an optimal CART tree set; (3-14) Combine the optimal CART tree set obtained in step (3-13) to generate a trained XGBoost prediction model.
4. The method for predicting birth weight of newborns based on the XGBoost model according to any one of claims 1 to 3, characterized in that: The ultrasound measurement data in step (3-1) are four routine ultrasound examination data of the mother of the newborn, namely, NT ultrasound examination from 11 to 14 weeks, level III ultrasound examination from 18 to 26 weeks, level II ultrasound examination from 29 to 31 weeks, level I ultrasound examination at 36 weeks, and ultrasound examination within 7 days before birth, involving indicators such as the newborn's head-to-hip length, biparietal diameter, head circumference, abdominal circumference, femur length, and newborn's heart rate; The parental attribute data of the newborn include the age, ethnicity, place of origin, gestational age at delivery, number of pregnancies and parities, pre-pregnancy weight, weight and height at each ultrasound examination of the newborn's mother, and weight and height of the newborn's father.
5. The method for predicting birth weight of newborns based on the XGBoost model according to claim 4, characterized in that: In the process of using data to construct K CART trees in step (3-3), first, the data of all newborns are extracted from the preprocessed training set, and then, according to the dimension of each newborn's data, a CART tree corresponding to the data is constructed, and all CART trees constitute a CART tree set.
6. The method for predicting birth weight of newborns based on the XGBoost model according to claim 5, characterized in that: The predicted value of the i-th data of the newborn at the t-th iteration in step (3-7) is: where f k ∈τ Among them, y i (t) represents the predicted value of the i-th data of the newborn at the t-th iteration, which corresponds to the i-th leaf node in each CART tree, x i (t) represents the i-th data of the newborn at the t-th iteration, τ represents the CART tree set obtained in step (3-3), and f k (x i (t) ) represents the kth CART tree corresponding to the i-th data of the newborn at the t-th iteration, f k represents the kth CART tree, and has i∈[1, the total number of data of the newborn obtained in step (3-1) n], k∈[1, K].
7. The method for predicting birth weight of newborns based on the XGBoost model according to claim 6, characterized in that: The objective function for the tth iteration in step (3-8) is obtained using the following formula: where y i represents the true value of the i-th data of the newborn at the t-th iteration, represents the mean square error (MSE) function, Ω(f t ) represents the regularization term. The mean square error function is specifically: The regularization term is: Where γ and λ are regularization parameters, ω i (t) Represents the prediction score of the i-th leaf node in the CART tree at the t-th iteration.
8. The method for predicting birth weight of newborns based on the XGBoost model according to claim 7, characterized in that: Steps (3-9) are performed using the following formula: Steps (3-10) are performed using the following formula: Where C is a constant term.
9. The method for predicting birth weight of newborns based on the XGBoost model according to claim 8, characterized in that: Step (3-11) is to use the following formula: Among them, g i and h i are the first and second order derivatives. Here, g i is the loss function right The first derivative of i is the loss function for The second derivative of .
10. A newborn birth weight prediction system based on XGBoost model, characterized in that: include: The first module is used to obtain ultrasound measurement data and parent attribute data of the newborn to be predicted. The second module is used to preprocess the ultrasound measurement data and the parent attribute data of the newborn to be predicted obtained by the first module, respectively, and splice the preprocessed ultrasound measurement data and the preprocessed parent attribute data to obtain a comprehensive feature vector of the newborn to be predicted; The third module is used to input the comprehensive feature vector of the newborn to be predicted obtained in the second module into the pre-trained XGBoost model to obtain the birth weight prediction result of the newborn to be predicted.
