Method for predicting overweight obesity of children based on machine learning
By constructing a machine learning model to screen key variables and integrating early pregnancy maternal physical examination and child follow-up data, and adopting a multi-stage modeling approach, the problem of insufficient accuracy and interpretability in predicting childhood obesity in existing technologies has been solved. This has enabled early identification of high-risk children and personalized intervention, and promoted data interconnection in the prevention and control of childhood obesity.
Patent Information
- Application Number
- CN202511304599.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-12
- Publication Date
- 2025-12-30
AI Technical Summary
Existing technologies struggle to simultaneously integrate early pregnancy maternal examination data with multiple follow-up information from children for phased modeling, resulting in insufficient predictive accuracy and interpretability. Furthermore, the lack of fusion analysis of multi-source heterogeneous data leads to missed diagnoses in high-risk groups.
A machine learning model was constructed to screen key variables. Maternal physical examination data in early pregnancy and child follow-up data were integrated. Random forest and extreme gradient boosting algorithms were used to evaluate feature importance. A child obesity risk scoring system was established by combining logistic regression, random forest, XGBoost and neural network algorithms. Key features were screened and risk prediction was performed through a multi-stage model.
It enables early identification and personalized intervention of the risk of overweight/obesity in children, breaking through the limitations of traditional single-time-point prediction, improving the accuracy and interpretability of prediction, meeting the requirements of Healthy China 2030 for the prevention and control of childhood obesity, and promoting the interconnection of data between prenatal care and child health care.
Smart Images
Figure CN121237403A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of artificial intelligence and public health, and in particular to a method for predicting childhood overweight and obesity based on machine learning. BACKGROUND
[0002] Childhood obesity, a complex metabolic disease triggered by multi-gene-environment interactions, has been listed by the World Health Organization as a serious challenge facing the global public health system in the 21st century. According to the data of the China Students' Physical Health Research, from 1985 to 2019, the obesity rate of children and adolescents aged 7 to 18 increased from 0.1% to 9.6%, an increase of 75.6 times. At the same time, data from the China National Nutrition Survey showed that the obesity rate of children and adolescents was 2.1% in 2002, and increased to 7.9% in 2016-2017. This change reflects the intensification of the obesity problem, and it is estimated that by 2030, the prevalence of childhood obesity will reach 15.1%, which will pose a huge challenge to China's health system. This makes the construction of an early warning system a core breakthrough in prevention and control strategies.
[0003] In current clinical practice, obesity risk assessment mainly relies on WHO Child Growth Standards Z-score and BMI percentile threshold method. However, such methods have three limitations: (1) Based on single time point anthropometric indicators (such as weight, height), they cannot integrate maternal metabolic parameters during pregnancy (such as early pregnancy OGTT glucose values) and dynamic development trajectories during infancy; (2) Traditional logistic regression models have insufficient modeling capabilities for non-linear interactions between variables (such as the synergistic effect of maternal BMI and complementary food introduction time), leading to missed diagnosis of high-risk groups; (3) Existing prediction systems mostly use a single data modality, lacking fusion analysis of multi-source heterogeneous data (antenatal electronic medical records, child health follow-up records, family environment questionnaires), making it difficult to capture the cross-stage transmission effects of key risk factors.
[0004] Although machine learning techniques have shown significant advantages in handling high-dimensional feature interactions, existing research still has methodological flaws: first, feature engineering is mostly limited to a single life stage, such as using only birth weight or school-age dietary data, ignoring the temporal association between maternal lipid metabolism indicators at 22 weeks of pregnancy and intestinal flora characteristics at 24 months of age; second, the model architecture lacks clinical interpretability, black box models such as deep neural networks can achieve a prediction performance of AUC 0.82, but cannot output a feature contribution ranking that can guide clinical intervention; third, there is a serious selection bias at the data governance level, existing public datasets mostly lack key covariates (such as paternal obesity history, preschool dietary structure), leading to performance degradation when externally validating the model. SUMMARY
[0005] The present application aims at the problem that it is difficult to simultaneously integrate maternal physical examination data in early pregnancy and multiple follow-up information of children for phased modeling in the prior art, and the prediction accuracy and interpretability are insufficient, and provides a method for predicting childhood overweight and obesity based on machine learning.
[0006] In order to solve the above technical problems, the present application is solved by the following technical solutions:
[0007] The method for predicting childhood overweight and obesity based on machine learning comprises:
[0008] Step 1: constructing a machine learning model for predicting whether the offspring is a large baby at birth, and screening key variables with predictive value based on the machine learning model of large babies;
[0009] Step 2: fusing the key variables with the childhood follow-up data to establish a prediction model of early childhood overweight risk;
[0010] Step 3: forming a childhood obesity risk scoring system based on the key variables of step 1 and the prediction model of step 2.
[0011] As preferred: it further comprises data collection and data processing;
[0012] The data collection collects maternal physical examination indicators in early pregnancy, birth information of children and childhood follow-up data through a big data platform system;
[0013] The data processing processes the collected data for missing values, abnormal values and duplicate records, and outputs the processed data set.
[0014] As preferred: in step 1, the machine learning model for predicting whether the offspring is a large baby at birth is constructed, and the key variables with predictive value are screened based on the machine learning model of large babies, which comprises:
[0015] The processed data is randomly divided into a training set and a test set;
[0016] Random forest and extreme gradient boosting XGBoost algorithm are used to construct a large baby prediction model on the training set;
[0017] The prediction performance of each model is evaluated on the test set by cross-validation calculation, and the importance score of the candidate features is extracted and quantified by calling plot_importance.
[0018] For feature importance ranking, the variables ranked in the top three in both random forest and XGBoost are selected as key variables.
[0019] As preferred: the key variables include maternal BMI in early pregnancy, gestational weeks and child gender.
[0020] As preferred: in step 2, the key variables are fused with the childhood follow-up data to establish a prediction model of early childhood overweight risk, including:
[0021] The fused data is randomly divided into a training set, a test set and a validation set;
[0022] In the training set, logistic regression, random forest, XGBoost and neural network algorithms are used respectively, and 10-fold cross-validation is used in the training process to complete model fitting;
[0023] The performance of the model is monitored in real time through the set indicators, and the performance is evaluated on the test set.
[0024] As preferred: the childhood follow-up data includes child sleep parameters, child feeding method information, child growth data and child growth environment information.
[0025] As preferred: it also includes verification of the child scoring system, and the scoring system formed in step 3 is verified by the child scoring system;
[0026] Threshold setting, statistical verification of the prediction probability distribution of the obese group samples in the validation set, extraction of the minimum value Pmin of the prediction probability distribution and the median Pmedian of the prediction probability distribution;
[0027] The minimum value Pmin of the probability distribution is used as the first dividing line, and all children with a prediction probability lower than the minimum value Pmin of the probability distribution are judged as non-obese in the validation set;
[0028] The median Pmedian of the prediction probability distribution is used as the second dividing line to divide the remaining individuals into medium risk and high risk.
[0029] In order to solve the above technical problems, the present application also provides a system for predicting childhood overweight and obesity based on machine learning, which is used to realize the method for predicting childhood overweight and obesity based on machine learning, comprising:
[0030] A key variable screening module for constructing a machine learning model for predicting whether the offspring is a large child at birth, and screening key variables with predictive value based on the machine learning model for large children;
[0031] An early overweight risk prediction model establishment module for fusing key variables with childhood follow-up data to establish a prediction model of early childhood overweight risk;
[0032] A child obesity risk scoring system generation module for forming a child obesity risk scoring system based on the key variables screened by the key variable screening module and the prediction model established by the early overweight risk prediction model establishment module.
[0033] The present application has the following technical effects:
[0034] The present application integrates health data of three key stages of early pregnancy (fetal period), birth (perinatal period) and 0-60 months (growth and development period), and breaks through the limitations of traditional single time point prediction.
[0035] The method used in the present application is used for early identification of high-risk overweight / obese children, and realizes early detection, early intervention and early prevention of children's overweight / obesity risk. It meets the requirements of health China 2030 child obesity prevention and control, can promote the data interconnection of prenatal care and child care, and has high clinical application value and popularization prospect.
[0036] The present application realizes early identification and individualized intervention of high-risk children by screening key features and risk prediction in turn through a multi-stage model, and helps precision medicine. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 is the overall technical flowchart of the child overweight prediction method based on machine learning of the present application.
[0038] Figure 2 is the feature importance ranking chart of the present application for constructing a large baby prediction model on the training set by using XGBoost algorithm.
[0039] Figure 3 is the large baby prediction model feature importance ranking chart of the present application based on the Gini impurity average reduction method of random forest algorithm to calculate feature weight.
[0040] Figure 4 is the performance comparison chart of XGBoost model (AUC=0.70) and random forest model (AUC=0.67) on the test set in the construction of the large baby model of the present application.
[0041] Figure 5 is the performance comparison chart of XGBoost model in the child follow-up (0-5 years old) obesity prediction model of the present application on the training set (AUC=0.85) and the test set (AUC=0.81).
[0042] Figure 6 is the SHAP feature importance ranking in the child follow-up (0-5 years old) obesity prediction model of the present application.
[0043] Figure 7 is the child obesity three-level risk score system chart constructed by the present application.
[0044] Figure 8 is the child obesity three-level risk score system chart constructed by the present application. DETAILED DESCRIPTION
[0045] The application will be further described in conjunction with the accompanying drawings and examples.
[0046] Examples
[0047] Method for predicting overweight and obesity of children based on machine learning, Figure 1 As shown in the figure, the overall workflow includes
[0048] Data collection and dataset construction: Collect maternal physical examination indicators in early pregnancy (9-14 weeks), child birth information and 0-5 year follow-up data, and construct an original sample set containing input features and output labels.
[0049] Data preprocessing: Null value marking, missing value imputation, outlier correction, discrete variable label encoding and continuous variable standardization are performed on the original data.
[0050] Phase one feature screening and modeling: Based on early pregnancy maternal physical examination and birth information, a model for predicting macrosomia is trained, and key variables are screened by combining feature importance evaluation methods.
[0051] Phase two risk prediction model construction: The key variables screened in phase one are combined with the 0-5 year follow-up information of children to train a model for predicting the risk of overweight / obesity of children during follow-up.
[0052] Result output and risk stratification: The target children are predicted, individualized risk scores are output, and expert opinions are combined to divide them into low, medium and high risk levels.
[0053] First, after completing data collection, dataset construction and data preprocessing, the child obesity status is labeled by body mass index age Z score (BMI-for-age Z score) according to the 0-5 year child growth standard published by the World Health Organization (WHO) in 2006.
[0054] Definition of overweight / obesity in children aged 0-5 years:
[0055] Obesity determination criteria: When the body mass index age Z score of a child exceeds +3 standard deviations (SD), it is determined to be obese.
[0056] Overweight determination criteria: When the body mass index age Z score of a child is in the +2SD to +3SD interval, it is determined to be overweight.
[0057] All individuals with a body mass index age Z score greater than +2SD are labeled as "overweight / obese children group" (Class = 1), and the rest are labeled as "non-overweight / obese children group" (Class = 0).
[0058] The cleaned data was randomly divided into training set (70%) and test set (30%); Random Forest and XGBoost algorithm were used to build the macrosomia prediction model on the training set, such as Figure 2 and Figure 3 The AUC (Area Under receiver operating characteristic Curve) was calculated by cross-validation, and the prediction performance of each model was evaluated on the test set, such as Figure 4 The importance score of the candidate features was extracted and quantified by calling plot_importance;
[0059] The feature importance ranking of the two algorithms was integrated, and the variables that ranked in the top three in both Random Forest and XGBoost—maternal BMI in early pregnancy, gestational age, and child gender—were selected as the core input for the construction of the second-stage prediction model;
[0060] Descriptive statistical analysis was performed on the distribution differences of the three core features—maternal BMI in early pregnancy, gestational age, and child gender—between the "overweight / obese children group" and the "non-overweight / obese children group" in the baseline cohort (n=12154) (Table 2). The continuous variables were represented by mean ± standard deviation (mean ± SD), and the categorical variables were presented by sample frequency (n) and corresponding percentage. There were significant differences between the two groups.
[0061] Table 1 Basic information of the born infants
[0062]
[0063] To verify the discriminant ability of these three variables between the macrosomia group and the non-macrosomia group, the inter-group difference test was performed for each variable. Welch's t-test, equal variance independent sample t-test, and chi-square test were used for analysis:
[0064] Due to the non-homogeneity of variance of the two types of samples in gestational age, Welch's t-test was used, and p=2.13x10 -70 ; Since the assumption of homogeneity of variance of maternal BMI in early pregnancy between the two groups was established, the independent sample t-test with var.equal=TRUE was used, and p=1.72x10 -29 ; Gender is a binary categorical variable, and chi-square test is used, and p=9.31x10 -44 The above tests are all highly statistically significant, further proving the discriminant ability and practical application value of these three indicators.
[0065] The second-stage model construction and evaluation were performed. Logistic regression, Random Forest, XGBoost, and neural network algorithms were used to predict the risk of obesity in children during follow-up:
[0066] The input variables of the model consist of two parts: one is the three core features screened in the first stage, namely, early pregnancy BMI, gestational age, and child gender; the other is the multi-dimensional lifestyle variables reflected in the childhood follow-up questionnaire, covering early childhood sleep patterns, feeding methods, and living environment, so as to comprehensively depict multiple factors affecting the risk of childhood obesity.
[0067] Based on the feature value summary between groups of obese and non-obese, the specific description is as follows:
[0068] Table 2 Key feature variable information
[0069]
[0070]
[0071] After the overall input framework of the second stage model was established, the data was randomly divided into a training set (70%), a test set (15%), and a validation set (15%);
[0072] In the training set, logistic regression, random forest, XGBoost, and neural network algorithms were used, and 10-fold cross-validation was implemented during the training process to complete model fitting;
[0073] During the training process, the model performance was monitored in real time in this embodiment by AUC, accuracy, and F1 score, and the overall performance was evaluated on the test set. The results showed that the AUC of the XGBoost model on the training set and the test set reached 0.85 and 0.81 (see Figure 5 ), respectively, and its accuracy (74%) and F1 score (0.73) were also better than those of other models, showing the most stable prediction ability;
[0074] To further quantify the contribution of each feature, the SHAP (SHapley Additive exPlanations) method was used for global and local interpretation of the final XGBoost model. By calculating the marginal contribution of each variable to the prediction result of a single sample, and based on the sample average absolute SHAP value, a global importance ranking was generated to quantify the contribution of each feature as Figure 6 .
[0075] In addition, to evaluate the independent predictive ability of each indicator, a classification model was built based on XGBoost for the obesity status of children (obese / non-obese) with only a single feature as input, respectively. The experimental results show that whether the body weight growth rate, feeding mode, or sleep parameters are used as a single feature for prediction, the model is significantly weaker in AUC, accuracy, and other evaluation indicators than the joint prediction model that takes all variables into account. The above comparison not only provides intuitive evidence for the performance improvement of the multivariate model, but also further verifies the necessity and effectiveness of integrating all candidate features into the model.
[0076] The statistical results of each variable are summarized as follows.
[0077] Table 3 Single feature prediction results
[0078]
[0079]
[0080] Finally, the child obesity risk scoring system was constructed. Based on the individualized risk probability output by the stage two model, combined with the suggestions of clinical experts, a low / medium / high three-level scoring system was constructed in the validation set (15%), aiming to ensure that no obese children appear in the low risk group, while accurately identifying individuals who need intensive intervention to improve the efficiency of precise intervention, such as Figure 7 and Figure 8 The specific steps are as follows:
[0081] Threshold setting: First, the prediction probability distribution of the "obese group" (Class = 1) samples in the validation set is counted, and the minimum value P_min and the median P_median are extracted;
[0082] Take P_min as the first dividing line to ensure that all children with a prediction probability below this value are judged as non-obese (Class = 0) in the validation set, thereby achieving 100% specificity (Negative Predictive Value);
[0083] Take P_median as the second dividing line to further distinguish the remaining individuals into medium and high risk. Finally, a "risk level report" is generated to provide a low / medium / high three-level warning for clinical and family, aiming to accurately identify high-risk individuals who need intensive intervention, while avoiding excessive intervention on truly low-risk children.
[0084] Finally, the effectiveness verification is performed:
[0085] The bar chart in the validation set confirmed that only non-obese individuals existed in the low-risk group (i.e., 0% of obese children); 9.4% of obese children in the medium-risk level, and 30.5% of obese children in the high-risk group, which was significantly higher than the baseline obesity rate (9.6%). The sensitivity, specificity, and negative predictive value of each risk level were calculated to further evaluate the stratification effect.
[0086] The process and method shown in this embodiment can be flexibly adjusted according to the needs of different populations and medical institutions to meet the application scenarios of precision medicine and public health management.
Claims
1. A method for predicting childhood overweight and obesity based on machine learning, the method comprising: Step 1: constructing a machine learning model for predicting whether the offspring is a large-for-date baby at birth, and screening key variables with predictive value based on the machine learning model for large-for-date babies; Step 2: fusing the key variables with childhood follow-up data to establish a prediction model for early childhood overweight risk; Step 3: forming a childhood obesity risk scoring system based on the key variables of Step 1 and the prediction model of Step 2. 2.The method of predicting childhood overweight and obesity based on machine learning according to claim 1, characterized in that: Also included are data collection and data processing; Data collection: collecting maternal early pregnancy physical examination indicators, child birth information and childhood follow-up data through a big data platform system; Data processing: processing missing values, outliers and duplicate records for the collected data, and outputting the processed data set. 3.The method of predicting childhood overweight and obesity based on machine learning according to claim 2, characterized in that: In Step 1, constructing a machine learning model for predicting whether the offspring is a large-for-date baby at birth, and screening key variables with predictive value based on the machine learning model for large-for-date babies includes: Randomly dividing the processed data into a training set and a test set; Using random forest and extreme gradient boosting XGBoost algorithm to construct a large-for-date baby prediction model on the training set; The predictive performance of each model is evaluated on the test set by computing, for each model, the mean absolute error (MAE) and the root mean squared error (RMSE) of the predictions on the test set. The MAE and RMSE are computed by calling plot - importance extracts and quantifies the importance scores of the candidate features; Selecting the top three variables in both random forest and XGBoost for feature importance ranking as key variables. 4.The method of predicting childhood overweight and obesity based on machine learning according to claim 1, characterized in that: The key variables include maternal early pregnancy BMI, gestational age and child gender. 5.The method of predicting childhood overweight and obesity based on machine learning according to claim 1, characterized in that: In Step 2, fusing the key variables with childhood follow-up data to establish a prediction model for early childhood overweight risk includes: Randomly dividing the fused data into a training set, a test set and a validation set; In the training set, respectively using logistic regression, random forest, XGBoost and neural network algorithm, and completing model fitting through 10-fold cross-validation during the training process; Real-time monitoring of model performance through set indicators, and performance evaluation on the test set. 6.The method of predicting childhood overweight and obesity based on machine learning according to claim 1, characterized in that: The childhood follow-up data includes child sleep parameters, child feeding method information, child growth data and child growth environment information. 7.The method of predicting childhood overweight and obesity based on machine learning according to claim 1, characterized in that: Also included is the verification of the child scoring system, which verifies the scoring system formed in Step 3; Threshold setting: statistical analysis of the prediction probability distribution of the obese group samples in the validation set, extraction of the minimum value Pmin of the prediction probability distribution and the median value Pmedian of the prediction probability distribution; Taking the minimum value Pmin of the probability distribution as the first dividing line, all children with a prediction probability lower than the minimum value Pmin of the probability distribution are judged as non-obese in the validation set; Taking the median value Pmedian of the prediction probability distribution as the second dividing line, the remaining individuals are divided into medium risk and high risk.
8. A system for predicting childhood overweight and obesity based on machine learning, characterized by: A method for predicting childhood overweight and obesity based on machine learning according to any one of claims 1-7, comprising: A key variable screening module for constructing a machine learning model for predicting whether the offspring is a large-for-date baby at birth, and screening key variables with predictive value based on the machine learning model for large-for-date babies; An early overweight risk prediction model establishment module for fusing the key variables with childhood follow-up data to establish a prediction model for early childhood overweight risk; The child obesity risk score system generation module forms a child obesity risk score system based on the key variables screened by the key variable screening module and the prediction model established by the prediction model establishment module. The child obesity risk score system generation module forms a child obesity risk score system based on the key variables screened by the key variable screening module and the prediction model established by the prediction model establishment module.