Machine learning diagnosis system fusing metabolome and clinical data and method thereof
Through a machine learning diagnostic system that fuses metabolomics and clinical data, using one-dimensional residual convolutional neural network and multi-task deep learning model Child-Net, the efficient and accurate diagnosis of precocious puberty and obesity in the existing technology is solved, and the diagnostic effects of high accuracy and cross-platform deployment are achieved.
Patent Information
- Application Number
- CN202510383597.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-07-11
AI Technical Summary
The multi-task learning method lacking the fusion of metabolomic and clinical data in the prior art is used to diagnose precocious puberty and obesity in children. The traditional diagnosis method is cumbersome and expensive, making it difficult to achieve efficient and accurate diagnosis.
Using a machine learning diagnostic system that integrates metabolomics and clinical data, a one-dimensional residual convolutional neural network and a multi-task deep learning model Child-Net are used to achieve efficient and accurate diagnosis through a graphical user interface.
It achieves high-accuracy diagnosis of precocious puberty and obesity in children, shortens the diagnostic process, reduces costs, improves diagnostic performance, and supports cross-platform deployment to provide early treatment basis.
Smart Images
Figure CN120299679A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the cross - field of the combination of artificial intelligence and clinical medical diagnosis, and specifically relates to a machine - learning diagnosis system and method that integrates metabolomics and clinical data, which is used for diagnosing precocious puberty and obesity and their complications in children. Background Art
[0002] Obesity and precocious puberty are two common diseases in current children, which seriously affect the normal growth and development of children. Childhood obesity not only affects brain development and mental health, but may also persist into adulthood, causing complications such as diabetes, hypertension, and cardiovascular diseases. Precocious puberty refers to the premature appearance of secondary sexual characteristics in children, which may lead to premature cessation of growth and increase the risk of type 2 diabetes, obesity, cardiovascular diseases, and breast cancer. Therefore, early and accurate diagnosis of childhood precocious puberty and obesity and timely treatment are crucial for the healthy development of children.
[0003] With the development of metabolomics technology, metabolomics has been widely used in biomarker identification and disease mechanism research, showing unique advantages in disease prediction and diagnosis. However, in the field of childhood precocious puberty and obesity, there is still a lack of research that combines metabolomics and clinical data. In addition, there are potential connections between precocious puberty and obesity at the physiological and metabolic levels, but there is currently no technology or method for diagnosis using common disease predictors.
[0004] Currently, the clinical diagnosis of childhood obesity mainly relies on body mass index (BMI), while the diagnosis of precocious puberty depends on the comprehensive analysis of testicular volume in boys, uterine B - ultrasound in girls, bone age, luteinizing hormone and follicle - stimulating hormone, gonadotropin - releasing hormone stimulation test, and brain magnetic resonance imaging. These methods are cumbersome, time - consuming, and costly.
[0005] The core idea of the multi - task learning network model is to improve the model performance and generalization ability by sharing information between tasks. However, there is currently no relevant report on the use of a multi - task learning method that integrates metabolomics and clinical data for diagnosing childhood precocious puberty and obesity. Summary of the Invention
[0006] The purpose of the present invention is to provide a machine - learning diagnosis system and method that integrates metabolomics and clinical data for the above - mentioned technical problems existing in the prior art, and to achieve efficient and accurate diagnosis of childhood precocious puberty and obesity through a deep - learning model.
[0007] To achieve the above purpose, the present invention provides the following technical solutions:
[0008] The present invention provides a machine - learning diagnosis system that integrates metabolomics and clinical data, including:
[0009] System Composition
[0010] Storage device: Used to store the characteristic data of the population, including clinical data and serum metabolome data after feature extraction by a one-dimensional residual convolutional neural network.
[0011] Data input device: Used to input the sample data of the subject, including serum metabolome data and clinical data.
[0012] Graphical user interface: Connected to the data input device and interacting with the diagnostic model, used to input sample data, display diagnostic results, and adjust parameters.
[0013] Diagnostic model: Based on the characteristic data in the storage device, a diagnostic model is established using machine learning methods, and it diagnoses whether the subject has precocious puberty, obesity, or both complications according to the sample data.
[0014] Model architecture
[0015] The diagnostic model includes the following modules:
[0016] Data preprocessing module: Processes missing values, data balancing, and segmentation of the serum metabolome data.
[0017] Feature extraction module: Contains a one-dimensional residual convolutional neural network, used to extract features from the metabolome data and simplify the number of features.
[0018] Feature fusion module: Fuses metabolome features with clinical data.
[0019] Diagnostic engine: Constructed based on a multi-task deep learning algorithm, and outputs the diagnostic probabilities of precocious puberty, obesity, or complications.
[0020] Result output module: Displays the diagnostic results and confidence levels through the graphical user interface.
[0021] Graphical user interface
[0022] The graphical user interface includes the following modules:
[0023] Initialization module: Loads and initializes the deep learning model.
[0024] Sample selection module: Supports the user to select single or multiple patient samples for diagnosis.
[0025] Preprocessing module: Preprocesses the input data and converts it into a format recognizable by the model.
[0026] Identification result display module: Visually displays the diagnostic results.
[0027] The graphical user interface supports the input of single or multiple patient samples and simultaneously outputs the diagnostic results of precocious puberty and obesity.
[0028] One-dimensional Residual Convolutional Neural Network
[0029] This network includes the following layers:
[0030] Convolutional layer: Extract local features through multiple convolutional kernels.
[0031] Residual connection layer: Alleviate the problem of gradient vanishing.
[0032] Pooling layer: Reduce the dimension and retain key information.
[0033] Dropout layer: Reduce overfitting and improve the generalization ability of the model.
[0034] Training process
[0035] Use the grid search algorithm to find the best parameters.
[0036] Evaluate the generalization ability of the model through five-fold cross-validation.
[0037] Use population feature data to train the model, and the optimization objectives include accuracy, precision, recall, F1 value, and maximizing the area under the receiver operating characteristic curve (AUC).
[0038] Output results
[0039] The output of the diagnostic model includes the probability of precocious puberty (P1), the probability of obesity (P2), and the diagnostic conclusion.
[0040] Data source
[0041] Serum metabolome data: Obtain spectral data by detecting serum samples through one-dimensional nuclear magnetic resonance hydrogen spectrum ( 1 H-NMR), and after phase adjustment, baseline correction, calibration, and segmented integration of the spectral data, 3356 chemical shift integral values are obtained as serum metabolome data.
[0042] Clinical data: Include age, height, weight, body mass index, and 24 blood routine indicators, a total of 28 features.
[0043] Specifically, the clinical data include the following 28 indicators: age, height, weight, body mass index (BMI), and 24 routine blood test indicators including white blood cell count (WBC), lymphocyte ratio (LY%), monocyte ratio (MO%), neutrophil percentage (NE%), eosinophil ratio (EO%), basophil ratio (BA%), lymphocyte count (LY), monocyte count (MO), neutrophil count (NE), eosinophil count (EO), basophil count (BA), red blood cell count (RBC), hemoglobin (HGB), hematocrit (HCT%), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), red blood cell distribution width (RDW-CV), red blood cell distribution width (RDW-SDF(FI)), platelet count (PLT), mean platelet volume (MPV), plateletcrit (PCT), large platelet ratio (P-LCR), and platelet distribution width (PDW).
[0044] Diagnostic model
[0045] The machine learning method can adopt the multi-task learning method. The diagnostic model can be designed by using the multi-task learning method
[0046] The present invention adopts a multi-task deep learning model based on the architecture of the Multi-gate Mixture-of-Experts (MMoE), named Child-Net. The model includes a feature fusion layer, an expert network, a gating network, and a task output layer, and can learn the specific features of precocious puberty and obesity and output the diagnostic probability.
[0047] Model performance
[0048] Compared with seven models such as the hard parameter sharing model (hard Share), single task learning model (Single Task), convolutional neural network model (CNN), deep neural network model (DNN), random forest model (RF), support vector machine model (SVM), and gradient boosting decision tree model (XGBoost), Child-Net is superior to other models in terms of model parameters and performance.
[0049] A machine learning diagnosis method that fuses metabolomics and clinical data, using the above system, the method includes the following steps:
[0050] 1) Collect the serum samples and clinical data of the subjects.
[0051] 2) Preprocess the metabolomics data and extract high-order features.
[0052] 3) Input the metabolomic features and clinical data into the Child-Net model.
[0053] 4) Output the diagnostic results and confidence levels.
[0054] In step 2), the preprocessing includes missing value handling, data balancing, and standardization, etc.; the high-order feature extraction is implemented by a one-dimensional residual convolutional neural network.
[0055] In step 3), the Child-Net model optimizes the parameters through grid search, and evaluates the generalization ability using five-fold cross-validation. The optimization objectives include accuracy, precision, recall, F1 value, and AUC value, etc.
[0056] Compared with the prior art, the present invention has the following outstanding technical effects and advantages:
[0057] 1. High accuracy: The diagnostic accuracies of the Child-Net model for precocious puberty and obesity reach 96.52% and 97.08% respectively, and the AUC values both exceed 0.98.
[0058] 2. High efficiency: Avoid traditional invasive examinations and shorten the diagnostic process.
[0059] 3. Generalization ability: Share potential associated features through multi-task learning, reduce the number of model parameters, and improve the diagnostic performance.
[0060] 4. Cross-platform deployment: Support hospital systems and mobile terminals, and provide real-time diagnostic support.
[0061] 5. Clinical value: Provide a more reliable basis for early treatment and have important clinical application value.
[0062] 6. The present invention provides an efficient and accurate solution for the diagnosis of precocious puberty and obesity in children by fusing metabolomic and clinical data and combining an advanced multi-task deep learning model. Description of the Drawings
[0063] Figure 1 It is a technical roadmap for the Child-Net network model to diagnose precocious puberty and obesity in children.
[0064] Figure 2 It is the model architecture of the Child-Net network model proposed by the present invention.
[0065] Figure 3 It is the ROC curve of the test sets for diagnosing precocious puberty and obesity in children by eight models in Example 2 of the present invention. Among them, A represents precocious puberty (PP); B represents obesity (Ob).
[0066] Figure 4 This is the confusion matrix of the Child-Net network model for diagnosing central precocious puberty and obesity in the test set of Example 2 of the present invention. Among them, A represents central precocious puberty (PP); B represents obesity (Ob).
[0067] Figure 5 This is a schematic diagram of a diagnostic system for applying machine learning that integrates metabolomics and clinical data to central precocious puberty and obesity in Example 3 of the present invention. Detailed implementation manners
[0068] In order to more clearly elaborate on the technical problems, technical solutions, and advantages solved by the present invention, the present invention will be described in detail below in conjunction with embodiments. These embodiments are only preferred solutions of the present invention for demonstrating its applications. Those skilled in the art should understand according to this specification that the disclosed embodiments of the present invention can be modified in various ways, and all other embodiments belong to the protection scope of the present invention without involving creative labor. Unless otherwise defined, the technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs. All documents cited in the present invention and the materials cited therein are incorporated by reference. Those skilled in the art will realize, or can learn through routine experiments, that many described embodiments have many equivalent technologies, and these equivalent technologies are also included in the claims. The experimental methods in the following embodiments are conventional methods unless otherwise specified; the instruments and equipment used are conventional laboratory equipment unless otherwise specified; the test materials used are purchased from conventional biochemical reagent stores unless otherwise specified.
[0069] Example 1: Acquisition of serum metabolomics data
[0070] Due to the certain difficulty in the identification of serum metabolites and the possible periodic differences in the metabolic characteristics of childhood obesity, the present invention uses one-dimensional nuclear magnetic resonance hydrogen spectrum ( 1 1H-NMR) to detect serum samples, obtain spectral data, and through phase adjustment, baseline correction, calibration, and segmented integration processing, finally obtain 3356 chemical shift integral values. The specific steps are as follows:
[0071] (I) Preparation of serum samples
[0072] Preparation of phosphate buffer solution: Accurately weigh disodium hydrogen phosphate and sodium dihydrogen phosphate, and use heavy water to prepare a 60 mM phosphate buffer solution with a pH of 7.4.
[0073] Preparation of blood samples: After pre-thawing the serum samples at 4°C, accurately pipette 400 μL of serum using a micropipette, and add 200 μL of phosphate buffer solution to form serum samples.
[0074] (II) Sample pretreatment
[0075] Place the mixed serum sample in a centrifuge and centrifuge it at 4 °C and 13,000 g for 10 minutes. Take 550 μL of the supernatant and transfer it to a 5 mm NMR tube for testing.
[0076] (III)Detection of the sample to be tested
[0077] Collect the prepared serum sample on a Bruker AV-600 nuclear magnetic resonance spectrometer (Bruker BioSpin, Germany) equipped with a cryoprobe. 1 1H-NMR spectra. The sampling parameters are as follows:
[0078] Proton resonance frequency: 600.13 MHz
[0079] Working temperature: 298.0 K
[0080] Pulse sequence: Presaturation water suppression CPMG sequence
[0081] Spin echo time: 70 ms
[0082] Spectral width: 12019.2 Hz
[0083] Sampling time: 1.36 s
[0084] Relaxation delay time: 4.0 s
[0085] Number of scans: 64 times
[0086] Number of sampling points: 16 K
[0087] (IV)Data processing
[0088] Perform Fourier transform on the collected data to convert it into a nuclear magnetic resonance spectrum. Perform phase adjustment, baseline correction, calibration, and piecewise integration on the spectrum, and finally obtain 3356 chemical shift integral values.
[0089] Example 2: Diagnostic model based on machine learning
[0090] In order to better diagnose precocious puberty and obesity in children, the present invention collected serum samples of healthy children, children with precocious puberty, children with obesity, and children with precocious puberty combined with obesity, obtained serum metabolome data by the method of Example 1, and analyzed them in combination with 28 clinical data (including age, height, weight, body mass index, and 24 blood routine indexes). The sample data is as follows:
[0091] Healthy children: 372 cases (130 boys and 242 girls)
[0092] Children with precocious puberty: 468 cases (71 boys and 397 girls)
[0093] Obese children: 252 cases (125 boys and 127 girls)
[0094] Children with precocious puberty complicated with obesity: 239 cases (72 boys and 167 girls)
[0095] Table 1 shows the clinical characteristic data of children in each group, including 28 clinical data such as age, weight, height, BMI, and blood routine indexes. These data provide data support for model training and intuitively reflect the distribution characteristics of physiological indexes in each group.
[0096] Table 1. Clinical Characteristic Data of Children in Each Group
[0097]
[0098] a Mean (Standard Deviation)
[0099] Figure 1 It is the technical roadmap for the Child-Network model to diagnose precocious puberty and obesity in children.
[0100] Model Construction and Evaluation
[0101] 1. Data collection
[0102] The data sources include metabolomics data (3356 chemical shift integral values obtained by 1 detecting serum samples by H-NMR) and clinical data (28 clinical indexes).
[0103] 2. Data preprocessing
[0104] Treatment of missing values and outliers: The mean method is used to fill in the missing values, and outliers are identified and corrected or removed through statistical tests.
[0105] Data balancing: The SMOTE algorithm is used to deal with the problem of unbalanced sample categories to ensure the balance of the proportion of samples in each category.
[0106] 3. Model construction
[0107] As Figure 2 shown, the model architecture of the Child-Network model proposed by the present invention includes:
[0108] (1) Batch Normalization: Standardize the metabolomics data to eliminate the differences in data distribution, accelerate the convergence of model training, and improve stability.
[0109] (2) One-dimensional convolution feature extraction and average pooling layer (Conv1D + Average Pooling): Data is input into the one-dimensional convolution layer, and the local features of the metabolome data are extracted through convolution sliding calculation to capture key metabolites. The average pooling layer is used to reduce the data dimension, retain the overall statistical information of the features, reduce the number of parameters and alleviate overfitting.
[0110] (3) Residual Connections: It contains two Conv1D layers to form a residual unit, which simplifies the number of feature inputs and retains the original input information through residual connections, thus alleviating the gradient vanishing problem of deep networks.
[0111] (4) Max Pooling + Flatten: The Max Pooling layer further extracts significant features and highlights key metabolome information. The Flatten layer then flattens the multidimensional data into a one-dimensional vector for subsequent processing.
[0112] (5) Dropout layer: The Dropout layer is introduced to randomly disconnect the connections between neurons during the training process, thereby reducing overfitting and improving the generalization ability of the network.
[0113] (6) Fully connected layer (Dense): Clinical data is directly input into the fully connected layer to implement nonlinear combination and spatial embedding of features of structured clinical information (such as age, BMI, blood routine indicators, etc.), thereby providing richer feature representation.
[0114] (7) Feature fusion layer (Concatenate): The metabolomics features are fused with the clinical data to form a unified input feature vector.
[0115] (8) Core network structure: A multi-task learning framework based on a multi-gated mixture of experts (MMoE) network. It learns common predictors of diseases by sharing the underlying network, and sets up independent task branches (precocious puberty diagnosis branch and obesity diagnosis branch) to capture disease-specific features. It is mainly composed of the following components:
[0116] Batch Normalization: Normalize the fused features again to make them uniformly distributed.
[0117] Expert network: Through three fully connected layers (Dense) plus Tanh loss function, it further learns the complex associations between features and explores the common and unique characteristics of precocious puberty and obesity.
[0118] Gating Network: The gating network consists of a single fully-connected layer (Dense) followed by a Softmax function, which determines the contribution of each expert's output to the final prediction.
[0119] Task Output Layer: The output structures for the two sub-tasks consist of a single fully-connected layer (Dense) followed by a Tanh loss function, a Dropout layer, and a single fully-connected layer (Dense) followed by a Softmax function. Dropout is inserted in each task branch to reduce the risk of parameter co-adaptation and improve the robustness of the model. The outputs of the two sub-tasks correspond to the diagnosis of precocious puberty and obesity, respectively. Among them, Output PP (Precocious Puberty Diagnosis) and Output Ob (Obesity Diagnosis) output the diagnosis results and confidence rates, respectively.
[0120] Child-Net takes into account both the microscopic features of metabolomic data and the macroscopic characteristics of clinical data to comprehensively characterize the disease state. Residual connections ensure the effectiveness of deep network training, and Dropout and batch normalization improve the generalization ability of the model to adapt to complex pediatric disease diagnosis scenarios. By sharing the underlying features and independent task branches, data information is efficiently utilized while improving the diagnostic accuracy of two types of diseases, precocious puberty and obesity.
[0121] 4. Model Performance Comparison
[0122] Evaluation Metrics: Accuracy, Precision, Recall, F1-score, and Area Under the ROC Curve (AUC).
[0123] Comparative Experiment: The Child-Net model is compared with models such as the Hard Parameter Sharing model (Hard Share), Single Task Learning model (Single Task), Convolutional Neural Network (CNN), Deep Neural Network (DNN), Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost).
[0124] Tables 2 and 3 show the performance comparison of the Child-Net model with other models on the training set, test set, and validation set. Figure 3 is the ROC curve for the test sets of diagnosing precocious puberty and obesity in children for each model, Figure 4 is the confusion matrix for the test sets of diagnosing precocious puberty and obesity in children by the Child-Net network model.
[0125] Table 2 Model Performance Comparison (Accuracy and AUC)
[0126]
[0127] Table 3 Model Performance Comparison (Precision, Recall, and F1-score)
[0128]
[0129] The experimental results show that the Child-Net model is superior to other models in terms of the accuracy rate, AUC value, precision rate, recall rate, and F1 value for the diagnosis of precocious puberty and obesity, demonstrating higher classification performance and diagnostic accuracy.
[0130] 5. Model Deployment:
[0131] Deploy the model to the GUI (Graphical User Interface) to implement functions such as sample data input, diagnostic result display, and parameter adjustment, facilitating practical clinical applications.
[0132] Example 3: Construction of a Diagnostic System
[0133] Based on the model in Example 2, the present invention constructs a complete diagnostic system, as Figure 5 shown. The system includes the following components:
[0134] Storage device: Used to store the characteristic data of the population.
[0135] Data input device: Used to input the sample data of the subject.
[0136] Diagnostic model: Based on the characteristic data in the storage device, a diagnostic model is established using machine learning methods, and it diagnoses whether the subject has precocious puberty, obesity, or a combination of both according to the sample data.
[0137] Graphical user interface: Connected to the data input device and the diagnostic model, used to input sample data, display diagnostic results and confidence levels. This interface supports the simultaneous input and output of multiple sample data and diagnostic results.
[0138] By integrating metabolomic and clinical data and using the multi-task deep learning model Child-Net, this system realizes the efficient and accurate diagnosis of childhood precocious puberty and obesity.
[0139] The above are only specific embodiments of the present invention, and those skilled in the art can make various modifications according to this specification. These modifications and equivalent forms all fall within the protection scope of the appended claims of this application.
Claims
1. A machine learning diagnosis system integrating metabolomics and clinical data, characterized in that, Including: Storage device: used to store the characteristic data of the population; Data input device: used to input the sample data of the subject; Graphical user interface: connected to the data input device and interacting with the diagnostic model, used to input sample data, display diagnostic results and adjust parameters; Diagnostic model: respectively connected to the storage device, data input device and graphical user interface. The diagnostic model is established by using machine learning methods based on the population characteristic data in the storage device, and realizes the diagnosis of whether the subject has precocious puberty, obesity or the combined disease of both based on the sample data; Among them, the characteristic data includes clinical data and serum metabolome data after feature extraction by a one-dimensional residual convolutional neural network, and the sample data includes serum metabolome data and clinical data.
2. The machine learning diagnosis system integrating metabolomics and clinical data according to claim 1, wherein The diagnostic model includes the following modules: Data preprocessing module: used to process missing values, data balancing and segmentation of serum metabolome data; Feature extraction module: contains a one-dimensional residual convolutional neural network, used to extract features from metabolome data and simplify the number of features; Feature fusion module: used to fuse metabolome features and clinical data; Diagnostic engine: built based on a multi-task deep learning algorithm, used to output the diagnostic probabilities of precocious puberty, obesity or combined disease; Result output module: used to display the diagnostic results and confidence levels through the graphical user interface.
3. The machine learning diagnosis system integrating metabolome and clinical data according to claim 1, characterized in that, The graphical user interface includes the following modules: Initialization module: loads and initializes the deep learning model; Sample selection module: supports users to select single or multiple patient samples for diagnosis; Preprocessing module: preprocesses the input data and converts it into a format recognizable by the model; Identification result display module: visually displays the diagnostic results.
4. The machine learning diagnosis system integrating metabolomics and clinical data according to claim 1, wherein The one-dimensional residual convolutional neural network includes the following layers: Convolutional layer: extracts local features through multiple convolutional kernels; Residual connection layer: alleviates the problem of gradient disappearance; Pooling layer: reduces the dimension and retains key information; Dropout layer: reduces overfitting and improves the generalization ability of the model.
5. The machine learning diagnosis system integrating metabolome and clinical data according to claim 4, wherein The training process of the one-dimensional residual convolutional neural network includes: Using the grid search algorithm to find the best parameters; Evaluating the generalization ability of the model through five-fold cross-validation; Training the model using population characteristic data, and the optimization objectives include accuracy, precision, recall, F1 value and maximizing the area under the receiver operating characteristic curve AUC value.
6. The machine learning diagnosis system integrating metabolomics and clinical data according to claim 1, characterized in that, The output of the diagnostic model includes the probability of precocious puberty P1, the probability of obesity P2 and the diagnostic conclusion.
7. The machine learning diagnosis system integrating metabolomics and clinical data according to claim 1, characterized in that The serum metabolome data: spectral data is obtained by detecting serum samples through one-dimensional nuclear magnetic resonance proton spectroscopy, and 3356 chemical shift integral values are obtained after phase adjustment, baseline correction, calibration and segmented integration processing.
8. The machine learning diagnosis system integrating metabolome and clinical data according to claim 1, characterized in that, The clinical data: includes age, height, weight, body mass index and 24 blood routine indexes, a total of 28 features.
9. The machine learning diagnosis system integrating metabolomics and clinical data according to claim 1, characterized in that, The diagnostic model is a multi-task deep learning model using a multi-gate mixture of experts network architecture, named Child-Net; this diagnostic model includes a feature fusion layer, an expert network, a gating network and a task output layer, and can learn the specific features of precocious puberty and obesity and output diagnostic probabilities.
10. A machine learning diagnosis method that integrates metabolomic and clinical data, characterized in that Adopt the system described in any one of claims 1-9, and the method includes the following steps: 1) Collect the serum samples and clinical data of the subject; 2) Preprocess the metabolome data and extract high-order features; 3) Input the metabolome features and clinical data into the Child-Net model; 4) Output the diagnosis result and the confidence level.