Chronic liver disease multi-mode dialectical model dynamic optimization method based on graph neural network

Abstract

The invention discloses a dynamic optimization method for a chronic liver disease multi-mode syndrome differentiation model based on a graph neural network, and belongs to the field of artificial intelligence and traditional Chinese medicine syndrome differentiation, and the method comprises the steps: obtaining multi-source heterogeneous data, collecting tongue images, pulse diagnosis data, biochemical indexes and clinical symptom data of chronic liver disease patients from a plurality of medical institutions, sublingual collateral vein fractal features are extracted through a second-order Laplacian algorithm; constructing a hierarchical syndrome association network, analyzing a symptom-syndrome association structure based on topological space mapping and a spectrogram theory, and learning an influence relationship between symptom nodes and syndrome nodes by using a graph neural network; the contribution degree of each symptom to syndrome judgment is evaluated through information entropy, the attention weight is dynamically adjusted, a hidden Markov chain is introduced to capture a syndrome transfer rule, model knowledge is distilled to realize lightweight deployment, multi-modal data is comprehensively utilized, and the complex relationship between the symptoms and the syndromes is accurately captured; and individualized syndrome differentiation and dynamic syndrome prediction are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and traditional Chinese medicine syndrome differentiation, and particularly relates to a chronic liver disease multi-modal syndrome differentiation model dynamic optimization method based on a graph neural network, which is applied to a traditional Chinese medicine intelligent syndrome differentiation system. BACKGROUND

[0002] Chronic liver disease is a common digestive system disease with a long course and a high recurrence rate, which seriously affects the quality of life of patients. Traditional Chinese medicine syndrome differentiation has unique advantages in the treatment of chronic liver disease. However, traditional Chinese medicine syndrome differentiation mainly relies on the experience and subjective judgment of doctors, and has certain subjectivity and uncertainty, which is difficult to realize standardization and standardization.

[0003] In recent years, artificial intelligence technology has been continuously applied in the medical field. In the prior art, there are some traditional Chinese medicine syndrome differentiation models based on machine learning, such as a syndrome identification system based on a support vector machine (SVM) and a traditional Chinese medicine syndrome differentiation model based on a decision tree. However, these methods have the following shortcomings: first, they mostly use single modal data, which is difficult to fully reflect the clinical characteristics of patients; second, they mostly use shallow models, which cannot effectively capture the complex nonlinear relationship between symptoms and syndromes; third, they are mostly static models, which cannot adapt to individual differences and disease evolution; fourth, the model complexity is high, which is difficult to popularize and apply in primary medical environment.

[0004] Therefore, there is an urgent need for a traditional Chinese medicine syndrome differentiation model that can comprehensively utilize multi-modal data, accurately capture the complex relationship between symptoms and syndromes, adapt to individual differences and disease evolution, and be convenient for primary popularization. SUMMARY

[0005] The purpose of the present application is to provide a chronic liver disease multi-modal syndrome differentiation model dynamic optimization method based on a graph neural network to solve the problems in the prior art.

[0006] The present application provides a chronic liver disease multi-modal syndrome differentiation model dynamic optimization method based on a graph neural network, which comprises:

[0007] Obtaining multi-source heterogeneous data, including: collecting tongue image, pulse diagnosis data, biochemical indicators and clinical symptom data of chronic liver disease patients from multiple medical institutions; preprocessing the tongue image to extract the sublingual collateral fractal feature; extracting features from the pulse diagnosis data to obtain the pulse fractal feature;

[0008] Constructing a hierarchical syndrome association network, including: based on the sublingual collateral fractal feature and the pulse fractal feature, constructing a topological space mapping of symptoms and syndromes; applying spectral graph theory to analyze the structure of the symptom-syndrome association network to obtain the spectral representation of the syndrome association; learning the influence relationship between the symptom nodes and the syndrome nodes by using the graph neural network to form the syndrome probability distribution;

[0009] The syndrome differentiation model dynamic optimization is realized, including: based on information entropy, evaluating the contribution degree of each symptom to syndrome judgment, dynamically adjusting the attention weight of symptom characteristics; introducing hidden Markov chain to capture the syndrome transition rule, predicting the evolution trend of syndrome; knowledge distillation is performed on the syndrome differentiation model to realize lightweight deployment.

[0010] As preferred, the acquisition of multi-source heterogeneous data specifically includes: collecting the diagnosis information of patients with chronic liver disease from multiple third-grade hospitals and community health service centers, including patient complaints, tongue image diagnosis and pulse diagnosis, laboratory test data, and historical diagnosis information; based on the BertNer Chinese named entity recognition algorithm, the clinical medical record information and tongue image report are analyzed, and the text and image are matched, and the feature words and index fields related to the disease are screened; the index fields with repeated records in the medical record information are integrated and normalized, and the complaint text record and tongue image data are retained.

[0011] As preferred, the extraction of sublingual collateral vessel fractal features specifically includes: establishing patient electronic medical record data set according to disease category; extracting tongue image data and preprocessing, using second-order Laplacian graph fractal algorithm to obtain tongue diagnosis fractal dimension feature, and obtaining sublingual collateral vessel fractal feature according to the depth distribution of sublingual collateral vessel; extracting features from pulse diagnosis pressure waveform data, and selecting pulse fractal features that can represent patient syndrome; based on the historical diagnosis record, the disease course, the disease duration, smoking, drinking, psychological symptoms, and test results of each patient with chronic liver disease are extracted.

[0012] As preferred, the construction of hierarchical syndrome association network specifically includes: establishing a tongue diagnosis and pulse tongue syndrome fusion graph network, constructing a symptom and syndrome feature fusion graph according to the contribution degree of clinical symptoms to syndrome, learning node and edge weight using graph neural network, and calculating the influence of symptom node on syndrome node; respectively calculating the contribution degree weight of symptoms to syndromes, and summing the graph model weight and the symptom feature weight to obtain the probability distribution of each syndrome type of chronic liver disease, and determining the syndrome category to which the patient belongs according to the decision threshold; taking the first record symptoms and tongue image of the patient as the input of the syndrome reasoning model, the model outputs the syndrome probability distribution of the patient, and the model calculates the predicted value and the syndrome probability distribution annotated by experts as the initialization parameters of hidden Markov chain; introducing hidden Markov chain to capture the clinical syndrome transition link of the patient, and determining the parameter matrix and the initial state according to the first diagnosis symptoms and the diagnosis results of the patient, updating the parameter matrix based on the forward and backward dynamic programming algorithm and combining the back propagation, so as to update the reasoning model graph network weight according to the medical record data; in the process of syndrome reasoning and sample learning, the graph model weight and parameters are iterated by using incremental learning method, and the model parameters are updated according to the evolution of patient syndrome.

[0013] Preferably, the graph network weight calculation comprises: generating a symmetrical graph according to the first diagnosis information of the patient as a node, and introducing a symptom node to learn the weight value according to the graph neural network learning weight feature vector; inputting the symptom feature into the graph convolution network for node fusion, multiplying the symptom feature and the weight value through the graph convolution layer parameter matrix, and then summing to obtain a diagnosis result vector; according to the training set label result, using the cross entropy function as the loss function, and updating the graph convolution layer weight parameters by using the back propagation algorithm.

[0014] Preferably, the symptom node input feature comprises: taking the first hospitalization record and tongue and pulse diagnosis information of each patient as diagnosis information, and establishing a weight vector of clinical symptoms based on a hierarchical graph neural network; calculating the fractal feature of the collected tongue and pulse diagnosis data, and calculating the tongue diagnosis information and pulse diagnosis information of a single visit based on the graph neural network to obtain a fractal feature weight vector highly related to the syndrome; extracting the feature of the historical visit record of each patient, and calculating the weight according to the contribution degree of the influencing syndrome factors, and fusing the historical visit and the symptom fractal weight vector; taking the constitution information of each patient as supplementary knowledge, fusing the constitution information with the historical syndrome evolution information feature to form a symptom feature vector matrix, and calculating the current syndrome probability distribution of the patient according to the parameter matrix of the hidden Markov chain.

[0015] Preferably, the knowledge distillation step in the implementation of the dynamic optimization of the syndrome differentiation model comprises: training a lightweight syndrome differentiation model, migrating the model feature layer mapping weight to a large mobile device pre-training model, extracting the model graph weight combined with the knowledge distillation method, and migrating the graph weight to a small embedded device; encapsulating the lightweight model based on the Docker container, establishing a distributed server and providing the model to a third party user; the third party application calls the chronic liver disease model service, initiates an API request to the container cluster, and returns the diagnosis result to the user intelligent terminal.

[0016] Preferably, the feature weight learning process in the implementation of the dynamic optimization of the syndrome differentiation model is based on deep learning and a knowledge graph to build a hierarchical graph neural network for node classification, wherein the feature matrix extracted from the tongue and pulse diagnosis tongue-biochemical-syndrome data set by the deep learning subnetwork is taken as input, and the syndrome feature distribution of the patient is learned in combination with the graph neural network.

[0017] As preferred, in the fractal feature extraction of the sublingual collateral vessels, the fractal calculation is performed by a second-order Laplacian operator, and the specific steps are as follows: based on a deep learning algorithm, the tongue image is converted into a binary image, the pixel values of the original image are processed into two types of binary images, the tongue contour is extracted by a threshold segmentation algorithm, the hollow area of the tongue body is filled, and a tongue binary image is obtained; the skeleton of the tongue body is obtained by using skeleton transformation, and the tongue information is extracted according to the skeleton length and the number of branches; the graph structure of the tongue image is established based on the second-order Laplacian operator, the distance curves in the horizontal and vertical directions of the tongue body are calculated respectively, and the fractal dimension of the image is calculated by using Euler's formula; the fractal dimension features of the tongue image are obtained, a database is established, the feature vector of the tongue-biochemical-syndrome set feature of the tongue image is calculated based on a deep learning model, the feature vector of the patient is mapped through a neural network, and the weight vector of the tongue-biochemical-syndrome of the tongue image is calculated by combining a graph neural network.

[0018] As preferred, the node graph neural network of the graph neural network uses a symmetric structure, including a graph convolution layer, a pooling layer, an activation layer, and a nonlinear function, wherein the dimension of the input layer feature vector is determined by the number of input layer nodes, the node features are learned by weight using a graph convolution layer matrix, each node vector is aggregated by a graph propagation layer, the fusion features are calculated according to the activation layer, the graph node probability vector is calculated by the nonlinear function, the cross-entropy is used as the loss function, the weight parameters of each layer are updated according to the back propagation algorithm, and the model node propagation process specifically includes the following steps: calculating the contribution degree of the symptom node to the diagnosis node, obtaining the symptom feature vector based on the weight mapping of the graph convolution layer, multiplying the symptom feature vector by the weight of each symptom node and summing to obtain the diagnosis node feature vector; classifying the nodes by the pooling layer, calculating the probability distribution of the diagnosis result according to the conditional entropy; updating the parameters by the BP algorithm and calculating the loss function gradient; optimizing the number of graph network nodes according to the size of the learning data set, adjusting the learning parameters and the number of activation function layers, and obtaining the generalization ability suitable for different sizes of data sets.

[0019] The present application has the following beneficial effects:

[0020] 1. Multi-modal data fusion: By integrating tongue image, pulse diagnosis data, biochemical indicators and clinical symptoms and other multi-source heterogeneous data, the overall representation of the patient's state is realized, and the comprehensiveness and accuracy of the syndrome differentiation are improved.

[0021] 2. Complex association capture: The hierarchical graph neural network and the spectral graph theory are adopted to accurately capture the complex nonlinear relationship between symptoms and syndromes, and the problem that the traditional method is difficult to express the complex relationship between symptom combination and syndrome is solved.

[0022] 3. Dynamic optimization capability: based on the adaptive attention mechanism of information entropy and the hidden Markov chain syndrome transition model, the adaptation to different patient individual differences and the dynamic prediction of syndrome evolution are realized, which realizes the leap from static syndrome differentiation to dynamic syndrome differentiation.

[0023] 4. Lightweight deployment: through the knowledge distillation technology, the knowledge transfer from complex model to lightweight model is realized, so that the complex model can be applied in the primary medical environment and the primary medical service ability is improved.

[0024] 5. Syndrome differentiation accuracy improvement: by comprehensively applying the above-mentioned technologies, the syndrome differentiation accuracy of chronic liver disease is significantly improved, especially the recognition ability of complex syndrome combination. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The whole flow chart of the dynamic optimization method of the chronic liver disease multi-modal syndrome differentiation model based on the graph neural network of the application;

[0026] Figure 2 The structural diagram of the multi-source heterogeneous data processing module of the application;

[0027] Figure 3 The fractal feature extraction flow chart of the sublingual collateral of the application;

[0028] Figure 4 The structural diagram of the hierarchical graph neural network syndrome association module of the application;

[0029] Figure 5 The working flow chart of the adaptive attention optimization module of the application;

[0030] Figure 6 The state transition schematic diagram of the hidden Markov syndrome transition module of the application;

[0031] Figure 7 The architecture diagram of the knowledge distillation and lightweight deployment module of the application;

[0032] Figure 8 The schematic diagram of the hierarchical syndrome association network of the application;

[0033] Figure 9 The structural schematic diagram of the symptom-syndrome topological space mapping of the application. DETAILED DESCRIPTION

[0034] Please refer to Figures 1-9 The application will be further described in detail below in combination with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0035] As Figure 1As shown, the present invention provides a dynamic optimization method for a multimodal syndrome differentiation model of chronic liver disease based on graph neural networks, including the following steps: acquiring multi-source heterogeneous data, constructing a hierarchical syndrome association network, and realizing dynamic optimization of the syndrome differentiation model.

[0036] In a preferred embodiment of the present invention, acquiring multi-source heterogeneous data includes collecting tongue images, pulse diagnosis data, biochemical indicators, and clinical symptom data of patients with chronic liver disease from multiple medical institutions; preprocessing the tongue images to extract sublingual vein fractal features; and extracting features from the pulse diagnosis data to obtain pulse fractal features.

[0037] The construction of a hierarchical syndrome association network includes: constructing a topological spatial mapping between symptoms and syndromes based on the fractal features of the sublingual veins and the fractal features of the pulse; applying spectral graph theory to perform structural analysis on the symptom-syndrome association network to obtain the spectral representation of syndrome association; and using graph neural networks to learn the influence relationship between symptom nodes and syndrome nodes to form a syndrome probability distribution.

[0038] Achieving dynamic optimization of the syndrome differentiation model includes evaluating the contribution of each symptom to syndrome judgment based on information entropy and dynamically adjusting the attention weight of symptom features; introducing hidden Markov chains to capture syndrome transfer patterns and predict syndrome evolution trends; and performing knowledge distillation on the syndrome differentiation model to achieve lightweight deployment.

[0039] like Figure 2 As shown, in another embodiment of the present invention, acquiring multi-source heterogeneous data specifically includes collecting medical information of patients diagnosed with chronic liver disease from multiple tertiary hospitals and community health service centers, including patient complaints, tongue and pulse examination findings, laboratory test data, and historical medical information. Preferably, the present invention employs a standardized data acquisition process to ensure the consistency of data quality. For example, tongue image acquisition uses uniform light source conditions (color temperature 5500K±500K) and a fixed shooting distance (15-20cm) to reduce the impact of differences in the acquisition environment on feature extraction.

[0040] Furthermore, this invention utilizes the BertNer Chinese named entity recognition algorithm to parse clinical medical record information and tongue image reports, performing entity matching between text and images to filter feature words and indicator fields associated with diseases. The BertNer algorithm is a fine-tuned algorithm based on the BERT pre-trained model for medical entity recognition tasks, achieving high accuracy in medical entity recognition. In this invention, the algorithm is primarily used to identify medical entities such as symptom descriptions, tongue features, and pulse features in medical record texts, achieving an accuracy rate of over 90%.

[0041] Meanwhile, the application integrates and normalizes the repeatedly recorded index fields in the medical record information, and retains the chief complaint text record and tongue image data. For data of different sources, the application adopts a Z-Score standardization method for normalization processing, and ensures that different indexes are compared on the same scale. The Z-Score standardization formula is as follows:

[0042] ,

[0043] wherein X is an original data value, μ is a data mean value, σ is a data standard deviation, and Z is a standardized data value. Through the method, data of different sources and different dimensions are converted to a standard normal distribution with a mean value of 0 and a standard deviation of 1, and the scale inconsistency problem in multi-modal data fusion is effectively solved.

[0044] As shown in FIG. 1, Figure 3 In another embodiment of the application, extracting the fractal features of the sublingual collateral vessels specifically includes establishing a patient electronic medical record data set according to disease categories. The application preferably classifies chronic liver disease patients into different syndrome categories such as liver stagnation and spleen deficiency type and liver stagnation and blood stasis type, each category containing clinical data of 100-500 patients to form a labeled training data set.

[0045] The application extracts tongue image data and pre-processes, acquires tongue diagnosis fractal dimension features by using a second-order Laplacian graph fractal algorithm, and obtains sublingual collateral vessel fractal features according to the depth distribution of the sublingual collateral vessels. The second-order Laplacian graph fractal algorithm is a fractal feature extraction method based on the local neighborhood structure of an image, and can effectively capture the texture and detail features of the tongue image. The algorithm first calculates the second-order Laplacian operator of the image, and then estimates the fractal dimension of the image by a box dimension method. The specific calculation formula is as follows:

[0046] ,

[0047] ,

[0048] wherein: is a second-order Laplacian operator of the image I at a point , and are the second-order partial derivatives of the image in the x and y directions, is a box dimension, N(ε) is the number of boxes with a side length of ε required to cover the image, ε is a box size, and lim represents a limit operation. Through the method, the application can convert the complex morphological features of the tongue image into quantifiable fractal dimension features, and provide an objective basis for subsequent syndrome identification.

[0049] In addition, the present application extracts features from pulse diagnosis pressure waveform data, and selects pulse image fractal features that can represent patient syndromes. The pulse diagnosis data is collected using a standardized pulse diagnosis instrument to obtain time series data of pulse pressure waveform. The present application uses a wavelet transform and Hurst index combination method to extract the fractal features of pulse image. Wavelet transform is used to decompose the multi-scale features of pulse signal, and Hurst index is used to quantify the long-range correlation of time series. The Hurst index calculation formula is as follows:

[0050]

[0051] is the Hurst index, the value range is 0 to 1, represents that the time series has a persistent trend, represents that the time series has an anti-persistent trend, represents that the time series is a random walk; is the range of the time series, that is, the difference between the maximum and minimum values; is the standard deviation of the time series; is the length of the time series, that is, the number of sampling points; log represents the logarithm with base 10. By this method, the fractal features of the pulse signal can be effectively extracted, which can be used to distinguish the pulse characteristics of different syndrome types.

[0052] Meanwhile, based on historical medical records, the present application extracts information of each patient, such as the course of chronic liver disease, the length of the disease, smoking, drinking, psychological symptoms, laboratory results, etc., to form a comprehensive patient feature representation. After encoding, these information are fused with tongue image and pulse image features to form a multi-modal feature vector.

[0053] As shown in Figure 4 , in another embodiment of the present application, the hierarchical syndrome association network is specifically constructed by establishing a tongue image diagnosis pulse tongue syndrome fusion graph network, constructing a symptom and syndrome feature fusion graph according to the contribution degree of clinical symptoms to syndromes, learning node and edge weights by using a graph neural network, and calculating the influence of symptom nodes on syndrome nodes.

[0054] In a preferred embodiment of the present application, the graph network is constructed using a heterogeneous graph network structure, which contains two types of nodes, symptom nodes and syndrome nodes. The symptom nodes represent tongue image features, pulse image features, biochemical indicators and clinical symptoms, etc., and the syndrome nodes represent liver stagnation and spleen deficiency syndrome, liver stagnation and blood stasis syndrome, etc. The edges between nodes represent the association strength between symptoms and syndromes, and the initial edge weight is determined based on expert knowledge and historical data statistics.

[0055] ​​In addition, the application calculates the contribution weight of symptoms to syndromes respectively, and obtains the probability distribution of each syndrome type of chronic liver disease by weighted summation of the graph model weight and the symptom characteristics, and determines the syndrome category to which the patient belongs according to the decision threshold. The contribution weight of symptoms to syndromes is calculated by a graph attention mechanism, which allows the model to dynamically adjust the importance of different symptoms. The graph attention calculation formula is as follows:

[0056] ,

[0057] ,

[0058] Among them: is the attention coefficient of node to node , indicating the importance of node to the feature update of node ; is the attention vector, which is a learnable model parameter; is the weight matrix, which is used for linear transformation of node features; and are the feature vectors of node and node respectively; represents the vector splicing operation; is the neighbor node set of node ; LeakyReLU is a leaky linear rectifier unit activation function, defined as represents the natural exponential function; is an activation function, usually selected as ReLU or is the updated node feature vector; represents the summation operation. Through the graph attention mechanism, the application can adaptively adjust the importance of different symptoms and improve the accuracy of syndrome differentiation.

[0059] At the same time, the first record symptoms and tongue of the patient are taken as the input of the syndrome reasoning model, the model outputs the syndrome probability distribution of the patient, the graph model weight vector is adjusted according to the symptom weight, and the predicted value calculated by the model and the syndrome probability distribution labeled by experts are taken as the initialization parameters of the hidden Markov chain. The syndrome probability distribution is calculated by a softmax function:

[0060] ,

[0061] Among them: is the probability that the patient belongs to syndrome under given input features ; is syndrome logit value of the probability of a syndrome, i.e. the unnormalized score output by the model; is the natural exponential function; is the total number of syndrome categories; represents the summation of the probabilities of all syndrome categories. The decision threshold is set to 0.5, i.e. when the probability of a syndrome exceeds 0.5, the patient is determined to belong to the syndrome category.

[0062] In addition, the present application introduces a hidden Markov chain to capture the clinical syndrome transition link of the patient, and determines the parameter matrix and initial state according to the first visit symptoms and diagnosis results of the patient, updates the parameter matrix based on the forward and backward dynamic programming algorithm and combined with the back propagation, and thus updates the inference model graph network weight according to the medical record data. The hidden Markov model is composed of an initial state distribution , a state transition matrix A and an emission probability matrix B, wherein:

[0063] ,

[0064] ,

[0065] ,

[0066] wherein: is the probability of the initial state being , is the initial state probability distribution vector, and the dimension is the number of states; is the probability of transitioning from state to state , is the state transition probability matrix, and the dimension is the number of states x the number of states; is the probability of observing in state , is the emission probability matrix, and the dimension is the number of states x the number of observations; is the state at time , and are possible states; is the observation value at time ; represents the probability. In the present application, the state corresponds to the syndrome type, and the observation value corresponds to the symptom combination.

[0067] Finally, in the process of syndrome inference and sample learning, the present application iterates the graph model weight and parameter in an incremental learning method, and updates the model parameter according to the syndrome evolution of the patient. The incremental learning enables the model to continuously learn from new data, and adapt to the update of disease knowledge and the change of patient characteristics.

[0068] In another embodiment of the present invention, the graph network weight calculation includes generating a symmetrical graph using the patient's initial diagnosis information as nodes based on the weight feature vector learned from the graph neural network, and introducing symptom nodes to learn weights. The present invention uses a Graph Convolutional Network (GCN) as the basic model structure, which can effectively utilize graph structure information for node feature learning.

[0069] Preferably, in this invention, symptom features are input into a graph convolutional network for node fusion. The symptom features are multiplied by weights and then summed using the graph convolutional layer parameter matrix to obtain the diagnostic result vector. The calculation formula for the graph convolutional layer is as follows:

[0070] ,

[0071] in: For the first The node feature matrix of the layer has the dimension of the number of nodes x the feature dimension; For the first The node feature matrix of the layer; For adjacency matrices with added self-connections, This is the original adjacency matrix. for The identity matrix, The number of nodes; Let be a degree matrix, which is a diagonal matrix where the elements on the diagonal are the degrees of the nodes; This represents the negative 1 / 2 power of the degree matrix, used for normalization; For the first The weight matrix of a layer is the learnable parameter of the model; The activation function is such as ReLU or tanh. By stacking multiple layers of graph convolution, this invention can extract high-order correlation features between symptoms and syndromes layer by layer.

[0072] Furthermore, based on the annotation results of the training set, this invention uses the cross-entropy function as the loss function and updates the weight parameters of the graph convolutional layer using the backpropagation algorithm. The calculation formula for the cross-entropy loss function is as follows:

[0073] ,

[0074] in: This represents the cross-entropy loss value. The number of samples; The number of syndrome categories; For the sample Category The true label (0 or 1) when the sample Category The value is 1 if it is true, and 0 otherwise. Predict samples for the model The probability belonging to the category , with a value range of [0, 1]; is the natural logarithm; represents a summation operation. By minimizing the cross-entropy loss, the present application can optimize the model parameters and improve the diagnosis accuracy.

[0075] In the training process of the present application, the Adam optimizer is preferably used, the initial learning rate is set to 0.001, the batch size is 32, and the number of training iterations is 100 rounds. To prevent overfitting, the present application also introduces an early stopping strategy. When the loss on the validation set does not decrease for 5 consecutive rounds, the training process is stopped.

[0076] In still another embodiment of the present application, the symptom node input features include the medical records of the first hospitalization of each patient and the tongue and pulse diagnosis data as diagnosis information, and a weight vector of clinical symptoms is established based on a hierarchical graph neural network. The present application designs a multi-level feature extraction and fusion mechanism to fully utilize the complementary information of multi-modal data.

[0077] Preferably, the present application calculates the fractal features of the collected tongue and pulse diagnosis data, and calculates the tongue diagnosis and pulse diagnosis information of a single visit based on a graph neural network to obtain a fractal feature weight vector highly related to syndromes. To ensure effective fusion of features from different sources, the present application designs a feature alignment and fusion module. First, each modal feature is converted to a feature vector of the same dimension through an independent feature extraction network, and then multi-modal feature fusion is achieved through a feature fusion network. The feature fusion adopts an attention mechanism, and the calculation formula is as follows:

[0078] ,

[0079] ,

[0080] wherein Z is the fused feature vector, with the same dimension as each modal feature vector; is the feature vector of the i-th modal; is the attention weight of the i-th modal, representing the importance of the modal in the fusion process, with a value range of [0, 1] and satisfying ; is a learnable weight vector, which is a model parameter; represents the transpose of ; M is the number of modalities; exp is the natural exponential function; represents a summation operation. Through this mechanism, the present application can adaptively adjust the importance of different modal features to achieve optimal feature fusion effect.

[0081] In addition, the application extracts the historical visit records of each patient, and calculates the weight according to the contribution degree of the influencing syndrome factors, and fuses the historical visit and symptom fractal weight vectors. The historical visit records contain the time sequence information of the disease evolution of the patient, and have important value for predicting the evolution of the syndrome. The application uses a long short-term memory network (LSTM) to process time sequence data, and the calculation formula of the LSTM unit is as follows:

[0082] ,

[0083] ,

[0084] ,

[0085] ,

[0086] ,

[0087] ,

[0088] Among them: is a forget gate, which controls the retention degree of the memory unit at the last moment, and the value range is [0, 1]; is an input gate, which controls the acceptance degree of the current input, and the value range is [0, 1]; is an output gate, which controls the output degree of the memory unit content, and the value range is [0, 1]; is the candidate memory unit content, and the value range is [-1, 1]; is the memory unit content at the current moment, and the value range is [-1, 1]; , , , is a learnable weight matrix; , , , is a bias vector; is a Sigmoid function, defined as , the value range is [0, 1]; tanh is a hyperbolic tangent activation function, defined as , the value range is [-1, 1]; * is element multiplication, that is, the corresponding elements are multiplied; · represents matrix multiplication; represents vector splicing. Through the LSTM network, the application can effectively extract the time sequence features in the historical visit records of the patient, and provide an important basis for syndrome identification.

[0089] Meanwhile, the constitution information of each patient is taken as supplementary knowledge, and is fused with the historical syndrome evolution information features to form a symptom feature vector matrix, and a current syndrome probability distribution of the patient is calculated according to a parameter matrix of a hidden Markov chain. The constitution information is an important reference basis for TCM differentiation, and the constitution type of the patient is acquired through questionnaire investigation and expert evaluation, and is coded into a feature vector, which is fused with other features to form a complete patient feature representation.

[0090] In still another embodiment of the present application, the knowledge distillation step in the dynamic optimization of the syndrome differentiation model comprises training a lightweight syndrome differentiation model, migrating model feature layer weights to a large mobile device pre-training model, extracting model graph weights in combination with a knowledge distillation method, and migrating the graph weights to a small embedded device. Knowledge distillation is a technology for transferring the knowledge of a complex model (teacher model) to a simple model (student model), which can greatly reduce the model complexity while maintaining the model performance.

[0091] Preferably, the present application adopts a temperature-scaled soft label knowledge distillation method, and the loss function thereof is composed of a soft label distillation loss and a hard label classification loss, and the calculation formula is as follows:

[0092] ,

[0093] Among them: is the total loss of knowledge distillation; is the soft label of the teacher model under temperature , and is calculated as , wherein is the logit value output by the teacher model; is the predicted probability of the student model under temperature , and is calculated as , wherein is the logit value output by the student model; is the real label, which is one-hot encoded; is the standard predicted probability of the student model, i.e. the predicted probability when ; is the number of categories; is a balance factor, which controls the proportion of the soft label loss and the hard label loss; is a temperature parameter, which controls the "softness" of the soft label; is the natural logarithm; denotes the summation operation. The temperature parameter controls the "softness" of the soft label, and is generally taken as 2-5, and the present application preferably takes a value of 3. The balance factor controls the proportion of the soft label loss and the hard label loss, and the present application preferably takes a value of 0.7.

[0094] In addition, the present application encapsulates lightweight models based on Docker containers, establishes distributed servers, and provides models to third-party users. Docker container technology provides a consistent runtime environment, solving the environmental dependency problem in model deployment. The present application designs standardized API interfaces, supporting RESTful and gRPC calling methods, facilitating third-party application integration. API interface design follows the principle of simplicity and clarity, mainly including the following interfaces:

[0095] 1. / api / v1 / diagnose: receives patient multi-modal data and returns syndrome identification results;

[0096] 2. / api / v1 / predict: based on patient historical data, predicts syndrome evolution trend;

[0097] 3. / api / v1 / feedback: receives clinical feedback for model optimization;

[0098] Finally, the present application realizes third-party application calling of chronic liver disease model service, initiates API request to container cluster, and obtains diagnosis results returned to user intelligent terminal. In order to ensure high availability and low delay of service, the present application adopts load balancing and service degradation strategy to ensure that the system can still run normally under high concurrency and unstable network conditions.

[0099] In another embodiment of the present application, the feature weight learning process in the dynamic optimization of syndrome differentiation model is realized, and a hierarchical graph neural network is built based on deep learning and knowledge graph for node classification. The feature matrix extracted from the tongue and pulse tongue-biochemical-syndrome data set by the deep learning subnetwork is taken as input, and the graph neural network is combined to learn the syndrome feature distribution of the patient.

[0100] Preferably, the present application constructs a TCM syndrome knowledge graph, formalizing symptoms, syndromes and their relationships in TCM theory into a knowledge graph structure. The entities in the knowledge graph include symptoms (such as bitter mouth, abdominal distension), syndromes (such as liver stagnation and spleen deficiency syndrome), etc., and the relationships include "symptom-indicates-syndrome", "syndrome-evolves-syndrome", etc. Through the construction of knowledge graph, the present application combines TCM theoretical knowledge with data-driven methods, improving the explainability and accuracy of the model.

[0101] In addition, the application designs a hierarchical graph neural network structure, including a bottom feature extraction network, a middle relationship learning network and a top syndrome reasoning network. The bottom feature extraction network is responsible for extracting basic features from multi-modal data, the middle relationship learning network learns the relationship between symptoms and between symptoms and syndromes, and the top syndrome reasoning network performs final syndrome reasoning based on the learned features and relationships. This hierarchical structure can extract and fuse multi-modal features layer by layer, realize the mapping from low-level features to high-level semantics, and improve the accuracy and explainability of the diagnosis.

[0102] In the model training process, the application adopts a multi-task learning strategy to simultaneously optimize the syndrome classification task and the syndrome transfer prediction task, fully utilizes the associated information between tasks, and improves the generalization ability of the model. The loss function of multi-task learning is as follows:

[0103] ,

[0104] Among them: is the total loss of multi-task learning; is the syndrome classification loss, which adopts cross-entropy loss; is the syndrome transfer prediction loss, which adopts mean square error loss; and are balance factors to control the weights of the task losses, and satisfy . The application preferably takes , , which ensures the accuracy of syndrome classification while considering the accuracy of syndrome transfer prediction.

[0105] In another embodiment of the application, in the extraction of the fractal features of the sublingual collateral vessels, a second-order Laplacian operator is used for fractal calculation, and the specific steps include converting the tongue image into a binary image based on a deep learning algorithm, processing the pixel values of the original image into two types of binary images, extracting the tongue contour using a threshold segmentation algorithm, filling the hollow area of the tongue, and obtaining a tongue binary image.

[0106] Preferably, the application uses an Otsu adaptive threshold algorithm for tongue segmentation, which can automatically determine the optimal segmentation threshold and effectively separate the tongue from the background. The principle of the Otsu algorithm is to maximize the inter-class variance, and its objective function is:

[0107] ,

[0108] Among them: is the inter-class variance, which represents the separation degree of the foreground and background gray value distribution; and are the pixel ratios of the foreground and background below the threshold , respectively, and satisfy ; and are the average gray values of foreground and background respectively; is the segmentation threshold, usually in the range of [0, 255]; represents the square of the difference between the average gray values of foreground and background. The optimal threshold is the threshold that makes the maximum;

[0109] ,

[0110] wherein: represents the independent variable that makes the objective function reach the maximum value. Through the Otsu algorithm, the tongue region can be accurately segmented, laying a foundation for subsequent feature extraction.

[0111] In addition, the skeleton of the tongue is obtained by skeleton transformation, and tongue image information is extracted according to the skeleton length and the number of bifurcations. Skeleton transformation is a morphological operation that can simplify the object in a binary image into a skeleton line with a single pixel width, preserving the topological structure and geometric features of the object. The Zhang-Suen thinning algorithm is used for skeleton extraction in the present application, which has the characteristics of high computational efficiency and smooth results by iteratively deleting boundary points to obtain the skeleton.

[0112] Meanwhile, the present application establishes the graph structure of the tongue image based on the second-order Laplacian operator, calculates the distance curve in the horizontal and vertical directions of the tongue body respectively, and calculates the fractal dimension of the image using Euler's formula. The second-order Laplacian operator can effectively detect the edges and details in the image, and its discrete form is:

[0113] ,

[0114] wherein is the gray value of the image at point , usually in the range of [0, 255]; is the result of the second-order Laplacian operator, representing the approximate value of the second-order derivative of the image at point ; are the gray values of the upper, lower, left and right four adjacent points of point . After applying the operator, the fractal dimension of the image is estimated by the box counting method, which principle is to calculate the number of boxes required to cover the image at different scales, and then estimate the fractal dimension through the slope of the double logarithmic graph.

[0115] Finally, this invention acquires the fractal dimension features of tongue images and establishes a database. Based on a deep learning model, it calculates the feature vectors of the tongue-pulse-biochemical-syndrome set features, maps the patient's feature vectors through a neural network, and calculates the weight vectors of the tongue-pulse-biochemical-syndrome set using a graph neural network. This invention constructs a comprehensive database containing tongue image fractal features, pulse fractal features, biochemical indicators, and syndrome types, providing data support for model training and validation.

[0116] In another embodiment of the present invention, the node graph neural network of the graph neural network uses a symmetric structure, including graph convolutional layers, pooling layers, activation layers, and nonlinear functions. The dimension of the input layer feature vector is determined by the number of input layer nodes. The graph convolutional layer matrix is ​​used to learn the weights of the node features. The node vectors are aggregated through the graph propagation layer. The fused features are calculated according to the activation layer. The graph node probability vector is calculated through the nonlinear function. Cross-entropy is used as the loss function. The weight parameters of each layer are updated according to the backpropagation algorithm.

[0117] Preferably, the graph neural network model of the present invention specifically includes three graph convolutional layers, each followed by a batch normalization layer, a ReLU activation function, and a Dropout layer (with a dropout rate of 0.3). The dimensions of the graph convolutional layers are [64, 32, 16], and finally, the probability distribution of the syndrome is output through a fully connected layer and a softmax function. This structural design balances the expressive power and computational complexity of the model, making it suitable for deployment on hardware devices of different scales.

[0118] In this invention, the model node propagation process specifically includes the following steps: calculating the contribution of symptom nodes to diagnostic nodes, obtaining symptom feature vectors based on graph convolutional layer weight mapping, and multiplying the symptom feature vectors by the weights of each symptom node and summing them to obtain the diagnostic node feature vectors. The core of the graph convolutional layer is the message passing and aggregation mechanism, and its calculation process can be divided into three steps: message calculation, message aggregation, and feature updating. The specific calculation formula is as follows:

[0119] ,

[0120] ,

[0121] ,

[0122] in: For nodes Send to node The message is a vector with the same dimension as the node features; This is a message computation function used to calculate the information transmitted between nodes; and They are nodes and nodes eigenvectors; For the edge The characteristics of nodes and nodes Characteristics of the relationship between them; For nodes The received aggregated message is a vector with the same dimensions as the node features; This is an aggregation function used to integrate messages from multiple neighboring nodes; For nodes The set of neighboring nodes, containing nodes All directly connected nodes; For the updated node A feature is a vector, and its dimensions may differ from those of the original feature. This is an update function used to generate new features by combining the original features and the aggregated message; Represents a node The set of messages sent by all neighboring nodes. In this invention, , and The specific form is as follows:

[0123] ,

[0124] ,

[0125] ,

[0126] in: and is the weight matrix, which contains the learnable parameters of the model; This is the bias vector, which is also a model parameter; For activation functions, such as ReLU; Indicates a node The summation is calculated from all neighboring nodes. Through this message-passing mechanism, the present invention can effectively learn the interaction relationships between nodes and improve the accuracy of symptom identification.

[0127] Furthermore, this invention classifies nodes using pooling layers and calculates the probability distribution of diagnostic results based on conditional entropy. Pooling operations are used to reduce feature dimensionality and extract the most important feature information. This invention employs graph attention pooling, performing a weighted average based on the importance of nodes, calculated using the following formula:

[0128] ,

[0129] ,

[0130] ,

[0131] in: For nodes The importance score is a scalar; The query vector is a learnable parameter. The weight matrix contains learnable parameters. The bias vector is a learnable parameter; For nodes eigenvectors; h is the hyperbolic tangent activation function, which maps the input to the interval [-1, 1]. express transpose; For nodes The attention weight represents the importance of the node in the pooling process, with a value range of [0,1] and satisfying the following conditions: It is a natural exponential function; The pooled graph-level features are vectors with the same dimensions as the node features. The total number of nodes; This represents a summation operation. Through graph attention pooling, this invention can adaptively focus on important nodes, improving the efficiency and accuracy of feature extraction.

[0132] Simultaneously, this invention updates parameters and calculates the gradient of the loss function using the BP (Backpropagation) algorithm. The BP algorithm is a standard algorithm for training neural networks, calculating the gradient of the loss function with respect to the parameters of each layer using the chain rule, and then updating the parameters using gradient descent. In this invention, the BP algorithm is implemented based on automatic differentiation technology, which can automatically calculate the gradient of complex models, simplifying the training process.

[0133] Finally, this invention optimizes the number of graph network nodes and adjusts the learning parameters and activation function layers according to the size of the learning dataset to obtain generalization ability suitable for datasets of different sizes. For small datasets (number of samples < 1000), this invention reduces the number of model parameters and increases the regularization strength (e.g., setting the L2 regularization coefficient to 0.01) to prevent overfitting; for large datasets (number of samples > 10000), this invention increases the model complexity (e.g., increasing the number of graph convolutional layers to 5 and increasing the feature dimensions to [128, 96, 64, 32, 16]) to improve the model's expressive power. This adaptive model complexity adjustment strategy enables this invention to achieve good performance on datasets of different sizes.

[0134] In summary, this invention provides a dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks. Through multi-source heterogeneous data fusion, hierarchical graph neural networks, adaptive attention mechanisms, hidden Markov chain syndrome transfer prediction, and knowledge distillation, it achieves accurate diagnosis and personalized treatment of chronic liver disease, which has significant theoretical value and application prospects.

[0135] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks, characterized in that, Includes the following steps: Acquiring multi-source heterogeneous data includes: collecting tongue images, pulse diagnosis data, biochemical indicators, and clinical symptom data of patients with chronic liver disease from multiple medical institutions; preprocessing the tongue images to extract sublingual vein fractal features; and extracting features from the pulse diagnosis data to obtain pulse fractal features. Constructing a hierarchical syndrome association network includes: constructing a topological spatial mapping between symptoms and syndromes based on the fractal features of the sublingual veins and the fractal features of the pulse; applying spectral graph theory to perform structural analysis on the symptom-syndrome association network to obtain the spectral representation of syndrome association; and using graph neural networks to learn the influence relationship between symptom nodes and syndrome nodes to form a syndrome probability distribution. To achieve dynamic optimization of the syndrome differentiation model, the following steps are taken: assessing the contribution of each symptom to syndrome judgment based on information entropy and dynamically adjusting the attention weight of symptom features; introducing hidden Markov chains to capture syndrome transfer patterns and predict syndrome evolution trends; and performing knowledge distillation on the syndrome differentiation model to achieve lightweight deployment.

2. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 1, characterized in that, The acquisition of multi-source heterogeneous data specifically includes: collecting medical information of patients diagnosed with chronic liver disease from multiple tertiary hospitals and community health service centers, including patient complaints, tongue and pulse diagnosis, laboratory test data, and historical medical information; parsing clinical medical record information and tongue image reports based on the BertNer Chinese named entity recognition algorithm, and performing entity matching between text and images to filter feature words and indicator fields associated with the disease; integrating and normalizing the indicator fields recorded repeatedly in the medical record information, and retaining the text records of complaints and tongue image data.

3. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 1, characterized in that, The extraction of sublingual vein fractal features specifically includes: establishing a patient electronic medical record dataset according to disease category; extracting and preprocessing tongue image data, using a second-order Laplacian graph fractal algorithm to obtain the fractal dimension features of tongue diagnosis images, and obtaining sublingual vein fractal features based on the depth distribution of sublingual veins; extracting features from pulse pressure waveform data and selecting pulse fractal features that can characterize the patient's syndrome; and extracting the course of chronic liver disease, duration of illness, smoking, drinking, psychological symptoms, and laboratory results for each patient based on historical medical records.

4. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 1, characterized in that, The construction of the hierarchical syndrome association network specifically includes: establishing a tongue and pulse diagnosis fusion graph network; constructing a symptom and syndrome feature fusion graph based on the contribution of clinical symptoms to the syndrome; using a graph neural network to learn node and edge weights; calculating the influence of symptom nodes on syndrome nodes; calculating the contribution weight of symptoms to the syndrome respectively; and weighting and summing the graph model weights and symptom features to obtain the probability distribution of each syndrome type in chronic liver disease, and determining the syndrome category to which the patient belongs based on a decision threshold; using the patient's initial recorded symptoms and tongue appearance as input to the syndrome inference model, and the model outputs the patient's syndrome probability distribution, and based on... The weight vector of the graph model is adjusted according to the symptom weights. The predicted values ​​calculated by the model and the probability distribution of syndrome types labeled by experts are used as the initial parameters of the Hidden Markov Chain. The Hidden Markov Chain is introduced to capture the clinical syndrome transfer links of patients. The parameter matrix and initial state are determined according to the patient's first visit symptoms and diagnosis results. The parameter matrix is ​​updated based on forward and backward forward dynamic programming algorithms combined with backpropagation, thereby updating the graph network weights of the inference model according to the medical record data. In the process of syndrome inference and sample learning, the graph model weights and parameters are iterated using an incremental learning method, and the model parameters are updated according to the evolution of the patient's syndrome.

5. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 4, characterized in that, The graph network weight calculation includes: learning weight feature vectors based on the graph neural network, generating a symmetrical graph using the patient's initial diagnosis information as nodes, and introducing symptom nodes to learn weights; inputting symptom features into the graph convolutional network for node fusion, multiplying the symptom features and weights through the graph convolutional layer parameter matrix and then summing to obtain the diagnosis result vector; and updating the graph convolutional layer weight parameters using the cross-entropy function as the loss function based on the training set annotation results and the backpropagation algorithm.

6. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 4, characterized in that, The symptom node input features include: using each patient's initial admission medical records and tongue and pulse diagnosis data as diagnostic information, and establishing a weight vector for clinical symptoms based on a hierarchical graph neural network; performing fractal feature calculations on the collected tongue and pulse diagnosis data, and calculating the tongue and pulse diagnosis information of a single visit based on a graph neural network to obtain a fractal feature weight vector highly correlated with the syndrome; extracting features from each patient's historical medical records, and calculating weights based on the contribution of factors influencing the syndrome, and fusing the historical medical records with the symptom fractal weight vectors; using each patient's constitution information as supplementary knowledge, fusing it with historical syndrome evolution information features to form a symptom feature vector matrix, and calculating the patient's current syndrome probability distribution based on the parameter matrix of a hidden Markov chain.

7. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 1, characterized in that, The knowledge distillation steps in the dynamic optimization of the dialectical model include: training a lightweight dialectical model, transferring the model feature layer mapping weights to a pre-trained model on a large mobile device, extracting the model graph weights using the knowledge distillation method, and transferring the graph weights to a small embedded device; encapsulating the lightweight model based on Docker containers, establishing a distributed server, and providing the model to third-party users; and having third-party applications call the chronic liver disease model service, initiate API requests to the container cluster, and obtain diagnostic results to return to the user's smart terminal.

8. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 1, characterized in that, The feature weight learning process in the dynamic optimization of the syndrome differentiation model is based on deep learning and knowledge graph to build a hierarchical graph neural network for node classification. The feature matrix extracted from the tongue diagnosis, pulse diagnosis, tongue-biochemistry-syndrome data set by the deep learning sub-network is used as input, and the graph neural network is combined to learn the syndrome feature distribution of the patient.

9. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 3, characterized in that, In the extraction of sublingual vein fractal features, fractal calculation is performed using a second-order Laplacian operator. The specific steps are as follows: Based on a deep learning algorithm, the tongue image is converted into a binary image. The pixel values ​​of the original image are processed into positive and negative binary images. The tongue contour is extracted using a threshold segmentation algorithm, and the hollow areas of the tongue are filled to obtain a binary image of the tongue. The skeleton of the tongue is obtained using skeleton transformation, and tongue image information is extracted based on the skeleton length and number of branches. The graphic structure of the tongue image is established based on the second-order Laplacian operator. The distance curves in the horizontal and vertical directions of the tongue are calculated respectively, and the fractal dimension of the image is calculated using Euler's formula. The fractal dimension features of the tongue image are obtained and a database is established. The feature vector of the tongue-pulse-biochemical-syndrome set features is calculated based on a deep learning model. The feature vector of the patient is mapped through a neural network, and the weight vector of the tongue-pulse-biochemical-syndrome set is calculated by combining a graph neural network.

10. The dynamic optimization method for a multimodal diagnostic model of chronic liver disease based on graph neural networks according to claim 5, characterized in that, The node-based graph neural network described uses a symmetric structure, including graph convolutional layers, pooling layers, activation layers, and nonlinear functions. The dimension of the input layer feature vector is determined by the number of input layer nodes. Graph convolutional layer matrices are used to learn the weights of node features. Graph propagation layers aggregate the node vectors, activation layers calculate fused features, and nonlinear functions calculate graph node probability vectors. Cross-entropy is used as the loss function. Backpropagation updates the weight parameters of each layer. The model node propagation process specifically includes the following steps: calculating the contribution of symptom nodes to diagnostic nodes, obtaining symptom feature vectors based on graph convolutional layer weight mapping, multiplying the symptom feature vectors by the weights of each symptom node and summing the results to obtain the diagnostic node feature vectors; classifying nodes through pooling layers and calculating the probability distribution of diagnostic results based on conditional entropy; updating parameters using the backpropagation algorithm and calculating the gradient of the loss function; optimizing the number of graph network nodes based on the size of the learning dataset, adjusting the learning parameters and the number of activation function layers to obtain generalization ability suitable for datasets of different sizes.

Images