Knowledge Graph-Based Laboratory Report Review and Clinical Meaning Generation Assistance System

By using a knowledge graph-based test report review system, combined with technologies such as dual-graph branching and residual graph convolutional networks, the problems of joint judgment of multiple indicators and identification of symmetrical lesions have been solved, achieving high-precision and highly interpretable test report review.

CN123072609APending Publication Date: 2026-11-10HUNAN YUNYING INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611388613.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-10

Smart Images

  • Figure CN123072609A_ABST
    Figure CN123072609A_ABST
Patent Text Reader

Abstract

This invention relates to the field of intelligent review technology for medical testing. It discloses a knowledge graph-based system for reviewing test reports and assisting in generating clinical meaning, comprising: a data input module for receiving data; a dual-graph branch construction module for constructing a first graph branch and a second graph branch; a residual graph convolution stacking module for performing neighbor aggregation on the dual branches to generate first and second aggregated features for each node; a node-level adaptive gating module for generating first and second weights to adjust the contribution of the two features to node updates; a symmetry-aware residual convolution module for calculating the feature differences between mirror pairs and applying symmetry constraints; a self-pairing regularization training module for flipping the second branch and applying consistency constraints to the features before and after the flip; a feature fusion module for concatenating features from different layers across layers to generate a fusion matrix; and a classification output module for mapping the fusion matrix to output the review results and clinical descriptions. This invention achieves high-precision review of test reports.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent review technology for medical testing, and in particular to a knowledge graph-based system for reviewing test reports and assisting in the generation of clinical significance. Background Technology

[0002] Laboratory report review is a crucial aspect of clinical laboratory quality management. Traditional review methods rely primarily on manual work by laboratory personnel, resulting in low efficiency, inconsistent interpretations, and lack of standardized criteria. With the rapid development of laboratory medicine, the number of daily laboratory reports generated by clinical laboratories continues to increase, posing significant challenges to the efficiency and accuracy of manual review.

[0003] To address these issues, automated review systems based on pre-defined rules are gradually being adopted in clinical laboratories. These systems allow staff to set review standards and logical rules, which are then embedded into the laboratory information system to automatically review test reports. However, rule-based review systems primarily issue results without obvious abnormalities. Rule engines struggle to cover situations involving combined judgments of multiple indicators, individualized difference analysis, and complex clinical logical reasoning. Furthermore, existing applications of knowledge graphs in test report review mainly focus on the construction and retrieval of test knowledge graphs, and still have shortcomings in terms of reasoning depth and interpretability.

[0004] In recent years, large language models and retrieval-enhanced generation technologies have shown potential in interpreting medical examination reports, enabling the identification of potential disease risks from test reports and the generation of interpretive text. However, existing methods mainly focus on the overall interpretation of medical examination reports, lacking systematic modeling of prior knowledge such as the correlation between test indicators and the symmetry of human anatomical structures. This makes it difficult to fully capture the differential characteristics of multi-indicator joint abnormal patterns and bilateral symmetrical structural lesions.

[0005] In summary, the existing technologies have the following shortcomings in the review of test reports: (1) Rule-based systems have difficulty in handling joint judgment of multiple indicators and complex clinical logical reasoning; (2) Existing schemes that combine knowledge graphs and LLM lack systematic modeling of prior knowledge of the relationship between test indicators and the symmetry of anatomical structures. Summary of the Invention

[0006] One objective of this invention is to propose a knowledge graph-based system for reviewing test reports and assisting in the generation of clinical meaning. This invention combines dual-graph branching, residual graph convolutional neural networks, node-level adaptive gating, symmetry-aware residual convolution, and self-pairing regularization training to describe the entire process from data input, dual-graph branching, feature aggregation and updating, dynamic gating adjustment, mirror node symmetry constraints, flip enhancement and consistency constraints, cross-layer feature fusion, to review classification and clinical text generation. It has the advantages of high review accuracy, accurate identification of symmetrical lesions, strong interpretability, and stable generalization performance.

[0007] The knowledge graph-based laboratory report review and clinical significance generation system according to an embodiment of the present invention includes: The data input module is used to receive test report data and prior anatomical data; The dual-graph branching module is used to construct the first graph branch based on the numerical similarity of the test report data, and to construct the second graph branch by dissecting the mirror node pairings in the prior data. In the second graph branch, each mirror node pair shares a symmetric constraint hyperedge. The residual graph convolutional layer stacking module, based on the residual graph convolutional neural network architecture, is used to perform neighbor aggregation on the dual graph branches and generate the first aggregation feature and the second aggregation feature of each node. The node-level adaptive gating module is used to generate a first gating weight and a second gating weight for each node. The first gating weight adjusts the contribution of the first aggregated feature to the node update, and the second gating weight adjusts the contribution of the second aggregated feature to the node update. The second gating weight is set to 0 or 1 according to the abnormal flag of the mirror node alignment index. It is set to 0 when one side is abnormal and to 1 when both sides are normal. The symmetry-aware residual convolution module is used to calculate the feature difference values ​​between mirror node pairs as symmetry consistency constraints, and the second gate weight adjusts the relaxation strength of the constraints. The self-pairing regularization training module is used to perform flip enhancement on the second graph branch and apply consistency constraints to the node feature output before and after the flip. The feature fusion module is used to perform cross-layer concatenation of the aggregated features output from each residual map convolutional layer to generate a fused feature matrix; The classification output module is used to map the fusion feature matrix to generate review results and clinical description text.

[0008] Optionally, the test report data and prior anatomical data in the data input module specifically include: The data in the test report includes: patient identification, patient age, patient gender, department visited, sample type, sample collection time, test item name, test item measurement value, test item unit, upper limit of the reference range for the test item, lower limit of the reference range for the test item, and abnormal markers for the test item. Prior anatomical data include: organ names, left-right symmetrical pairing of organs in standard anatomical positions, and anatomical distance thresholds between mirror node pairs.

[0009] Optionally, the construction of the first graph branch and the second graph branch by the dual-graph branch construction module specifically includes: The first graph branch is constructed by using the test items in the test report data as nodes of the first graph branch and the numerical similarity between two test items as the weight of the edge between nodes in the first graph branch. Numerical similarity is determined by the co-occurrence frequency of two test items in the same patient and the same batch of visits in historical test reports and the correlation of measured values. The correlation coefficient of measured values ​​is calculated using the Pearson correlation coefficient. After normalizing the numerical similarity to the [0,1] interval, it is used as the value of the corresponding position in the first graph branch adjacency matrix. The nodes of the second graph branch are the test items in the test report data, the anatomical structures in the anatomical prior data, and the diseases that are clinically related to the test items. The left-right symmetrical pairing relationship of organs recorded in the anatomical prior data under the standard anatomical position is used as the basis for mirror node pairing. The anatomical distance between mirror nodes is less than a preset threshold as the condition for determining the validity of the pairing. The pairing relationship of mirror nodes is used as the edge connection to construct the second graph branch. In the second graph branch, each pair of mirror nodes is connected to the same symmetric constraint hyperedge. The symmetric constraint hyperedge connects both the mirror node pair and the common associated disease node corresponding to the mirror node pair. The common associated disease nodes corresponding to mirror node pairs are the intersection or union of the disease nodes associated with the nodes on either side of the mirror node pair and the disease nodes associated with the nodes on the other side. When there are disease nodes in the intersection, the disease nodes are added to the symmetric constraint hyperedge as common associated disease nodes. When there are no disease nodes in the intersection, all disease nodes in the union are added to the symmetric constraint superedge as commonly associated disease nodes.

[0010] Optionally, the generation of the first aggregated feature and the second aggregated feature of each node in the residual graph convolutional layer stacking module specifically includes: Receive the first graph branch and the second graph branch. For each node in the first graph branch, use the weighted sum of the node's own feature vector and the features of the neighboring nodes recorded in the adjacency matrix as the message passing result. After mapping the message passing result through a non-linear activation function, perform residual connection with the node's own feature vector to obtain the first aggregated feature output by this layer. For each node in the second graph branch, the node's own feature vector is weighted and summed with the adjacency matrix constructed by the mirror pairing relationship in the second graph branch. The summation result is mapped by a non-linear activation function and then residually connected with the node's own feature vector to obtain the second aggregated feature output by this layer. For each node, the first and second aggregated features output by the current residual graph convolutional layer are used as the input features of the next residual graph convolutional layer, respectively. After N residual graph convolutional layers are stacked, each node obtains the first and second aggregated features updated by N residual graph convolutional operations.

[0011] Optionally, the generation of the first gating weight and the second gating weight for each node in the node-level adaptive gating module specifically includes: The node-level adaptive gating module receives the first and second aggregate features of each node output by the residual graph convolutional layer stacking module. For each node, the first aggregate feature is mapped to the [0,1] interval through the first fully connected layer to output the first gating weight, and the second aggregate feature is mapped to the [0,1] interval through the second fully connected layer to output the second gating weight. The sum of the first gating weight and the second gating weight is 1. When each node updates its features in each residual graph convolutional layer, the product of the first aggregated feature and the first gate weight is taken as the first contribution component, and the product of the second aggregated feature and the second gate weight is taken as the second contribution component. The input features of the previous layer of the node are calculated with the first contribution component and the second contribution component respectively to obtain the updated features of the node in the current layer.

[0012] Optionally, setting the second gating weight to 0 or 1 based on the anomaly flag of the mirror node alignment index in the node-level adaptive gating module specifically includes: The node-level adaptive gating module reads the abnormal markers of the inspection items corresponding to each node in the inspection report data and the mirror node. The abnormal markers are generated by comparing the measured values ​​of the inspection items with the preset normal value range. When the measured value is lower than the lower limit of the reference range or higher than the upper limit of the reference range, it is marked as abnormal. When the measured value is between the lower limit and the upper limit of the reference range, it is marked as normal. When the test item corresponding to any one node in the mirror node pair is marked as abnormal while the test item corresponding to the other node is marked as normal, the second gating weight corresponding to the mirror node pair is assigned to 0, and the first gating weight is assigned to 1. The node update depends only on the first aggregated feature. When the abnormal flags of the test items corresponding to the two nodes in the mirror node pair are both normal, the second gating weight corresponding to the mirror node pair is assigned a value of 1, and the first gating weight is assigned a value of 0. The node update depends only on the second aggregated feature. The node-level adaptive gating module multiplies the assigned first and second gating weights by their respective aggregated features before participating in the node feature update of the current residual graph convolutional layer.

[0013] Optionally, the calculation of feature difference values ​​between mirror node pairs as symmetry consistency constraints by the symmetry-aware residual convolution module specifically includes: Extract the second aggregated feature of each node from each residual graph convolutional layer. For each mirror node pair in the second graph branch, calculate the Euclidean distance between the second aggregated feature of the left node and the second aggregated feature of the right node as the feature difference value. Map the feature difference value to the [0,1] interval and use it as the original symmetric consistency constraint value of the mirror node pair in the current layer. The symmetry-aware residual convolution module receives the second gating weights output by the node-level adaptive gating module, multiplies the original symmetry consistency constraint value with the second gating weights, and obtains the weighted symmetry consistency constraint value. The symmetry-aware residual convolution module takes the average of the weighted symmetry consistency constraint values ​​of all mirror node pairs across all residual map convolutional layers, and uses this as the global symmetry consistency constraint loss term. This global symmetry consistency constraint loss term is then added to the main loss function of the classification output module to form a joint optimization objective.

[0014] Optionally, the self-pairing regularization training module's flipping enhancement of the second graph branch specifically includes: The self-pairing regularization training module performs horizontal flip enhancement on the second graph branch during the training phase. Based on the mirror node pairing relationship, it swaps the features of the left node and the right node in each mirror node pair in the second graph branch, while keeping the edge connection relationship and hyperedge connection relationship between the nodes unchanged, and generates the flipped second graph branch. The self-pairing regularization training module inputs the original second graph branch and the flipped second graph branch into the residual graph convolutional layer stacking module to obtain the second aggregated features of each node corresponding to the original second graph branch and the second aggregated features of each node corresponding to the flipped second graph branch. The self-pairing regularization training module locates the corresponding mirror pairing node for each node in the original second graph branch and the mirror pairing node in the flipped second graph branch. It calculates the mean squared error between the second aggregated feature of the node in the original second graph branch and the second aggregated feature of the mirror pairing node in the flipped second graph branch. The mean squared error of all nodes is taken as the consistency constraint loss term, and the consistency constraint loss term is added to the main loss function of the classification output module.

[0015] Optionally, the feature fusion module performs cross-layer concatenation of the aggregated features output from each residual map convolutional layer to generate a fused feature matrix, specifically including: The feature fusion module extracts the first and second aggregate features of each node from each residual map convolutional layer to obtain N sets of feature matrices, each set containing the first and second aggregate features of all nodes in the current layer; The feature fusion module concatenates N sets of feature matrices along the feature dimension. Using the original input feature vector of each node as the reference, the first aggregated feature of the same node in layers 1 to N is arranged sequentially after the reference feature, and the second aggregated feature of the same node in layers 1 to N is arranged sequentially after the first aggregated feature, forming the cross-layer fusion feature vector of the node. The feature fusion module arranges the cross-layer fusion feature vectors of all nodes in node order to form a fusion feature matrix.

[0016] Optionally, the classification output module's mapping of the fusion feature matrix to generate the review results and clinical description text specifically includes: The classification output module performs global average pooling on the fused feature matrix to obtain a fixed-length global feature vector. The global feature vector is then input into the first fully connected classifier, which is activated by softmax and outputs the probability distribution of each preset review category. The category corresponding to the maximum probability is used as the review classification result. The review categories include three categories: normal, abnormal pending review, and critical value warning. The classification output module concatenates the global feature vector with the category embedding vector corresponding to the review classification result. The concatenated vector is then input into a Transformer-based decoder. The decoder generates clinical description text word by word in an autoregressive manner. The clinical description text includes a list of abnormal test items, a description of the degree of deviation of abnormal indicators, the comparison results with mirror node pairs, and clinical prompts based on the comparison results. The classification output module outputs the review classification results and clinical description text as system outputs. The review classification results are output in the form of structured tags, and the clinical description text is output in the form of natural language.

[0017] The beneficial effects of this invention are: (1) This invention uses a dual-graph branch construction module to separate and model the numerical similarity association of test indicators and the prior knowledge of anatomical structure mirror pairing, so that the model can explicitly distinguish between functional association and structural association. Compared with the test report review method with a single graph structure, it solves the problem of decreased reasoning accuracy caused by mixed modeling of multiple types of medical associations and improves the sensitivity in the detection of test abnormalities related to bilateral symmetrical organs.

[0018] (2) The present invention uses a node-level adaptive gating module to set the second gating weight to 0 or 1 according to the two states of unilateral abnormality or bilateral normality in the mirror node pair. This allows the second graph branch to fully transmit symmetric prior constraints when both sides are normal and to completely remove constraints when one side is abnormal in order to retain unilateral lesion features. Compared with the fixed weight multi-branch fusion method, this invention solves the problem of information interference between normal side features and abnormal side features.

[0019] (3) In this invention, the feature difference value between mirror node pairs is calculated by the symmetry-aware residual convolution module as a symmetry consistency constraint term, and the relaxation strength of the constraint term is adjusted by the second gate weight, so that the network explicitly learns the inherent law of the bilateral symmetric structure during the training process. Compared with graph convolutional networks without symmetric priors, the feature difference value between mirror node pairs is reduced, and the output features of bilateral normal samples tend to be consistent.

[0020] (4) The present invention performs flip enhancement on the second graph branch through the self-pairing regularization training module and applies consistency constraints to the node feature output before and after flipping, so that the model internalizes the position invariance of the bilateral symmetric structure during training. Compared with the model that only depends on the original data distribution, the output feature deviation on the input data after mirror flipping is reduced and the model generalization ability is improved.

[0021] (5) This invention uses the residual map convolutional layer stacking module and cross-layer feature fusion to generate a fusion feature matrix by cross-layer splicing of the aggregated features output by each layer, so that shallow local structural information and deep global semantic information can participate in the review decision in the classification output module. Compared with the method that only uses the last layer features, the classification accuracy is improved. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the module of the knowledge graph-based test report review and clinical significance generation system proposed in this invention; Figure 2 This is a schematic diagram of the node-level adaptive gating and dual-graph branching collaborative mechanism of the knowledge graph-based test report review and clinical significance generation system proposed in this invention; Figure 3 This is a schematic diagram of the symmetry-enhanced residual graph convolutional network for the knowledge graph-based test report review and clinical significance generation system proposed in this invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figures 1-3 A knowledge graph-based system for reviewing and generating clinically relevant laboratory reports includes: Environmental data acquisition module: used for gamma-ray dose rate data, wind speed data, wind direction data and geographical location information, and to construct a nuclear radiation environment dataset; Dose rate prediction module: used to predict future trends in gamma-ray dose rate data based on a sparse recurrent neural network model; This invention achieves data perception, intelligent modeling, and collaborative analysis throughout the entire nuclear radiation early warning process by constructing six functional modules: environmental data acquisition, dose rate prediction, anomaly detection, diffusion area modeling, grid risk evolution, and fusion judgment. The closed-loop data flow and complementary functional logic among the modules effectively solve the problems of information isolation, incomplete modeling, and delayed response in existing technologies. The system boasts advantages such as clear structure, well-defined division of labor, high processing efficiency, and strong scalability, significantly improving the intelligence level and practical performance of nuclear radiation monitoring and early warning.

[0025] In this embodiment, the modules are connected through the following method: The data input module is used to receive test report data and prior anatomical data; The dual-graph branching module is used to construct the first graph branch based on the numerical similarity of the test report data, and to construct the second graph branch by dissecting the mirror node pairings in the prior data. In the second graph branch, each mirror node pair shares a symmetric constraint hyperedge. The residual graph convolutional layer stacking module, based on the residual graph convolutional neural network architecture, is used to perform neighbor aggregation on the dual graph branches and generate the first aggregation feature and the second aggregation feature of each node. The node-level adaptive gating module is used to generate a first gating weight and a second gating weight for each node. The first gating weight adjusts the contribution of the first aggregated feature to the node update, and the second gating weight adjusts the contribution of the second aggregated feature to the node update. The second gating weight is set to 0 or 1 according to the abnormal flag of the mirror node alignment index. It is set to 0 when one side is abnormal and to 1 when both sides are normal. The symmetry-aware residual convolution module is used to calculate the feature difference values ​​between mirror node pairs as symmetry consistency constraints, and the second gate weight adjusts the relaxation strength of the constraints. The self-pairing regularization training module is used to perform flip enhancement on the second graph branch and apply consistency constraints to the node feature output before and after the flip. The feature fusion module is used to perform cross-layer concatenation of the aggregated features output from each residual map convolutional layer to generate a fused feature matrix; The classification output module is used to map the fusion feature matrix to generate review results and clinical description text.

[0026] In this embodiment, the test report data and prior anatomical data in the data input module specifically include: The data in the test report includes: patient identification, patient age, patient gender, department visited, sample type, sample collection time, test item name, test item measurement value, test item unit, upper limit of the reference range for the test item, lower limit of the reference range for the test item, and abnormal markers for the test item. Prior anatomical data include: organ names, left-right symmetrical pairing of organs in standard anatomical positions, and anatomical distance thresholds between mirror node pairs.

[0027] In this embodiment, the construction of the first graph branch and the second graph branch by the dual-graph branch construction module specifically includes: The first graph branch is constructed by using the test items in the test report data as nodes of the first graph branch and the numerical similarity between two test items as the weight of the edge between nodes in the first graph branch. Numerical similarity is determined by the co-occurrence frequency of two test items in the same patient and the same batch of visits in historical test reports and the correlation of measured values. The correlation coefficient of measured values ​​is calculated using the Pearson correlation coefficient. After normalizing the numerical similarity to the [0,1] interval, it is used as the value of the corresponding position in the first graph branch adjacency matrix. The nodes of the second graph branch are the test items in the test report data, the anatomical structures in the anatomical prior data, and the diseases that are clinically related to the test items. The left-right symmetrical pairing relationship of organs recorded in the anatomical prior data under the standard anatomical position is used as the basis for mirror node pairing. The anatomical distance between mirror nodes is less than a preset threshold as the condition for determining the validity of the pairing. The pairing relationship of mirror nodes is used as the edge connection to construct the second graph branch. In the second graph branch, each pair of mirror nodes is connected to the same symmetric constraint hyperedge. The symmetric constraint hyperedge connects both the mirror node pair and the common associated disease node corresponding to the mirror node pair. The common associated disease nodes corresponding to mirror node pairs are the intersection or union of the disease nodes associated with the nodes on either side of the mirror node pair and the disease nodes associated with the nodes on the other side. When there are disease nodes in the intersection, the disease nodes are added to the symmetric constraint hyperedge as common associated disease nodes. When there are no disease nodes in the intersection, all disease nodes in the union are added to the symmetric constraint superedge as commonly associated disease nodes.

[0028] The method for determining the preset threshold, which uses the anatomical distance between mirror nodes being less than a preset threshold as a criterion for determining pairing validity, is as follows: Obtain the normal measurement range of the symmetrical anatomical structure of the target organ in the standard anatomical position, including the coordinate range of the left and right lobes in each anatomical position and the normal separation distance range between the two lobes. The spatial distance between the geometric center points of the two lateral lobes was used as the basic anatomical distance of the mirror node pair. The mean of this distance in the normal population plus 1.96 times the standard deviation was used as the statistical upper limit. This statistical upper limit corresponds to the anatomical distance range of about 95% of individuals in the normal population. The maximum observed distance between corresponding anatomical landmarks on both lobes in normal individuals was used as the initial value for the anatomical distance threshold. Based on this initial value, and according to the tolerance requirements for symmetrical structure pairing in actual clinical applications, an adjustable relaxation coefficient is added. The value of this relaxation coefficient ranges from 5% to 15% of the initial value. The sum of the initial value and the relaxation coefficient is used as the final anatomical distance threshold.

[0029] The specific steps for connecting each pair of mirror nodes in the second graph branch to the same symmetric constraint hyperedge are as follows: Each pair of mirror nodes that has passed the anatomical distance threshold is used as the basic connection node. The left and right nodes in the mirror node pair are used as the main connection nodes of the symmetric constraint hyperedge. At the same time, disease nodes that have clinical relevance to any side node in the mirror node pair are included in the connection node set of the symmetric constraint hyperedge. When establishing symmetric constraint hyperedges, a first initial weight is assigned to the main connected nodes, and a second initial weight is assigned to the associated disease nodes. The sum of the weights of the main connected nodes is not less than the sum of the weights of the associated disease nodes. Each disease node is assigned a second initial weight in proportion to its correlation strength with the test index. The aggregation method for symmetric constraint hyperedge pair information is as follows: the features of the left and right nodes in the mirror node pair are added element by element to generate joint features; the features of the associated disease nodes are weighted and summed according to the second initial weight to generate disease aggregate features; the joint features are mapped to the disease feature space by the first linear transformation matrix and then added element by element with the disease aggregate features to complete feature alignment. In each residual graph convolutional layer, the symmetric constraint hyperedge performs bidirectional information transfer. The joint features are transferred to the disease node side along the symmetric constraint hyperedge, and the disease aggregate features are transferred in the opposite direction to the mirror node side along the symmetric constraint hyperedge. The bidirectional transfer is mapped using the first linear transformation matrix and the second linear transformation matrix, respectively. The dimension of the two transformation matrices is the same as the dimension of the hidden layer features. Symmetric constraint hyperedges impose two constraints when transmitting information. The first constraint is that the Euclidean distance between the features of the left and right nodes in a mirror node pair after each layer output is less than the distance threshold set for the current layer. This threshold gradually decreases as the number of network layers increases. The second constraint is that the cosine similarity between the joint features and the disease aggregate features after mapping by their respective transformation matrices is not less than the similarity threshold. The connection weights of the symmetric constraint hyperedge are adaptively adjusted after the feature updates of each layer are completed. The change magnitude of the feature vector of each node before and after the update is calculated. When the change magnitude is greater than a preset first threshold, the connection weight of the node is increased by the first adjustment magnitude in the next layer. When the change magnitude is less than a preset second threshold, the connection weight is decreased by the second adjustment magnitude in the next layer. After adjustment, the connection weights of all nodes in the same symmetric constraint hyperedge are normalized.

[0030] The first and second thresholds are set as follows: Both the first and second thresholds are set during the system initialization configuration phase, based on the statistical distribution characteristics of feature changes of all nodes in the training dataset after the first round of forward propagation. Specifically, after the system completes the first round of forward propagation, it calculates the rate of change of the L2 norm of the feature vectors of all nodes in the training dataset before and after the current layer update. The average of the rate of change of all nodes is taken as the baseline rate of change. The first threshold is the baseline rate of change multiplied by a first proportional coefficient, and the second threshold is the baseline rate of change multiplied by a second proportional coefficient. The value of the first proportional coefficient is greater than that of the second proportional coefficient. For scenarios with a small training dataset, the first and second thresholds are empirically fixed values. The value of the first threshold ranges from 0.1 to 0.2, and the value of the second threshold ranges from 0.02 to 0.05. The specific values ​​are automatically adapted by the system according to the feature dimension of the hidden layer. The specific values ​​of the first and second thresholds are determined before the start of training and remain unchanged during subsequent training processes, without being dynamically adjusted with the increase of training rounds.

[0031] The first adjustment range and the second adjustment range are respectively: The first adjustment range refers to the percentage increase in the connection weight when the change in node features exceeds a preset first threshold. This percentage is usually set to 5% to 10% of the current value of the connection weight, and is used to enhance the influence of the node on the aggregation of hyperedge information when the node features change significantly. The second adjustment magnitude refers to the percentage by which the connection weights are adjusted downwards when the change in node features is less than a preset second threshold. This percentage is set to range from 2% to 5% of the current value of the connection weights, and the first magnitude is greater than the second magnitude to ensure that the model pays more attention to adjusting the weights of nodes with drastic feature changes. The specific value is dynamically determined based on the convergence speed during training. In the early stages of training, a larger first magnitude is used to accelerate weight updates, while in the later stages of training, a smaller first magnitude is used to avoid weight oscillations.

[0032] Specifically, the bidirectional transfer uses the first linear transformation matrix and the second linear transformation matrix respectively: The first and second linear transformation matrices are essentially learnable weight parameter matrices. Their elements are continuously updated during training through backpropagation and gradient descent, eventually converging to a numerical state that can effectively complete the feature space mapping.

[0033] Specifically, if the feature dimension of the hidden layer is D, then the first linear transformation matrix and the second linear transformation matrix are both D×D square matrices; Each element in the matrix (i.e., the parameter value corresponding to row i and column j) represents the contribution strength of the j-th dimension in the source feature vector to the i-th dimension in the target feature vector; In the initial training phase, all elements are assigned random initial values ​​according to a uniform distribution or a truncated normal distribution to ensure an unbiased initial mapping relationship between each dimension; During the forward propagation process, the source feature vector is multiplied on the left by the transformation matrix to obtain the mapped target feature vector. Each row of the matrix is ​​equivalent to a weight vector, which is generated by the inner product of the source feature vector to produce the corresponding dimension of the target feature vector. During backpropagation, the gradients of the classification loss function and the symmetric consistency constraint loss function are backpropagated layer by layer to each element of the transformation matrix. The optimizer applies updates to each element in the corresponding direction and magnitude based on the gradient values. After multiple rounds of iterative training, the matrix elements gradually converge—the first linear transformation matrix learns the optimal linear mapping relationship from the mirror node to the joint feature space to the disease aggregate feature space, and its element values ​​reflect the contribution weight of each dimension of the joint features to each dimension of the disease aggregate features. The second linear transformation matrix learns the optimal linear mapping relationship from the disease aggregate feature space to the mirror node to the joint feature space. The values ​​of its elements reflect the contribution weights of each dimension of the disease aggregate features to each dimension of the joint features. After training, the elements of both transformation matrices are fixed values ​​and directly participate in the feature mapping calculation during the inference stage, without changing.

[0034] The symmetric constraint hyperedge connects both the mirror node pair and the common associated disease node corresponding to the mirror node pair. The common associated disease node is: For each pair of mirror nodes in the second graph branch, obtain all disease nodes associated with the left node and all disease nodes associated with the right node, and take the intersection of the two disease node sets as the common associated disease nodes. When the intersection is not empty, the disease nodes in the intersection are used as mirror nodes to connect the corresponding common associated disease nodes to the symmetric constraint hyperedge. When there are no disease nodes in the intersection, all disease nodes in the union of the two disease node sets are used as the common associated disease nodes and connected to the symmetric constraint superedge.

[0035] In this embodiment, the generation of the first and second aggregate features of each node by the residual map convolutional layer stacking module specifically includes: Receive the first graph branch and the second graph branch. For each node in the first graph branch, use the weighted sum of the node's own feature vector and the features of the neighboring nodes recorded in the adjacency matrix as the message passing result. After mapping the message passing result through a non-linear activation function, perform residual connection with the node's own feature vector to obtain the first aggregated feature output by this layer. For each node in the second graph branch, the node's own feature vector is weighted and summed with the adjacency matrix constructed by the mirror pairing relationship in the second graph branch. The summation result is mapped by a non-linear activation function and then residually connected with the node's own feature vector to obtain the second aggregated feature output by this layer. For each node, the first and second aggregated features output by the current residual graph convolutional layer are used as the input features of the next residual graph convolutional layer, respectively. After N residual graph convolutional layers are stacked, each node obtains the first and second aggregated features updated by N residual graph convolutional operations.

[0036] In this process, the message passing result is mapped using a non-linear activation function and then residually concatenated with the node's own feature vector to obtain the first aggregated feature output of this layer. The non-linear activation function here is: For the first half of the stacked convolutional layers in the residual map, the ReLU function is used as the nonlinear activation function; for the second half of the layers, the LeakyReLU function is used as the nonlinear activation function, and the negative slope coefficient of LeakyReLU is set to 0.2. When the number of stacked residual graph convolutional layers exceeds 4, ReLU and LeakyReLU are used alternately every other layer; the specific type of non-linear activation function used in each layer is automatically configured and fixed according to the total number of layers N during system initialization and remains unchanged during training and inference.

[0037] The process of performing residual connections to obtain the first aggregated feature of this layer's output includes the following steps: The node features obtained after the current residual graph convolutional layer is processed by neighbor aggregation and nonlinear activation function mapping are used as the first candidate features, and the node features before the current residual graph convolutional layer is used as the second candidate features. The first candidate features and the second candidate features are added element by element to obtain the original residual fusion features. The value range boundary constraint is applied to the original residual fusion features to compress each element value of the original residual fusion features to between the negative boundary value and the positive boundary value. The original residual fusion features after applying the range boundary constraints are used as the first aggregated features of the current residual graph convolutional layer output. The dimension of the first aggregated feature is consistent with the dimension of the node features before the input of the current residual graph convolutional layer. The first aggregated feature output by the current residual map convolutional layer is used as one of the input features of the next residual map convolutional layer to participate in subsequent residual map convolution calculations; The specific steps for applying the domain boundary constraints here are as follows: Define a positive boundary value. For each element in the feature vector, if the element value is greater than the positive boundary value, then the element value is truncated to the positive boundary value; if the element value is less than the negative boundary value, then the element value is truncated to the negative boundary value; if the element value is between the negative boundary value and the positive boundary value, then the element value remains unchanged.

[0038] The nonlinear activation function used in the second branch is the same as that used in the first branch.

[0039] The specific implementation steps, involving N stacked residual map convolutional layers, are as follows: The first graph branch and the second graph branch output by the dual graph branch construction module are used as the dual-branch input of the first residual graph convolutional layer. The dual-branch input contains both the original feature vector of each node and the adjacency matrix of the dual graph branches. Starting from the first residual graph convolutional layer, each residual graph convolutional layer receives the first and second aggregated features output from the previous layer as the input features of the current layer. Neighbor aggregation operations are performed on the first graph branch and the second graph branch respectively to obtain the first and second candidate features of each node. After the first and second candidate features are mapped by nonlinear activation functions, they are added element-wise to the first and second aggregated features before the input of the current residual map convolutional layer. The result of the element-wise addition is constrained by the range boundary and then used as the first and second aggregated features output by the current residual map convolutional layer. The first and second aggregated features output by the current residual graph convolutional layer are directly used as input features for the next residual graph convolutional layer, participating in the neighbor aggregation operation of the next residual graph convolutional layer; Repeat the neighbor aggregation, nonlinear activation mapping, residual connection and range boundary constraint operations described above until the Nth residual graph convolutional layer is calculated. Use the first and second aggregated features output by the Nth residual graph convolutional layer as the final output of the residual graph convolutional layer stacking module.

[0040] The neighbor aggregation operation specifically includes: For each node in the first graph branch, the node's own feature vector is weighted and summed with the features of its neighboring nodes recorded in the adjacency matrix. The non-zero elements in each row of the adjacency matrix correspond to the weights of the features of each neighboring node in the aggregation result of that node. The original aggregation result of the first graph branch is obtained by multiplying each neighboring node feature by its corresponding weight and then summing the results. For each node in the second graph branch, the node's own feature vector is weighted and summed with the adjacency matrix constructed by the mirror pairing relationship in the second graph branch. The non-zero elements in each row of the adjacency matrix correspond to the weights of the features of each mirror neighbor node on the aggregation result of the node. The features of each mirror neighbor node are multiplied by their corresponding weights and then summed to obtain the original aggregation result of the second graph branch. The adjacency matrices of the first and second graph branches do not participate in gradient updates during training, and their weight values ​​are fixed to the similarity values ​​or pairing relationship values ​​determined during construction. Each node's feature vector undergoes layer normalization before being input into each residual graph convolutional layer. The layer normalization operation is applied to each dimension of the feature vector, making the mean of the node's feature vector zero and the variance one. The neighbor aggregation operations of the first graph branch and the second graph branch both use the same normalization method and weighted summation method. The two are executed independently and do not interfere with each other.

[0041] In this embodiment, the generation of the first gating weight and the second gating weight for each node in the node-level adaptive gating module specifically includes: The node-level adaptive gating module receives the first and second aggregate features of each node output by the residual graph convolutional layer stacking module. For each node, the first aggregate feature is mapped to the [0,1] interval through the first fully connected layer to output the first gating weight, and the second aggregate feature is mapped to the [0,1] interval through the second fully connected layer to output the second gating weight. The sum of the first gating weight and the second gating weight is 1. When each node updates its features in each residual graph convolutional layer, the product of the first aggregated feature and the first gate weight is taken as the first contribution component, and the product of the second aggregated feature and the second gate weight is taken as the second contribution component. The input features of the previous layer of the node are added element-wise to the first contribution component and the second contribution component to obtain the updated features of the node in the current layer.

[0042] The first aggregated feature is mapped to the [0,1] interval through the first fully connected layer to output the first gate weight. The specific implementation steps are as follows: The input dimension of the first fully connected layer is equal to the length of the first aggregated feature vector of each node, and the output dimension is set to 1. The first fully connected layer uses learnable weight vectors and bias scalars as trainable parameters. The first aggregated feature vector and the weight vector are multiplied together and then the bias scalar is added to obtain the original scalar output. The original scalar output is mapped to the [0,1] interval through the Sigmoid activation function to obtain the first gating weight of the node, where the Sigmoid activation function compresses any real value input into an open interval between 0 and 1; The weight vector and bias scalar of the first fully connected layer participate in gradient descent updates synchronously with the weight parameters of the residual graph convolutional layer stacked module during training, and use the same optimizer and learning rate as the residual graph convolutional layer.

[0043] The second gating weight is the same as above.

[0044] In this embodiment, setting the second gating weight to 0 or 1 based on the abnormal marker of the mirror node alignment index in the node-level adaptive gating module specifically includes: The node-level adaptive gating module reads the abnormal markers of the inspection items corresponding to each node in the inspection report data and the mirror node. The abnormal markers are generated by comparing the measured values ​​of the inspection items with the preset normal value range. When the measured value is lower than the lower limit of the reference range or higher than the upper limit of the reference range, it is marked as abnormal. When the measured value is between the lower limit and the upper limit of the reference range, it is marked as normal. When the test item corresponding to any one node in the mirror node pair is marked as abnormal while the test item corresponding to the other node is marked as normal, the second gating weight corresponding to the mirror node pair is assigned to 0, and the first gating weight is assigned to 1. The node update depends only on the first aggregated feature. When the abnormal flags of the test items corresponding to the two nodes in the mirror node pair are both normal, the second gating weight corresponding to the mirror node pair is assigned a value of 1, and the first gating weight is assigned a value of 0. The node update depends only on the second aggregated feature. The node-level adaptive gating module multiplies the assigned first and second gating weights by their respective aggregated features before participating in the node feature update of the current residual graph convolutional layer.

[0045] The abnormality marker is determined by comparing the measured value of the test item with the preset normal value range. The preset normal value range is set in the following way: The preset normal range is defined by the lower limit and upper limit of the reference range. The lower limit and upper limit are derived from publicly available clinical laboratory reference interval standards or from historical data statistics of the laboratory where the test is performed.

[0046] The node update relies solely on the first aggregated feature. The specific implementation steps are as follows: After the node-level adaptive gating module determines that the mirror node pair is in a one-sided abnormal state, it resets the second gating weight corresponding to the mirror node pair to 0, and at the same time resets the first gating weight corresponding to the mirror node pair to 1. During the node feature update process of the current residual graph convolutional layer, the product of the first aggregated feature and the first gate weight (value 1) is taken as the first contribution component, and the product of the second aggregated feature and the second gate weight (value 0) is taken as the second contribution component. The second contribution component is a zero vector. The input features of the previous layer of the node are added element by element to the first contribution component to obtain the updated features of the node in the current layer. The second aggregated features and their corresponding second contribution components do not participate in the current layer update calculation of the node. After the node update is completed, the updated features are used as the output features of the current residual graph convolutional layer for forward propagation of subsequent layers. However, the second aggregated features are still retained in the output of the residual graph convolutional layer stacking module for use by the feature fusion module when stitching across layers.

[0047] The node-level adaptive gating module multiplies the assigned first and second gating weights by their corresponding aggregated features before participating in the node feature update of the current residual graph convolutional layer. The specific implementation steps are as follows: The node-level adaptive gating module receives the first and second aggregate features of each node output by the residual graph convolutional layer stacking module in the current layer, and also receives the first and second gating weight values ​​of each node that have been assigned values ​​within the module. For each node, the first aggregated feature is multiplied element-wise by the first gate weight value to obtain the first weighted aggregated feature, and the second aggregated feature is multiplied element-wise by the second gate weight value to obtain the second weighted aggregated feature. The first weighted aggregate feature and the second weighted aggregate feature are added element by element to obtain a weighted fusion feature. The dimension of the weighted fusion feature is the same as that of the first aggregate feature and the second aggregate feature. The weighted fused features are added element-wise to the node features before the current residual map convolutional layer input to obtain the original updated features; Apply a domain boundary constraint to the original updated feature, compress each element value of the original updated feature to between the preset negative boundary value and the positive boundary value, and use the feature with the domain boundary constraint as the updated feature of the node in the current residual map convolutional layer. The updated feature is directly used as one of the input features of the next residual map convolutional layer, participating in the neighbor aggregation operation and nonlinear activation mapping of the next residual map convolutional layer.

[0048] The specific steps for applying range boundary constraints to the original updated feature are as follows: Obtain the original updated feature map, which is a three-dimensional tensor containing spatial height, spatial width, and number of channels, with each element value being a floating-point number; Obtain the preset lower threshold and the preset upper threshold, which are predetermined floating-point constants; Iterate through each element in the feature map and compare the value of that element with a preset lower threshold. If the value of the element is less than the preset lower threshold, then the value of the element is reset to the preset lower threshold. If the value of the element is greater than the preset upper limit threshold, then the value of the element is reset to the preset upper limit threshold. If the value of the element is greater than or equal to the preset lower threshold and less than or equal to the preset upper threshold, then the value of the element remains unchanged. After traversing all elements, output the constrained feature map. The spatial height, spatial width, and number of channels of the constrained feature map are completely consistent with the feature map before constraint. The preset lower threshold and the preset upper threshold are determined by the learnable quantization step size. The preset lower threshold is equal to the negative value of the learnable quantization step size multiplied by a preset multiple, and the preset upper threshold is equal to the positive value of the learnable quantization step size multiplied by a preset multiple. As another possible implementation, the preset lower threshold and the preset upper threshold are determined by the statistical characteristics of the numerical distribution of the feature map, the standard deviation of the numerical distribution of the feature map is obtained, the preset lower threshold is set as the product of the negative standard deviation and the preset multiple, and the preset upper threshold is set as the product of the positive standard deviation and the preset multiple.

[0049] In this embodiment, the calculation of feature difference values ​​between mirror node pairs by the symmetry-aware residual convolution module as a symmetry consistency constraint term specifically includes: Extract the second aggregated features of each node from each residual graph convolutional layer. For each mirror node pair in the second graph branch, calculate the Euclidean distance between the second aggregated features of the left node and the second aggregated features of the right node as the feature difference value. Map the feature difference value to the [0,1] interval and use it as the original symmetric consistency constraint value of the mirror node pair in the current layer. The symmetry-aware residual convolution module receives the second gating weights output by the node-level adaptive gating module, multiplies the original symmetry consistency constraint value with the second gating weights, and obtains the weighted symmetry consistency constraint value. The symmetry-aware residual convolution module takes the average of the weighted symmetry consistency constraint values ​​of all mirror node pairs across all residual map convolutional layers, and uses this as the global symmetry consistency constraint loss term. This global symmetry consistency constraint loss term is then added to the main loss function of the classification output module to form a joint optimization objective.

[0050] In this embodiment, the self-pairing regularization training module performs flipping enhancement on the second graph branch, specifically including: The self-pairing regularization training module performs horizontal flip enhancement on the second graph branch during the training phase. Based on the mirror node pairing relationship, it swaps the features of the left node and the right node in each mirror node pair in the second graph branch, while keeping the edge connection relationship and hyperedge connection relationship between the nodes unchanged, and generates the flipped second graph branch. The self-pairing regularization training module inputs the original second graph branch and the flipped second graph branch into the residual graph convolutional layer stacking module to obtain the second aggregated features of each node corresponding to the original second graph branch and the second aggregated features of each node corresponding to the flipped second graph branch. The self-pairing regularization training module locates the corresponding mirror pairing node for each node in the original second graph branch and the mirror pairing node in the flipped second graph branch. It calculates the mean squared error between the second aggregated feature of the node in the original second graph branch and the second aggregated feature of the mirror pairing node in the flipped second graph branch. The mean squared error of all nodes is taken as the consistency constraint loss term, and the consistency constraint loss term is added to the main loss function of the classification output module.

[0051] Specifically, the mirror pairing node corresponding to the node located in the second branch of the flipped graph is: Obtain the spatial coordinates of the target node in the first graph branch, which includes the row index and column index; obtain the flip transformation parameters of the second graph branch, which include the flip direction indicator and the position of the flip center line; perform coordinate transformation calculation on the spatial coordinates of the target node according to the flip transformation parameters, and generate the transformed row index and transformed column index after transformation, respectively. The transformed row index is equal to twice the row position of the flip center line minus the row index of the target node, and the transformed column index is equal to twice the column position of the flip center line minus the column index of the target node.

[0052] In the second graph branch, search for nodes whose spatial coordinates are completely consistent with the transformed row index and column index. If a unique node is found, it is determined as the paired node of the target node. If multiple nodes are found, calculate the Euclidean distance between the spatial coordinates of each node and the transformed row index and column index, and select the node with the smallest distance as the paired node. If no node is found, expand the search radius and search for nodes whose spatial coordinates are less than a preset distance threshold within the neighborhood of the transformed row index and column index, and select the closest node as the paired node.

[0053] The node index of the paired node is associated with the node index of the target node to form a pairing mapping table.

[0054] The preset distance threshold for searching nodes whose spatial coordinates are less than a preset distance threshold within the neighborhood of the transformed row index and column index is set as follows: The preset distance threshold is used to limit the neighborhood search range in the transformed two-dimensional coordinate space, and only other nodes whose spatial coordinates are less than the Euclidean distance between the current node coordinates and the spatial coordinates are included in the neighborhood node set to participate in subsequent aggregation calculations; The preset distance threshold is determined based on the spatial distribution density of nodes on the two-dimensional plane after coordinate transformation. Specifically, the average distribution density of all nodes in the transformed two-dimensional coordinate space is calculated. A circle is drawn with each node as the center and the initial search radius is used. The average number of nodes contained in the circle is calculated. The search radius is adjusted so that the average number of nodes in the circle reaches the preset target number. The adjusted search radius is used as the initial value of the preset distance threshold. The preset distance threshold gradually decreases as the number of layers in the residual map convolutional layer stacking module increases. The initial layer uses a larger distance threshold to capture global spatial structure information, while subsequent layers use a smaller distance threshold to focus on local fine features. The preset distance threshold is set differently for different anatomical regions. Nodes inside the organ boundary use a smaller distance threshold, while nodes near the organ boundary use a larger distance threshold, in order to maintain the correlation transmission across organ boundaries.

[0055] In this embodiment, the feature fusion module performs cross-layer concatenation of the aggregated features output from each residual map convolutional layer to generate a fused feature matrix, specifically including: The feature fusion module extracts the first and second aggregate features of each node from each residual map convolutional layer to obtain N sets of feature matrices, each set containing the first and second aggregate features of all nodes in the current layer; The feature fusion module concatenates N sets of feature matrices along the feature dimension. Using the original input feature vector of each node as the reference, the first aggregated feature of the same node in layers 1 to N is arranged sequentially after the reference feature, and the second aggregated feature of the same node in layers 1 to N is arranged sequentially after the first aggregated feature, forming the cross-layer fusion feature vector of the node. The feature fusion module arranges the cross-layer fusion feature vectors of all nodes in node order to form a fusion feature matrix.

[0056] In this embodiment, the classification output module's mapping of the fusion feature matrix to generate review results and clinical description text specifically includes: The classification output module performs global average pooling on the fused feature matrix to obtain a fixed-length global feature vector. The global feature vector is then input into the first fully connected classifier, which is activated by softmax and outputs the probability distribution of each preset review category. The category corresponding to the maximum probability is used as the review classification result. The review categories include three categories: normal, abnormal pending review, and critical value warning. The classification output module concatenates the global feature vector with the category embedding vector corresponding to the review classification result. The concatenated vector is then input into a Transformer-based decoder. The decoder generates clinical description text word by word in an autoregressive manner. The clinical description text includes a list of abnormal test items, a description of the degree of deviation of abnormal indicators, the comparison results with mirror node pairs, and clinical prompts based on the comparison results. The classification output module outputs the review classification results and clinical description text as system outputs. The review classification results are output in the form of structured tags, and the clinical description text is output in the form of natural language.

[0057] The output includes the probability distribution of each preset review category after softmax activation. The specific steps for preset are as follows: The preset review categories are three: normal, abnormal and pending review, and critical value warning. The preset process for these three categories is completed during the system initialization configuration phase, and the specific steps are as follows: The criteria for determining the normal category are configured as follows: the measured values ​​of all test items in the test report data are within their respective preset normal value ranges, and the feature difference values ​​of the two sides of the mirror node pair are less than the distance threshold set for the current layer. The criteria for judging the abnormal pending review category are configured as follows: the measured value of at least one test item in the test report data exceeds the corresponding preset normal value range but does not reach the critical value judgment threshold, or the feature difference value of a single node in the mirror node pair is greater than the distance threshold set in the current layer but less than the critical value judgment threshold. The criteria for determining the critical value warning category are configured as follows: the measured value of any test item in the test report data reaches or exceeds the preset critical value judgment threshold. The critical value judgment threshold is entered together with the preset normal value range during the system configuration stage. The critical value judgment threshold for each test item is set independently by clinical laboratory experts based on the clinical significance and risk level of the test item. The category embedding vectors corresponding to the three preset review categories are initialized as independent learnable vectors during the system training phase. The dimension of the category embedding vectors is consistent with the dimension of the global feature vectors, and they are updated synchronously with the weight parameters of the residual map convolutional layer stacking module during the training process. After the classification output module performs global average pooling on the fused feature matrix to obtain the global feature vector, it inputs the global feature vector into the first fully connected classifier. The output dimension of the first fully connected classifier is 3, and the output values ​​correspond to the original scores of the three categories: normal category, abnormal category to be reviewed category, and critical value warning category. The three original scores are mapped to three probability values ​​through the softmax function, and the category with the highest probability value is used as the review classification result output.

[0058] The classification output module embeds the global feature vector with the category embedding vector corresponding to the review classification result. Specifically, the category embedding vector is: During system initialization, an independent category embedding vector is generated for each of the normal category, the abnormal category awaiting review category, and the critical value warning category. The dimensions of the three category embedding vectors are the same as the dimensions of the global feature vector. The category embedding vector participates in gradient updates as a trainable parameter during the training phase. Its initial value is randomly generated by a truncated normal distribution and is optimized synchronously with the weight parameters of the residual graph convolutional layer stacking module and the classification output module during training. After training, the category embedding vectors converge to a fixed value, which represents the position of the semantic features of the category in the feature space. During the inference phase, the classification output module first determines the review classification result through the first fully connected classifier and the softmax function, and then extracts the corresponding category embedding vector from the pre-trained category embedding vector set according to the category index corresponding to the review classification result. The global feature vector and the category embedding vector are concatenated along the feature dimension to form a joint feature vector. The joint feature vector is then input into a Transformer-based decoder to participate in the subsequent word-by-word generation process of clinical descriptive text.

[0059] Example 1: The laboratory department of a tertiary general hospital handles a large volume of thyroid function tests daily, processing approximately 400 reports per day. Under the traditional review model, laboratory personnel must verify each indicator in the report individually and make a comprehensive judgment based on the patient's clinical presentation. In practice, the hospital found that a significant proportion of the review difficulties in thyroid function test reports stemmed from the interpretation of indicators related to bilateral thyroid structure—for example, the clinical significance of thyroglobulin antibody (TgAb) and thyroid peroxidase antibody (TPOAb) differs considerably between patients with unilateral thyroid nodules and those with bilateral diffuse thyroid lesions, making it difficult to accurately distinguish between them using a rule-based engine alone.

[0060] In practical application of this laboratory report review system, the workflow is as follows: The system first receives laboratory report data and anatomical prior data through the data input module. The laboratory report data includes fields such as patient identification, age and gender, department visited, sample type and collection time, names and measured values ​​of thyroid function-related tests, upper and lower limits of reference ranges for each test, and abnormality markers. The anatomical prior data pre-injects structural knowledge such as the symmetrical pairing relationship of the left and right lobes of the thyroid gland and anatomical distance thresholds into the system.

[0061] Next, the dual-graph branch construction module begins operation. On one hand, the first graph branch is constructed using the numerical similarity between the test indicators as the edge weights. For example, in historical data, the correlation coefficient and co-occurrence frequency of the measured values ​​of thyroid-stimulating hormone (TSH) and free thyroxine (FT4) are high, so the edge weight between them is set to a large value. On the other hand, the system constructs the second graph branch based on the mirror pairing relationship between the left and right lobes of the thyroid gland in the anatomical prior. Each mirror node pair shares a symmetric constraint hyperedge, which connects the mirror node pair and its commonly associated thyroid disease nodes. The system then feeds the two branches into the residual graph convolutional layer stacking module, which performs neighbor aggregation operations on the two branches respectively, generating a first aggregated feature and a second aggregated feature for each node. The node-level adaptive gating module then intervenes: when the TSH value of the left node in a mirror node pair is high while the TSH value of the right node is normal, the system reads the anomaly marker and sets the second gating weight corresponding to the mirror node pair to 0, while setting the first gating weight to 1. At this time, the node update only depends on the aggregated features of the first graph branch. When both sides of the mirror node's indicators are within the normal range, the second gating weight is set to 1, the first gating weight is set to 0, and the node update becomes dependent on the aggregation features of the second graph branch.

[0062] The system also continuously calculates the Euclidean distance of the second aggregated features between mirror node pairs as a symmetry consistency constraint value through a symmetry-aware residual convolution module. The self-pairing regularization training module performs horizontal flip enhancement on the second graph branch during the training phase, forcing the node feature outputs before and after the flip to remain consistent, ensuring the model has symmetry-awareness capabilities that maintain left and right positions. After the outputs of all residual graph convolutional layers are concatenated across layers by the feature fusion module to form a fused feature matrix, the classification output module finally outputs the review classification results and clinical description text.

[0063] To verify the system's effectiveness, the research team used 8,920 thyroid function test reports accumulated by the hospital over the past six months as a test sample. This included 432 clinically diagnosed cases of Hashimoto's thyroiditis, 198 cases of Graves' disease, 265 cases of thyroid nodules, 112 cases of post-thyroid cancer follow-up, and several normal control samples. The test was conducted in three groups: the first group used a traditional rule engine for review, the second group used a conventional graph neural network model without symmetric priors, and the third group used the system of this invention. The results are as follows: Table 1: Comparison of the auditing performance of three auditing methods in thyroid function test reports

[0064] The test data shows that in 8920 thyroid function test reports, the overall accuracy rate of the traditional rule engine was 78.3%. In diseases such as Hashimoto's thyroiditis and Graves' disease, which often present as bilateral diffuse lesions, the review logic relies excessively on the threshold judgment of a single indicator, leading to a large number of samples near the borderline being misclassified as normal or abnormal, resulting in low accuracy. Although the conventional graph neural network model captures some correlations between test indicators through neighbor aggregation, improving the accuracy to 86.7%, this system does not model the mirror symmetry prior of the left and right lobes of the thyroid gland. When faced with special samples such as unilateral thyroid nodules with normal indicators on the contralateral side, the model still tends to pull both indicators towards the mean, causing the normal indicators on the contralateral side to be incorrectly "transmitted" as abnormal features, resulting in a false positive rate as high as 16.2%. This invention is significantly superior to the above two methods. Specifically, among 432 patients with Hashimoto's thyroiditis, this invention correctly identified 406 cases, with a sensitivity of 93.9%; among 198 patients with Graves' disease, it correctly identified 184 cases, with a sensitivity of 92.9%; and among 265 patients with thyroid nodules, this invention successfully identified 237 unilateral nodule cases, of which only 17 cases failed to be accurately identified due to small nodule volume and insignificant changes in indicators, with a detection sensitivity of 89.4%. More notably, in 112 follow-up samples after thyroid cancer surgery, this invention correctly identified 108 cases as normal follow-up status, and only 4 cases were marked as "re-examination recommended" due to mild TSH suppression, with a specificity of 96.4%. In contrast, conventional graph neural network models misclassified as suspicious abnormalities in as many as 19 cases in follow-up samples after thyroid cancer surgery, indicating that structures without symmetry priors are prone to misinterpreting mild fluctuations in indicators of unilateral residual tissue as signs of recurrence, while this invention effectively eliminates this interference. In terms of review efficiency, this invention takes an average of 9.1 seconds per report, a significant improvement over the 12.3 seconds of conventional graph neural network models. This is attributed to the node-level adaptive gating mechanism, which sets the second gating weight to 1 when both sides are normal, and allows the second graph branch to directly dominate node updates, greatly reducing redundant computation. Furthermore, when individually analyzing the 27 critical value reports from all test samples, this invention missed only 1 report, a miss rate of 3.7%, while the conventional graph neural network model missed 5 reports, a miss rate of 18.5%, and the traditional rule engine missed 11 reports, a miss rate as high as 40.7%. Based on the above data, this invention's system demonstrates significant advantages in review accuracy, sensitivity in detecting symmetrical structural lesions, control of unilateral nodule false alarm rate, review efficiency, and critical value early warning capabilities.

[0065] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based system for reviewing laboratory reports and assisting in generating clinical significance, characterized in that: include: The data input module is used to receive test report data and prior anatomical data; The dual-graph branching module is used to construct the first graph branch based on the numerical similarity of the test report data, and to construct the second graph branch by dissecting the mirror node pairings in the prior data. In the second graph branch, each mirror node pair shares a symmetric constraint hyperedge. The residual graph convolutional layer stacking module, based on the residual graph convolutional neural network architecture, is used to perform neighbor aggregation on the dual graph branches and generate the first aggregation feature and the second aggregation feature of each node. The node-level adaptive gating module is used to generate a first gating weight and a second gating weight for each node. The first gating weight adjusts the contribution of the first aggregated feature to the node update, and the second gating weight adjusts the contribution of the second aggregated feature to the node update. The second gating weight is set to 0 or 1 according to the abnormal flag of the mirror node alignment index. It is set to 0 when one side is abnormal and to 1 when both sides are normal. The symmetry-aware residual convolution module is used to calculate the feature difference values ​​between mirror node pairs as symmetry consistency constraints, and the second gate weight adjusts the relaxation strength of the constraints. The self-pairing regularization training module is used to perform flip enhancement on the second graph branch and apply consistency constraints to the node feature output before and after the flip. The feature fusion module is used to perform cross-layer concatenation of the aggregated features output from each residual map convolutional layer to generate a fused feature matrix; The classification output module is used to map the fusion feature matrix to generate review results and clinical description text.

2. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The test report data and prior anatomical data in the data input module specifically include: The data in the test report includes: patient identification, patient age, patient gender, department visited, sample type, sample collection time, test item name, test item measurement value, test item unit, upper limit of the reference range for the test item, lower limit of the reference range for the test item, and abnormal markers for the test item. Prior anatomical data include: organ names, left-right symmetrical pairing of organs in standard anatomical positions, and anatomical distance thresholds between mirror node pairs.

3. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The construction of the first graph branch and the second graph branch by the dual-graph branch construction module specifically includes: The first graph branch is constructed by using the test items in the test report data as nodes of the first graph branch and the numerical similarity between two test items as the weight of the edge between nodes in the first graph branch. Numerical similarity is determined by the co-occurrence frequency of two test items in the same patient and the same batch of visits in historical test reports and the correlation of measured values. The correlation coefficient of measured values ​​is calculated using the Pearson correlation coefficient. After normalizing the numerical similarity to the [0,1] interval, it is used as the value of the corresponding position in the first graph branch adjacency matrix. The nodes of the second graph branch are the test items in the test report data, the anatomical structures in the anatomical prior data, and the diseases that are clinically related to the test items. The left-right symmetrical pairing relationship of organs recorded in the anatomical prior data under the standard anatomical position is used as the basis for mirror node pairing. The anatomical distance between mirror nodes is less than a preset threshold as the condition for determining the validity of the pairing. The pairing relationship of mirror nodes is used as the edge connection to construct the second graph branch. In the second graph branch, each pair of mirror nodes is connected to the same symmetric constraint hyperedge. The symmetric constraint hyperedge connects both the mirror node pair and the common associated disease node corresponding to the mirror node pair. The common associated disease nodes corresponding to mirror node pairs are the intersection or union of the disease nodes associated with the nodes on either side of the mirror node pair and the disease nodes associated with the nodes on the other side. When there are disease nodes in the intersection, the disease nodes are added to the symmetric constraint hyperedge as common associated disease nodes. When there are no disease nodes in the intersection, all disease nodes in the union are added to the symmetric constraint superedge as commonly associated disease nodes.

4. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The generation of the first and second aggregate features of each node by the residual graph convolutional layer stacking module specifically includes: Receive the first graph branch and the second graph branch. For each node in the first graph branch, use the weighted sum of the node's own feature vector and the features of the neighboring nodes recorded in the adjacency matrix as the message passing result. After mapping the message passing result through a non-linear activation function, perform residual connection with the node's own feature vector to obtain the first aggregated feature output by this layer. For each node in the second graph branch, the node's own feature vector is weighted and summed with the adjacency matrix constructed by the mirror pairing relationship in the second graph branch. The summation result is mapped by a non-linear activation function and then residually connected with the node's own feature vector to obtain the second aggregated feature output by this layer. For each node, the first and second aggregated features output by the current residual graph convolutional layer are used as the input features of the next residual graph convolutional layer, respectively. After N residual graph convolutional layers are stacked, each node obtains the first and second aggregated features updated by N residual graph convolutional operations.

5. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The generation of the first gating weight and the second gating weight for each node in the node-level adaptive gating module specifically includes: The node-level adaptive gating module receives the first and second aggregate features of each node output by the residual graph convolutional layer stacking module. For each node, the first aggregate feature is mapped to the [0,1] interval through the first fully connected layer to output the first gating weight, and the second aggregate feature is mapped to the [0,1] interval through the second fully connected layer to output the second gating weight. The sum of the first gating weight and the second gating weight is 1. When each node updates its features in each residual graph convolutional layer, the product of the first aggregated feature and the first gate weight is taken as the first contribution component, and the product of the second aggregated feature and the second gate weight is taken as the second contribution component. The input features of the previous layer of the node are calculated with the first contribution component and the second contribution component respectively to obtain the updated features of the node in the current layer.

6. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The node-level adaptive gating module's setting of the second gating weight to 0 or 1 based on the abnormal marker of the mirror node alignment index specifically includes: The node-level adaptive gating module reads the abnormal markers of the inspection items corresponding to each node in the inspection report data and the mirror node. The abnormal markers are generated by comparing the measured values ​​of the inspection items with the preset normal value range. When the measured value is lower than the lower limit of the reference range or higher than the upper limit of the reference range, it is marked as abnormal. When the measured value is between the lower limit and the upper limit of the reference range, it is marked as normal. When the test item corresponding to any one node in the mirror node pair is marked as abnormal while the test item corresponding to the other node is marked as normal, the second gating weight corresponding to the mirror node pair is assigned to 0, and the first gating weight is assigned to 1. The node update depends only on the first aggregated feature. When the abnormal flags of the test items corresponding to the two nodes in the mirror node pair are both normal, the second gating weight corresponding to the mirror node pair is assigned a value of 1, and the first gating weight is assigned a value of 0. The node update depends only on the second aggregated feature. The node-level adaptive gating module multiplies the assigned first and second gating weights by their respective aggregated features before participating in the node feature update of the current residual graph convolutional layer.

7. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The symmetry-aware residual convolution module calculates the feature difference values ​​between mirror node pairs as symmetry consistency constraints, specifically including: Extract the second aggregated features of each node from each residual graph convolutional layer. For each mirror node pair in the second graph branch, calculate the Euclidean distance between the second aggregated features of the left node and the second aggregated features of the right node as the feature difference value. Map the feature difference value to the [0,1] interval and use it as the original symmetric consistency constraint value of the mirror node pair in the current layer. The symmetry-aware residual convolution module receives the second gating weights output by the node-level adaptive gating module, multiplies the original symmetry consistency constraint value with the second gating weights, and obtains the weighted symmetry consistency constraint value. The symmetry-aware residual convolution module takes the average of the weighted symmetry consistency constraint values ​​of all mirror node pairs across all residual map convolutional layers, and uses this as the global symmetry consistency constraint loss term. This global symmetry consistency constraint loss term is then added to the main loss function of the classification output module to form a joint optimization objective.

8. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The self-pairing regularization training module performs flipping enhancement on the second graph branch, specifically including: The self-pairing regularization training module performs horizontal flip enhancement on the second graph branch during the training phase. Based on the mirror node pairing relationship, it swaps the features of the left node and the right node in each mirror node pair in the second graph branch, while keeping the edge connection relationship and hyperedge connection relationship between the nodes unchanged, and generates the flipped second graph branch. The self-pairing regularization training module inputs the original second graph branch and the flipped second graph branch into the residual graph convolutional layer stacking module to obtain the second aggregated features of each node corresponding to the original second graph branch and the second aggregated features of each node corresponding to the flipped second graph branch. The self-pairing regularization training module locates the corresponding mirror pairing node for each node in the original second graph branch and calculates the mean square error between the second aggregated feature of the node in the original second graph branch and the second aggregated feature of the mirror pairing node in the flipped second graph branch. The mean squared error of all nodes is taken as the consistency constraint loss term, and the consistency constraint loss term is added to the main loss function of the classification output module.

9. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The feature fusion module performs cross-layer concatenation of the aggregated features output from each residual map convolutional layer to generate a fused feature matrix, specifically including: The feature fusion module extracts the first and second aggregate features of each node from each residual map convolutional layer to obtain N sets of feature matrices, each set containing the first and second aggregate features of all nodes in the current layer; The feature fusion module concatenates N sets of feature matrices along the feature dimension. Using the original input feature vector of each node as the reference, the first aggregated feature of the same node in layers 1 to N is arranged sequentially after the reference feature, and the second aggregated feature of the same node in layers 1 to N is arranged sequentially after the first aggregated feature, forming the cross-layer fusion feature vector of the node. The feature fusion module arranges the cross-layer fusion feature vectors of all nodes in node order to form a fusion feature matrix.

10. The knowledge graph-based laboratory report review and clinical significance generation system according to claim 1, characterized in that, The classification output module maps the fusion feature matrix to generate review results and clinical description text, specifically including: The classification output module performs global average pooling on the fused feature matrix to obtain a fixed-length global feature vector. The global feature vector is then input into the first fully connected classifier, which is activated by softmax and outputs the probability distribution of each preset review category. The category corresponding to the maximum probability is used as the review classification result. The review categories include three categories: normal, abnormal pending review, and critical value warning. The classification output module concatenates the global feature vector with the category embedding vector corresponding to the review classification result. The concatenated vector is then input into a Transformer-based decoder. The decoder generates clinical description text word by word in an autoregressive manner. The clinical description text includes a list of abnormal test items, a description of the degree of deviation of abnormal indicators, the comparison results with mirror node pairs, and clinical prompts based on the comparison results. The classification output module outputs the review classification results and clinical description text as system outputs. The review classification results are output in the form of structured tags, and the clinical description text is output in the form of natural language.