Automatic scoring system for individual case security report based on large model

Abstract

The invention relates to the technical field of medicine safety monitoring, in particular to an individual case safety report automatic scoring system based on a large model. The system comprises a report receiving module, a deep semantic understanding module, a dynamic graph structure construction module, a graph neural network scoring module and an interpretability output module. Deep semantic analysis is carried out through a large language model pre-trained in the medical field, a multi-dimensional semantic graph with time sequence characteristics is constructed, a graph neural network based on an attention mechanism is utilized to carry out multi-round iterative reasoning to generate scores, and reasoning process interpretation based on a critical path is provided. According to the method, medical report semantics can be deeply understood, the dynamic relation between entities is accurately captured, the scoring accuracy and interpretability are improved, and reliable technical support is provided for drug safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of drug safety monitoring technology, and more specifically, to an automatic scoring system for individual case safety reports based on a large model. Background Technology

[0002] Individual safety reports are a crucial component of the drug safety monitoring system, and their quality directly impacts the identification of drug safety signals and the formulation of regulatory decisions. With the rapid increase in drug safety data, traditional manual scoring methods are no longer sufficient to meet the demands for efficient and accurate processing of large volumes of reports. Establishing an automated scoring system is of significant practical importance for improving the efficiency of drug safety monitoring and ensuring scoring consistency.

[0003] Existing technologies suffer from the following main shortcomings: Traditional rule-based automated systems heavily rely on expert experience to construct scoring rules, making them ill-suited to complex and ever-changing natural language descriptions. They also have limited understanding of implicit information and contextual relationships within medical texts, and rule maintenance costs are high. Systems employing traditional machine learning methods require extensive manual feature engineering annotation, separating the feature extraction process from the scoring task and failing to learn the deep semantic information of reports end-to-end. Most existing deep learning-based systems use standard text classification architectures, failing to fully consider the unique structured information and complex logical relationships within medical reports, and lacking specialized processing for key elements such as causal relationships and time series within the reports. Some studies attempting to use graph neural networks primarily employ static graph structures, which cannot adapt to the dynamic characteristics of different reports and neglect the depth of semantic understanding during graph structure construction. Furthermore, existing systems generally lack interpretability; the reasoning process for scoring results is opaque, making it difficult to gain the trust and acceptance of medical professionals. These technical limitations restrict the accuracy and practicality of existing scoring systems.

[0004] Therefore, this paper proposes an automatic scoring system for individual safety reports based on a large model to address the above problems. The technical issues to be solved include: how to deeply understand the complex medical semantics in safety reports, how to effectively capture the dynamic relationships between entities within the report, how to improve the accuracy and interpretability of the scoring process, and how to adapt to the structural differences and content changes of different reports. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an automatic scoring system for individual case security reports based on a large model, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an automatic scoring system for individual case security reports based on a large model, comprising:

[0007] The report receiving module is used to receive individual security reports containing natural language descriptions and structured data;

[0008] The deep semantic understanding module is connected to the report receiving module. It uses a large language model pre-trained in the medical field to perform deep semantic analysis on the report. The large language model is pre-trained on more than 1 million medical documents and clinical reports and uses a multi-task learning approach to simultaneously perform entity recognition, relation extraction and logical reasoning.

[0009] The dynamic graph structure construction module is connected to the deep semantic understanding module. Based on the semantic parsing results, it constructs a multi-dimensional semantic graph with temporal characteristics. The semantic graph includes drug nodes, adverse event nodes, time nodes, and patient status nodes. The nodes are connected by causal relationship edges, temporal relationship edges, and logical relationship edges, and each edge contains a confidence weight.

[0010] The graph neural network scoring module is connected to the dynamic graph structure construction module. It uses a graph neural network based on an attention mechanism to perform multi-round iterative reasoning on the semantic graph and generates a comprehensive score by aggregating node features and edge features. The graph neural network contains at least 3 graph attention layers, and the number of neurons in each layer is between 256 and 1024.

[0011] An interpretability output module is connected to the graph neural network scoring module to output scoring results and an explanation of the reasoning process based on the semantic graph key path.

[0012] Preferably, the deep semantic understanding module adopts a Transformer-based encoder-decoder architecture, wherein the number of encoder layers is between 12 and 24, the hidden layer dimension is between 768 and 1024, the number of attention heads is between 12 and 16, and the model adopts a dual pre-training task of masked language modeling and next sentence prediction during training.

[0013] Preferably, the deep semantic understanding module includes a multi-granularity attention mechanism, specifically word-level attention, sentence-level attention, and document-level attention. Word-level attention captures local semantic features, sentence-level attention analyzes the logical relationships between sentences, and document-level attention grasps the overall narrative structure. The outputs of the three attention mechanisms are fused through a gating mechanism.

[0014] Preferably, the dynamic graph structure construction module adopts an incremental graph construction method. When a new report is received, the graph structure difference detection algorithm identifies the subgraph regions that need to be updated, and only the affected nodes and edges are locally updated, with the update response time controlled within 500ms.

[0015] Preferably, the dynamic graph structure construction module further includes a graph structure verification unit for detecting logical consistency in the semantic graph, including time sequence verification, causal relationship verification, and dosage rationality verification. The time sequence verification ensures that adverse events occur after medication, the causal relationship verification eliminates contradictory medical logic, and the dosage rationality verification checks that the drug dosage is within a safe range.

[0016] Preferably, the graph neural network scoring module adopts a multi-scale graph convolution architecture, including local subgraph convolution, global graph convolution, and hierarchical graph pooling. Local subgraph convolution extracts microscopic semantic features, global graph convolution captures macroscopic structural features, and hierarchical graph pooling retains important topological information through multi-level downsampling.

[0017] Preferably, the graph neural network scoring module further includes an adversarial training mechanism, which generates adversarial samples during the training process to enhance the robustness of the model by maximizing the model loss function. The proportion of adversarial training is controlled between 5% and 15% of the training data.

[0018] Preferably, the interpretability output module employs an attention-based critical path extraction algorithm to identify the nodes and edge sequences that have the greatest impact on the score from the semantic graph, and generates an explanation of the reasoning chain in natural language form, including key evidence extraction, logical reasoning process, and uncertainty explanation.

[0019] Preferably, the interpretability output module further includes a visualization unit, which uses a force-directed graph layout algorithm to display the semantic graph structure, uses different colors and sizes to distinguish the importance of nodes, highlights key paths and abnormal areas, and provides an interactive scoring and exploration interface.

[0020] Preferably, the system further includes a continuous learning module that incrementally updates model parameters through an online learning mechanism. When expert feedback or new labeled data is received, an elastic weight consolidation method is used to prevent catastrophic forgetting. The update frequency is set to update the model once every 1,000 reports processed.

[0021] The technical effects and advantages of this invention are as follows:

[0022] Compared to existing technologies, this invention employs a large-scale language model pre-trained in the medical field, combined with a multi-task learning mechanism. It is pre-trained on over one million medical documents and clinical reports, simultaneously performing entity recognition, relation extraction, and logical reasoning tasks. This approach enables the model to deeply understand the specific meanings and contextual relationships of medical terms, accurately capture the implicit medical logic within reports, effectively enhance the semantic understanding depth of complex medical texts, significantly improve the accuracy of key information extraction, and provide a reliable semantic foundation for subsequent scoring.

[0023] Compared to existing technologies, this invention constructs a multi-dimensional semantic graph structure with temporal characteristics, representing elements such as drugs, adverse events, time points, and patient status in the report as nodes, and establishing connections through causal relationship edges, temporal relationship edges, and logical relationship edges with confidence weights. This dynamic graph construction method can completely preserve the temporal information and logical connections in the report, and the visualization of the graph structure makes the scoring criteria clearer and more explicit, providing medical staff with an intuitive reference for understanding the scoring results, and enhancing the credibility and practicality of the system.

[0024] Compared to existing technologies, this invention employs an attention-based graph neural network for multi-round iterative reasoning. The network contains at least three graph attention layers, each with 256 to 1024 neurons, and generates a comprehensive score by aggregating node and edge features. This architecture simultaneously considers local semantic features and global structural information, continuously optimizing feature representations through multiple iterations. It effectively captures complex dependencies in the report, significantly improving scoring accuracy while maintaining processing efficiency.

[0025] Compared to existing technologies, this invention designs an attention-based critical path extraction algorithm and a visualization unit to identify the nodes and edge sequences that have the greatest impact on scoring from the semantic graph, generate an explanation of the reasoning chain in natural language, and use a force-directed graph layout algorithm to visually display critical paths and abnormal regions. This method makes the scoring process transparent and traceable, helps users understand the system's decision-making basis, facilitates the discovery and correction of potential errors, and improves the system's usability and acceptability. Attached Figure Description

[0026] Figure 1 This is the overall system architecture and data flow diagram of the present invention.

[0027] Figure 2 This is a detailed diagram of the deep semantic understanding module of the present invention.

[0028] Figure 3 This is a diagram illustrating the construction and verification of the dynamic graph structure of the present invention.

[0029] Figure 4 This is the graph neural network scoring and interpretable output graph of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1

[0032] As attached Figures 1 to 4 The system shown is an automated scoring system for individual safety reports based on a large model. This system achieves automated quality assessment and decision support for drug safety reports by deeply integrating natural language processing, dynamic graph structure analysis and interpretable machine learning technologies.

[0033] The system adopts a modular design, with its core architecture including a report receiving module, a deep semantic understanding module, a dynamic graph structure construction module, a graph neural network scoring module, and an interpretable output module. These modules are connected through standardized interfaces to form a complete data processing pipeline, capable of effectively handling individual safety reports from medical institutions, pharmaceutical companies, and regulatory authorities.

[0034] The report receiving module, serving as the system entry point, is responsible for receiving multi-source, heterogeneous safety report data. This module supports various data formats, including JSON, XML, and structured text, and includes a data validation unit to perform integrity checks and format standardization on the input reports. Validation rules are based on international standards such as the ICHE2B guidelines, ensuring the integrity of required fields such as patient identification, drug details, event description, and time information.

[0035] Data preprocessing includes removing personally identifiable information to protect privacy, correcting spelling errors, and standardizing medical terminology using the MedDRA dictionary. The processed data is then converted to a unified internal representation format, providing clean and standardized input for subsequent modules. This module also includes a data quality assessment function, generating a quality score for each input report for prioritizing subsequent processing flows.

[0036] The deep semantic understanding module undertakes the task of deep parsing of the report content. This module is built on a large-scale language model pre-trained in the medical field and adopts a Transformer encoder-decoder structure. The encoder has 12 layers, the hidden layer dimension is 768, and the number of attention heads is 12. These parameters have been optimized through extensive experiments to achieve a balance between computational efficiency and semantic understanding depth.

[0037] The decoder employs an autoregressive generative approach, simultaneously optimizing both masked language modeling and next-sentence prediction during training. The loss function for masked language modeling is defined as:

[0038]

[0039] Where x represents the input sequence, and D is the training dataset. \i This represents the masked sequence. The next sentence prediction task uses the cross-entropy loss function, with a negative sampling ratio set to 50%.

[0040] This dual training objective enables the model to learn semantic representations at both the lexical and sentence levels, enhancing its understanding of medical texts. Through pre-training on a large-scale medical corpus and domain-adaptive fine-tuning, the module can accurately identify key entities such as drug names, adverse event types, and clinical symptoms, and extract semantic relationships between entities.

[0041] Furthermore, the deep semantic understanding module includes a multi-granularity attention mechanism, a key design feature for achieving accurate semantic analysis. Specifically, word-level attention employs a local window attention mechanism, with a window size set to 5 words, calculated as follows:

[0042]

[0043] Where Q, K, and V represent the query, key, and value matrices, respectively, and d k It is the dimension of the key vector.

[0044] This design allows the model to focus on local semantic units, accurately capturing subtle differences in medical terminology. Sentence-level attention combines a bidirectional LSTM with an attention mechanism; the LSTM hidden layer dimension is set to 512, and a gating mechanism is used to fuse forward and backward states to analyze the logical relationships between sentences. Document-level attention employs a hierarchical attention network, aggregating word-level and sentence-level features layer by layer to grasp the overall narrative structure.

[0045] The fusion of the three attention outputs employs a gated recurrent unit, with the update and reset gates computed based on the current input and the previous state. This multi-granularity design enables the model to simultaneously handle local details and global context, effectively parsing complex semantic structures commonly found in medical reports, such as nested descriptions and implicit logical relationships.

[0046] Furthermore, the dynamic graph structure construction module employs an incremental graph construction method, a core innovation for addressing real-time data processing needs. This method is based on a graph structure difference detection algorithm; upon receiving new report information, it first calculates the similarity between the current graph structure and the new structure.

[0047]

[0048] Where V(G) and E(G) represent the node set and edge set of graph G, respectively.

[0049] A similarity threshold of 0.8 is set, triggering a local update when the similarity falls below this threshold. The update process employs a transaction mechanism to ensure data consistency while maintaining the graph structure's version history. This incremental construction method significantly reduces computational overhead, enabling the system to process report updates in real time while maintaining the temporal consistency of the graph structure.

[0050] This design works closely with the deep semantic understanding module to transform semantic parsing results into a dynamic graph structure representation, providing a structured data foundation for subsequent analysis. The graph structure storage employs a combination of adjacency lists and attribute matrices, supporting efficient graph traversal and feature lookup.

[0051] Furthermore, the dynamic graph structure construction module includes a graph structure verification unit, a crucial component for ensuring the quality of the semantic graph. This unit implements a multi-dimensional verification rule engine, where time-sequence verification employs temporal logical reasoning, and the verification conditions are expressed as follows:

[0052]

[0053] Where E event E represents the set of adverse events. drug This represents the set of drug exposures, and t(·) returns the event timestamp.

[0054] Causal relationship verification is based on a medical knowledge graph, using a path ranking algorithm to calculate the confidence level of causal relationships. Dosage rationality verification employs anomaly detection based on a Gaussian distribution, calculating the Mahalanobis distance between the current dose and historical doses. The verification rule base contains over 1000 medical logic rules and supports dynamic updates.

[0055] This unit works in conjunction with an incremental graph construction mechanism to verify the graph logic in real time during the construction process, ensuring that the generated semantic graph conforms to medical common sense and domain knowledge. The verification results are stored in the edge attributes in the form of confidence scores, providing a reference for subsequent scoring.

[0056] Furthermore, the graph neural network scoring module employs a multi-scale graph convolutional architecture, a key technology for achieving accurate scoring. This architecture comprises three components: local subgraph convolution, global graph convolution, and hierarchical graph pooling. The local subgraph convolution uses a graph attention network, with attention coefficients calculated as follows:

[0057]

[0058] Where h i Let W represent the features of node i, W be the weight matrix, a be the attention vector, and N(i) be the set of neighbors of node i.

[0059] Global graph convolution employs a graph isomorphic network, iteratively updating node representations using the Weisfeiler-Lehman algorithm. Hierarchical graph pooling uses Top-k pooling, retaining the node with the highest importance score in each layer. This multi-scale design enables the model to learn both local structure and global features simultaneously, fully utilizing the semantic graph information output by the dynamic graph structure building module to improve scoring accuracy.

[0060] The network was trained end-to-end, and the loss function combined mean squared error and contrastive learning loss to ensure that the scoring was both accurate and discriminative. The model parameters were optimized using the backpropagation algorithm, and the Adam optimizer was used for parameter updates, with the learning rate set to 0.001.

[0061] Furthermore, the graph neural network scoring module incorporates an adversarial training mechanism, a crucial measure to enhance model robustness. During training, adversarial examples are generated using the projective gradient descent method.

[0062]

[0063] Where x is the original input, y is the label, J is the loss function, and θ is the model parameters. The disturbance amplitude is controlled within 0.01.

[0064] The adversarial training loss function combines the original loss and the adversarial loss, with dynamically adjusted weights. This training mechanism enables the model to better handle noisy data and anomalous inputs, improving the stability of the scoring results. This mechanism works in conjunction with a multi-scale graph convolutional architecture to ensure the model's reliability in complex real-world environments.

[0065] Adversarial examples were generated taking into account the characteristics of medical reports, introducing perturbations while maintaining semantic plausibility, enabling the model to learn more generalizable feature representations. During training, the proportion of adversarial examples started at 5% and gradually increased to 15% with each training epoch.

[0066] Furthermore, the interpretability output module employs an attention-based critical path extraction algorithm, a core design element for achieving transparent decision-making. This algorithm first calculates the attention score of nodes in the graph, then uses an improved Dijkstra's algorithm to find the path with the highest importance. The path score is calculated as follows:

[0067]

[0068] Where p represents the path, α i,i+1 It is the edge attention weight, deg(v) i ) represents the degree centrality of nodes.

[0069] Natural language interpretation generation employs a template-based approach. The template library contains various medical reasoning patterns, supporting dynamic parameter binding and logical chain generation. This interpretation mechanism is tightly integrated with the graph neural network scoring module, transforming complex graph computations into understandable reasoning processes, helping users comprehend the basis for their decisions.

[0070] During the interpretation generation process, the system identifies key evidence nodes, extracts relevant text fragments, and generates natural language descriptions that conform to medical reporting conventions. Each interpretation includes an assessment of the strength of evidence and an explanation of uncertainty, providing users with comprehensive decision support information.

[0071] Furthermore, the interpretability output module includes a visualization unit, a crucial component for enhancing the user experience. This unit employs a force-directed graph layout algorithm, calculating node positions based on a physical simulation model.

[0072]

[0073] Where q i q j It is the nodal charge, r ij It is the distance between nodes, k e and k s It is the elasticity coefficient.

[0074] Node importance is represented visually, color mapping uses the HSL color space, and size mapping is based on node degree centrality. Interactive features support node dragging, zooming, and detailed viewing, providing an immersive analysis experience. This design, combined with a critical path extraction algorithm, presents scoring criteria in an intuitive way, lowering the barrier to understanding for users.

[0075] The visualization system employs a responsive design, adapting to display devices of various sizes. Users can explore the graph structure, view node details, and trace reasoning paths through interactive operations. The system also provides multiple view modes, including chronological views, causal relationship views, and evidence strength views, to meet diverse analytical needs.

[0076] Furthermore, the system includes a continuous learning module, a key mechanism for adapting to environmental changes. This module employs an elastic weight consolidation algorithm, adding regularization terms during model updates.

[0077]

[0078] Where L new It is the loss function for the new task, F. i These are the diagonal elements of the Fisher information matrix. It is an important parameter for the old task, and λ is the regularization coefficient.

[0079] The model updates employ mini-batch gradient descent with dynamically adjusted learning rates, and the update cycle is automatically optimized based on the frequency of data inflow. This continuous learning mechanism enables the system to adapt to changes in data distribution while avoiding catastrophic forgetting, ensuring long-term stability.

[0080] During continuous learning, the system monitors model performance metrics and automatically triggers a model update process when performance degradation is detected. The update process employs a gradual strategy, first fine-tuning on a small batch of new data, and then validating on the complete dataset to ensure the safety and effectiveness of the update.

[0081] Example 2

[0082] The following explanation uses a specific application scenario. In this scenario, the system is deployed on a drug safety monitoring platform to process individual safety reports from medical institutions. The report receiving module receives JSON-formatted report data via a RESTful API, performing data validation and format conversion.

[0083] Assume the received report includes the patient's medication history, adverse event descriptions, and laboratory test results. The data validation unit checks the completeness of required fields, such as patient age, drug name, event occurrence time, and severity level. After validation, the data is converted into a uniform internal representation format, including structured fields and text descriptions.

[0084] The deep semantic understanding module parses the report text. First, it performs word segmentation, and then generates word vector representations using a pre-trained language model. The entity recognition unit uses a conditional random field model to label entity boundaries and types, recognizing the drug name "Drug A", the adverse event "abnormal liver function", and the clinical symptom "jaundice".

[0085] The relation extraction unit uses a multilayer perceptron classifier to determine the semantic relationships between entities, identifying a causal relationship between "drug A" and "abnormal liver function," and a symptom relationship between "jaundice" and "abnormal liver function." During the parsing process, a multi-granularity attention mechanism considers both local context and global structure simultaneously, ensuring accurate understanding of complex medical descriptions.

[0086] The dynamic graph structure construction module builds a multidimensional semantic graph based on the semantic parsing results. Nodes include entities such as "drug A", "abnormal liver function", and "jaundice", and edges represent semantic connections such as causal relationships and temporal relationships. An incremental construction algorithm calculates the similarity between the new graph and the existing graph. If the similarity is found to be 0.75, which is below the threshold of 0.8, a local update is triggered.

[0087] The graph structure verification unit performs time sequence verification to confirm that the adverse event occurred after the medication administration time; it performs causal relationship verification to check the rationality of the medical logic; and it performs dosage rationality verification to compare the current dosage with the safe range. After passing the verification, the semantic graph is updated and stored in the graph database.

[0088] The graph neural network scoring module loads a pre-trained model and analyzes the constructed semantic graph. Local subgraph convolution focuses on the "Drug A - Abnormal Liver Function" subgraph, calculating attention weights; global graph convolution analyzes the structural features of the entire graph; hierarchical graph pooling preserves important nodes. An adversarial training mechanism ensures the model's robustness to data noise.

[0089] The scoring process involves multiple iterations, ultimately generating a quantitative score of 85 out of 100, indicating a high-quality report. During the scoring process, the model comprehensively considers multiple factors such as node features, edge weights, and graph structure complexity to ensure the comprehensiveness and accuracy of the score.

[0090] The interpretability output module extracts key reasoning paths and generates a natural language explanation: "The score is mainly based on the strong causal relationship between drug A and abnormal liver function, complete temporal information, and consistent clinical manifestations." The visualization unit displays the semantic graph structure, highlighting the key path "drug A → abnormal liver function → jaundice," with node size and color depth reflecting the degree of importance.

[0091] Users can view detailed explanations and understand the scoring criteria through the interactive interface. The system also provides a detailed display of scoring information, including scores for each scoring dimension and suggestions for improvement, helping users to fully understand the report's quality.

[0092] The entire processing workflow adopts a pipelined architecture, with asynchronous communication between modules via message queues. The system supports real-time processing mode, with a single report processing time controlled within 2 seconds; it also supports batch processing mode for processing historical data. Through these technologies, the system provides an accurate, efficient, and reliable automated scoring solution for drug safety monitoring, capable of adapting to data processing needs of different scales and meeting the practical application scenarios of regulatory agencies and pharmaceutical companies.

[0093] The system also includes monitoring and management functions, tracking the operational status of each module in real time, recording processing logs, and providing system performance reports. Administrators can configure system parameters, adjust scoring rules, and view processing statistics through the management interface to ensure stable system operation and continuous optimization.

[0094] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0095] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0096] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic scoring system for individual case security reports based on a large model, characterized in that, include: The report receiving module is used to receive individual security reports containing natural language descriptions and structured data; The deep semantic understanding module is connected to the report receiving module. It uses a large language model pre-trained in the medical field to perform deep semantic analysis on the report. The large language model is pre-trained on more than 1 million medical documents and clinical reports and uses a multi-task learning approach to simultaneously perform entity recognition, relation extraction and logical reasoning. The dynamic graph structure construction module is connected to the deep semantic understanding module. Based on the semantic parsing results, it constructs a multi-dimensional semantic graph with temporal characteristics. The semantic graph includes drug nodes, adverse event nodes, time nodes, and patient status nodes. The nodes are connected by causal relationship edges, temporal relationship edges, and logical relationship edges, and each edge contains a confidence weight. The graph neural network scoring module is connected to the dynamic graph structure construction module. It uses a graph neural network based on an attention mechanism to perform multi-round iterative reasoning on the semantic graph and generates a comprehensive score by aggregating node features and edge features. The graph neural network contains at least 3 graph attention layers, and the number of neurons in each layer is between 256 and 1024. An interpretability output module is connected to the graph neural network scoring module to output scoring results and an explanation of the reasoning process based on the semantic graph key path.

2. The automatic scoring system for individual case security reports based on a large model according to claim 1, characterized in that, The deep semantic understanding module adopts a Transformer-based encoder-decoder architecture, in which the number of encoder layers is between 12 and 24, the hidden layer dimension is between 768 and 1024, the number of attention heads is between 12 and 16, and the model adopts a dual pre-training task of masked language modeling and next sentence prediction during training.

3. The automatic scoring system for individual case security reports based on a large model according to claim 2, characterized in that, The deep semantic understanding module includes a multi-granularity attention mechanism, specifically word-level attention, sentence-level attention, and document-level attention. Word-level attention captures local semantic features, sentence-level attention analyzes the logical relationships between sentences, and document-level attention grasps the overall narrative structure. The outputs of the three attention mechanisms are fused through a gating mechanism.

4. The automatic scoring system for individual case security reports based on a large model according to claim 1, characterized in that, The dynamic graph structure construction module adopts an incremental graph construction method. When a new report is received, the graph structure difference detection algorithm identifies the subgraph regions that need to be updated and only locally updates the affected nodes and edges, with the update response time controlled within 500ms.

5. The automatic scoring system for individual case security reports based on a large model according to claim 4, characterized in that, The dynamic graph structure construction module also includes a graph structure verification unit for detecting logical consistency in the semantic graph, including time sequence verification, causal relationship verification, and dosage rationality verification. The time sequence verification ensures that adverse events occur after medication, the causal relationship verification eliminates contradictory medical logic, and the dosage rationality verification checks that the drug dosage is within the safe range.

6. The automatic scoring system for individual case security reports based on a large model according to claim 1, characterized in that, The graph neural network scoring module adopts a multi-scale graph convolution architecture, including local subgraph convolution, global graph convolution, and hierarchical graph pooling. Local subgraph convolution extracts micro-semantic features, global graph convolution captures macro-structural features, and hierarchical graph pooling retains important topological information through multi-level downsampling.

7. The automatic scoring system for individual case security reports based on a large model according to claim 6, characterized in that, The graph neural network scoring module also includes an adversarial training mechanism, which generates adversarial samples during training and enhances the robustness of the model by maximizing the model loss function. The proportion of adversarial training is controlled between 5% and 15% of the training data.

8. The automatic scoring system for individual case security reports based on a large model according to claim 1, characterized in that, The interpretability output module employs an attention-based critical path extraction algorithm to identify the nodes and edge sequences that have the greatest impact on the score from the semantic graph, and generates an explanation of the reasoning chain in natural language form, including key evidence extraction, logical reasoning process, and uncertainty explanation.

9. The automatic scoring system for individual case security reports based on a large model according to claim 8, characterized in that, The interpretability output module also includes a visualization unit, which uses a force-directed graph layout algorithm to display the semantic graph structure, uses different colors and sizes to distinguish the importance of nodes, highlights key paths and abnormal areas, and provides an interactive scoring and exploration interface.

10. The automatic scoring system for individual case security reports based on a large model according to claim 1, characterized in that, The system also includes a continuous learning module that incrementally updates model parameters through an online learning mechanism. When expert feedback or new labeled data is received, an elastic weight consolidation method is used to prevent catastrophic forgetting. The update frequency is set to update the model once every 1,000 reports processed.

Images