Intelligent early warning method and system for tumor risk based on natural language processing

Abstract

This invention relates to an intelligent early warning method and system for tumor risk based on natural language processing (NLP). Specifically, it relates to the field of NLP, which involves deep extraction and standardization of tumor risk entities and relationships from multi-source heterogeneous medical texts to construct a dynamically evolving domain knowledge graph. A heterogeneous graph attention network is then used to fuse multi-source evidence to quantify risk confidence. Finally, continuous collaborative optimization with privacy protection is achieved within a federated learning framework. The effect is to significantly improve the accuracy and credibility of risk association mining, ensure the interpretability of early warning results, and overcome data silo limitations. Under the premise of strictly protecting medical data privacy, the system possesses the ability to autonomously discover new knowledge and continuously evolve, providing reliable technical support for accurate, real-time early warning and proactive discovery of tumor risks.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and more specifically, to a method and system for intelligent early warning of tumor risk based on natural language processing. Background Technology

[0002] With the acceleration of medical informatization and the in-depth advancement of the health big data strategy, early warning and intelligent assessment of tumor risk have become key issues in the fields of public health and clinical medicine. Currently, hospital information systems, regional health management platforms, and disease control and prevention centers have accumulated massive amounts of diverse medical text data in their daily operations and scientific research analysis. These data sources include, but are not limited to, structured diagnostic codes and laboratory indicators in electronic health records, unstructured clinical course records and surgical reports, research abstracts and reviews in medical literature databases, and self-reported symptom texts and health questionnaires from patients. These data exhibit significant heterogeneity in format, ranging from fully structured database tables to highly free natural language descriptions. At the same time, due to the development of the medical profession itself, differences in terminology habits among different medical institutions, and the existence of regional expressions, the same clinical concepts are often expressed in multiple terms, abbreviations, and even colloquial variations. Furthermore, under the dual requirements of policy-driven data interconnection and increasingly stringent privacy and security regulations, multi-source medical data across systems and institutions faces enormous practical challenges in compliant sharing and integrated utilization. The aforementioned complex application scenarios constitute the data ecosystem upon which the intelligent early warning system for tumor risk depends for its survival and operation.

[0003] However, existing methods for automatically extracting tumor risk factors from multi-source heterogeneous medical texts using natural language processing (NLP) technology face severe technical bottlenecks. The core problem lies in the insufficient deep semantic understanding of medical knowledge by existing NLP models, particularly the lack of mechanisms for effectively aligning and fusing semantics between data from different sources. Specifically, mainstream methods largely rely on template matching based on manually configured rules or shallow machine learning models. These techniques struggle to accurately capture the complex contextual semantic relationships and subtle differences in terminology within medical texts. For instance, when faced with terms like "metastasis" and "differentiation," which have multiple medical and general meanings, the system is highly prone to misjudgments due to contextual misunderstandings. At the level of medical text annotation, the scarcity of high-quality medical text annotation data severely restricts the performance of supervised learning models. More importantly, as an important tool supporting semantic understanding, the construction and maintenance of existing knowledge graphs rely heavily on human intervention, resulting in long update cycles and limited coverage. This makes it difficult to integrate new knowledge from multiple channels such as electronic health records and the latest literature in real time, thus failing to form a dynamically evolving unified medical semantic framework. This situation of semantic barriers between multi-source data and lagging knowledge representation directly leads to problems such as incomplete extraction, incorrect association, or low confidence when extracting key risk factors such as family history of cancer, clinical symptoms, and lifestyle habits from complex texts, which seriously restricts the accuracy and reliability of the early warning system. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a tumor risk intelligent early warning method and system based on natural language processing, thereby solving the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a tumor risk intelligent early warning method based on natural language processing, specifically including the following steps: Step S1: For the acquired multi-source heterogeneous medical text data stream, a deep language model that has been further pre-trained on medical domain corpus is used for feature extraction. Then, using the multi-head attention mechanism and conditional random field joint decoding layer integrated at the top of the model, named entity recognition and preliminary relation classification are performed simultaneously, and an initial entity-relation pair set consisting of preliminary risk entities and their preliminary semantic relations is output. Step S2: Input the initial entity-relation pair set into the dynamic oncology domain knowledge graph, which has a built-in terminology library of hybrid vector space and symbolic logic. Through a voting decision algorithm based on context graph structure, calculate the hybrid similarity between each preliminary risk entity and its context and the standard concept nodes in the dynamic oncology domain knowledge graph. Map the diverse expressions in the multi-source heterogeneous medical text data stream to the unified unambiguous standard concept nodes and output the semantically aligned standardized entity-relation triplet set. Step S3: Based on the data source and relation type attributes associated with each triple in the standardized entity relation triple set, construct a heterogeneous graph from the standardized entity relation triple set, and input the heterogeneous graph into a multi-layer heterogeneous graph attention network. This multi-layer heterogeneous graph attention network learns and aggregates evidence of the same risk association from different data sources through its cross-source attention fusion layer, quantifies the weight and confidence of each risk association, and outputs a tumor risk subgraph with quantified confidence. Step S4: Deploy the tumor risk subgraph with quantified confidence and the multi-layer heterogeneous graph attention network that generated it in a federated learning framework. Each participating node in the federated learning framework uses local private medical text data to fine-tune the multi-layer heterogeneous graph attention network and only uploads the model parameters to the central server for secure aggregation. At the same time, a graph incremental update triggering mechanism is set up. When the global model obtained after secure aggregation identifies a high-frequency new risk pattern higher than a preset threshold, the stream processing pipeline is triggered to re-execute steps S1 to S2 on the newly acquired multi-source heterogeneous medical text data related to the high-frequency new risk pattern. After confirmation, the new knowledge is updated to the dynamic tumor domain knowledge graph in an incremental manner. In a preferred embodiment, the specific operation of feature extraction in step S1 using a deep language model that has been further pre-trained with medical domain corpus is as follows: Each text sequence in the acquired multi-source heterogeneous medical text data stream is input into a pre-loaded medical domain corpus for further pre-training of a deep language model. The multi-source heterogeneous medical text data stream consists of text sequences from electronic health records, clinical notes, and medical literature. The deep language model encodes each input text sequence to generate a deep context representation sequence. The text sequence consists of multiple sequentially arranged lexical units, and the deep context representation sequence consists of vectors corresponding to each lexical unit arranged in the same order.

[0006] In a preferred embodiment, the specific process of simultaneously performing named entity recognition and preliminary relation classification using the multi-head attention mechanism and conditional random field joint decoding layer integrated at the top layer of the model is as follows: First, the deep context representation sequence is input into a multi-head attention-enhanced feature refining unit to generate a refined feature representation sequence; Next, based on the refined feature representation sequence, on the one hand, a linear chain conditional random field is used to decode the entity label sequence to identify the preliminary risk entities and their types in the text sequence. Each identified preliminary risk entity contains its corresponding text content fragment and entity type information. On the other hand, relation classification decoding is performed simultaneously on the candidate entity pairs formed by all identified preliminary risk entities. The relation classification decoding employs a relation-aware attention pooling operation. This attention pooling operation performs the following process for each candidate entity pair: The candidate entity pair contains a preliminary risk entity as the subject and a preliminary risk entity as the object; the set of all lexical positions covered by the preliminary risk entity as the subject and the preliminary risk entity as the object in the text sequence is obtained; for each position in the set of lexical positions covered by the preliminary risk entity as the subject, a first contribution weight is calculated, which represents the semantic contribution of the lexical at that position to the semantic contribution of the preliminary risk entity as the subject in the current candidate entity pair relation; for each position in the set of lexical positions covered by the preliminary risk entity as the object, a second contribution weight is calculated, which represents the semantic contribution of the lexical at that position to the semantic contribution of the preliminary risk entity as the object in the current candidate entity pair relation; the first contribution weight and the second contribution weight are determined by a set of lexical positions covered by the preliminary risk entity as the subject and the preliminary risk entity as the object. The third vector representation corresponding to all lexical positions in the refined feature representation sequence is calculated by the dedicated attention network and normalized within the lexical position sets of the preliminary risk entities as subjects and objects, respectively. Then, using all the calculated first contribution weights, the third vector representation corresponding to each lexical position covered by the preliminary risk entities as subjects is weighted, and all weighted results are summed to obtain a subject entity relation feature vector. Simultaneously, using all the calculated second contribution weights, the third vector representation corresponding to each lexical position covered by the preliminary risk entities as objects is weighted, and all weighted results are summed to obtain an object entity relation feature vector. Finally, the subject entity relation feature vector and the object entity relation feature vector are concatenated or added to generate a specific relation representation vector for candidate entity pairs. This specific relation representation vector is input into the classifier to determine the preliminary semantic relation category between the preliminary risk entities as subjects and objects. By synchronously executing entity label sequence decoding and relation classification decoding in a linear chain conditional random field, an initial entity-relation pair set consisting of zero or one or more triples is output, where each triple contains a preliminary risk entity as the subject, a preliminary risk entity as the object, and a preliminary semantic relation category between the two.

[0007] In a preferred embodiment, the specific process of calculating the mixed similarity between each preliminary risk entity and its context and standard concept nodes in the dynamic tumor domain knowledge graph in step S2 is as follows: First, for each preliminary risk entity in the initial entity-relationship pair set, multiple candidate standard concept nodes with similar semantics are retrieved from the dynamic tumor domain knowledge graph to form a refined candidate standard concept node set. Subsequently, for each candidate standard concept node in the refined candidate standard concept node set, semantic similarity and context graph structure similarity are calculated respectively. The process of calculating semantic similarity is as follows: calculate the cosine similarity value between the vector space representation of the preliminary risk entity and the vector space representation of the candidate standard concept node; The process of calculating the context graph structure similarity is as follows: From the initial entity-relation pair set, extract all other preliminary risk entities that are directly connected to the current preliminary risk entity through preliminary semantic relations, forming the context entity set of the current preliminary risk entity; for each context entity in this context entity set, determine its corresponding mapping candidate node in the dynamic tumor domain knowledge graph; obtain the set of direct neighbor nodes of the current candidate standard concept node in the dynamic tumor domain knowledge graph; then, for each context entity in the context entity set, calculate the cosine similarity between the vector space representation of its mapping candidate node and the vector space representation of each node in the neighbor node set of the current candidate standard concept node, and take the maximum value; finally, take the arithmetic mean of the maximum values ​​corresponding to all context entities, and the average value is the context graph structure similarity. Based on the size of the context entity set of the preliminary risk entity and the number of connection edges of the candidate standard concept nodes in the dynamic tumor domain knowledge graph, a fusion weight is dynamically determined. Using this fusion weight, the semantic similarity and the context graph structural similarity are linearly weighted and summed to obtain the final mixed similarity.

[0008] In a preferred embodiment, the specific operation of outputting the semantically aligned standardized entity relation triplet set is as follows: For each preliminary risk entity in the initial entity-relationship pair set, select the candidate standard concept node with the maximum mixed similarity from the refined candidate standard concept node set, and use it as the unambiguous standard concept node for its final mapping. For each triple in the initial entity-relation pair set, the initial risk entity as the subject and the initial risk entity as the object are replaced with their respective mapped unambiguous standard concept nodes, and the initial semantic relationship category between the subject entity and the object entity is mapped to a predefined standardized relationship type in the dynamic tumor domain knowledge graph by querying a preset relationship mapping table. This generates a new triplet consisting of a standardized subject concept node, a standardized relation type, and a standardized object concept node. Finally, the entire set of all newly generated triplets is defined and output as a set of semantically aligned standardized entity relation triplets.

[0009] In a preferred embodiment, the specific process of constructing the standardized entity relation triplet set into a heterogeneous graph in step S3 is as follows: Identify the data source associated with each triple in the set of semantically aligned standardized entity relation triples, and expand each triple into a quadruple containing a data source identifier; Using standard medical concept nodes in the dynamic tumor domain knowledge graph as entity nodes and unique data source identifiers as source nodes, a set of nodes in a heterogeneous graph is formed. The edge set of the heterogeneous graph is constructed based on the expanded set of quadruples. The edge set includes relation edges connecting two entity nodes and source edges connecting the source node and the relation edge. Then, initialized vector representations are assigned to the nodes and edges in the heterogeneous graph: Each entity node is assigned an initial vector representation, which is directly adopted from the vector representation of the entity node in the dynamic tumor domain knowledge graph. Assign a trainable vector representation to each source node as the initial vector representation; Each relation edge is assigned an initial vector representation, which is composed of a preset embedding vector corresponding to its relation type and an encoded value representing the frequency of occurrence of the risk association corresponding to the relation edge in the semantically aligned standardized entity relation triplet set. This led to the construction and completion of the initial heterogeneous graph.

[0010] In a preferred embodiment, the specific process of outputting the tumor risk submap with quantified confidence is as follows: Construct a multi-layer heterogeneous graph attention network containing L consecutive computational layers. Each layer of the multi-layer heterogeneous graph attention network iteratively updates the vector representation of the relation edges in the heterogeneous graph in sequence. In the computation of each layer, the cross-relation edge information propagation operation is first performed: for each relation edge in the heterogeneous graph, the vector representations of the entity nodes at both ends in the previous layer of the current layer and the vector representation of the relation edge in the previous layer of the current layer are aggregated, and the updated vector representation of the relation edge in the current layer is generated through a linear transformation and a nonlinear activation function. For each relation edge in a heterogeneous graph, the multilayer heterogeneous graph attention network calculates the attention weight of that relation edge for each source node connected to it via a source edge through its cross-source attention fusion layer. The calculation process for this attention weight is as follows: First, for the combination of a relation edge and a source node, the vector representation of the relation edge in the current layer, the vector representation of the source node in the previous layer, and a source-edge collaborative signal vector generated for the combination are concatenated to obtain a combined vector representation. Then, this combined vector representation is multiplied by a trainable attention parameter vector and passed through a LeakyReLU nonlinear activation function to obtain an unnormalized attention score. Finally, the unnormalized attention scores corresponding to all source nodes connected to the relation edge are exponentially normalized to obtain the normalized attention weight corresponding to each source node. The process of generating the source-edge collaborative signal vector is as follows: First, obtain the vector representation of the source node in the previous layer of the current layer; then, calculate the average vector of the vector representations of all other relation edges connected to the source node through the source edge in the heterogeneous graph in the current layer; then, concatenate the vector representation of the source node with the average vector, and transform it through a trainable neural network unit. The output is the source-edge collaborative signal vector. Then, the source aggregation update operation of the edge features is performed: using the calculated normalized attention weights, the vector representations from different source nodes are weighted and fused; specifically, for each relation edge, the vector representation of each source node connected to the edge in the previous layer of the current layer is multiplied by its corresponding normalized attention weight, and then multiplied by a trainable weight matrix corresponding to the current layer to obtain the weighted contribution of the source node; the weighted contributions of all connected source nodes are added together to obtain the aggregated vector representation; finally, the aggregated vector representation is added element-wise to the vector representation of the relation edge in the current layer, and the addition result is subjected to layer normalization to obtain the final vector representation of the relation edge in the current layer after source aggregation update; In the final layer of the multi-layer heterogeneous graph attention network, a confidence prediction function is applied to the final vector representation of each relation edge after the update, which is mapped to a scalar value as the quantified confidence of the risk association represented by the relation edge. Finally, all relation edges in the heterogeneous graph, along with their two-end entity nodes, relation types, calculated quantitative confidence scores, and a list of data source identifiers connected to each relation edge through the source edge, are extracted to form a tumor risk subgraph with quantitative confidence scores.

[0011] In a preferred embodiment, in step S4, the specific process by which each participating node in the federated learning framework fine-tunes the multi-layer heterogeneous graph attention network using local private medical text data and only uploads model parameter updates to the central server for secure aggregation is as follows: The central server distributes the current version of the global multi-layer heterogeneous graph attention network to each participating node; Each participating node independently executes steps S1 and S2 locally using its private medical text data, generates a set of locally standardized entity relationship triples, and constructs a local heterogeneous graph based on this and the current version of the tumor risk subgraph. Each node uses the received global model to process the local heterogeneous graph, calculates the loss value and updates the model parameters to obtain the local model parameter update amount. The local model parameter update amount or the result after adding noise perturbation to the local model parameter update amount is encrypted, and only the encrypted result is uploaded to the central server. After collecting the encrypted local model parameter updates uploaded by all participating nodes, the central server first uses a secure aggregation algorithm to aggregate the encrypted data uploaded by each node, then decrypts the aggregation result to obtain the aggregated value of the model parameter updates of all nodes, and then applies this aggregated value to the current global model parameters, thereby obtaining a new round of optimized global multi-layer heterogeneous graph attention network.

[0012] In a preferred embodiment, the specific process of setting a map incremental update triggering mechanism is as follows: After the central server completes the global model security aggregation for each round of federated learning, the newly optimized global multilayer heterogeneous graph attention network is used to perform inference on a retained validation dataset. For each candidate risk association triple generated by the model processing on the validation dataset, it is checked whether it already exists in the current version of the tumor risk subgraph with quantified confidence. If it does not exist, it is marked as a candidate new risk pattern. For all marked candidate new risk patterns, their frequency of occurrence in the current batch of the validation dataset is counted, and the prediction of the global model for the triple in each candidate new risk pattern is recorded. The reliability is calculated by averaging these confidence levels and calculating the variance of these confidence levels across multiple consecutive inference batches. When a candidate new risk pattern exists whose frequency of occurrence exceeds a first preset frequency threshold, whose average confidence level exceeds a second preset reliability threshold, and whose variance of confidence level across multiple consecutive batches is lower than a third preset variance threshold, the pattern is determined to be a high-frequency new risk pattern to be confirmed. The central server then broadcasts a lightweight trigger signal containing key feature information of the high-frequency new risk pattern to all participating nodes. Upon receiving the trigger signal, each participating node starts its local stream processing pipeline. The pipeline continuously monitors newly generated medical text data streams on the local nodes and independently re-executes step S1 (entity and relation extraction) and step S2 (semantic alignment) only for segments of text content that match the key features of the trigger signal, generating a new set of candidate standardized entity-relation triplets related to the trigger pattern. Each node anonymizes these new candidate standardized entity-relation triplet sets and uploads them to the central server. The central server aggregates candidates from all nodes, performs deduplication, frequency statistics, and consistency verification based on the frequency and confidence of the same candidates reported by multiple nodes. The verified new knowledge is then... Triples are incrementally merged into the dynamic oncology domain knowledge graph by adding new nodes or new relation edges. Based on the updated dynamic oncology domain knowledge graph, the tumor risk subgraph with quantified confidence is updated synchronously. After the synchronous update is completed, the central server distributes the updated tumor risk subgraph with quantified confidence and its corresponding global multi-layer heterogeneous graph attention network, which has been optimized by this round of federated learning, as the shared knowledge base and initial model for the next round of iteration computing cycle to all participating nodes, so as to start a new round of data and knowledge co-evolution cycle that includes model optimization and knowledge discovery.

[0013] This application also provides a tumor risk intelligent early warning system based on natural language processing, specifically including: a medical text entity relationship joint extraction module, a multi-source semantic alignment and knowledge graph mapping module, a multi-source evidence fusion and risk quantification module, and a privacy-protected collaborative evolution module connected in sequence; wherein; Medical text entity relation joint extraction module: It is used to receive multi-source heterogeneous medical text data streams, encode them by calling a deep language model that has been pre-trained on medical domain corpus, and drive the multi-head attention mechanism and conditional random field joint decoding layer at the top of the model to perform synchronous decoding, and output an initial entity-relation pair set consisting of preliminary risk entities and their preliminary semantic relations. Multi-source semantic alignment and knowledge graph mapping module: Connects to the medical text entity relation joint extraction module. It receives the initial entity-relation pair set, accesses the dynamic tumor domain knowledge graph of the terminology library with built-in vector space and symbolic logic hybrid representation, and executes a voting decision algorithm based on context graph structure to perform hybrid similarity calculation and mapping decision, and outputs a set of semantically aligned standardized entity relation triples. Multi-source evidence fusion and risk quantification module: Connects the multi-source semantic alignment and knowledge graph mapping module. It is used to receive a standardized entity relationship triplet set, construct a heterogeneous graph based on its data source and relationship type attributes, and perform reasoning and information aggregation on the heterogeneous graph through the cross-source attention fusion layer in the multi-layer heterogeneous graph attention network, and output a tumor risk subgraph with quantified confidence. The privacy-preserving co-evolution module connects the multi-source evidence fusion and risk quantification modules, and is used to deploy a federated learning framework. The central server in the framework distributes the tumor risk subgraph and the corresponding network model to the participating nodes. Each node uses local private data to fine-tune the model and uploads parameter updates. The central server performs secure aggregation. At the same time, the module has a built-in graph incremental update triggering mechanism and stream processing pipeline. When the aggregated global model detects a high-frequency new risk pattern, the pipeline is triggered to perform re-extraction and alignment of the relevant new text data, and the confirmed new knowledge increment is updated to the dynamic tumor domain knowledge graph.

[0014] The beneficial effects of this invention are as follows: by deeply extracting and standardizing tumor risk entities and relationships from multi-source heterogeneous medical texts, a dynamically evolving domain knowledge graph is constructed. Furthermore, a heterogeneous graph attention network is used to fuse multi-source evidence to quantify risk confidence. Finally, under a federated learning framework, continuous collaborative optimization with privacy protection is achieved. The effect is to significantly improve the accuracy and credibility of risk association mining, ensure the interpretability of early warning results, and break through the limitations of data silos. Under the premise of strictly protecting medical data privacy, the system possesses the ability to autonomously discover new knowledge and continuously evolve, providing reliable technical support for accurate, real-time early warning and proactive discovery of tumor risks. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1 This paper presents an intelligent early warning method for tumor risk based on natural language processing, which specifically includes the following steps: Step S1: For the acquired multi-source heterogeneous medical text data stream, a deep language model that has been further pre-trained on medical domain corpus is used for feature extraction. Then, using the multi-head attention mechanism and conditional random field joint decoding layer integrated at the top of the model, named entity recognition and preliminary relation classification are performed simultaneously, and an initial entity-relation pair set consisting of preliminary risk entities and their preliminary semantic relations is output. Step S2: Input the initial entity-relation pair set into the dynamic oncology domain knowledge graph, which has a built-in terminology library of hybrid vector space and symbolic logic. Through a voting decision algorithm based on context graph structure, calculate the hybrid similarity between each preliminary risk entity and its context and the standard concept nodes in the dynamic oncology domain knowledge graph. Map the diverse expressions in the multi-source heterogeneous medical text data stream to the unified unambiguous standard concept nodes and output the semantically aligned standardized entity-relation triplet set. Step S3: Based on the data source and relation type attributes associated with each triple in the standardized entity relation triple set, construct a heterogeneous graph from the standardized entity relation triple set, and input the heterogeneous graph into a multi-layer heterogeneous graph attention network. This multi-layer heterogeneous graph attention network learns and aggregates evidence of the same risk association from different data sources through its cross-source attention fusion layer, quantifies the weight and confidence of each risk association, and outputs a tumor risk subgraph with quantified confidence. Step S4: Deploy the tumor risk subgraph with quantified confidence and the multi-layer heterogeneous graph attention network that generated it in a federated learning framework. Each participating node in the federated learning framework uses local private medical text data to fine-tune the multi-layer heterogeneous graph attention network and only uploads the model parameter updates to the central server for secure aggregation. At the same time, a graph incremental update triggering mechanism is set up. When the global model obtained after secure aggregation identifies a high-frequency new risk pattern higher than a preset threshold, the stream processing pipeline is triggered to re-execute steps S1 to S2 on the newly acquired multi-source heterogeneous medical text data related to the high-frequency new risk pattern. After confirmation, the new knowledge is updated to the dynamic tumor domain knowledge graph in an incremental manner.

[0020] In this embodiment, it is specifically necessary to explain the following steps in step S1: the feature extraction using a deep language model that has been further pre-trained with medical domain corpus. Each text sequence from the acquired multi-source heterogeneous medical text data stream is input into a pre-loaded medical domain corpus for further pre-training of a deep language model. The multi-source heterogeneous medical text data stream consists of text sequences from electronic health records, clinical notes, and medical literature sources. The pre-trained medical domain deep language model is a model based on a deep bidirectional Transformer architecture and has undergone domain-adaptive training on a large-scale corpus containing medical literature summaries and clinical records. For example, a BERT model or its variants (such as BioBERT or ClinicalBERT) pre-trained on PubMed summaries and the MIMIC-III clinical note corpus can be used to optimize its encoder. The parameters contain rich prior knowledge of medical semantics. This deep language model encodes each input text sequence to generate a deep context representation sequence. The text sequence consists of multiple sequentially arranged lexical units, and the deep context representation sequence consists of vectors corresponding to each lexical unit arranged in the same order. Each vector is a point in a d-dimensional real space, where d represents the feature dimension of the hidden layer of the deep language model. The generation of the deep context representation sequence ensures that the vector representation of each lexical unit in the sequence incorporates contextual information from the entire sentence, laying a solid foundation for accurate identification of entity boundaries and semantic relationships between entities. This overcomes the shortcomings of traditional bag-of-words models or static word vectors in handling polysemy (such as medical abbreviations). The specific process of simultaneously performing named entity recognition and preliminary relation classification using the multi-head attention mechanism integrated at the top layer of this model and the joint decoding layer of conditional random fields is as follows: First, the deep context representation sequence is input into a multi-head attention-enhanced feature refinement unit. This unit, through its multi-head attention mechanism, calculates the interaction weights between the first vector representation at any position in the deep context representation sequence and the second vector representations at all positions in the sequence. It then performs a weighted summation of all second vector representations based on these interaction weights, thereby generating a refined third vector representation for the first vector position. The multi-head attention mechanism allows the model to focus on semantic information at different positions in the sequence from multiple different representation subspaces in parallel, making it particularly suitable for capturing long-range modification relationships (such as descriptions of symptoms and body parts) or negation ranges in medical texts. Through this process, a corresponding third vector representation is generated for each position in the deep context representation sequence, forming a refined feature representation sequence that explicitly captures long-range semantic dependencies within the context of medical risk. Next, based on the refined feature representation sequence, on the one hand, a linear chain conditional random field is used to decode the globally optimal entity label sequence, identifying preliminary risk entities and their types in the text sequence. Each identified preliminary risk entity includes its corresponding text content fragment and entity type information. The linear chain conditional random field effectively alleviates the label inconsistency problem that may be caused by local predictions by modeling the transition probability between adjacent entity labels and considering the global optimality of the label sequence during decoding (such as mistakenly adding an "I-symptom" label after a "B-disease" label). Entity types can include, but are not limited to, categories related to tumor risk such as diseases, symptoms, examinations, treatments, genes, lifestyle habits, and family history. On the other hand, all identified entities are simultaneously processed. The candidate entity pairs, composed of the identified initial risk entities, are subjected to relation classification decoding. Relation classification decoding employs a relation-aware attention pooling operation, which performs the following process for each candidate entity pair: The candidate entity pair contains an initial risk entity as the subject and an initial risk entity as the object; the complete set of lexical positions covered by the initial risk entity as the subject and the initial risk entity as the object in the text sequence is obtained; for each position in the set of lexical positions covered by the initial risk entity as the subject, a first contribution weight is calculated, which indicates the contribution of the lexical at that position to the representation of the initial risk entity as the subject in the current candidate entity pair relation. The first and second contribution weights are calculated for each position in the set of lexical positions covered by the initial risk entity as the object. This second contribution weight represents the semantic contribution of the lexical at that position to the representation of the initial risk entity as the object in the current candidate entity pair relation. The first and second contribution weights are calculated by a dedicated attention network that takes as input the third vector representation of all lexical positions covered by the initial risk entities as the subject and object in the refined feature representation sequence, and are normalized within the lexical position sets of the initial risk entities as the subject and object, respectively. For example, for entity pairs... In the sequence “Smoking (subject) - leads to - lung cancer (object)”, among the lexical units covered by the subject entity “smoking”, the lexical unit “long-term” may receive a higher first contribution weight; among the lexical units covered by the object entity “lung cancer”, the lexical unit “familial” may receive a higher second contribution weight. This relation-aware pooling method, compared to the simple method of averaging the first and last word vectors of an entity, can more precisely capture which lexical units within an entity are more critical to expressing a specific relationship. Then, using all the calculated first contribution weights, the third vector representation corresponding to each lexical position covered by the preliminary risk entity as the subject is weighted, and all weighted results are summed to obtain a subject entity relation feature vector.Simultaneously, using all the calculated second contribution weights, the third vector representation corresponding to each word position covered by the preliminary risk entity as the object is weighted, and all weighted results are summed to obtain an object entity relationship feature vector. Finally, the subject entity relationship feature vector and the object entity relationship feature vector are concatenated or added to generate a dedicated relationship representation vector for the candidate entity pair. This dedicated relationship representation vector is input into the classifier to determine the preliminary semantic relationship category between the preliminary risk entity as the subject and the preliminary risk entity as the object. The classifier can be a fully connected neural network layer followed by a Softmax function, outputting the probability distribution of the relationship category. The relationship category can include "cause", "symptom", "inspection indicator", "treatment", "concurrency", etc. This "joint decoding" architecture, because entity recognition and relationship classification share the underlying refined feature representation and are optimized synchronously, allows relationship classification to make full use of the intermediate results and global context of entity recognition, avoiding the error propagation problem of relationship classification failure caused by entity recognition errors in traditional pipeline models, and improving the overall robustness and accuracy of extraction. By synchronously executing entity label sequence decoding and relation classification decoding in a linear chain conditional random field, an initial entity-relation pair set consisting of zero or one or more triples is output. Each triple contains a preliminary risk entity as the subject, a preliminary risk entity as the object, and a preliminary semantic relation category between them. The initial entity-relation pair set constitutes a preliminary structured representation of the original unstructured medical text and serves as the direct input for subsequent semantic standardization and knowledge fusion.

[0021] In this embodiment, it is particularly important to explain that in step S2, the dynamic tumor domain knowledge graph consists of a set of standard medical concept nodes and a set of standardized relationships between nodes. Each standard medical concept node is represented using a hybrid approach that combines vector space representation and symbolic logic representation. The vector space representation can be learned, for example, through graph embedding technology, and is used to calculate semantic similarity. The symbolic logic representation can include the definition of the concept's type, attributes, and relationships with other concepts, and is used for logical consistency verification. This hybrid representation combines the representational capabilities of deep learning with the interpretability and constraints of symbolic knowledge. The specific process for calculating the mixed similarity between each initial risk entity and its context and standard concept nodes in the dynamic tumor domain knowledge graph is as follows: The initial entity-relation pair set output in step S1 is used as the input in step S2. The initial entity-relation pair set consists of zero or one or more triples, where each triple contains a preliminary risk entity as the subject, a preliminary risk entity as the object, and a preliminary semantic relationship category between the subject entity and the object entity. Each preliminary risk entity contains its corresponding text content fragment and entity type information. First, for each preliminary risk entity in the initial entity-relationship pair set, multiple candidate standard concept nodes with semantically similar meanings are retrieved from the dynamic tumor domain knowledge graph; the specific retrieval process includes: Based on the text content fragments of the initial risk entity, a vector space representation is calculated; for example, this can be obtained by calculating the average of the deep context representation vectors of all words in the entity's text fragment. In the set of standard medical concept nodes in the dynamic oncology domain knowledge graph, a predetermined number of nodes are found that are closest to the vector space representation of the initial risk entity in terms of cosine distance, and these nodes are used as the initial candidate standard concept node set. The predetermined number K can be set as an integer between 3 and 10, aiming to achieve a balance between recall and computational efficiency. Next, the entity type information of the initial risk entity is compared with the category information, represented by symbolic logic, associated with each node in the initial candidate standard concept node set. If incompatibility is found, the corresponding node is removed from the initial candidate standard concept node set, thus obtaining a filtered and refined candidate standard concept node set. For example, if the type of the initial risk entity is "symptoms" and the logical category of the candidate node is "drugs," it is considered incompatible and filtered. This step effectively reduces the number of candidate nodes calculated subsequently and improves accuracy. Subsequently, for each candidate standard concept node in the refined candidate standard concept node set, semantic similarity and context graph structure similarity are calculated respectively. The process of calculating semantic similarity is as follows: calculate the cosine similarity value between the vector space representation of the preliminary risk entity and the vector space representation of the candidate standard concept node; The process of calculating the similarity of the context graph structure is as follows: From the initial entity-relation pair set, extract all other preliminary risk entities that are directly connected to the currently processed preliminary risk entity through preliminary semantic relations, forming the context entity set of the current preliminary risk entity; for each context entity in this context entity set, determine its corresponding mapping candidate node in the dynamic oncology domain knowledge graph; the process of determining the mapping candidate node can adopt the same or simplified retrieval and filtering logic as the current preliminary risk entity; obtain the set of direct neighbor nodes of the current candidate standard concept node in the dynamic oncology domain knowledge graph; then, for each context entity in the context entity set, calculate its mapping candidate node. The cosine similarity between the vector space representation of a point and the vector space representation of each node in the set of neighboring nodes of the current candidate standard concept node is calculated, and the maximum value is taken. Finally, the arithmetic mean of the maximum value corresponding to all context entities is taken, and the average value is the context graph structure similarity. This design cleverly utilizes the topological information of the graph structure. Its effect is that a correct mapping not only requires the entity itself to be similar to the concept node, but also requires that the nodes mapped by its "context neighbors" have high semantic similarity to the "topological neighbors" of the target concept node in the graph. This can effectively solve ambiguity problems such as "Apple (Fruit) Company" and "Apple (Company)" that can only be distinguished by context. Based on the size of the context entity set of the initial risk entity and the number of connection edges of the candidate standard concept node in the dynamic tumor domain knowledge graph, a fusion weight ranging from zero to one is dynamically determined. Specifically, the larger the context entity set (the richer the context) and the more connection edges the candidate node has (the more central and extensive the connection in the graph), the higher the weight assigned to the context graph structure similarity (i.e., the fusion weight) should be. A feasible calculation method is to perform linear or non-linear combination after normalization. Using this fusion weight, the semantic similarity and the context graph structure similarity are linearly weighted and summed. The semantic similarity is multiplied by one minus the difference of the fusion weight, and the context graph structure similarity is multiplied by the fusion weight. The two products are added together, and the result is the final mixed similarity between the initial risk entity and the candidate standard concept node. This dynamic weighting mechanism enables the algorithm to adaptively rely on the most reliable evidence source, improving the robustness of the mapping decision. The specific operation for outputting the semantically aligned standardized entity relation triplet set is as follows: For each preliminary risk entity in the initial entity-relationship pair set, select the candidate standard concept node with the highest mixed similarity with the preliminary risk entity from the refined candidate standard concept node set, and use it as the unambiguous standard concept node for the final mapping target of the preliminary risk entity; if the maximum mixed similarity is lower than a preset confidence threshold (e.g., 0.6), the preliminary risk entity can be considered unmapped, and the entity can be discarded or marked for manual review. Subsequently, for each triple in the initial entity-relation pair set consisting of a preliminary risk entity as the subject, a preliminary semantic relation category between the subject entity and the object entity, and a preliminary risk entity as the object, a transformation operation is performed: the preliminary risk entity as the subject is replaced with the unambiguous standard concept node mapped to it; the preliminary risk entity as the object is replaced with the unambiguous standard concept node mapped to it; and the preliminary semantic relation category between the subject entity and the object entity is mapped to a predefined standardized relation type in the dynamic oncology domain knowledge graph by querying a predefined relation mapping table or calculating the similarity between its description vector and the relation description vector in the standard relation set. For example, the preliminary relation "leads to" is mapped to... The standard relation "causes" is mapped to "co-occurs_with". This relation mapping ensures that relation representations from different sources are unified into a standardized relation system, thereby generating a new triple consisting of standardized subject concept nodes, standardized relation types, and standardized object concept nodes. Finally, the entire set of all new triples obtained through the transformation operation is defined and output as a set of semantically aligned standardized entity relation triples. All entities in this set of standardized entity relation triples have been mapped to unified knowledge graph concept nodes, and all relations have been standardized. This completely eliminates the heterogeneity and ambiguity in the representation of texts from different sources, providing a clean and consistent structured input for the subsequent steps of unified fusion and quantitative analysis of cross-source information.

[0022] In this embodiment, the specific process of constructing the standardized entity relation triplet set into a heterogeneous graph in step S3 is as follows: First, the data source associated with each triple in the semantically aligned standardized entity relation triple set is identified, expanding each element in the set into a quadruple consisting of a standardized subject concept node, a standardized relation type, a standardized object concept node, and a data source identifier. The data source identifier uniquely identifies the data origin, such as the electronic health record system code of different hospitals, the unique identifier of a document digital object (DOI) in a publicly available medical literature database, or a specific clinical trial registration number. Each standard medical concept node in the dynamic oncology knowledge graph is used as an entity node, and each identified unique data source identifier is used as a source node, together forming the node set of the heterogeneous graph. Next, the edge set of the heterogeneous graph is constructed based on the expanded semantically aligned standardized entity relation triple set. This edge set includes two types: relation edges and source edges. Each relation edge connects two entity nodes, and its type is determined by the standardized relation type in the quadruple, with each relation edge corresponding to one quadruple. Each source edge connects a source node and a relation edge, indicating that the risk association evidence corresponding to the relation edge originates from the data source represented by the source node. Then, initialized vector representations are assigned to the nodes and edges in the heterogeneous graph: Each entity node is assigned an initial vector representation, which is directly adopted from the vector representation of the entity node in the dynamic tumor domain knowledge graph. Each source node is assigned a trainable vector representation as an initial vector representation. The initial value of the initial vector is randomly generated or set according to the preset credibility level of the data source. For example, data sources from authoritative medical journals can be initialized with a vector of higher norm, while data sources from patient self-report forums can be initialized with a vector of lower norm, so as to incorporate prior credibility knowledge in the early stage of model training. Each relation edge is assigned an initial vector representation, which is composed of a preset embedding vector corresponding to its relation type and an encoded value representing the frequency of occurrence of the risk association corresponding to the relation edge in the semantically aligned standardized entity relation triplet set. The relation type embedding vector is a fixed-dimensional vector that is pre-trained or randomly initialized, and the frequency encoded value can be a logarithmic frequency value, which is used to provide the model with an initial signal of the amount of evidence for the association. The heterogeneous graph constructed and initialized in this way explicitly expresses the various types of risk associations that may exist between the same pair of entity nodes, as well as the multi-source knowledge fusion scenario in which the same risk association may be supported by evidence from multiple different data sources. This heterogeneous graph structure models the source information of evidence as a first-class citizen, providing a direct and structured input for subsequent multi-source evidence fusion and credibility measurement based on graph neural networks. The specific process for outputting a tumor risk subplot with quantified confidence levels is as follows: Construct a multi-layer heterogeneous graph attention network with L consecutive computational layers, where L is a positive integer; the value of L can be, for example, 2 to 6. By stacking multiple layers of the network, information can be propagated in multiple hops between nodes and edges in the heterogeneous graph, thereby capturing more complex dependencies. Each layer of the multi-layer heterogeneous graph attention network iteratively updates the vector representation of the relation edges in the heterogeneous graph in sequence. In the computation of each layer, the cross-relation edge information propagation operation is first performed: for each relation edge in the heterogeneous graph, the vector representations of the entity nodes at both ends in the previous layer and the vector representation of the relation edge in the previous layer are aggregated. Through a linear transformation and a nonlinear activation function, the updated vector representation of the relation edge in the current layer is generated. The nonlinear activation function can be, for example, the ReLU function. This operation enables the representation of the relation edge to incorporate the semantic information of the entities at both ends, realizing the transfer of knowledge on the graph structure. For each relation edge in a heterogeneous graph, the multilayer heterogeneous graph attention network calculates the attention weight of that relation edge for each source node connected to it via a source edge through its cross-source attention fusion layer. The calculation process for this attention weight is as follows: First, for a combination of a relation edge and a source node, the vector representation of the relation edge in the current layer, the vector representation of the source node in the previous layer, and a source-edge cooperative signal vector generated for this combination are concatenated to obtain a combined vector representation. Then, this combined vector representation is multiplied by a trainable attention parameter vector and passed through a LeakyReLU nonlinear activation function to obtain an unnormalized attention score. Finally, the unnormalized attention scores corresponding to all source nodes connected to the relation edge are exponentially normalized to obtain the normalized attention weight corresponding to each source node. This attention mechanism enables the model to dynamically evaluate the support strength of different data sources for a specific risk association. For example, for a gene-disease association, evidence from top journals may receive a higher attention weight than evidence from popular science websites. The process of generating the source-edge collaborative signal vector is as follows: First, obtain the vector representation of the source node in the previous layer of the current layer; then, calculate the average vector of the vector representations of all other relation edges connected to the source node through the source edge in the heterogeneous graph in the current layer; then, concatenate the vector representation of the source node with the average vector, and transform it through a small trainable neural network unit. The output is the source-edge collaborative signal vector. The small trainable neural network unit can be a single-layer or multi-layer fully connected network. This design enables the model to learn the "evidence preference pattern" of each data source. For example, one data source may tend to support the "gene-pathogenicity" type association, while another may focus more on the "environmental factor-related" type association. Then, the source aggregation update operation of the edge features is performed: using the calculated normalized attention weights, the vector representations from different source nodes are weighted and fused; specifically, for each relation edge, the vector representation of each source node connected to the edge in the previous layer of the current layer is multiplied by its corresponding normalized attention weight, and then multiplied by a trainable weight matrix corresponding to the current layer to obtain the weighted contribution of the source node; the weighted contributions of all connected source nodes are added together to obtain the aggregated vector representation; finally, the aggregated vector representation is added element-wise to the vector representation of the relation edge in the current layer, and the addition result is subjected to layer normalization to obtain the final vector representation of the relation edge in the current layer after source aggregation update; this operation realizes the adaptive fusion of multi-source evidence information, and the fused relation edge representation contains not only the semantic information of the association itself, but also encodes the comprehensive strength and quality of the multi-source evidence supporting the association; In the final layer of the multi-layer heterogeneous graph attention network, a confidence prediction function is applied to the final vector representation of each updated relation edge, mapping it to a scalar value between zero and one, which serves as the quantified confidence of the risk association represented by that relation edge. The confidence prediction function can be a single-layer neural network that takes the final layer relation edge vector representation as input, followed by a sigmoid activation function. This scalar value can be directly interpreted as the probability of the risk association being valid. This quantified confidence integrates the following three aspects: the semantic rationality of the risk association itself at the level of its corresponding entity nodes and relation types; the strength and quality of the evidence supporting the risk association from different data sources; and the contextual information of the risk association in the topological structure of the heterogeneous graph. For example, an association that is frequently reported by multiple high-quality sources and is closely connected to confirmed high-risk nodes in the graph will obtain a high confidence; conversely, an association that is only mentioned by a single low-quality source and is located at the edge of the graph will have a low confidence. This comprehensive quantification method greatly improves the interpretability and reliability of risk assessment. Finally, all relation edges in the heterogeneous graph, along with their two-end entity nodes, relation types, calculated quantitative confidence scores, and a list of data source identifiers connected to each relation edge via source edges, are extracted to form a tumor risk subgraph with quantitative confidence scores. This tumor risk subgraph can be directly used for downstream applications. For example, a confidence score threshold (such as 0.7) can be set to filter out high-confidence risk associations for early warning; or associations can be sorted according to confidence scores to help medical researchers focus on high-value research clues.

[0023] In this embodiment, it is specifically necessary to explain the process in step S4, where each participating node in the federated learning framework uses its local private medical text data to fine-tune the multi-layer heterogeneous graph attention network and only uploads the model parameter updates to the central server for secure aggregation. First, the federated learning framework is initialized and the model is deployed: a federated learning architecture is established, consisting of a central server and N participating nodes, where N is a positive integer. Each participating node holds a local, private, non-shared medical text dataset. The tumor risk subgraph with quantified confidence output from step S3 is used as the initial shared knowledge base, and the multilayer heterogeneous graph attention network that generates the tumor risk subgraph with quantified confidence is used as the initial global model. The central server distributes the initial shared knowledge base and the initial global model to all participating nodes, which then deploy them locally. The initial deployment ensures that all nodes have the same risk assessment benchmark and inference capabilities at the start of collaborative learning. Subsequently, the iterative computation cycle of federated learning begins: In each computation cycle, the central server distributes the current version of the global multilayer heterogeneous graph attention network and its parameters to all participating nodes; the parameter transmission of the model can use an encrypted communication channel to ensure transmission security; each participating node performs the following operations locally: First, using its own private medical text data, it independently executes steps S1 and S2 to generate a locally unique set of standardized entity relation triples; this process is completed entirely locally on the node, and the original medical text data does not need to leave the node, fundamentally protecting patient privacy and data security; then, based on this locally standardized entity relation triple set and the current version synchronized from the server with quantitative... A local heterogeneous graph is constructed by quantifying the confidence level of the tumor risk subgraph. During construction, for knowledge graph concept nodes involved in the local triples, existing node vector representations in the shared subgraph can be directly referenced. For nodes not included in the subgraph, default or random initialization methods can be used, and they can be gradually absorbed through federated learning and graph update mechanisms. Then, the received global model is used to process the local heterogeneous graph, and the difference between the confidence level predicted by the global model for the relation edges in the local heterogeneous graph and the supervision signal is calculated as the loss value. The supervision signal comes from the existing high-confidence associations or local data annotations in the current version of the tumor risk subgraph with quantified confidence level; for example, the confidence level in the shared subgraph can be higher than 0.The associations of 8 are used as positive samples, and a portion of low-confidence or absent entity pairs are randomly sampled as negative samples to construct a contrastive learning or binary classification training objective. If there is a small amount of labeled local data, it can be used first. Based on the calculated loss value, the received global model parameters are adjusted using a gradient descent algorithm to reduce the loss value, resulting in updated local model parameters. The gradient descent algorithm can use Adam or SGD optimizers, and a small number of iterations (e.g., 1-5 epochs) are performed locally. Finally, the participating node calculates the difference between the updated local model parameters and the received global model parameters to obtain the local model parameter update. This local model parameter update, or the result after adding noise perturbation to the local model parameter update, is encrypted, and only the encrypted result is uploaded to the central server. The server adds noise perturbation based on differential privacy technology to further obscure the contribution of individual node data to model updates, enhancing privacy protection. After collecting the encrypted local model parameter updates uploaded by all participating nodes, the central server first uses a secure aggregation algorithm to aggregate the encrypted data uploaded by each node, then decrypts the aggregation result to obtain the aggregated value of the model parameter updates of all nodes. This aggregated value is then applied to the current global model parameters, resulting in a new, optimized global multi-layer heterogeneous graph attention network. The secure aggregation algorithm can complete the aggregation without decrypting the update of individual nodes, ensuring that the server cannot spy on the update content of individual nodes. The aggregation strategy can adopt FedAvg, which is a weighted average of node update amounts, and the weights can be set according to the amount of local data or update quality of each node. The specific process for setting up a graph incremental update trigger mechanism is as follows: After the central server completes the global model secure aggregation for each round of federated learning, a newly optimized global multilayer heterogeneous graph attention network is used to perform inference on a retained validation dataset. This validation dataset consists of medical texts from multiple data sources, but personally identifiable information has been removed. The secure aggregation algorithm can complete the aggregation without decrypting the update amounts of individual nodes, ensuring that the server cannot snoop on the update content of individual nodes. The aggregation strategy can adopt FedAvg, which is a weighted average of node update amounts, and the weights can be set according to the local data volume or update quality of each node. For each candidate risk association triple generated by the model processing of the validation dataset, it is checked whether it already exists in the current version of the tumor risk subgraph with quantified confidence; if it does not exist, it is marked. For each labeled candidate new risk pattern, its frequency of occurrence in the current batch of the validation dataset is calculated. The confidence level of the triple prediction for each candidate new risk pattern by the global model is recorded. The average of these confidence levels is calculated, and the variance of these confidence levels in multiple consecutive inference batches is calculated. When a candidate new risk pattern exists whose frequency of occurrence exceeds a first preset frequency threshold, whose average confidence level exceeds a second preset confidence threshold, and whose variance of confidence level in multiple consecutive batches is lower than a third preset variance threshold, then the pattern is determined to be a high-frequency new risk pattern to be confirmed. The first preset frequency threshold can be set to occur once per minute or five times per batch; the second preset confidence threshold can be set to 0.85; and the third preset variance threshold can be set to 0.01. To ensure that the model's judgment of new patterns is stable and highly confident, rather than random noise, these thresholds can be adjusted according to the tolerance for new discovery sensitivity and false alarm rate in practical applications. The central server then broadcasts a lightweight trigger signal containing key feature information of the high-frequency new risk pattern to all participating nodes. The trigger signal can be a hash value, root word, or standardized concept code of the key entity name involved in the pattern, with a very small data volume, designed to guide nodes to perform local data filtering. After receiving the trigger signal, each participating node starts its local stream processing pipeline, which continuously monitors the newly generated medical text data stream on the node's local machine, and independently re-executes the entity and relation extraction in step S1 and the semantic alignment in step S2 only for the fragments in the text content that match the key features of the trigger signal, generating a new set of candidate standardized entity-relation triples related to the trigger pattern. This "precise triggering" mechanism avoids the huge overhead of full data processing, making the knowledge discovery process efficient and scalable. The stream processing pipeline can use Apache Flink or Spark. The system is implemented using frameworks such as Streaming. Each node uploads these new candidate standardized entity relation triples to the central server after de-identifying them. The de-identification process includes removing any text fragments, timestamps, locations, or other identifying information that may be associated with specific patients, retaining only the triple structure itself and its statistical frequency within the node. The central server aggregates candidates from all nodes, performs deduplication, frequency statistics, and consistency verification based on the frequency and confidence of the same candidates reported by multiple nodes. For example, a candidate triple is required to be reported by at least two different nodes, and its average confidence across nodes must exceed a verification threshold (e.g., 0.7) to pass verification. This utilizes the idea of ​​"majority consensus" in federated learning to improve the reliability of new knowledge discovery. The verified new knowledge triples are incrementally merged into the dynamic oncology domain knowledge graph by adding new nodes or new relation edges, and based on... The updated dynamic oncology knowledge graph synchronously updates a tumor risk subgraph with quantified confidence scores. The graph update operation is atomic, ensuring system availability during the update process. For newly added nodes, graph embedding technology is used to update their vector representations and integrate them into the existing graph topology. After the synchronous update is complete, the central server distributes the updated tumor risk subgraph with quantified confidence scores and its corresponding globally multi-layered heterogeneous graph attention network, optimized by this round of federated learning, as the shared knowledge base and initial model for the next iteration. This initiates a new round of data and knowledge co-evolution, encompassing model optimization and knowledge discovery. This closed-loop mechanism enables the system to continuously and automatically enhance model reasoning capabilities and knowledge base coverage without centralized data, forming a truly privacy-preserving, continuously learning intelligent system.

[0024] Example 2 This embodiment provides, for example Figure 2 The present invention discloses a tumor risk intelligent early warning system based on natural language processing, which specifically includes: a medical text entity relationship joint extraction module, a multi-source semantic alignment and knowledge graph mapping module, a multi-source evidence fusion and risk quantification module, and a privacy-preserving collaborative evolution module, which are connected in sequence. Medical text entity relation joint extraction module: It is used to receive multi-source heterogeneous medical text data streams, encode them by calling a deep language model that has been pre-trained on medical domain corpus, and drive the multi-head attention mechanism and conditional random field joint decoding layer at the top of the model to perform synchronous decoding, and output an initial entity-relation pair set consisting of preliminary risk entities and their preliminary semantic relations. Multi-source semantic alignment and knowledge graph mapping module: Connects to the medical text entity relation joint extraction module. It receives the initial entity-relation pair set, accesses the dynamic tumor domain knowledge graph of the terminology library with built-in vector space and symbolic logic hybrid representation, and executes a voting decision algorithm based on context graph structure to perform hybrid similarity calculation and mapping decision, and outputs a set of semantically aligned standardized entity relation triples. Multi-source evidence fusion and risk quantification module: Connects the multi-source semantic alignment and knowledge graph mapping module. It is used to receive a standardized entity relationship triplet set, construct a heterogeneous graph based on its data source and relationship type attributes, and perform reasoning and information aggregation on the heterogeneous graph through the cross-source attention fusion layer in the multi-layer heterogeneous graph attention network, and output a tumor risk subgraph with quantified confidence. The privacy-preserving co-evolution module connects the multi-source evidence fusion and risk quantification modules, and is used to deploy a federated learning framework. The central server in the framework distributes the tumor risk subgraph and the corresponding network model to the participating nodes. Each node uses local private data to fine-tune the model and uploads parameter updates. The central server performs secure aggregation. At the same time, the module has a built-in graph incremental update triggering mechanism and stream processing pipeline. When the aggregated global model detects a high-frequency new risk pattern, the pipeline is triggered to perform re-extraction and alignment of the relevant new text data, and the confirmed new knowledge increment is updated to the dynamic tumor domain knowledge graph.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A tumor risk intelligent early warning method based on natural language processing, characterized in that, Specifically, the following steps are included: Step S1: For the acquired multi-source heterogeneous medical text data stream, a deep language model that has been further pre-trained on medical domain corpus is used for feature extraction. Then, using the multi-head attention mechanism and conditional random field joint decoding layer integrated at the top of the model, named entity recognition and preliminary relation classification are performed simultaneously, and an initial entity-relation pair set consisting of preliminary risk entities and their preliminary semantic relations is output. Step S2: Input the initial entity-relation pair set into the dynamic oncology domain knowledge graph, which has a built-in terminology library of hybrid vector space and symbolic logic. Through a voting decision algorithm based on context graph structure, calculate the hybrid similarity between each preliminary risk entity and its context and the standard concept nodes in the dynamic oncology domain knowledge graph. Map the diverse expressions in the multi-source heterogeneous medical text data stream to the unified unambiguous standard concept nodes and output the semantically aligned standardized entity-relation triplet set. Step S3: Based on the data source and relation type attributes associated with each triple in the standardized entity relation triple set, construct a heterogeneous graph from the standardized entity relation triple set, and input the heterogeneous graph into a multi-layer heterogeneous graph attention network. This multi-layer heterogeneous graph attention network learns and aggregates evidence of the same risk association from different data sources through its cross-source attention fusion layer, quantifies the weight and confidence of each risk association, and outputs a tumor risk subgraph with quantified confidence. Step S4: Deploy the tumor risk subgraph with quantified confidence and the multi-layer heterogeneous graph attention network that generated it in a federated learning framework. Each participating node in the federated learning framework uses local private medical text data to fine-tune the multi-layer heterogeneous graph attention network and only uploads the model parameter updates to the central server for secure aggregation. At the same time, a graph incremental update triggering mechanism is set up. When the global model obtained after secure aggregation identifies a high-frequency new risk pattern higher than a preset threshold, the stream processing pipeline is triggered to re-execute steps S1 to S2 on the newly acquired multi-source heterogeneous medical text data related to the high-frequency new risk pattern. After confirmation, the new knowledge is updated to the dynamic tumor domain knowledge graph in an incremental manner.

2. The intelligent early warning method for tumor risk based on natural language processing according to claim 1, characterized in that: In step S1, the specific operation of feature extraction using a deep language model that has been further pre-trained with medical domain corpus is as follows: Each text sequence in the acquired multi-source heterogeneous medical text data stream is input into a pre-loaded medical domain corpus for further pre-training of a deep language model. The multi-source heterogeneous medical text data stream consists of text sequences from electronic health records, clinical notes, and medical literature. The deep language model encodes each input text sequence to generate a deep context representation sequence. The text sequence consists of multiple sequentially arranged lexical units, and the deep context representation sequence consists of vectors corresponding to each lexical unit arranged in the same order.

3. The intelligent early warning method for tumor risk based on natural language processing according to claim 2, characterized in that: The specific process of simultaneously performing named entity recognition and preliminary relation classification by utilizing the multi-head attention mechanism and conditional random field joint decoding layer integrated at the top layer of the model is as follows: First, the deep context representation sequence is input into a multi-head attention-enhanced feature refining unit to generate a refined feature representation sequence; Next, based on the refined feature representation sequence, on the one hand, a linear chain conditional random field is used to decode the entity label sequence to identify the preliminary risk entities and their types in the text sequence. Each identified preliminary risk entity contains its corresponding text content fragment and entity type information. On the other hand, relation classification decoding is performed simultaneously on the candidate entity pairs formed by all identified preliminary risk entities. Relation classification decoding employs a relation-aware attention pooling operation, which performs the following process for each candidate entity pair: the candidate entity pair contains a preliminary risk entity as the subject and a preliminary risk entity as the object; Obtain the set of all word positions covered by the initial risk entity as the subject and the initial risk entity as the object in the text sequence; For each position in the set of lexical positions covered by the initial risk entity as the subject, a first contribution weight is calculated. This first contribution weight represents the semantic contribution of the lexical at that position to the representation of the initial risk entity as the subject in the current candidate entity pair relation. For each position in the set of lexical positions covered by the initial risk entity as the object, a second contribution weight is calculated. This second contribution weight represents the semantic contribution of the lexical at that position to the representation of the initial risk entity as the object in the current candidate entity pair relation. The first and second contribution weights are calculated by a dedicated attention network that takes the third vector representation corresponding to all lexical positions covered by the initial risk entity as the subject and the initial risk entity as the object in the refined feature representation sequence as input. The second contribution weight is calculated for each position in the set of lexical positions covered by the initial risk entity as the subject and the initial risk entity as the object. The initial risk entity's lexical position set is normalized. Then, using all calculated first contribution weights, the third vector representation corresponding to each lexical position covered by the initial risk entity as the subject is weighted, and all weighted results are summed to obtain a subject entity relation feature vector. Simultaneously, using all calculated second contribution weights, the third vector representation corresponding to each lexical position covered by the initial risk entity as the object is weighted, and all weighted results are summed to obtain an object entity relation feature vector. Finally, the subject entity relation feature vector and the object entity relation feature vector are concatenated or added to generate a candidate entity pair's exclusive relation representation vector. This exclusive relation representation vector is input into a classifier to determine the initial semantic relation category between the initial risk entity as the subject and the initial risk entity as the object. By synchronously executing entity label sequence decoding and relation classification decoding in a linear chain conditional random field, an initial entity-relation pair set consisting of zero or one or more triples is output, where each triple contains a preliminary risk entity as the subject, a preliminary risk entity as the object, and a preliminary semantic relation category between the two.

4. The intelligent early warning method for tumor risk based on natural language processing according to claim 3, characterized in that: In step S2, the specific process for calculating the mixed similarity between each preliminary risk entity and its context and standard concept nodes in the dynamic tumor domain knowledge graph is as follows: First, for each preliminary risk entity in the initial entity-relationship pair set, multiple candidate standard concept nodes with similar semantics are retrieved from the dynamic tumor domain knowledge graph to form a refined candidate standard concept node set. Subsequently, for each candidate standard concept node in the refined candidate standard concept node set, semantic similarity and context graph structure similarity are calculated respectively. The process of calculating semantic similarity is as follows: calculate the cosine similarity value between the vector space representation of the preliminary risk entity and the vector space representation of the candidate standard concept node; The process of calculating the similarity of the context graph structure is as follows: from the initial entity-relation pair set, extract all other initial risk entities that are directly connected to the current initial risk entity through initial semantic relations, and form the context entity set of the current initial risk entity; For each context entity in the set of context entities, determine its corresponding mapping candidate node in the dynamic tumor domain knowledge graph; obtain the set of direct neighbor nodes of the current candidate standard concept node in the dynamic tumor domain knowledge graph. Then, for each context entity in the context entity set, calculate the cosine similarity between the vector space representation of its mapped candidate node and the vector space representation of each node in the neighbor node set of the current candidate standard concept node, and take the maximum value among them; finally, take the arithmetic mean of the maximum value corresponding to all context entities, and the average value is the context graph structure similarity. Based on the size of the context entity set of the preliminary risk entity and the number of connection edges of the candidate standard concept nodes in the dynamic tumor domain knowledge graph, a fusion weight is dynamically determined. Using this fusion weight, the semantic similarity and the context graph structural similarity are linearly weighted and summed to obtain the final mixed similarity.

5. The intelligent early warning method for tumor risk based on natural language processing according to claim 4, characterized in that: The specific operation for outputting the semantically aligned standardized entity relation triplet set is as follows: For each initial risk entity in the initial entity-relation pair set, select the candidate standard concept node with the highest mixed similarity from the refined candidate standard concept node set, and use it as the unambiguous standard concept node for its final mapping. For each triple in the initial entity-relation pair set, the initial risk entity as the subject and the initial risk entity as the object are replaced with their respective mapped unambiguous standard concept nodes, and the initial semantic relationship category between the subject entity and the object entity is mapped to a predefined standardized relationship type in the dynamic tumor domain knowledge graph by querying the preset relationship mapping table. This generates a new triplet consisting of a standardized subject concept node, a standardized relation type, and a standardized object concept node. Finally, the entire set of all newly generated triplets is defined and output as a set of semantically aligned standardized entity relation triplets.

6. The intelligent early warning method for tumor risk based on natural language processing according to claim 5, characterized in that: In step S3, the specific process of constructing the standardized entity relation triplet set into a heterogeneous graph is as follows: Identify the data source associated with each triple in the set of semantically aligned standardized entity relation triples, and expand each triple into a quadruple containing a data source identifier; Using standard medical concept nodes in the dynamic oncology domain knowledge graph as entity nodes and unique data source identifiers as source nodes, a set of nodes in a heterogeneous graph is formed. The edge set of the heterogeneous graph is constructed based on the expanded set of quadruples. This edge set includes relation edges connecting two entity nodes and source edges connecting the source node and the relation edge. Then, initialized vector representations are assigned to the nodes and edges in the heterogeneous graph: Each entity node is assigned an initial vector representation, which is directly adopted from the vector representation of the entity node in the dynamic tumor domain knowledge graph. Assign a trainable vector representation to each source node as the initial vector representation; Each relation edge is assigned an initial vector representation, which is composed of a preset embedding vector corresponding to its relation type and an encoded value representing the frequency of occurrence of the risk association corresponding to the relation edge in the semantically aligned standardized entity relation triplet set. This led to the construction and completion of the initial heterogeneous graph.

7. The intelligent early warning method for tumor risk based on natural language processing according to claim 6, characterized in that: The specific process of outputting the tumor risk submap with quantified confidence is as follows: Construct a multi-layer heterogeneous graph attention network containing L consecutive computational layers. Each layer of the multi-layer heterogeneous graph attention network iteratively updates the vector representation of the relation edges in the heterogeneous graph in sequence. In the computation of each layer, the cross-relation edge information propagation operation is first performed: for each relation edge in the heterogeneous graph, the vector representations of the entity nodes at both ends in the previous layer of the current layer and the vector representation of the relation edge in the previous layer of the current layer are aggregated, and the updated vector representation of the relation edge in the current layer is generated through a linear transformation and a nonlinear activation function. For each relation edge in a heterogeneous graph, the multilayer heterogeneous graph attention network calculates the attention weight of that relation edge for each source node connected to it via a source edge through its cross-source attention fusion layer. The calculation process for this attention weight is as follows: First, for the combination of a relation edge and a source node, the vector representation of the relation edge in the current layer, the vector representation of the source node in the previous layer, and a source-edge collaborative signal vector generated for the combination are concatenated to obtain a combined vector representation. Then, this combined vector representation is multiplied by a trainable attention parameter vector and passed through a LeakyReLU nonlinear activation function to obtain an unnormalized attention score. Finally, the unnormalized attention scores corresponding to all source nodes connected to the relation edge are exponentially normalized to obtain the normalized attention weight corresponding to each source node. The process of generating the source-edge collaborative signal vector is as follows: First, obtain the vector representation of the source node in the previous layer of the current layer; then, calculate the average vector of the vector representations of all other relation edges connected to the source node through the source edge in the heterogeneous graph in the current layer; then, concatenate the vector representation of the source node with the average vector, and transform it through a trainable neural network unit. The output is the source-edge collaborative signal vector. Then, the source aggregation update operation of the edge features is performed: using the calculated normalized attention weights, the vector representations from different source nodes are weighted and fused; specifically, for each relation edge, the vector representation of each source node connected to the edge in the previous layer of the current layer is multiplied by its corresponding normalized attention weight, and then multiplied by a trainable weight matrix corresponding to the current layer to obtain the weighted contribution of the source node; the weighted contributions of all connected source nodes are added together to obtain the aggregated vector representation; finally, the aggregated vector representation is added element-wise to the vector representation of the relation edge in the current layer, and the addition result is subjected to layer normalization to obtain the final vector representation of the relation edge in the current layer after source aggregation update; In the final layer of the multi-layer heterogeneous graph attention network, a confidence prediction function is applied to the final vector representation of each relation edge after the update, which is mapped to a scalar value as the quantified confidence of the risk association represented by the relation edge. Finally, all relation edges in the heterogeneous graph, along with their two-end entity nodes, relation types, calculated quantitative confidence scores, and a list of data source identifiers connected to each relation edge through the source edge, are extracted to form a tumor risk subgraph with quantitative confidence scores.

8. The intelligent early warning method for tumor risk based on natural language processing according to claim 7, characterized in that: In step S4, the specific process by which each participating node in the federated learning framework fine-tunes the multi-layer heterogeneous graph attention network using local private medical text data and only uploads model parameter updates to the central server for secure aggregation is as follows: The central server distributes the current version of the global multi-layer heterogeneous graph attention network to each participating node; Each participating node independently executes steps S1 and S2 locally using its private medical text data, generates a set of locally standardized entity relationship triples, and constructs a local heterogeneous graph based on this and the current version of the tumor risk subgraph. Each node uses the received global model to process the local heterogeneous graph, calculates the loss value and updates the model parameters to obtain the local model parameter update amount. The local model parameter update amount or the result after adding noise perturbation to the local model parameter update amount is encrypted, and only the encrypted result is uploaded to the central server. After collecting the encrypted local model parameter updates uploaded by all participating nodes, the central server first uses a secure aggregation algorithm to aggregate the encrypted data uploaded by each node, then decrypts the aggregation result to obtain the aggregated value of the model parameter updates of all nodes, and then applies this aggregated value to the current global model parameters, thereby obtaining a new round of optimized global multi-layer heterogeneous graph attention network.

9. The intelligent early warning method for tumor risk based on natural language processing according to claim 8, characterized in that: The specific process of setting up a map incremental update triggering mechanism is as follows: After the central server completes the global model security aggregation for each round of federated learning, the newly optimized global multilayer heterogeneous graph attention network is used to infer a retained validation dataset. For each candidate risk association triple generated by the model processing of the validation dataset, it is checked whether it already exists in the tumor risk subgraph with quantified confidence in the current version. If it does not exist, mark it as a candidate new risk pattern; For all labeled candidate new risk patterns, count their frequency of occurrence in the current batch of the validation dataset, record the confidence level of the triple prediction of the global model for each candidate new risk pattern, calculate the average of these confidence levels, and calculate the variance of these confidence levels in multiple consecutive inference batches. When a candidate new risk pattern exists, and its frequency of occurrence exceeds a first preset frequency threshold, its average confidence level exceeds a second preset confidence level threshold, and its variance of confidence level in multiple consecutive batches is lower than a third preset variance threshold, then the pattern is determined to be a high-frequency new risk pattern to be confirmed. The central server then broadcasts a lightweight trigger signal containing key feature information of the high-frequency new risk pattern to all participating nodes. After receiving the trigger signal, each participating node starts its local stream processing pipeline, which continuously monitors the newly generated medical text data stream locally and independently re-executes the entity and relation extraction in step S1 and the semantic alignment in step S2 for only the fragments in the text content that match the key features of the trigger signal, generating a new set of candidate standardized entity and relation triples related to the trigger pattern. Each node uploads these new candidate standardized entity relation triples to the central server after de-identifying them. The central server aggregates the candidates from all nodes, performs deduplication, frequency statistics, and consistency verification based on the frequency and confidence of the same candidates reported by multiple nodes. The new knowledge triples that pass the verification are incrementally merged into the dynamic oncology domain knowledge graph in the form of adding new nodes or new relation edges. Based on the updated dynamic oncology domain knowledge graph, the tumor risk subgraph with quantified confidence is updated synchronously. After the synchronization update is completed, the central server will distribute the updated tumor risk subgraph with quantified confidence and its corresponding global multi-layer heterogeneous graph attention network, which has been optimized by this round of federated learning, to all participating nodes as the shared knowledge base and initial model for the next round of iteration computing cycle, so as to start a new round of data and knowledge co-evolution cycle that includes model optimization and knowledge discovery.

10. A tumor risk intelligent early warning system based on natural language processing, applied to the tumor risk intelligent early warning method based on natural language processing as described in any one of claims 1-9, characterized in that: Specifically, it includes: The module consists of a medical text entity relationship joint extraction module, a multi-source semantic alignment and knowledge graph mapping module, a multi-source evidence fusion and risk quantification module, and a privacy-protected collaborative evolution module, which are connected sequentially. Medical text entity relation joint extraction module: It is used to receive multi-source heterogeneous medical text data streams, encode them by calling a deep language model that has been pre-trained on medical domain corpus, and drive the multi-head attention mechanism and conditional random field joint decoding layer at the top of the model to perform synchronous decoding, and output an initial entity-relation pair set consisting of preliminary risk entities and their preliminary semantic relations. Multi-source semantic alignment and knowledge graph mapping module: Connects to the medical text entity relation joint extraction module. It receives the initial entity-relation pair set, accesses the dynamic tumor domain knowledge graph of the terminology library with built-in vector space and symbolic logic hybrid representation, and executes a voting decision algorithm based on context graph structure to perform hybrid similarity calculation and mapping decision, and outputs a set of semantically aligned standardized entity relation triples. Multi-source evidence fusion and risk quantification module: Connects the multi-source semantic alignment and knowledge graph mapping module. It is used to receive a standardized entity relationship triplet set, construct a heterogeneous graph based on its data source and relationship type attributes, and perform reasoning and information aggregation on the heterogeneous graph through the cross-source attention fusion layer in the multi-layer heterogeneous graph attention network, and output a tumor risk subgraph with quantified confidence. The privacy-preserving co-evolution module connects the multi-source evidence fusion and risk quantification modules, and is used to deploy a federated learning framework. The central server in the framework distributes the tumor risk subgraph and the corresponding network model to the participating nodes. Each node uses local private data to fine-tune the model and uploads parameter updates. The central server performs secure aggregation. At the same time, the module has a built-in graph incremental update triggering mechanism and stream processing pipeline. When the aggregated global model detects a high-frequency new risk pattern, the pipeline is triggered to perform re-extraction and alignment of the relevant new text data, and the confirmed new knowledge increment is updated to the dynamic tumor domain knowledge graph.

Images