Chronic wound nursing scheme recommendation method and system based on clustering and mapping knowledge domain

Abstract

The invention discloses a chronic wound nursing scheme recommendation method and system based on clustering and a knowledge graph, and relates to the technical field of chronic wound nursing, chronic wound nursing associated data is collected to obtain chronic wound nursing multi-source data, and the chronic wound nursing multi-source data is preprocessed; and analyzing the preprocessed chronic wound nursing multi-source data, clustering the core demand problem and the nursing scheme data, and establishing a correlation database between the core demand problem and the nursing scheme. Through a technical path from multi-source data fusion to semantic cluster generation by layer clustering to cluster-level map construction to dynamic semantic matching, three major pain points of knowledge fragmentation, inaccurate demand matching and lack of recommendation basis in chronic wound nursing are solved. Compared with a traditional method, image volume accumulation captures multi-dimensional association, and semantic noise is reduced through cluster mapping; the evidence chain runs through the whole process of problems, schemes, wound types and literature sources.

Description

Technical Field

[0001] The present invention relates to the technical field of chronic wound care, and in particular to a method and system for recommending chronic wound care plans based on clustering and knowledge graphs. Background Art

[0002] Chronic wounds (such as diabetic foot ulcers, pressure injuries, and venous ulcers) are characterized by long course, high recurrence rates, and complex care. Treatment and care plans must be precisely tailored to the wound type, individual patient differences, and dynamic needs. Current chronic wound care plans rely primarily on healthcare professionals' experience, clinical guidelines, and fragmented literature, resulting in significant deficiencies: Existing nursing knowledge is scattered across medical literature, electronic medical records, online health community Q&A, and patient survey data. This lacks effective linkage between heterogeneous data sources (e.g., text, labels, and case studies). This makes it difficult for medical staff to quickly access complete, structured care plans tailored to specific wound types and patients' core needs. Manual sifting and piecing together information is often inefficient and prone to missing key links.

[0003] Traditional methods (such as keyword-based databases or simple question-and-answer systems) have limited understanding of the complex and diverse core needs raised by patients and their families (e.g., how to manage exudate from diabetic foot ulcers? How to determine the frequency of home dressing changes?). This inability to deeply parse the semantics of questions makes it difficult to accurately match best-practice care solutions hidden within large amounts of data, resulting in poorly relevant and practical recommendations.

[0004] Existing systems often rely on static rules or shallow database associations (such as simple question-answer mappings), failing to effectively explore deep semantic connections and clustering relationships between entities such as core needs, care plans, wound types, and evidence sources. The integration of new data (such as the latest clinical research and real-world nursing case studies) requires manual reconstruction of the association logic, making it difficult for the system to automatically update and dynamically optimize knowledge.

[0005] Therefore, there is an urgent need to design a chronic wound care plan recommendation method and related system hardware based on clustering and knowledge graphs to improve the care plan that can provide accurate and explainable recommendations based on users' chronic wound care needs. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a method and system for recommending chronic wound care solutions based on clustering and knowledge graph. The following technical solutions are adopted: The chronic wound care plan recommendation method based on clustering and knowledge graph includes the following steps: Step 1: collecting chronic wound care related data to obtain chronic wound care multi-source data, and preprocessing the chronic wound care multi-source data; Step 2: Analyze the pre-processed multi-source chronic wound care data to obtain the core needs and answers of chronic wound patients and their families, perform clustering on the core needs and care plan data, and establish a database of the correlation between the core needs and care plans. Step 3: Construct a knowledge graph for chronic, difficult-to-heal wound care. The knowledge graph associates wound type clusters, problem clusters, nursing plan clusters, and evidence source clusters. Step 4: Build a chronic wound care plan recommendation platform based on the knowledge graph of chronic difficult-to-heal wound care. The chronic wound care plan recommendation platform includes a question input module, a question semantic analysis module, a core demand question and care plan association analysis module, and a care plan output module. In step 5, the user's smart terminal raises core demand questions about chronic difficult-to-heal wounds to the chronic wound nursing plan recommendation platform through the question entry module. The question semantic analysis module analyzes the core demand questions. The core demand question nursing plan association analysis module outputs nursing plan data based on the semantic analysis results. The nursing plan output module parses the nursing plan data and feeds back the graphic and text results corresponding to the nursing plan results to the user's smart terminal.

[0007] By adopting the above technical solutions, multi-source heterogeneous data are deeply integrated: integrating four types of heterogeneous data sources: online Q&A, medical literature, electronic medical records, and patient questionnaires. Through cleaning and structuring within a unified preprocessing framework, the problems of incomplete knowledge coverage and sample bias caused by traditional solutions relying on a single data source are solved, thus building a comprehensive and reliable nursing knowledge base.

[0008] Hierarchical clustering (based on text similarity) and graph convolutional clustering networks (based on heterogeneous relationships) work together: Hierarchical clustering is first used to initially aggregate semantically similar questions and solutions to reduce noise; Then, through the graph convolutional network, the explicit associations (such as applicability) and implicit associations (such as semantic similarity) of problem-solution-wound-evidence are integrated to generate fine-grained, multi-dimensional associated clusters (such as the exudate treatment problem cluster associated with the dressing change solution cluster), breaking through the limitations of traditional single clustering in modeling complex relationships.

[0009] By upgrading discrete entities to cluster units to build a knowledge graph (e.g., problem cluster ID-P001 is associated with nursing solution cluster ID-C020), we achieve three major innovations: User questions do not need to strictly match existing question statements; they only need to be assigned to the correct question cluster to trigger recommendations; Cross-entity reasoning is achieved through inter-cluster relationship chains (e.g., problem cluster → solution cluster → wound type cluster → evidence cluster); Recommendations can be traced back to cluster labels (e.g., diabetic foot exudate management cluster) and associated evidence sources.

[0010] When the output module parses the nursing plan cluster, it synchronously returns the associated wound type cluster description (such as applicable to venous ulcers) and the evidence source cluster summary (such as from the 5th edition of the "International Guidelines for Wound Care"), forming a complete evidence chain of problem-solution-basis, enhancing the credibility of the results and clinical operability.

[0011] A closed loop is formed from data collection, cluster analysis, graph construction to recommended applications: The interrelated database is indexed by cluster ID, supporting efficient retrieval and incremental updates; The platform's modular design (semantic analysis → association matching → result generation) ensures process scalability and facilitates the integration of new data sources or optimization algorithms.

[0012] Optionally, in step 1, the specific method for collecting multi-source data on chronic wound care is: We used crawler tools to obtain question-and-answer instance data related to chronic wound care from online health networks, and obtained literature data related to chronic wound care from medical paper databases. We analyzed the literature data using literature analysis methods to obtain data on wound types, corresponding nursing plans, and evidence sources recorded in the literature. Collect real nursing case data from the hospital's electronic medical record system and analyze it to obtain wound type data and corresponding nursing plan data; A questionnaire survey was conducted among patients with chronic difficult-to-heal wounds and their families, and the questionnaire data were sorted to obtain core demand issue data.

[0013] By adopting the above technical solutions, four-dimensional data are collected simultaneously, including online questions and answers (patients' actual confusion), medical literature (authoritative solutions), electronic medical records (real nursing practices), and patient questionnaires (core needs and pain points), to ensure that the knowledge base has academic rigor, clinical practicality, and demand-orientedness, avoiding the bias of a single data source.

[0014] The scientific nature and evidence-based level of the medical literature assurance program (e.g., systematic reviews, clinical guidelines); Online Q&A and electronic medical records provide real-time updates on the latest nursing practices and difficult scenarios; The patient questionnaire directly addresses unmet home care needs, forming a dynamic knowledge loop.

[0015] Real nursing case data in electronic medical records provide operational details in specific situations (such as dressing change techniques and complication management), allowing recommended plans to break away from the theoretical framework and be clinically feasible.

[0016] Directly extract the core needs of patients and their families from questionnaires and Q&A (such as "How to reduce the pain of dressing changes?" and "How to deal with nighttime exudate?") to ensure that subsequent clustering and recommendations always focus on the user's real pain points rather than the subjective assumptions of medical staff.

[0017] The same nursing knowledge is cross-verified in different data sources (such as dressing selection theory in literature + actual application effects in medical records + home adaptation skills in questions and answers), automatically identifying high-confidence solutions and filtering invalid or controversial content.

[0018] Through multi-dimensional data complementarity, real-world scenario penetration, and demand source capture, a reliable knowledge base covering the entire "theory-practice-demand" chain has been built, providing high-quality, highly relevant, and strongly evidence-supported input guarantees for subsequent precise recommendations.

[0019] Optionally, in step 2, a hierarchical clustering framework and a graph convolutional clustering network are used to cluster the question and answer instance data, literature data, and real nursing case data, respectively. Based on the clustering results, an association relationship is established between the core demand question data and the nursing plan data, and a structured interrelated database is constructed. The interrelated database stores the wound type cluster ID, the question cluster ID, and the nursing plan cluster ID.

[0020] Alternatively, the specific method of clustering and clustering using the hierarchical clustering framework and graph convolution clustering network is: Step a: Clean the collected question-answering instance data, literature data, and real nursing case data; Step b: Extract key features. The key features of the question-answering instance data include the core requirement question text, nursing plan text, and wound type label; the key features of the literature data include wound type, corresponding nursing plan text, and evidence source. Step c, converting key feature text data into numerical feature vectors; Step d, calculate the similarity between all demand-question pairs to form a similarity matrix; Step e: construct a dendrogram. Initially, each demand problem is considered as a separate cluster. The two most similar separate clusters are merged iteratively until all demand problems are clustered into one demand problem cluster to obtain the hierarchical clustering result. Step f: Define different types of nodes, including core demand nodes, nursing plan nodes, wound type nodes, and evidence source nodes. Each node contains its key features. Define edges based on explicit associations and implicit similarities between key features to construct a heterogeneous graph. Explicit association edges clearly record that nursing plan A is applicable to wound type B, the solution to problem C is nursing plan D, and nursing plan E cites evidence F. Implicit similarity edges are established between core demand problem nodes based on hierarchical clustering results or direct calculation of semantic similarity, and semantic similarity edges are established between nursing plan nodes.

[0021] Optionally, in step d, cosine similarity, Jaccard similarity, or Euclidean distance is used to calculate the similarity between all demand-question pairs.

[0022] By adopting the above technical solution, we can aggregate demand issues and nursing plans with different expressions but similar semantics based on text similarity (such as cosine similarity), solve the problem of multiple synonymous expressions (such as "how to deal with exudate" and "how to care for wound exudate" are grouped together), form coarse-grained semantic units, and reduce data sparsity.

[0023] By fusing explicit associations (such as "Plan A is suitable for diabetic foot wounds") with implicit associations (such as the semantic similarity between different nursing plans) through heterogeneous graphs, complex cross-entity relationships are captured, and fine-grained, multi-dimensional associated clusters are generated (such as the "pain management problem cluster" associated with the "analgesic dressing plan cluster").

[0024] Explicit edges directly reuse structured knowledge (such as the "plan-wound type" relationship verified in the literature) to ensure that authoritative associations are not lost; Implicit edges are dynamically established based on semantic similarity (such as interconnecting nursing plan nodes that are not explicitly recorded but are highly similar), exploring potential knowledge associations and expanding the recommendation coverage.

[0025] Hierarchical clustering pre-filters outliers (such as irrelevant words in spoken questions), and graph convolutional networks smooth noise (such as incorrect labels in a single medical record) by aggregating neighbor nodes. These two mechanisms improve tolerance for low-quality real data.

[0026] Output a structured, interrelated database (with cluster ID as key) to achieve: Efficient retrieval: directly obtain related entities by cluster ID (e.g., problem cluster P001 → solution cluster C020 → wound type cluster W005); Newly added data only requires local re-clustering, without the need to reconstruct global associations.

[0027] The generated heterogeneous graph nodes (problems / solutions / wounds / evidence) and edge relationships are directly converted into entities and relationships in the knowledge graph, avoiding the complex annotation process of manually constructing the graph.

[0028] Optionally, in step 3, constructing a knowledge graph in the field of chronic, difficult-to-heal wound care includes the following steps: In step h, the constructed heterogeneous graph is input into the graph convolutional network. The graph convolutional network uses a multi-layer message passing mechanism to allow each node to aggregate the type and relationship information of its neighboring nodes. Step j: clustering is performed in the embedding space to establish cluster-level association relationships between problem clusters and nursing plan clusters, nursing plan clusters and wound type clusters, and nursing plan clusters and evidence source clusters; Step k: List of specific entity IDs contained in each cluster and the association relationship table between clusters to form a structured, mutually related database with cluster ID as the keyword.

[0029] By adopting the above technical solution, the graph convolutional network (GCN) enables nodes to fuse multiple neighbor information (such as wound types, evidence sources, and similar problem nodes aggregated and associated with nursing plan nodes) through multi-layer message passing, and mines implicit cross-entity associations that traditional methods cannot capture (such as the indirect connection between "hydrocolloid dressing plan" and "low pain demand").

[0030] Cluster nodes in the embedding space (such as K-means clustering semantic vectors) and upgrade discrete entities to semantic clusters (such as the "diabetic foot ulcer care plan cluster") to achieve: new problems / solutions only need to be matched to existing clusters to reuse associations; local data deviations do not affect the stability of cluster-level relationships.

[0031] Build an associative database with cluster ID as the key: Directly obtaining the association chain (e.g., problem cluster → solution cluster → evidence cluster) through cluster ID supports millisecond-level response to complex queries (e.g., "retrieve all solution clusters applicable to venous ulcers with a high level of evidence").

[0032] New nodes (e.g., the latest clinical protocol) are automatically integrated into neighboring clusters through GCN; The cluster relationship table is dynamically expanded to avoid global reconstruction.

[0033] Tracing the basis for recommendations through cluster association chains (e.g., wound type cluster labels associated with protocol clusters); View specific cases through the entity list within the cluster (such as the 10 medical records included in a certain plan cluster).

[0034] Optionally, in step j, the learned node embedding vector is used as a new feature representation, and the K-means or DBSCAN clustering algorithm is applied to nodes of the same type. The clustering results are the final refined problem clusters, nursing plan clusters, wound type nodes, and evidence source nodes for grouping or classification.

[0035] Optionally, in step 5, the method for the question semantic analysis module to analyze the core requirement question is: using a pre-trained Sentence-BERT model to convert the user question into a semantic vector.

[0036] Optionally, the method for the core demand problem nursing plan association analysis module to output nursing plan data based on the semantic analysis results is to calculate the cosine similarity of the center vector of the problem cluster in the semantic vector knowledge graph, match the most relevant problem cluster ID, and based on the matched problem cluster ID, dynamically retrieve the corresponding nursing plan cluster, wound type cluster and evidence source cluster through the knowledge graph relationship chain.

[0037] A chronic wound nursing plan recommendation system based on clustering and knowledge graph is used to implement a chronic wound nursing plan recommendation method based on clustering and knowledge graph. The system includes a network data acquisition module, a data analysis module, a chronic wound nursing plan recommendation platform and a user intelligent terminal. The network data acquisition module is used to collect question and answer instance data, literature data and real nursing case data. The data analysis module is communicated with the network data acquisition module. The chronic wound nursing plan recommendation platform includes an interrelated database, a question entry module, a question semantic analysis module, a core demand problem nursing plan association analysis module and a nursing plan output module; the interrelated database is communicated with the data analysis module, the question entry module is communicated with the user intelligent terminal, the question semantic analysis module is communicated with the question entry module, the core demand problem nursing plan association analysis module is communicated with the interrelated database and the question semantic analysis module, the nursing plan output module is communicated with the core demand problem nursing plan association analysis module, and communicates with the user intelligent terminal to feedback the graphic results corresponding to the nursing plan results.

[0038] In summary, the present invention includes at least one of the following beneficial technical effects: This invention provides a method and system for recommending chronic wound care solutions based on clustering and knowledge graphs. By integrating multi-source data, generating semantic clusters through hierarchical clustering, and then constructing cluster-level graphs for dynamic semantic matching, this method addresses the three major pain points in chronic wound care: fragmented knowledge, inaccurate demand matching, and a lack of evidence for recommendations. Compared to traditional methods, this method offers the following advantages: graph convolution clustering captures multi-dimensional associations, while cluster mapping reduces semantic noise; it achieves deep alignment between natural language requirements and structured knowledge; and it provides a chain of evidence that spans the entire process from problem to solution, wound type, and literature source. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of a method for recommending chronic wound care solutions based on clustering and knowledge graphs of the present invention; DETAILED DESCRIPTION

[0040] The present invention will be further described in detail below with reference to the accompanying drawings.

[0041] The embodiments of the present invention disclose a method and system for recommending chronic wound care plans based on clustering and knowledge graphs.

[0042] Reference Figure 1 Example 1: A method for recommending chronic wound care solutions based on clustering and knowledge graph, comprising the following steps: Step 1: collecting chronic wound care related data to obtain chronic wound care multi-source data, and preprocessing the chronic wound care multi-source data; Step 2: Analyze the pre-processed multi-source chronic wound care data to obtain the core needs and answers of chronic wound patients and their families, perform clustering on the core needs and care plan data, and establish a database of the correlation between the core needs and care plans. Step 3: Construct a knowledge graph for chronic, difficult-to-heal wound care. The knowledge graph associates wound type clusters, problem clusters, nursing plan clusters, and evidence source clusters. Step 4: Build a chronic wound care plan recommendation platform based on the knowledge graph of chronic difficult-to-heal wound care. The chronic wound care plan recommendation platform includes a question input module, a question semantic analysis module, a core demand question and care plan association analysis module, and a care plan output module. In step 5, the user's smart terminal raises core demand questions about chronic difficult-to-heal wounds to the chronic wound nursing plan recommendation platform through the question entry module. The question semantic analysis module analyzes the core demand questions. The core demand question nursing plan association analysis module outputs nursing plan data based on the semantic analysis results. The nursing plan output module parses the nursing plan data and feeds back the graphic and text results corresponding to the nursing plan results to the user's smart terminal.

[0043] Deep fusion of multi-source heterogeneous data: Integrate four types of heterogeneous data sources: online Q&A, medical literature, electronic medical records, and patient questionnaires. Through cleaning and structuring within a unified pre-processing framework, this solves the problems of incomplete knowledge coverage and sample bias caused by traditional solutions relying on a single data source, and builds a comprehensive and reliable nursing knowledge base.

[0044] Hierarchical clustering (based on text similarity) and graph convolutional clustering networks (based on heterogeneous relationships) work together: Hierarchical clustering is first used to initially aggregate semantically similar questions and solutions to reduce noise; Then, through the graph convolutional network, the explicit associations (such as applicability) and implicit associations (such as semantic similarity) of problem-solution-wound-evidence are integrated to generate fine-grained, multi-dimensional associated clusters (such as the exudate treatment problem cluster associated with the dressing change solution cluster), breaking through the limitations of traditional single clustering in modeling complex relationships.

[0045] By upgrading discrete entities to cluster units to build a knowledge graph (e.g., problem cluster ID-P001 is associated with nursing solution cluster ID-C020), we achieve three major innovations: User questions do not need to strictly match existing question statements; they only need to be assigned to the correct question cluster to trigger recommendations; Cross-entity reasoning is achieved through inter-cluster relationship chains (e.g., problem cluster → solution cluster → wound type cluster → evidence cluster); Recommendations can be traced back to cluster labels (e.g., diabetic foot exudate management cluster) and associated evidence sources.

[0046] When the output module parses the nursing plan cluster, it synchronously returns the associated wound type cluster description (such as applicable to venous ulcers) and the evidence source cluster summary (such as from the 5th edition of the "International Guidelines for Wound Care"), forming a complete evidence chain of problem-solution-basis, enhancing the credibility of the results and clinical operability.

[0047] A closed loop is formed from data collection, cluster analysis, graph construction to recommended applications: The interrelated database is indexed by cluster ID, supporting efficient retrieval and incremental updates; The platform's modular design (semantic analysis → association matching → result generation) ensures process scalability and facilitates the integration of new data sources or optimization algorithms.

[0048] In Example 2, in step 1, the specific method for collecting multi-source data on chronic wound care is: We used crawler tools to obtain question-and-answer instance data related to chronic wound care from online health networks, and obtained literature data related to chronic wound care from medical paper databases. We analyzed the literature data using literature analysis methods to obtain data on wound types, corresponding nursing plans, and evidence sources recorded in the literature. Collect real nursing case data from the hospital's electronic medical record system and analyze it to obtain wound type data and corresponding nursing plan data; A questionnaire survey was conducted among patients with chronic difficult-to-heal wounds and their families, and the questionnaire data were sorted to obtain core demand issue data.

[0049] Simultaneously collect four-dimensional data: online Q&A (patients' actual confusion), medical literature (authoritative solutions), electronic medical records (real nursing practices), and patient questionnaires (core needs and pain points), to ensure that the knowledge base is academically rigorous, clinically practical, and demand-oriented, avoiding bias from a single data source.

[0050] The scientific nature and evidence-based level of the medical literature assurance program (e.g., systematic reviews, clinical guidelines); Online Q&A and electronic medical records provide real-time updates on the latest nursing practices and difficult scenarios; The patient questionnaire directly addresses unmet home care needs, forming a dynamic knowledge loop.

[0051] Real nursing case data in electronic medical records provide operational details in specific situations (such as dressing change techniques and complication management), allowing recommended plans to break away from the theoretical framework and be clinically feasible.

[0052] Directly extract the core needs of patients and their families from questionnaires and Q&A (such as "How to reduce the pain of dressing changes?" and "How to deal with nighttime exudate?") to ensure that subsequent clustering and recommendations always focus on the user's real pain points rather than the subjective assumptions of medical staff.

[0053] The same nursing knowledge is cross-verified in different data sources (such as dressing selection theory in literature + actual application effects in medical records + home adaptation skills in questions and answers), automatically identifying high-confidence solutions and filtering invalid or controversial content.

[0054] Through multi-dimensional data complementarity, real-world scenario penetration, and demand source capture, a reliable knowledge base covering the entire "theory-practice-demand" chain has been built, providing high-quality, highly relevant, and strongly evidence-supported input guarantees for subsequent precise recommendations.

[0055] In Example 3, in step 2, a hierarchical clustering framework and a graph convolution clustering network are used to cluster and cluster the question and answer instance data, literature data, and real nursing case data, respectively. Based on the clustering and clustering results, an association relationship is established between the core demand question data and the nursing plan data, and a structured interrelated database is constructed. The interrelated database stores the wound type cluster ID, the question cluster ID, and the nursing plan cluster ID.

[0056] In Example 4, the specific method for clustering and clustering using the hierarchical clustering framework and graph convolution clustering network is: Step a: Clean the collected question-answering instance data, literature data, and real nursing case data; Step b: Extract key features. The key features of the question-answering instance data include the core requirement question text, nursing plan text, and wound type label; the key features of the literature data include wound type, corresponding nursing plan text, and evidence source. Step c, converting key feature text data into numerical feature vectors; Step d, calculate the similarity between all demand-question pairs to form a similarity matrix; Step e: construct a dendrogram. Initially, each demand problem is considered as a separate cluster. The two most similar separate clusters are merged iteratively until all demand problems are clustered into one demand problem cluster to obtain the hierarchical clustering result. Step f: Define different types of nodes, including core demand nodes, nursing plan nodes, wound type nodes, and evidence source nodes. Each node contains its key features. Define edges based on explicit associations and implicit similarities between key features to construct a heterogeneous graph. Explicit association edges clearly record that nursing plan A is applicable to wound type B, the solution to problem C is nursing plan D, and nursing plan E cites evidence F. Implicit similarity edges are established between core demand problem nodes based on hierarchical clustering results or direct calculation of semantic similarity, and semantic similarity edges are established between nursing plan nodes.

[0057] In Example 5, in step d, the similarity between all pairs of requirements and questions is calculated using cosine similarity, Jaccard similarity or Euclidean distance.

[0058] Based on text similarity (such as cosine similarity), we aggregate demand issues and nursing plans with different expressions but similar semantics to solve the problem of multiple synonymous expressions (for example, "how to deal with exudate" and "how to care for wound exudate" are grouped together), forming coarse-grained semantic units and reducing data sparsity.

[0059] By fusing explicit associations (such as "Plan A is suitable for diabetic foot wounds") with implicit associations (such as the semantic similarity between different nursing plans) through heterogeneous graphs, complex cross-entity relationships are captured, and fine-grained, multi-dimensional associated clusters are generated (such as the "pain management problem cluster" associated with the "analgesic dressing plan cluster").

[0060] Explicit edges directly reuse structured knowledge (such as the "plan-wound type" relationship verified in the literature) to ensure that authoritative associations are not lost; Implicit edges are dynamically established based on semantic similarity (such as interconnecting nursing plan nodes that are not explicitly recorded but are highly similar), exploring potential knowledge associations and expanding the recommendation coverage.

[0061] Hierarchical clustering pre-filters outliers (such as irrelevant words in spoken questions), and graph convolutional networks smooth noise (such as incorrect labels in a single medical record) by aggregating neighbor nodes. These two mechanisms improve tolerance for low-quality real data.

[0062] Output a structured, interrelated database (with cluster ID as key) to achieve: Efficient retrieval: directly obtain related entities by cluster ID (e.g., problem cluster P001 → solution cluster C020 → wound type cluster W005); Newly added data only requires local re-clustering, without the need to reconstruct global associations.

[0063] The generated heterogeneous graph nodes (problems / solutions / wounds / evidence) and edge relationships are directly converted into entities and relationships in the knowledge graph, avoiding the complex annotation process of manually constructing the graph.

[0064] In Example 6, in step 3, constructing a knowledge graph in the field of chronic and difficult-to-heal wound care includes the following steps: In step h, the constructed heterogeneous graph is input into the graph convolutional network. The graph convolutional network uses a multi-layer message passing mechanism to allow each node to aggregate the type and relationship information of its neighboring nodes. Step j: clustering is performed in the embedding space to establish cluster-level association relationships between problem clusters and nursing plan clusters, nursing plan clusters and wound type clusters, and nursing plan clusters and evidence source clusters; Step k: List of specific entity IDs contained in each cluster and the association relationship table between clusters to form a structured, mutually related database with cluster ID as the keyword.

[0065] Graph convolutional networks (GCNs) use multi-layer message passing to enable nodes to fuse multiple neighbor information (such as wound types, evidence sources, and similar problem nodes aggregated and associated with nursing plan nodes), and to mine implicit cross-entity associations that traditional methods cannot capture (such as the indirect connection between "hydrocolloid dressing plan" and "low pain demand").

[0066] Cluster nodes in the embedding space (such as K-means clustering semantic vectors) and upgrade discrete entities to semantic clusters (such as the "diabetic foot ulcer care plan cluster") to achieve: new problems / solutions only need to be matched to existing clusters to reuse associations; local data deviations do not affect the stability of cluster-level relationships.

[0067] Build an associative database with cluster ID as the key: Directly obtaining the association chain (e.g., problem cluster → solution cluster → evidence cluster) through cluster ID supports millisecond-level response to complex queries (e.g., "retrieve all solution clusters applicable to venous ulcers with a high level of evidence").

[0068] New nodes (e.g., the latest clinical protocol) are automatically integrated into neighboring clusters through GCN; The cluster relationship table is dynamically expanded to avoid global reconstruction.

[0069] Tracing the basis for recommendations through cluster association chains (e.g., wound type cluster labels associated with protocol clusters); View specific cases through the entity list within the cluster (such as the 10 medical records included in a certain plan cluster).

[0070] In Example 7, in step j, the learned node embedding vector is used as a new feature representation, and the K-means or DBSCAN clustering algorithm is applied to nodes of the same type. The clustering results are the final refined problem clusters, nursing plan clusters, wound type nodes, and evidence source nodes for grouping or classification.

[0071] In Example 8, in step 5, the method for the question semantic analysis module to analyze the core demand question is: using the pre-trained Sentence-BERT model to convert the user question into a semantic vector.

[0072] Example 9: The method by which the core demand problem nursing plan association analysis module outputs nursing plan data based on the semantic analysis results is to calculate the cosine similarity of the center vector of the problem cluster in the semantic vector knowledge graph, match the most relevant problem cluster ID, and based on the matched problem cluster ID, dynamically retrieve the corresponding nursing plan cluster, wound type cluster and evidence source cluster through the knowledge graph relationship chain.

[0073] Example 10, a chronic wound nursing plan recommendation system based on clustering and knowledge graph, is used to implement a chronic wound nursing plan recommendation method based on clustering and knowledge graph. The system includes a network data acquisition module, a data analysis module, a chronic wound nursing plan recommendation platform and a user smart terminal. The network data acquisition module is used to collect question and answer instance data, literature data and real nursing case data. The data analysis module is communicated with the network data acquisition module. The chronic wound nursing plan recommendation platform includes an interrelated database, a question entry module, a question semantic analysis module, a core demand problem nursing plan association analysis module and a nursing plan output module; the interrelated database is communicated with the data analysis module, the question entry module is communicated with the user smart terminal, the question semantic analysis module is communicated with the question entry module, the core demand problem nursing plan association analysis module is communicated with the interrelated database and the question semantic analysis module, the nursing plan output module is communicated with the core demand problem nursing plan association analysis module, and communicates with the user smart terminal to feedback graphic results corresponding to the nursing plan results.

[0074] The following specific embodiments are used to illustrate the implementation principle of the present invention: Mr. Zhang, a diabetic patient, developed a difficult-to-heal ulcer with exudate on his foot. His family asked the system a question via a smart terminal: "The wound on the sole of my foot keeps oozing yellow fluid. How can I treat it at home?" Technical solution implementation process: 1. Data Collection and Preprocessing (Step 1) The system integrates four types of data sources: Online Q&A: Crawl community discussions on diabetic foot exudate management (e.g., "How to clean thick exudate"); Medical literature: Extract dressing selection options for exudate management from the Guidelines for Care of Diabetic Foot Ulcers; Electronic medical records: Analysis of 50 similar cases in the hospital (e.g., use of alginate dressing to absorb exudate); Patient questionnaire: "home exudate treatment" is summarized as a high-frequency demand keyword.

[0075] Clean data: filter advertising text and standardize medical terminology (e.g., "yellow water" → "highly exudative exudate").

[0076] 2. Clustering and association construction (step 2); Hierarchical clustering: Similar expressions such as “yellow water flowing,” “continuous exudation,” and “wet wound” were grouped into the “high exudation treatment problem cluster” (ID: P088); The “alginate dressing,” “daily dressing change,” and “normal saline cleaning” protocols were grouped into the “exudate absorption protocol cluster” (ID: C102).

[0077] Graph convolution clustering: Explicit association: linking the “C102 regimen cluster” and the “diabetic foot ulcer wound type cluster” (ID: W003) (derived from literature and medical records); Implicit association: Based on semantic similarity, the "C102 regimen cluster" is associated with the "pain relief regimen cluster" (ID: C105) (because exudate is often accompanied by pain).

[0078] 3. Cluster-level knowledge graph generation (step 3); Graph convolutional network aggregation node information: The “C102 protocol cluster” node is fused with associated wound types (W003), evidence sources (guideline document ID: E205), and similar problem clusters (P088); Embedding Space Clustering: A cluster-level association chain P088→C102→W003→E205 is formed and stored in the association database.

[0079] 4. Dynamic recommendation and output (steps 4-5) Question semantic analysis: Sentence-BERT converts the user question "The wound on the sole of my foot keeps leaking yellow water" into a semantic vector; Calculate the similarity with the center vector of the problem cluster and match it to the P088 cluster (similarity 98%).

[0080] Correlation analysis and output: Retrieve the relationship chain P088→C102→W003→E205 through the knowledge graph; Output graphic results: Nursing plan: 1. Cleaning: Irrigate the wound with normal saline twice a day (from electronic medical record case CT2031); 2. Dressing: Use alginate dressing to absorb exudate and change it every 24 hours (from Chapter 5.2 of the Guidelines); 3. Pain management: Apply lidocaine spray before dressing change (associated cluster C105).

[0081] Applicable wound type: diabetic foot ulcer (W003 cluster) Evidence source: International Guidelines for Diabetic Foot Care (Cluster E205) The value of technical solutions 1. Accurate matching: The colloquial description "flowing yellow water" is accurately mapped to the professional question cluster P088; 2. Integration of multi-source evidence: The plan should include literature guidelines (authoritativeness), medical records and practices (operability), and related protocols (comprehensiveness); 3. Explainability: Clearly label the source of the protocol (guidelines / medical records) and the associated wound type to enhance user trust; 4. Improved efficiency: From asking a question to outputting a structured solution, the entire process takes less than 1 second, eliminating the need for manual review of multiple sources of information.

[0082] Through the full-process closed loop of "multi-source data → two-layer clustering → cluster-level graph → semantic matching", accurate, immediate and evidence-based nursing decision support is provided for complex chronic wound problems.

[0083] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for recommending chronic wound care plans based on clustering and knowledge graph, characterized in that: The following steps are involved: Step 1: collecting chronic wound care related data to obtain chronic wound care multi-source data, and preprocessing the chronic wound care multi-source data; Step 2: Analyze the pre-processed multi-source chronic wound care data to obtain the core needs and answers of chronic wound patients and their families, perform clustering on the core needs and care plan data, and establish a database of the correlation between the core needs and care plans. Step 3: Construct a knowledge graph for chronic, difficult-to-heal wound care. The knowledge graph associates wound type clusters, problem clusters, nursing plan clusters, and evidence source clusters. Step 4: Build a chronic wound care plan recommendation platform based on the knowledge graph of chronic difficult-to-heal wound care. The chronic wound care plan recommendation platform includes a question input module, a question semantic analysis module, a core demand question and care plan association analysis module, and a care plan output module. In step 5, the user's smart terminal raises core demand questions about chronic difficult-to-heal wounds to the chronic wound nursing plan recommendation platform through the question entry module. The question semantic analysis module analyzes the core demand questions. The core demand question nursing plan association analysis module outputs nursing plan data based on the semantic analysis results. The nursing plan output module parses the nursing plan data and feeds back the graphic and text results corresponding to the nursing plan results to the user's smart terminal.

2. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 1, characterized in that: In step 1, the specific method for collecting multi-source data on chronic wound care is: We used crawler tools to obtain question-and-answer instance data related to chronic wound care from online health networks, and obtained literature data related to chronic wound care from medical paper databases. We analyzed the literature data using literature analysis methods to obtain data on wound types, corresponding nursing plans, and evidence sources recorded in the literature. Collect real nursing case data from the hospital's electronic medical record system and analyze it to obtain wound type data and corresponding nursing plan data; A questionnaire survey was conducted among patients with chronic difficult-to-heal wounds and their families, and the questionnaire data were sorted to obtain core demand issue data.

3. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 2, characterized in that: In step 2, a hierarchical clustering framework and a graph convolutional clustering network are used to cluster the question-answering instance data, literature data, and real nursing case data, respectively. Based on the clustering results, the association relationship between the core demand question data and the nursing plan data is established, and a structured interrelated database is constructed. The interrelated database stores the wound type cluster ID, question cluster ID, and nursing plan cluster ID.

4. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 3, characterized in that: The specific method of clustering and clustering using the hierarchical clustering framework and graph convolution clustering network is: Step a: Clean the collected question-answering instance data, literature data, and real nursing case data; Step b: Extract key features. The key features of the question-answering instance data include the core requirement question text, nursing plan text, and wound type label; the key features of the literature data include wound type, corresponding nursing plan text, and evidence source. Step c, converting key feature text data into numerical feature vectors; Step d, calculate the similarity between all demand-question pairs to form a similarity matrix; Step e: construct a dendrogram. Initially, each demand problem is considered as a separate cluster. The two most similar separate clusters are merged iteratively until all demand problems are clustered into one demand problem cluster to obtain the hierarchical clustering result. Step f: define different types of nodes, including core demand problem nodes, nursing plan nodes, wound type nodes, and evidence source nodes. Each node contains its key features. Constructing heterogeneous graphs based on explicit associations and implicit similarities between key features to define edges; Explicit association edges clearly record that nursing plan A is applicable to wound type B, the solution to problem C is nursing plan D, and nursing plan E cites evidence F. Implicit similarity edges are established between core demand problem nodes based on hierarchical clustering results or direct calculation of semantic similarity, and semantic similarity edges are established between nursing plan nodes.

5. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 4, characterized in that: In step d, the similarity between all requirement-question pairs is calculated using cosine similarity, Jaccard similarity, or Euclidean distance.

6. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 4, characterized in that: In step 3, building a knowledge graph for chronic, difficult-to-heal wound care includes the following steps: In step h, the constructed heterogeneous graph is input into the graph convolutional network. The graph convolutional network uses a multi-layer message passing mechanism to allow each node to aggregate the type and relationship information of its neighboring nodes. Step j: clustering is performed in the embedding space to establish cluster-level association relationships between problem clusters and nursing plan clusters, nursing plan clusters and wound type clusters, and nursing plan clusters and evidence source clusters; Step k: List of specific entity IDs contained in each cluster and the association relationship table between clusters to form a structured, mutually related database with cluster ID as the keyword.

7. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 6, characterized in that: In step j, the learned node embedding vector is used as the new feature representation, and the K-means or DBSCAN clustering algorithm is applied to the nodes of the same type. The clustering results are the final refined problem clusters, nursing plan clusters, wound type nodes, and evidence source nodes for grouping or classification.

8. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 7, characterized in that: In step 5, the question semantic analysis module analyzes the core requirement questions by using the pre-trained Sentence-BERT model to convert user questions into semantic vectors.

9. The method for recommending chronic wound care solutions based on clustering and knowledge graph according to claim 8, characterized in that: The core demand problem nursing plan association analysis module outputs nursing plan data based on the semantic analysis results by calculating the cosine similarity of the center vector of the problem cluster in the semantic vector knowledge graph, matching the most relevant problem cluster ID, and based on the matched problem cluster ID, dynamically retrieving the corresponding nursing plan cluster, wound type cluster, and evidence source cluster through the knowledge graph relationship chain.

10. A chronic wound care plan recommendation system based on clustering and knowledge graph, characterized by: A method for recommending chronic wound nursing solutions based on clustering and knowledge graphs according to any one of claims 1 to 9, comprising a network data acquisition module, a data analysis module, a chronic wound nursing solution recommendation platform, and a user intelligent terminal, wherein the network data acquisition module is used to collect question-and-answer instance data, literature data, and real nursing case data, the data analysis module is communicatively connected to the network data acquisition module, and the chronic wound nursing solution recommendation platform comprises an interrelated database, a question entry module, a question semantic analysis module, a core demand question nursing solution association analysis module, and a nursing solution output module; The interrelated database is communicatively connected to the data analysis module, the problem entry module is communicatively connected to the user smart terminal, the problem semantic analysis module is communicatively connected to the problem entry module, the core demand problem nursing plan association analysis module is communicatively connected to the interrelated database and the problem semantic analysis module, the nursing plan output module is communicatively connected to the core demand problem nursing plan association analysis module, and communicates with the user smart terminal to feedback the graphic and text results corresponding to the nursing plan results.

Images