A nursing process information management system

Abstract

This invention discloses a nursing process information management system, relating to the field of nursing process data processing technology. It preprocesses raw medical business data to obtain static tree topology structured data, dynamic graph topology structured data, and corresponding dynamic graphs. A static tree topology database is constructed based on the static tree topology structured data. The dynamic graph topology structured data is then constructed and stored using spectral decomposition to obtain a dynamic graph topology database. The static tree topology database is encoded to obtain a static embedding matrix. The spectral domain representation in the dynamic graph topology database is then enhanced with spectral domain features and reconstructed to obtain a dynamic embedding matrix. The static and dynamic embedding matrices are aligned, and a fusion function is used for information balancing and fusion to generate a fused embedding matrix. This matrix is ​​then applied to obtain a task recommendation list, optimized nursing pathways, and nurse scheduling suggestions.

Description

Technical Field

[0001] This invention relates to the field of nursing process data processing technology, and in particular to a nursing process information management system. Background Technology

[0002] In recent years, with the continuous deepening of medical informatization, nursing process management, as a core link in the quality of medical services, directly affects the hospital's operational efficiency and patients' medical experience through its intelligent data processing level. The data generated by modern nursing operations exhibits significant heterogeneity and complexity, including not only static hierarchical relationships between departments, nurses' qualifications, and hospital areas, departments, and wards, but also a temporal interactive network composed of patients, nursing tasks, nurse nodes, and dynamic execution relationships. Effective analysis of this multi-source heterogeneous data can provide key decision support for personalized task recommendations, nursing pathway optimization, and human resource scheduling, thereby improving the accuracy of nursing work and the overall management level.

[0003] However, existing nursing data processing technologies generally have significant limitations. Traditional methods often treat static hierarchical structures and dynamic interactive networks separately, failing to effectively uncover the intrinsic connections between the two. This results in static information about nurses' qualifications not being able to work synergistically with real-time task execution. For dynamic graph topology data, conventional processing methods often remain at a shallow level of statistical aggregation, lacking in-depth mining of the structural features of the graph domain. This makes it difficult to capture the complex spectral patterns and long-range dependencies in the nursing task network. Furthermore, due to the heterogeneity of static and dynamic embeddings in the feature space, simple splicing or weighted fusion often leads to information redundancy or semantic conflicts, resulting in a fusion representation that cannot balance consistency and diversity. This, in turn, affects the accuracy and interpretability of downstream tasks, such as scheduling suggestions and path generation. Summary of the Invention

[0004] The technical problem addressed by this invention is that existing nursing data processing technologies generally have significant limitations. Traditional methods often treat static hierarchical structures and dynamic interactive networks separately, failing to effectively uncover the intrinsic connections between the two. This results in static information about nurses' qualifications not being able to dynamically synergize with real-time task execution. For dynamic graph topology data, conventional processing methods often remain at a shallow level of statistical aggregation, lacking in-depth mining of the structural features of the graph domain. This makes it difficult to capture the complex spectral patterns and long-range dependencies in the nursing task network. Furthermore, due to the heterogeneity of static and dynamic embeddings in the feature space, simple splicing or weighted fusion often causes information redundancy or semantic conflicts, resulting in a fusion representation that cannot balance consistency and diversity. This, in turn, affects the accuracy and interpretability of downstream tasks, such as scheduling suggestions and path generation.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a nursing process information management system, comprising a processing module, a construction module, an encoding module, an integration module, and an application module;

[0006] The processing module preprocesses the raw medical business data to obtain static tree topology structured data, dynamic graph topology structured data, and dynamic graphs corresponding to the dynamic graph topology structured data.

[0007] The module constructs a static tree topology database based on the static tree topology structured data, and constructs a spectral domain representation of the dynamic graph topology structured data through spectral decomposition and stores it to obtain the dynamic graph topology database.

[0008] The encoding module encodes the static tree topology database to obtain a static embedding matrix. After enhancing the spectral domain features of the spectral domain representation in the dynamic graph topology database, it reconstructs and encodes the dynamic embedding matrix.

[0009] The fusion module aligns the static and dynamic embedding matrices, performs information balancing fusion using a fusion function, and generates a fused embedding matrix based on the total loss function.

[0010] The application module processes the fusion embedding matrix to obtain a task recommendation list, optimized nursing pathways, and nurse scheduling suggestions.

[0011] As a preferred embodiment of the nursing process information management system described in this invention, the original medical business data includes static tree topology data and dynamic graph topology data.

[0012] Static tree topology data includes department attributes, nurse qualifications, and affiliation relationships;

[0013] The affiliation relationship refers to the affiliation between hospital areas, departments, wards, and nurses;

[0014] The dynamic graph topology data includes nodes of nursing tasks, execution relationships, patient information, nursing task descriptions, nurse execution records, and historical nursing path sequences for each patient;

[0015] The historical nursing pathway sequence consists of task nodes performed by the patient in chronological order;

[0016] The nodes of a nursing task include patient nodes, nursing task nodes, and nurse nodes, and the execution relationship is the execution relationship between the nodes of the nursing task.

[0017] Preprocessing of raw medical data includes constructing a tree structure based on static tree topology data;

[0018] The tree structure includes nodes such as hospital area, department, ward, and nurse, and the edges of the tree structure represent the membership relationship. The tree node features of each node in the tree structure are extracted to obtain static tree topology structured data.

[0019] The tree node features include department attributes and nurse qualifications;

[0020] A dynamic graph is constructed based on the dynamic graph topology data. The nodes of the dynamic graph include patient nodes, nursing task nodes, and nurse nodes, and the edges of the dynamic graph represent the execution relationships.

[0021] The execution relationships in the dynamic graph are aggregated according to a preset time window to obtain the graph Laplace matrix of the dynamic graph;

[0022] Extract the graph node features of each node in the dynamic graph to construct a graph node feature matrix;

[0023] Dynamic graph topology structured data is composed of the graph Laplacian matrix and the graph node feature matrix.

[0024] The graph node features include patient information, nursing task descriptions, and nurse execution records.

[0025] As a preferred embodiment of the nursing process information management system described in this invention, the static tree topology structured data is stored using a MySQL database. The table in the MySQL database used to store the tree structure includes node ID, parent node ID, node type, and level fields.

[0026] Here, the parent node ID represents the current node's direct parent node in the hierarchy;

[0027] The tree node features of the tree structure are stored as node feature vectors and indexed to obtain a static tree topology database.

[0028] As a preferred embodiment of the nursing process information management system described in this invention, the dynamic graph topology structured data is stored using a Neo4j database, and a time index is established to obtain a Neo4j database with a time index.

[0029] The spectral decomposition of the Laplacian matrix of the dynamic graph is then performed to obtain the orthogonal eigenvector basis and the diagonal eigenvalue matrix. The expression for the diagonal eigenvalue matrix is ​​as follows: ,and ;

[0030] in, The graph Laplacian matrix represents a dynamic graph. Describes the basis of orthogonal eigenvectors. This represents the transpose of the orthogonal eigenvector basis. Represents an eigenvalue diagonal matrix. Represents the identity matrix;

[0031] Mapping the graph node eigenvalue matrix to the spectral domain spanned by orthogonal eigenvector bases yields the eigencoefficient matrix, which is expressed as follows: ;

[0032] in, Represents the characteristic coefficient matrix, The node feature matrix representing a dynamic graph;

[0033] Will , , The spectral domain representation of the dynamic graph is stored in the graph database. During storage, the topological structured data of the dynamic graph is retained, resulting in a dynamic graph topology database that includes both the topological structured data and the spectral domain representation.

[0034] As a preferred embodiment of the nursing process information management system described in this invention, the spectral domain representation in the dynamic graph topology database is enhanced by masked topology enhancement (MTA) and cross-channel feature enhancement (CFA) in the spectral domain, and then reconstructed and encoded to obtain a dynamic embedding matrix.

[0035] For a static tree topology database, read the tree node features of each node to form a tree node feature matrix;

[0036] The tree structure and tree node feature matrix in the static tree topology database are processed by a bidirectional tree neural network. Specifically, the process includes: aggregating the child node information of each node from bottom to top to obtain the intermediate representation of each node; passing the context information of the parent node from top to bottom; fusing the context information of the parent node with the intermediate representation to generate the static embedding of each node; the child node is the node whose parent node ID points to the node; and the static embeddings of all nodes constitute the static embedding matrix.

[0037] As a preferred embodiment of the nursing process information management system described in this invention, feature enhancement processing is performed on the spectral domain representation in the dynamic graph topology database, wherein the spectral domain representation includes an eigenvalue diagonal matrix, an eigencoefficient matrix, and an orthogonal eigenvector basis.

[0038] Feature enhancement processing of the eigenvalue diagonal matrix includes processing the diagonal element sequence of the eigenvalue diagonal matrix using a mask topology enhancement (MTA). The MTA maps the diagonal element sequence into an embedding vector through a feature embedding layer. The embedding vector is then input into a multi-head self-attention encoder for self-attention processing to obtain an encoded representation of the eigenvalue sequence. Finally, the encoded representation of the eigenvalue sequence is input into a decoder for decoding processing to obtain the enhanced eigenvalue diagonal matrix.

[0039] Feature enhancement processing of the feature coefficient matrix includes processing the feature coefficient matrix using cross-channel feature enhancement CFA, wherein the cross-channel feature enhancement CFA performs a one-dimensional convolution on the feature coefficient matrix to obtain the enhanced feature coefficient matrix.

[0040] Multiplying the orthogonal eigenvector basis with the enhanced eigenvalue diagonal matrix and the transpose of the orthogonal eigenvector basis yields the graph Laplacian matrix of the enhanced dynamic graph.

[0041] Multiplying the orthogonal eigenvector basis with the enhanced eigencoefficient matrix yields the graph node feature matrix of the enhanced dynamic graph.

[0042] The graph Laplacian matrix and the graph node feature matrix of the enhanced dynamic graph are encoded using a graph neural network to obtain the dynamic embedding matrix.

[0043] As a preferred embodiment of the nursing process information management system described in this invention, the alignment process includes, for a static embedding matrix and a dynamic embedding matrix, wherein the static embedding matrix is ​​composed of nodes of a tree structure in a static tree topology database, and the dynamic embedding matrix is ​​composed of nodes of a dynamic graph in a dynamic graph topology database.

[0044] Obtain common nodes in the tree structure that exist simultaneously in the static tree topology database and the nodes in the dynamic graph in the dynamic graph topology database. Align the embeddings of the common nodes to obtain the aligned static embedding matrix and dynamic embedding matrix.

[0045] As a preferred embodiment of the nursing process information management system described in this invention, the aligned static embedding matrix and dynamic embedding matrix are subjected to information balancing fusion processing using a fusion function, and a fusion embedding matrix is ​​generated based on the total loss function. The fusion function is a multilayer perceptron.

[0046] The expression in the total loss function is,

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] in, Represents the total loss function. Indicates the downstream task loss item. Represents consistency constraints. Represents the diversity constraint. This represents a sufficiency constraint. Represents the fusion embedding matrix, Represents a static tree topology. This represents the dynamic graph topology, where F represents the Frobenius norm. This represents the aligned static embedding matrix. This represents the aligned dynamic embedding matrix. , and All represent preset balance coefficients. Nursing task-related tags.

[0052] As a preferred embodiment of the nursing process information management system described in this invention, the application processing includes: extracting the embedding vector corresponding to the current patient node as the current patient node embedding from the fusion embedding matrix, extracting the set of embedding vectors corresponding to the task nodes to be assigned as the task node embedding set, and calculating the similarity score between the patient and each task through cosine similarity.

[0053] Based on the similarity score, nurse qualifications are extracted from the aligned static embedding matrix. The nurse qualifications are then matched with the preset qualification requirements of each task. Tasks that do not meet the preset qualification requirements are filtered out to obtain a list of tasks that meet the qualification requirements.

[0054] The preset qualification requirements include the required nursing professional title and the required department affiliation;

[0055] Sort the list of tasks that meet the qualification requirements in descending order according to similarity score, and output the Top-K task recommendation list;

[0056] For the fusion embedding matrix, extract the patient node embedding and the corresponding patient's historical nursing path sequence, input the historical task sequence into the long short-term memory network to obtain the hidden state sequence of each time step, and use it as the sequence embedding;

[0057] The attention weight of each hidden state in the hidden state sequence relative to the patient node embedding is calculated using an attention mechanism, and the hidden state sequence is weighted and summed based on the attention weights to obtain the context vector.

[0058] The context vector and the patient node embedding are fused to obtain the fused features.

[0059] As a preferred embodiment of the nursing process information management system described in this invention, the fused features are input into the fully connected layer for prediction to obtain the prediction result of the next nursing task, and the prediction result is added to the historical nursing path sequence to generate an optimized nursing path.

[0060] Extract the embedding vectors corresponding to the nurse nodes from the fusion embedding matrix to form the nurse node embedding set;

[0061] Extract nurse qualifications and department attributes from the aligned static embedding matrix as nurse static attributes;

[0062] Extract nurse execution records from the aligned dynamic embedding matrix as nurse dynamic execution records;

[0063] Based on the nurse node embedding set, nurse static attributes, and nurse dynamic execution records, the matching degree between each nurse and each task is calculated, and a nurse scheduling suggestion is generated based on the matching degree.

[0064] The beneficial effects of this invention are as follows: Addressing the technical bottlenecks commonly found in existing nursing management systems, such as the separation of static structural information from dynamic execution data and insufficient mining of spectral domain features in complex graph data, this invention proposes an innovative dual-track information processing architecture. It constructs a dual-modal database of static tree topology and dynamic graph topology, and maps the dynamic graph to the spectral domain through spectral decomposition, solving the problem that traditional methods struggle to capture long-range dependencies and deep structural features in nursing networks. In the encoding stage, Masked Topology Enhancement (MTA) and Cross-Channel Feature Enhancement (CFA) are introduced to target and enhance spectral domain eigenvalues ​​and feature coefficient matrices, effectively overcoming the oversmoothing and feature annihilation defects that easily occur in existing graph neural networks when processing sparse medical data. By combining alignment processing and fusion functions with multi-constraint loss functions, the problem of difficult fusion of heterogeneous static and dynamic embeddings is solved. Through these improvements, not only is efficient fusion of static information such as nurse qualifications and affiliations with dynamic information such as task execution sequences and patient pathways achieved, but the accuracy and interpretability of task recommendation, path optimization, and scheduling suggestions are also effectively improved, significantly enhancing the intelligence level and scientific decision-making of nursing process management. Attached Figure Description

[0065] Figure 1 This is a basic flowchart of a nursing process information management system provided in one embodiment of the present invention. Detailed Implementation

[0066] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0067] Example, refer to Figure 1 As an embodiment of the present invention, a nursing process information management system is provided, including a processing module, a construction module, an encoding module, a fusion module and an application module;

[0068] The processing module preprocesses the raw medical business data to obtain static tree topology structured data, dynamic graph topology structured data, and dynamic graphs corresponding to the dynamic graph topology structured data.

[0069] The module constructs a static tree topology database based on the static tree topology structured data, and constructs a spectral domain representation of the dynamic graph topology structured data through spectral decomposition and stores it to obtain the dynamic graph topology database.

[0070] The encoding module encodes the static tree topology database to obtain a static embedding matrix. After enhancing the spectral domain features of the spectral domain representation in the dynamic graph topology database, it reconstructs and encodes the dynamic embedding matrix.

[0071] The fusion module aligns the static and dynamic embedding matrices, performs information balancing fusion using a fusion function, and generates a fused embedding matrix based on the total loss function.

[0072] The application module processes the fusion embedding matrix to obtain a task recommendation list, optimized nursing pathways, and nurse scheduling suggestions.

[0073] In one embodiment, a processing module extracts static hierarchical relationships and dynamic execution records from the original medical business data into static tree topology structured data and dynamic graph topology structured data, respectively, unifying heterogeneous information into a computable topological form. Subsequently, a construction module builds a static tree topology database to store hierarchical attributes based on the static tree topology, and simultaneously performs spectral decomposition on the dynamic graph topology to obtain a deep representation in the spectral domain, which is then stored in the dynamic graph topology database. An encoding module uses the static tree topology database to generate a static embedding matrix, and after applying masking topology enhancement and cross-channel feature enhancement to the dynamic graph spectral domain representation, reconstructs and encodes it to obtain a dynamic embedding matrix. This effectively overcomes the oversmoothing and other issues that traditional graph neural networks tend to encounter when processing sparse dynamic medical data. The feature annihilation defect is significantly improved, enhancing the expressive power of dynamic features. The fusion module solves the problem of effectively integrating heterogeneous static and dynamic features by aligning common nodes in static and dynamic embeddings and using a fusion function combined with a total loss function that considers consistency, diversity, and sufficiency constraints to achieve balanced information fusion. This generates a fusion embedding matrix that balances complementarity and robustness. The application module performs task recommendation, nursing pathway optimization, and nurse scheduling based on this fusion embedding matrix, outputting accurate and interpretable decision suggestions. Through a closed-loop process from data preprocessing, multimodal topology construction, spectral domain enhancement coding to alignment fusion, it achieves deep integration and intelligent application of nursing process information, significantly improving the accuracy and efficiency of nursing management.

[0074] Raw data for medical operations includes static tree topology data and dynamic graph topology data;

[0075] Static tree topology data includes department attributes, nurse qualifications, and affiliation relationships;

[0076] The affiliation relationship refers to the affiliation between hospital areas, departments, wards, and nurses;

[0077] The dynamic graph topology data includes nodes of nursing tasks, execution relationships, patient information, nursing task descriptions, nurse execution records, and historical nursing path sequences for each patient;

[0078] Historical nursing pathway sequences consist of task nodes performed by patients in chronological order;

[0079] The nodes of a nursing task include patient nodes, nursing task nodes, and nurse nodes, and the execution relationship is the execution relationship between the nodes of the nursing task.

[0080] Preprocessing of raw medical data includes constructing a tree structure based on static tree topology data;

[0081] The tree structure includes nodes such as hospital area, department, ward, and nurse. The edges of the tree structure represent membership relationships. The tree node features of each node in the tree structure are extracted to obtain static tree topology structured data.

[0082] Tree node features include department attributes and nurse qualifications;

[0083] A dynamic graph is constructed based on the dynamic graph topology data. The nodes of the dynamic graph include patient nodes, nursing task nodes, and nurse nodes, and the edges of the dynamic graph represent execution relationships.

[0084] The execution relationships in the dynamic graph are aggregated according to a preset time window to obtain the graph Laplace matrix of the dynamic graph;

[0085] Extract the graph node features of each node in the dynamic graph to construct a graph node feature matrix;

[0086] Dynamic graph topology structured data is composed of the graph Laplacian matrix and the graph node feature matrix.

[0087] Graph node features include patient information, nursing task descriptions, and nurse execution records.

[0088] In one embodiment, to transform the static hierarchical relationship into a structured form that facilitates subsequent coding, a tree structure is first constructed based on the static tree topology data in the original medical business data. The nodes of this tree structure include hospital areas, departments, wards, and nurses, with edges corresponding to the hierarchical relationships between hospital areas and departments, departments and wards, and wards and nurses. Tree node features are extracted from each node. Department attributes include department name, type (e.g., internal medicine, surgery), number of beds, and nursing unit category (e.g., general ward, intensive care unit). Nurse qualifications include professional title (e.g., head nurse, nurse), practice category (e.g., critical care nursing, general nursing), and years of service. The tree node features are mapped to fixed-dimensional vectors through an embedding layer. For example, by mapping professional titles to a 5-dimensional vector and normalizing years of service to a double-closed interval of 0-1, static tree topology structured data is obtained. This static tree topology structured data fully preserves the hierarchical relationships between hospital areas, departments, wards, and nurses, as well as the departmental attributes and nurse qualifications of each node. To capture the temporal patterns and structural characteristics of dynamic interactions in the nursing process, a dynamic graph is constructed based on the dynamic graph topology data. Nodes in the dynamic graph include patient nodes, nursing task nodes, and nurse nodes, and edges represent execution relationships, such as a nurse performing a nursing task or a patient receiving a nursing task. Each edge is accompanied by an execution timestamp. The execution relationships in the dynamic graph are aggregated according to a preset time window, which is set to 8 hours based on the nursing shift system, i.e., early shift, mid-shift ... For both afternoon and evening shifts, the frequency of execution relationships between nodes within each window is used as edge weights to construct a weighted adjacency matrix. Simultaneously, the dynamic characteristics of each node within that window are summarized. Specifically, patient information includes age, gender, and diagnosis code; nursing task descriptions include task type and required qualification codes; task types, such as intravenous infusion and vital sign measurement; and required qualification codes, which correspond to nurse professional titles (e.g., head nurse corresponds to code 3, nurse corresponds to code 2). Nurse execution records include the number of task executions, average time taken, and operation results (success marked as 1, exception marked as 0). This forms a graph node feature matrix, with each row corresponding to one node. The feature dimension is determined by embedding categorical features into a 16-dimensional vector and concatenating it with numerical features. The data is compressed into a unified 64-dimensional structure using a fully connected network layer. Classification features include gender and diagnosis codes from patient information, task type and required qualification codes from nursing task descriptions, and operation results from nurse execution records. Although age is a numerical feature, it is often binned by age group and treated as a classification feature. The number of task executions and average execution time are directly used as numerical features in the concatenation. Subsequently, a symmetric-normalized graph Laplacian matrix is ​​calculated for the weighted adjacency matrix within each time window. The graph Laplacian matrix effectively reflects the graph's topology and the strength of connections between nodes, and its spectral decomposition results are used for subsequent spectral domain feature enhancement. Finally, the graph Laplacian matrix of each time window and the graph node feature matrix are combined to form dynamic graph topology-structured data.Raw medical data was transformed into well-structured static tree topology and dynamic graph time-series data, laying a solid data foundation for subsequent spectral decomposition, feature enhancement, and multimodal fusion. Simultaneously, the introduction of an 8-hour time window enabled the dynamic graph topology structured data to align with clinical shift scheduling, effectively capturing the dynamic evolution of nursing processes across different shifts.

[0089] For static tree topology structured data, a MySQL database is used for storage. The tables in the MySQL database used to store the tree structure include node ID, parent node ID, node type, and level fields.

[0090] Here, the parent node ID represents the current node's direct parent node in the hierarchy;

[0091] The tree node features of the tree structure are stored as node feature vectors and indexed to obtain a static tree topology database.

[0092] The dynamic graph topology structured data is stored in the Neo4j database, and a time index is created to obtain a Neo4j database with a time index.

[0093] The spectral decomposition of the Laplacian matrix of the dynamic graph is then performed to obtain the orthogonal eigenvector basis and the diagonal eigenvalue matrix. The expression for the diagonal eigenvalue matrix is ​​as follows: ,and ;

[0094] in, The graph Laplacian matrix represents a dynamic graph. Describes the basis of orthogonal eigenvectors. This represents the transpose of the orthogonal eigenvector basis. Represents an eigenvalue diagonal matrix. Represents the identity matrix;

[0095] Mapping the graph node eigenvalue matrix to the spectral domain spanned by orthogonal eigenvector bases yields the eigencoefficient matrix, which is expressed as follows: ;

[0096] in, Represents the characteristic coefficient matrix, The node feature matrix representing a dynamic graph;

[0097] Will , , The spectral domain representation of the dynamic graph is stored in the graph database. During storage, the topological structured data of the dynamic graph is retained, resulting in a dynamic graph topology database that includes both the topological structured data and the spectral domain representation.

[0098] In one embodiment, to store and query the static hierarchical structure, the construction module first stores the static tree topology structured data using a MySQL relational database. A table is designed in MySQL to store the tree structure, including node ID (primary key), parent node ID, node type (e.g., hospital area, department, ward, nurse), and hierarchy field (e.g., 1 represents hospital area, 2 represents department, 3 represents ward, 4 represents nurse). The parent node ID represents the direct parent node of the current node in the hierarchy; for example, the node ID of the ward where a nurse in internal medicine works is used as its parent node ID. Subsequently, the tree node features of each node in the tree structure are stored as node feature vectors in the feature field of this table, and a B+ tree index is built for the node ID and parent node ID. To accelerate hierarchical traversal queries, a static tree topology database is obtained. This database fully preserves hierarchical relationships and supports fast retrieval, providing a directly callable structured data foundation for subsequent static embedding encoding. For flexible storage and spectral domain analysis of dynamic interactive networks, the module further stores the dynamic graph topology structured data using the Neo4j graph database. Each 8-hour time window of the dynamic graph is modeled as a subgraph. Each node in the graph (e.g., patient, nursing task, nurse) corresponds to a graph database node, and each edge (execution relationship) corresponds to a relationship. The execution frequency is recorded in the relationship attributes as the edge weight. A time index is also created for each subgraph, with the index key being the start timestamp of the time window. For example, 2025-03-04 08:00:00 represents the morning shift, resulting in a Neo4j database with a time index. To extract deep features of the dynamic graph in the spectral domain for subsequent targeted feature enhancement, the module further performs spectral decomposition on the graph Laplacian matrix of each time window. Specifically, it performs spectral decomposition on the symmetric normalized graph Laplacian matrix of each 8-hour time window. (The dimension is N×N, where N is the total number of nodes in the window) Perform eigenvalue decomposition to obtain the orthogonal eigenvector basis. (N×N matrix) and diagonal eigenvalue matrix (N×N diagonal matrix) satisfying ,and , where the eigenvalue diagonal matrix diagonal elements Arranged from smallest to largest, this represents the smoothness of the dynamic graph across different frequency components; subsequently, the feature matrices of the graph nodes within the same time window are... (An N×F matrix, where F is the feature dimension of the nodes, F=64 here) is mapped to the spectral domain spanned by orthogonal eigenvector bases to obtain the eigencoefficient matrix. = (N×F matrix) Each row corresponds to a characteristic coefficient in the spectral domain frequency component, reflecting the energy distribution of the signal in each frequency component; finally, the 8-hour time windows are... , , Along with the dynamic graph topology structured data (node ​​attributes, edge relationships, timestamps), it is stored in the Neo4j database. By associating the corresponding windows through time indexing, a dynamic graph topology database is obtained, which includes the dynamic graph topology structured data and the spectral domain representation of the dynamic graph topology structured data. The construction module not only realizes efficient storage and time-series tracing of complex dynamic networks, but more importantly, it transforms the graph data from the spatial domain to the spectral domain. This allows the subsequent encoding module to directly perform targeted enhancement on the feature values ​​(representing the frequency characteristics of the graph structure) and feature coefficients (representing the spectral distribution of the graph signal). This provides spectral domain operation space to overcome the oversmoothing problem that traditional graph neural networks are prone to in sparse medical data. At the same time, the coexistence of spectral decomposition results and original graph data also lays the data foundation for multi-angle feature fusion.

[0099] For the spectral domain representation in the dynamic graph topology database, after spectral domain feature enhancement using masked topology enhancement (MTA) and cross-channel feature enhancement (CFA), reconstruction encoding is performed to obtain the dynamic embedding matrix.

[0100] For a static tree topology database, read the tree node features of each node to form a tree node feature matrix;

[0101] The tree structure and tree node feature matrix in the static tree topology database are processed by a bidirectional tree neural network. Specifically, the process includes: aggregating the child node information of each node from bottom to top to obtain the intermediate representation of each node; passing the context information of the parent node from top to bottom; fusing the context information of the parent node with the intermediate representation to generate the static embedding of each node; the child node is the node whose parent node ID points to the node; and the static embeddings of all nodes constitute the static embedding matrix.

[0102] Feature enhancement processing is performed on the spectral domain representations in the dynamic graph topology database. The spectral domain representations include eigenvalue diagonal matrices, eigencoefficient matrices, and orthogonal eigenvector bases.

[0103] Feature enhancement processing of the eigenvalue diagonal matrix includes processing the diagonal element sequence of the eigenvalue diagonal matrix using Mask Topology Enhancement (MTA). The MTA maps the diagonal element sequence into an embedding vector through a feature embedding layer. The embedding vector is then input into a multi-head self-attention encoder for self-attention processing to obtain the encoded representation of the eigenvalue sequence. Finally, the encoded representation of the eigenvalue sequence is input into a decoder for decoding to obtain the enhanced eigenvalue diagonal matrix.

[0104] Feature enhancement processing of the feature coefficient matrix includes processing the feature coefficient matrix using cross-channel feature enhancement CFA, and performing one-dimensional convolution on the feature coefficient matrix using cross-channel feature enhancement CFA to obtain the enhanced feature coefficient matrix.

[0105] Multiplying the orthogonal eigenvector basis with the enhanced eigenvalue diagonal matrix and the transpose of the orthogonal eigenvector basis yields the graph Laplacian matrix of the enhanced dynamic graph.

[0106] Multiplying the orthogonal eigenvector basis with the enhanced eigencoefficient matrix yields the graph node feature matrix of the enhanced dynamic graph.

[0107] The graph Laplacian matrix and the graph node feature matrix of the enhanced dynamic graph are encoded using a graph neural network to obtain the dynamic embedding matrix.

[0108] In one embodiment, to extract node representations including hierarchical structures from a static tree topology database, the encoding module first reads the tree node features of each node in the static tree topology database to form a tree node feature matrix. Each row of the matrix corresponds to the feature vector of a node, and the feature dimension is the fixed dimension after the embedding layer mapping mentioned earlier, such as 64 dimensions. Subsequently, a bidirectional tree neural network is used to encode the tree structure. Bottom-up aggregation enables the representation of each node to incorporate the information of its subordinate child nodes, thereby capturing the inductive characteristics in the organizational structure. Top-down propagation enables each node to perceive the global context of its superior, thereby enhancing the semantic consistency of the representation. The specific steps are as follows: First, the tree structure is traversed from bottom to top. For each non-leaf node, the intermediate representations of all its child nodes are aggregated through mean pooling. Initially, the tree node features of the child nodes are used, and then combined with the tree node features of the node itself. After feature concatenation, the intermediate representation of the node is obtained through a fully connected network (with ReLU activation function). This process is repeated until the root node, obtaining the intermediate representations of all nodes. Then, a top-down traversal is performed, starting from the root node. The context representation of the parent node (initially the root node's intermediate representation) is transformed through a fully connected network and then element-wise added to the intermediate representation of the current node to generate the current node's context representation as its static embedding. The static embeddings of all nodes constitute a static embedding matrix. The static embedding matrix retains the tree node features of each node and incorporates the hierarchical relationships between hospital areas, departments, wards, and nurses, providing a structural alignment basis for subsequent fusion with dynamic information. To extract node representations including spectral domain information from the dynamic graph topology database, the encoding module further enhances and reconstructs the spectral domain representation through feature enhancement and reconstruction encoding, by modifying the eigenvalue diagonal matrix. The diagonal elements (i.e., eigenvalues ​​representing the smoothness of the dynamic graph across different frequency components) are masked for topology enhancement (MTA), which can adaptively recalibrate the importance of different frequency components, highlighting key structural patterns for subsequent task recommendation, path optimization, and scheduling suggestions; by applying the feature coefficient matrix... (i.e., graph node feature matrix) The projection in the spectral domain reflects the energy distribution of the graph signal across its frequency components. Cross-channel feature enhancement (CFA) can strengthen the feature interaction between different frequency channels and improve the discriminative power of the graph signal. The specific steps are as follows: First, the diagonal matrix of the eigenvalues ​​for each time window in the dynamic graph topology database is processed. The sequence of diagonal elements of an N×N diagonal matrix (Sorted in ascending order) The process involves masked topology enhancement (MTA), mapping the diagonal element sequence to an embedding vector sequence through a learnable feature embedding layer with an embedding dimension of 64. This embedding vector sequence is then fed into a Transformer encoder consisting of two 8-layer multi-head self-attention encoders for self-attention processing, yielding an encoded representation of the feature value sequence. This encoded representation is then input into a decoder consisting of two fully connected layers for reconstruction, outputting the enhanced feature value sequence, which is restored to its diagonal matrix form. This process captures the long-range dependencies between different frequency components through a self-attention mechanism, enabling the enhanced eigenvalues ​​to more accurately reflect the key frequency bands of the graph structure. Subsequently, the eigenvalue coefficient matrix... (N×64 matrix) Employing cross-channel feature enhancement CFA, for Perform a one-dimensional convolution with a kernel size of 3, padding of 1, 64 input channels, 64 output channels, and ReLU activation function to obtain the enhanced feature coefficient matrix. This convolution operation achieves information mixing between adjacent frequency channels, enhancing the local correlation of feature coefficients, and then the orthogonal eigenvector basis is... With the enhanced eigenvalue diagonal matrix and Multiplying by the transpose of the matrix yields the enhanced dynamic graph Laplacian matrix. ;Will With the enhanced eigenvalue matrix Multiplying and reconstructing yields an enhanced dynamic graph node feature matrix. Finally, a graph neural network is used to encode the enhanced graph data: and The input is fed into a 2-layer graph convolutional network (GCN), with 128 hidden units and 64 output units per layer, to obtain the dynamic embedding of each node. The dynamic embeddings of all nodes constitute a dynamic embedding matrix, which is then processed by an encoding module. The encoding module not only extracts hierarchical static embeddings from the static tree structure, but also enhances key structural and signal features from the dynamic graph domain through MTA and CFA, effectively overcoming the information loss caused by the sparsity of the original graph data and providing a higher quality and more complementary embedding representation for the subsequent fusion module.

[0109] The alignment process includes static embedding matrices and dynamic embedding matrices. Static embedding matrices are composed of nodes of a tree structure in a static tree topology database, while dynamic embedding matrices are composed of nodes of a dynamic graph in a dynamic graph topology database.

[0110] Obtain the common nodes in the tree structure that exist simultaneously in the static tree topology database and the nodes in the dynamic graph topology database. Align the embeddings of the common nodes to obtain the aligned static embedding matrix and dynamic embedding matrix.

[0111] The aligned static and dynamic embedding matrices are subjected to information balancing fusion processing using a fusion function, and a fused embedding matrix is ​​generated based on the total loss function. The fusion function is a multilayer perceptron.

[0112] The expression in the total loss function is,

[0113] ;

[0114] ;

[0115] ;;

[0116] ;

[0117] in, Represents the total loss function. Indicates the downstream task loss item. Represents consistency constraints. Represents the diversity constraint. This represents a sufficiency constraint. Represents the fusion embedding matrix, Represents a static tree topology. This represents the dynamic graph topology, where F represents the Frobenius norm. This represents the aligned static embedding matrix. This represents the aligned dynamic embedding matrix. , and All represent preset balance coefficients. Nursing task-related tags.

[0118] In one embodiment, to map the static and dynamic embedding matrices to a unified representation space for subsequent fusion, the fusion module first aligns the static and dynamic embedding matrices. The static and dynamic embedding matrices originate from the topology of a tree structure and a dynamic graph, respectively. The nodes in the static embedding matrix (i.e., hospital area, department, ward, nurse) and the nodes in the dynamic embedding matrix (i.e., patient, nursing task, nurse) differ in their structure, and the representations of the same node (e.g., nurse) reside in different feature spaces due to different modalities. Direct fusion would lead to semantic misalignment. By forcibly aligning common nodes, the representational differences between modalities can be eliminated, laying the foundation for subsequent fusion. Specifically, the steps are as follows: First, nodes that exist simultaneously in the static tree topology database and the dynamic graph topology database are obtained as common nodes. These nodes include nurse nodes and some ward nodes that can serve as static attribute anchors. For example, a certain intensive care unit ward is a layer in the static tree. Level nodes, in the dynamic graph, serve as context nodes for nurses performing tasks. Subsequently, the static and dynamic embedding matrices are input into two pre-configured fully connected networks. The parameters of these two networks are predetermined by minimizing the embedding distance of common nodes in the output space. Aligned static and dynamic embedding matrices are obtained through forward computation. At this point, the two matrices have semantically consistent representations at the same node positions, providing aligned input for subsequent fusion. Based on alignment, to fully integrate static tree topology information and dynamic graph topology information and achieve information complementarity, the fusion module further employs a fusion function for information balancing fusion processing. Simple concatenation or weighted averaging cannot adaptively balance the information contributions of the two modalities and is prone to introducing redundancy or conflicts. Fusion is performed using a multilayer perceptron combined with multiple constraint objectives, which can promote consistency, diversity, and sufficiency between modalities while preserving their unique information. Specifically, the aligned static embedding matrix... With dynamic embedding matrix The concatenated data is then input into a multilayer perceptron, which consists of two hidden layers, each with 128 units and a ReLU activation function. The output layer has 64 units. The initial fusion embedding matrix H is obtained through forward propagation. Subsequently, the fusion process is constrained based on a total loss function, which includes downstream task loss terms. (For example, the mean squared error of path prediction, the specific value of which depends on the current application task), consistency constraints. Diversity constraints and sufficiency constraints ,in By calculating the difference between HT' and H, as well as the difference between HG' and H, the fused embedding is forced to approximate the original embeddings of both modalities simultaneously, thereby preserving common information; Through calculation transpose and The square of the Frobenius norm of the product is used to measure the diversity of the two modalities, prompting them to retain their unique information and avoiding homogenization after fusion; The comparison between the fusion embedding H and the nursing task-related label Y is calculated using InfoNCE loss to ensure that the fusion embedding includes sufficient information to predict downstream tasks. Each loss is weighted by a preset balance coefficient α, β, and γ, where α, β, and γ are all set to 0.1 to keep the magnitudes of each loss term similar. The network parameters of the multilayer perceptron are determined through joint optimization under the objective of minimizing the total loss. The fusion embedding matrix H is obtained through forward propagation. Each row of this matrix corresponds to the fusion embedding vector of a node. Through the fusion module, the alignment mapping and fusion network are directly utilized to achieve the alignment and fusion of static and dynamic information. This allows the fusion embedding to retain the shared knowledge of the two modalities while highlighting their respective unique contributions, providing a rich and discriminative fusion representation for subsequent application modules.

[0119] The application processing includes extracting the embedding vector corresponding to the current patient node from the fused embedding matrix as the current patient node embedding, and extracting the set of embedding vectors corresponding to the task nodes to be assigned as the task node embedding set, and calculating the similarity score between the patient and each task through cosine similarity.

[0120] Based on similarity scores, nurse qualifications are extracted from the aligned static embedding matrix. The nurse qualifications are then matched with the preset qualification requirements of each task, and tasks that do not meet the preset qualification requirements are filtered out to obtain a list of tasks that meet the qualification requirements.

[0121] The pre-defined qualification requirements include the required nursing professional title and the required department affiliation;

[0122] Sort the list of tasks that meet the qualification requirements in descending order according to similarity score, and output the Top-K task recommendation list;

[0123] For the fusion embedding matrix, extract the patient node embedding and the corresponding patient's historical nursing path sequence, input the historical task sequence into the long short-term memory network to obtain the hidden state sequence of each time step, and use it as the sequence embedding;

[0124] The attention weight of each hidden state in the hidden state sequence relative to the patient node embedding is calculated using an attention mechanism, and the hidden state sequence is weighted and summed based on the attention weights to obtain the context vector.

[0125] The context vector and the patient node embedding are fused to obtain the fused features.

[0126] The fused features are input into the fully connected layer for prediction to obtain the prediction result of the next nursing task. The prediction result is added to the historical nursing path sequence to generate the optimized nursing path.

[0127] Extract the embedding vectors corresponding to the nurse nodes from the fusion embedding matrix to form the nurse node embedding set;

[0128] Extract nurse qualifications and department attributes from the aligned static embedding matrix as nurse static attributes;

[0129] Extract nurse execution records from the aligned dynamic embedding matrix as nurse dynamic execution records;

[0130] Based on the nurse node embedding set, nurse static attributes, and nurse dynamic execution records, the matching degree between each nurse and each task is calculated, and a nurse scheduling suggestion is generated based on the matching degree.

[0131] In one embodiment, to apply the fused embedding matrix to a real-world nursing scenario, the application module performs three processes: task recommendation, path optimization, and scheduling suggestion. By leveraging the static tree topology information and dynamic graph topology information included in the fused embedding, it directly supports clinical decision-making, improving the accuracy and efficiency of nursing work. In task recommendation, the embedding vector corresponding to the current patient node and the set of embedding vectors corresponding to all task nodes to be assigned are first extracted from the fused embedding matrix. A similarity score is obtained by calculating the cosine similarity between the patient's embedding vector and each task's embedding vector. This similarity score reflects the degree of matching between the patient's current condition and the task requirements. Subsequently, based on the similarity score, the embedding vectors are extracted from the aligned static embedding matrix. The nurse qualifications corresponding to each node include title vectors and practice category vectors. These qualifications are compared item by item with the preset qualification requirements of each task. For example, if a task may require a head nurse title and belong to internal medicine, tasks with titles below the requirement or mismatched department affiliations are filtered out, resulting in a list of tasks that meet the qualification requirements. Finally, the tasks in the list are sorted in descending order by similarity score, and the top K tasks are output as a recommendation list. K is preset to 5 based on actual clinical needs to balance the number of recommendations and accuracy, enabling nurses to quickly focus on the most urgent nursing tasks. In path optimization, the current patient node embedding and its corresponding historical nursing path sequence are first extracted from the fusion embedding matrix. This sequence is generated by the patient according to the time... The task node IDs executed sequentially are used to construct an embedding sequence. Each task ID is mapped to a corresponding task embedding. This sequence is then input into a Long Short-Term Memory (LSTM) network with 128 hidden units and 2 layers to obtain the hidden state sequence at each time step as the sequence embedding. Subsequently, an attention mechanism is used to calculate the attention weight of each hidden state relative to the patient node embedding. Specifically, an additive attention approach is adopted, first calculating the compatibility score between each hidden state and the patient node embedding, then normalizing it to obtain the weight. Based on these weights, the hidden state sequence is weighted and summed to obtain the context vector. The context vector is concatenated with the patient node embedding and input into a fully connected network with an input dimension of 256 and an output dimension of [missing information]. The number of nodes in the output layer is 128, and the activation function is ReLU to obtain the fused features. Finally, the fused features are input to the output layer, and the number of nodes in the output layer is the same as the number of nursing task categories, for example, 500 nursing tasks. The predicted probability of each task is obtained through the softmax function, and the task with the highest probability is taken as the prediction result of the next nursing task. This prediction result is added to the historical nursing path sequence to generate an optimized nursing path, providing nurses with forward-looking task planning. In the scheduling suggestion, the embedding vector set corresponding to all nurse nodes is first extracted from the fused embedding matrix; the static attributes of the corresponding nurses are extracted from the aligned static embedding matrix, including the professional title vector, practice category vector, and department vector, as the nurse's static attributes.The dynamic execution records of the corresponding nurses are extracted from the aligned dynamic embedding matrix, including the vector of the number of tasks executed, the vector of average execution time, and the vector of operation results, as the nurse's dynamic execution record. Then, the matching degree between each nurse and each task is calculated using a weighted fusion method: first, the cosine similarity between the nurse's embedding and the task's embedding is calculated as the embedding matching degree; then, the matching score between the nurse's static attributes and the task's static requirements is calculated, for example, 1 for matching professional title and 0 otherwise, 1 for matching department and 0 otherwise. The scores are then weighted and summed. Finally, the matching score between the nurse's dynamic execution record and the task's execution features is calculated, for example, the normalized similarity between the number of executions and the average number of executions of the task is used as the score. Finally, the embedding matching degree, static matching score, and dynamic matching score are weighted at 0.4, 0, and 1 respectively. The total matching degree is obtained by weighting and summing the values ​​of 0.3 and 0.3. This weight is set empirically to balance the three information sources. Based on the matching degree of each nurse with each task, combined with shift requirements (e.g., 5 nurses for the morning shift, 4 for the afternoon shift, and 3 for the evening shift), a greedy algorithm is used to sequentially select the nurses with the highest matching degree for each shift until the shift's staffing requirements are met, generating a suggested nurse scheduling table. This is then processed by the application module, which transforms the fusion embedding matrix into decision outputs that directly support clinical work. Task recommendation enables nurses to quickly locate the patient's most needed nursing procedures, path optimization provides personalized treatment process planning for patients, and scheduling suggestions achieve intelligent allocation of human resources, thereby comprehensively improving the informatization and intelligence level of the nursing process.

[0132] This invention addresses the technical bottlenecks of existing nursing management systems, such as the separation of static structural information and dynamic execution data, and insufficient mining of spectral domain features in complex graph data. It proposes an innovative dual-track information processing architecture, constructing a dual-modal database of static tree topology and dynamic graph topology. By mapping the dynamic graph to the spectral domain through spectral decomposition, it solves the problem of traditional methods failing to capture long-range dependencies and deep structural features in nursing networks. In the encoding stage, Masked Topology Enhancement (MTA) and Cross-Channel Feature Enhancement (CFA) are introduced to target and enhance spectral domain eigenvalues ​​and feature coefficient matrices, effectively overcoming the oversmoothing and feature annihilation defects that easily occur in existing graph neural networks when processing sparse medical data. By combining alignment processing and fusion functions with multi-constraint loss functions, the problem of difficult fusion of heterogeneous static and dynamic embeddings is solved. These improvements not only achieve efficient fusion of static information such as nurse qualifications and affiliations with dynamic information such as task execution sequences and patient pathways, but also effectively improve the accuracy and interpretability of task recommendation, path optimization, and scheduling suggestions, significantly enhancing the intelligence level and scientific decision-making of nursing process management.

[0133] 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 implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. 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.

[0134] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A nursing process information management system, characterized in that, It includes a processing module, a building module, an encoding module, a fusion module, and an application module; The processing module preprocesses the raw medical business data to obtain static tree topology structured data, dynamic graph topology structured data, and dynamic graphs corresponding to the dynamic graph topology structured data. The module constructs a static tree topology database based on the static tree topology structured data, and constructs a spectral domain representation of the dynamic graph topology structured data through spectral decomposition and stores it to obtain the dynamic graph topology database. The encoding module encodes the static tree topology database to obtain a static embedding matrix. After enhancing the spectral domain features of the spectral domain representation in the dynamic graph topology database, it reconstructs and encodes the dynamic embedding matrix. The fusion module aligns the static and dynamic embedding matrices, performs information balancing fusion using a fusion function, and generates a fused embedding matrix based on the total loss function. The application module processes the fusion embedding matrix to obtain a task recommendation list, optimized nursing pathways, and nurse scheduling suggestions.

2. The nursing process information management system as described in claim 1, characterized in that, Raw data for medical operations includes static tree topology data and dynamic graph topology data; Static tree topology data includes department attributes, nurse qualifications, and affiliation relationships; The affiliation relationship refers to the affiliation between hospital areas, departments, wards, and nurses; The dynamic graph topology data includes nodes of nursing tasks, execution relationships, patient information, nursing task descriptions, nurse execution records, and historical nursing path sequences for each patient; The historical nursing pathway sequence consists of task nodes performed by the patient in chronological order; The nodes of a nursing task include patient nodes, nursing task nodes, and nurse nodes, and the execution relationship is the execution relationship between the nodes of the nursing task. Preprocessing of raw medical data includes constructing a tree structure based on static tree topology data; The tree structure includes nodes such as hospital area, department, ward, and nurse, and the edges of the tree structure represent the membership relationship. The tree node features of each node in the tree structure are extracted to obtain static tree topology structured data. The tree node features include department attributes and nurse qualifications; A dynamic graph is constructed based on the dynamic graph topology data. The nodes of the dynamic graph include patient nodes, nursing task nodes, and nurse nodes, and the edges of the dynamic graph represent the execution relationships. The execution relationships in the dynamic graph are aggregated according to a preset time window to obtain the graph Laplace matrix of the dynamic graph; Extract the graph node features of each node in the dynamic graph to construct a graph node feature matrix; Dynamic graph topology structured data is composed of the graph Laplacian matrix and the graph node feature matrix. The graph node features include patient information, nursing task descriptions, and nurse execution records.

3. The nursing process information management system as described in claim 2, characterized in that, For static tree topology structured data, a MySQL database is used for storage. The tables in the MySQL database used to store the tree structure include node ID, parent node ID, node type, and level fields. The parent node ID represents the direct parent node of the current node in the hierarchy. The tree node features of the tree structure are stored as node feature vectors and indexed to obtain a static tree topology database.

4. The nursing process information management system as described in claim 3, characterized in that, The dynamic graph topology structured data is stored in the Neo4j database, and a time index is created to obtain a Neo4j database with a time index. The spectral decomposition of the Laplacian matrix of the dynamic graph is then performed to obtain the orthogonal eigenvector basis and the diagonal eigenvalue matrix. The expression for the diagonal eigenvalue matrix is ​​as follows: ,and ; in, The graph Laplacian matrix represents a dynamic graph. Describes the basis of orthogonal eigenvectors. This represents the transpose of the orthogonal eigenvector basis. Represents an eigenvalue diagonal matrix. Represents the identity matrix; Mapping the graph node eigenvalue matrix to the spectral domain spanned by orthogonal eigenvector bases yields the eigencoefficient matrix, which is expressed as follows: ; in, Represents the characteristic coefficient matrix, The node feature matrix representing a dynamic graph; Will , , The spectral domain representation of the dynamic graph is stored in the graph database. During storage, the topological structured data of the dynamic graph is retained, resulting in a dynamic graph topology database that includes both the topological structured data and the spectral domain representation.

5. The nursing process information management system as described in claim 4, characterized in that, For the spectral domain representation in the dynamic graph topology database, after spectral domain feature enhancement using masked topology enhancement (MTA) and cross-channel feature enhancement (CFA), reconstruction encoding is performed to obtain the dynamic embedding matrix. For a static tree topology database, read the tree node features of each node to form a tree node feature matrix; The tree structure and tree node feature matrix in the static tree topology database are processed by a bidirectional tree neural network. Specifically, the process includes: aggregating the child node information of each node from bottom to top to obtain the intermediate representation of each node; passing the context information of the parent node from top to bottom; fusing the context information of the parent node with the intermediate representation to generate the static embedding of each node; the child node is the node whose parent node ID points to the node; and the static embeddings of all nodes constitute the static embedding matrix.

6. The nursing process information management system as described in claim 5, characterized in that, Feature enhancement processing is performed on the spectral domain representations in the dynamic graph topology database, wherein the spectral domain representations include eigenvalue diagonal matrices, eigencoefficient matrices, and orthogonal eigenvector bases; Feature enhancement processing of the eigenvalue diagonal matrix includes processing the diagonal element sequence of the eigenvalue diagonal matrix using a mask topology enhancement (MTA). The MTA maps the diagonal element sequence into an embedding vector through a feature embedding layer. The embedding vector is then input into a multi-head self-attention encoder for self-attention processing to obtain an encoded representation of the eigenvalue sequence. Finally, the encoded representation of the eigenvalue sequence is input into a decoder for decoding processing to obtain the enhanced eigenvalue diagonal matrix. Feature enhancement processing of the feature coefficient matrix includes processing the feature coefficient matrix using cross-channel feature enhancement CFA, wherein the cross-channel feature enhancement CFA performs a one-dimensional convolution on the feature coefficient matrix to obtain the enhanced feature coefficient matrix. Multiplying the orthogonal eigenvector basis with the enhanced eigenvalue diagonal matrix and the transpose of the orthogonal eigenvector basis yields the graph Laplacian matrix of the enhanced dynamic graph. Multiplying the orthogonal eigenvector basis with the enhanced eigencoefficient matrix yields the graph node feature matrix of the enhanced dynamic graph. The graph Laplacian matrix and the graph node feature matrix of the enhanced dynamic graph are encoded using a graph neural network to obtain the dynamic embedding matrix.

7. The nursing process information management system as described in claim 6, characterized in that, The alignment process includes static embedding matrices and dynamic embedding matrices, wherein the static embedding matrix is ​​composed of nodes of a tree structure in a static tree topology database, and the dynamic embedding matrix is ​​composed of nodes of a dynamic graph in a dynamic graph topology database. Obtain common nodes in the tree structure that exist simultaneously in the static tree topology database and the nodes in the dynamic graph in the dynamic graph topology database. Align the embeddings of the common nodes to obtain the aligned static embedding matrix and dynamic embedding matrix.

8. The nursing process information management system as described in claim 7, characterized in that, The aligned static and dynamic embedding matrices are subjected to information balancing fusion processing using a fusion function, and a fused embedding matrix is ​​generated based on the total loss function. The fusion function is a multilayer perceptron. The expression in the total loss function is, ； ; ; ; in, Represents the total loss function. Indicates the downstream task loss item. Represents consistency constraints. Represents the diversity constraint. This represents a sufficiency constraint. Represents the fusion embedding matrix, Represents a static tree topology. This represents the dynamic graph topology, where F represents the Frobenius norm. This represents the aligned static embedding matrix. This represents the aligned dynamic embedding matrix. , and All represent preset balance coefficients. Nursing task-related tags.

9. The nursing process information management system as described in claim 8, characterized in that, The application processing includes extracting the embedding vector corresponding to the current patient node from the fused embedding matrix as the current patient node embedding, and extracting the set of embedding vectors corresponding to the task nodes to be assigned as the task node embedding set, and calculating the similarity score between the patient and each task through cosine similarity. Based on the similarity score, nurse qualifications are extracted from the aligned static embedding matrix. The nurse qualifications are then matched with the preset qualification requirements of each task. Tasks that do not meet the preset qualification requirements are filtered out to obtain a list of tasks that meet the qualification requirements. The preset qualification requirements include the required nursing professional title and the required department affiliation; Sort the list of tasks that meet the qualification requirements in descending order according to similarity score, and output the Top-K task recommendation list; For the fusion embedding matrix, extract the patient node embedding and the corresponding patient's historical nursing path sequence, input the historical task sequence into the long short-term memory network to obtain the hidden state sequence of each time step, and use it as the sequence embedding; The attention weight of each hidden state in the hidden state sequence relative to the patient node embedding is calculated using an attention mechanism, and the hidden state sequence is weighted and summed based on the attention weights to obtain the context vector. The context vector and the patient node embedding are fused to obtain the fused features.

10. The nursing process information management system as described in claim 1, characterized in that, The fused features are input into the fully connected layer for prediction to obtain the prediction result of the next nursing task. The prediction result is added to the historical nursing path sequence to generate the optimized nursing path. Extract the embedding vectors corresponding to the nurse nodes from the fusion embedding matrix to form the nurse node embedding set; Extract nurse qualifications and department attributes from the aligned static embedding matrix as nurse static attributes; Extract nurse execution records from the aligned dynamic embedding matrix as nurse dynamic execution records; Based on the nurse node embedding set, nurse static attributes, and nurse dynamic execution records, the matching degree between each nurse and each task is calculated, and a nurse scheduling suggestion is generated based on the matching degree.

Images