Nursing risk intervention decision system and method based on knowledge graph

Abstract

The present invention discloses a nursing risk intervention decision system and method based on knowledge graph, which relates to the field of medical information technology, including: constructing a dynamic medical nursing knowledge graph based on multi-source nursing data and deep learning models, using a bidirectional long short-term memory network and a conditional random field for entity recognition; capturing the evolutionary characteristics of entity relationships through a marginal attention mechanism and a temporal graph convolutional network, optimizing the knowledge graph structure based on connectivity constraints and relationship weight thresholds; using a graph optimization mechanism driven by reinforcement learning to dynamically adjust the knowledge structure, optimize the relationship weights between entities in real time, and construct a multi-level risk characterization model, and realize accurate identification and early warning of nursing risks through local risk feature extraction, global risk information aggregation, and temporal risk evolution analysis. The present invention deeply integrates the complex association relationship between medical entities with nursing risk decisions, constructs multi-dimensional medical entity features into an entity association network, and realizes early risk identification.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and in particular to a nursing risk intervention decision system and method based on a knowledge graph. Background Art

[0002] With the continuous growth of demand for medical services and the increasing complexity of the medical environment, medical institutions face many challenges in providing high-quality nursing services. The existing nursing management model mainly relies on manual experience and fixed processes, and has lags and limitations in risk identification, early warning and intervention. Although a large amount of clinical practice data is constantly accumulating, the lack of effective analysis and utilization mechanisms makes it difficult to play the role of data value in supporting nursing decisions. Traditional risk prevention and control methods often adopt a passive response approach and cannot meet the needs of intelligent and precise nursing. At the same time, the complex associations and dynamic changes between medical entities also bring new challenges to nursing work. It is difficult to effectively integrate information between various nursing links, which affects the continuity and coordination of nursing work. Nursing staff need to deal with a large number of tedious observation, recording and decision-making tasks in their daily work, which not only increases the workload but also may affect the quality of nursing.

[0003] Most of the current intelligent solutions focus on the realization of a single function, lacking systematic and holistic considerations. Existing technical solutions often ignore the professional characteristics and complexity of the medical field, and it is difficult to accurately grasp the subtle changes and potential risks in clinical practice. At the same time, due to the lack of in-depth exploration and utilization of the relationship between medical entities, the performance of existing systems in risk prediction and decision support still has a lot of room for improvement. In addition, the fragmentation problem of medical information systems is still prevalent, and the data sharing and business collaboration between subsystems are not ideal, which affects the improvement of overall nursing management efficiency.

[0004] Therefore, how to use modern information technology to improve the level of nursing management, realize the intelligent transformation of nursing work, and build an intelligent system that can organically integrate various medical information, accurately predict potential risks, and provide accurate decision-making support has become an important issue that needs to be solved in the current medical field. This is not only related to the improvement of nursing quality and the protection of patient safety, but also an important measure to promote the modernization of medical services. Summary of the invention

[0005] In view of the problems existing in the existing nursing risk intervention decision-making method based on knowledge graph, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is the lack of intelligent analysis and prediction of the association relationship of medical entities in the existing nursing management, which leads to technical problems such as passive risk prevention and control and insufficient decision support.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In the first aspect, an embodiment of the present invention provides a nursing risk intervention decision-making method based on a knowledge graph, which includes: constructing a dynamic medical nursing knowledge graph based on multi-source nursing data and a deep learning model, and using a bidirectional long short-term memory network and a conditional random field for entity recognition; capturing the dynamic expression and evolution characteristics of entity relationships through a marginal attention mechanism and a temporal graph convolutional network, and introducing medical field knowledge constraints to optimize entity semantic similarity calculations, and optimizing the knowledge graph structure based on connectivity constraints and relationship weight thresholds; using a reinforcement learning-driven graph optimization mechanism to dynamically adjust the knowledge structure, and by defining a graph state space, action space, and reward function, optimizing the relationship weights between entities in real time, while cooperating with a graph pruning strategy to maintain the simplicity and effectiveness of the knowledge structure; based on the optimized knowledge graph, constructing a multi-level risk characterization model, and realizing accurate identification and early warning of nursing risks through local risk feature extraction, global risk information aggregation, and temporal risk evolution analysis.

[0009] As a preferred solution of the nursing risk intervention decision-making method based on knowledge graph described in the present invention, the multi-source nursing data includes nursing text data, patient physiological index data and expert knowledge base data; the nursing text data includes nursing records, doctor's order execution records, nursing assessment records, nursing handover records, nursing adverse event records and nursing quality monitoring records; the patient's physiological index data includes real-time monitoring data of body temperature, blood pressure, heart rate, respiratory rate, blood oxygen saturation, blood sugar, weight, urine output, bowel movement frequency and consciousness state; the expert knowledge base data includes standard nursing specifications, expert experience summary, nursing case library, nursing guidelines, nursing technical operating procedures, nursing quality evaluation standards and nursing risk warning rules; the nursing text data is segmented and medical entity labeled, and each word is processed by the labeling function, which is expressed as:

[0010] ;

[0011] in, Represents the input word, Indicates the label category of the first position, Indicates the annotation category of the second position, Indicates The labeling categories of the positions include the medical entity types of "disease", "symptom", "treatment", "examination", "drug", "surgery", "instrument", "nursing measures", "vital signs", "abnormal events", "risk level", "nursing outcome", "nursing time", "nursing staff" and "nursing part"; the initial nursing knowledge entity set is generated, which is expressed as:

[0012] ;

[0013] in, Represents the initial nursing knowledge entity set, represents the first nursing knowledge entity, Represents the second nursing knowledge entity, Indicates Nursing knowledge entities, is the total number of entities;

[0014] Each entity Both contain the entity name, entity type, entity occurrence location, entity frequency, and entity timestamp.

[0015] As a preferred solution of the nursing risk intervention decision method based on knowledge graph described in the present invention, a feature vector matrix is ​​constructed based on the initial nursing knowledge entity set. , extracting semantic features through context windows and introducing medical domain knowledge constraints to improve The model calculates the entity semantic similarity of each entity, expressed as:

[0016] ;

[0017] in, and Represent entities and The characteristic vector of represents the vector norm;

[0018] Organize the semantic similarities between all entity pairs into a matrix form and construct a × Dimensional matrix, namely, entity semantic association matrix ;

[0019] Improvement of optimization objective function by adding medical knowledge constraints Model, expressed as:

[0020] ;

[0021] in, For the original Loss function, is a set of medical knowledge constraints, is the balance parameter, Represents the total loss function;

[0022] The entity semantic association matrix Input into the marginal attention network, adaptively learn the importance weights between entities through the attention mechanism, and calculate the initial association weight matrix between nursing knowledge entities ;

[0023] Based on the initial association weight matrix A temporal graph convolution model is established to capture the knowledge evolution characteristics by combining time information and graph structure information. The feature extraction formula of the temporal graph convolution model is expressed as:

[0024] ;

[0025] in, , is the adjacency matrix after adding the self-loop, is the original adjacency matrix; for dimensional identity matrix, Indicates The node feature matrix of the layer, Indicates The node feature matrix of the layer, for The corresponding degree matrix, Indicates The learnable weight matrix of the layer, represents the activation function, is the time coding matrix;

[0026] Furthermore, the time coding matrix The elements in are represented as:

[0027] ;

[0028] in, Represents the time coding matrix No. OK Column elements, For Node The corresponding timestamp, is the frequency parameter, is the phase parameter;

[0029] Extract the original features of the entity directly from the nursing documents to form the initial feature matrix , the initial feature matrix Input to the temporal graph convolution model for feature extraction and compare with the temporal encoding matrix Perform temporal convolution to obtain the output feature matrix after being processed by the temporal graph convolution model ;

[0030] The output feature matrix The data is input to the named entity recognition module, which uses a bidirectional long short-term memory network structure to improve the accuracy of entity recognition through forward and backward information flow.

[0031] The final output layer of the bidirectional long short-term memory network uses a conditional random field Perform sequence labeling, expressed as:

[0032] ;

[0033] in, Indicates that given an input sequence Under the condition, output the label sequence The conditional probability of is the normalization factor, Represents the potential function, calculating the current annotation and the previous annotation The transition probability, represents the sequence length, express Marking of time, express Marking of time, represents the input sequence;

[0034] pass Perform sequence annotation to identify new entities from the text. The newly identified entities will be added to the original nursing knowledge entity set to form a new nursing knowledge entity set. .

[0035] As a preferred solution of the nursing risk intervention decision method based on knowledge graph of the present invention, based on the initial nursing knowledge entity set , Added nursing knowledge entity set and the initial association weight matrix Building the initial knowledge graph , where the node set , represents the nursing knowledge entity; the constructed initial knowledge graph satisfies the connectivity constraint:

[0036] ;

[0037] in, represents any two nodes in the initial knowledge graph, Represents a collection of nodes. Indicates that there is a path between nodes;

[0038] By relationship weight threshold Control the density of the knowledge graph, expressed as:

[0039] ;

[0040] in, Representation Node and The edge between represents the relationship weight between any two nodes, Represents the weight threshold.

[0041] As a preferred solution of the nursing risk intervention decision method based on knowledge graph of the present invention, the number of entity nodes of the initial knowledge graph is , weight of relationship between entities And the graph density is mapped to a state vector, expressed as:

[0042] ;

[0043] in, represents the state vector, represents the total number of initial knowledge graph nodes, is the weight threshold for controlling the density of the graph;

[0044] Importance scoring of entity nodes based on nursing entities Clinical relevance Constructing the action space , the importance score calculation formula is:

[0045] ;

[0046] in, Representation Node The entity importance score of Represents an entity node The degree characteristics of Representation Node The centrality characteristics of Represents an entity node The frequency of occurrence in clinical data, , and is the weight coefficient of each feature; the action space Also includes relationship weight adjustment and node addition and deletion operations, It is expressed as:

[0047] ;

[0048] in, is the weight adjustment step size, Representation Node The entity importance score of , and is the weight coefficient, Indicates clinical relevance;

[0049] Based on the semantic similarity between the new entities and the existing entities in the nursing documents Establish a reward function, expressed as:

[0050] ;

[0051] in, Indicates that at time step The total reward value, is the indicative function, is the similarity threshold, Indicates a new entity With existing entities The semantic similarity of and are positive and negative reward values ​​respectively; based on the calculation result of the reward function, through the depth The network calculates the value of each action, expressed as:

[0052] ;

[0053] in, Indicates the current state Next action The estimated value of and is the weight parameter of temporal difference learning and satisfies + = 1; To perform actions The next state after Represents the set of optional actions for the next state; the network loss function is defined as:

[0054] ;

[0055] in, Represents the loss value of network training, represents the expected calculation, Indicates the target value, Indicates the current forecast Value; based on current state choose The action with the largest value executes the selected action and obtains the new state and rewards , according to the loss function renew The network parameters are iteratively optimized until the graph structure is stable, and the optimized relationship weight matrix is ​​obtained. ; Based on the accessibility between nursing entities Prune the graph and define the path score, which is expressed as:

[0056] ;

[0057] in, Represents the entity arrive The path quality score is is the weight of the edge on the path, Represents a connected entity and A complete path of is the path length, Represents a single edge on a path;

[0058] When the path length exceeds the preset threshold If there is an alternative path, delete Minimum redundant edges, preset threshold The calculation is expressed as:

[0059] ;

[0060] in, is the maximum acceptable path length threshold, is the total number of nodes in the current graph.

[0061] As a preferred solution of the nursing risk intervention decision-making method based on knowledge graph described in the present invention, a multi-level risk characterization model is constructed based on the optimized knowledge graph, including an attention layer, a knowledge fusion layer, a graph neural network layer, a feature aggregation layer and a prediction layer; the attention layer is based on the optimized knowledge graph. The collection of nursing entity nodes in and the relationship weight matrix , for each entity node A multi-layer attention mechanism is applied to consider the local semantic environment and clinical attributes of the node to obtain the local risk feature vector, which is expressed as:

[0062] ;

[0063] in, Representation Node The local risk feature vector of Representation Node The set of neighboring nodes of and Represents the current central node and neighboring nodes The entity importance score of represents the feature transformation matrix, is the attention weight coefficient;

[0064] For the optimized knowledge graph Each entity node in Build Hop-Neighborhood Subgraph , which contains possible risk propagation paths, combined with the updated relationship weight matrix , the structured risk pattern vector is extracted through a multi-layer graph convolutional network, which is expressed as:

[0065] ;

[0066] in, Representation Node In the The risk pattern vector of the layer, For Node of The set of hop neighbor nodes, is the normalization coefficient, For Node of The set of hop neighbor nodes, For the The learnable weight matrix of the layer, is the activation function, Indicated in In the layer, nodes The risk pattern vector of

[0067] go through The final risk pattern vector obtained after layer graph convolution , the final risk pattern vector is concatenated with the local risk feature vector to form an enhanced risk feature vector, which is expressed as:

[0068] ;

[0069] in, represents the enhanced risk feature vector;

[0070] Based on the acquired enhanced risk feature vector Construct a multi-layer risk propagation network and use a multi-head attention mechanism to model the complex correlation between different risk types. Attention heads are used to calculate the correlation strength matrix, which can be expressed as:

[0071] ;

[0072] in, Indicates The output matrix of the attention head is yes The query matrix represents The query matrix of the attention heads; yes The key matrix represents The key matrix of the attention heads; yes The value matrix represents The value matrix of the attention heads; is the scaling factor;

[0073] The outputs of multiple heads are integrated by splicing, expressed as:

[0074] ;

[0075] in, represents the final output matrix, Represents the output matrix of each attention head, Represents the output projection matrix.

[0076] As a preferred solution of the nursing risk intervention decision-making method based on knowledge graph described in the present invention, the knowledge fusion layer combines the clinical expert score set Adjust the strength of the association:

[0077] ;

[0078] in, represents the association strength matrix after expert knowledge adjustment, represents the set of clinical expert ratings;

[0079] The graph neural network layer is based on the adjusted association strength matrix Construct a risk propagation network and model the risk propagation process as message passing in a graph neural network. The risk transfer of the layer is expressed as:

[0080] ;

[0081] in, For Node In the The risk propagation status of the layer, For the The parameter matrix of the layer; Representation Node The set of neighboring nodes of is the association strength after adjustment by expert knowledge, indicating that the node and The intensity of risk transmission between Represents the neighboring nodes The risk propagation status in the previous layer;

[0082] The feature aggregation layer integrates risk information from different layers through skip connections, expressed as:

[0083] ;

[0084] in, is the learnable aggregation matrix, is the number of network layers, Representation Node The final aggregate risk representation of

[0085] The prediction layer calculates the final score based on the aggregated risk vector, expressed as:

[0086] ;

[0087] in, and is the prediction layer parameter, is the weight matrix, is the bias term, is the activation function, Representation Node The final risk score is obtained; decision intervention is made based on the output results of the prediction layer, including: if the prediction output shows that the correlation between symptom aggravation and treatment response is abnormal, the system will associate symptom observation records, match the execution of the treatment plan, and push corresponding nursing measures to adjust the decision; if the prediction output shows that the disease progression is associated with changes in vital signs, the severity of the disease is analyzed, the dynamic changes of vital signs are tracked, and the trend of nursing outcomes is further predicted; if the prediction output identifies abnormal correlations with examination results after medication, the medication records are traced, the changes in examination results are associated, and abnormal event warnings are generated; if the prediction output captures the correlation between surgical operations and instrument use timing, the nursing time nodes are planned, the instrument preparation list is matched, and a perioperative nursing plan is formulated; if the prediction output reflects the correlation between nursing operations and site risks, the responsibilities of nursing staff are assigned, key nursing sites are identified, and risk prevention measures are determined.

[0088] In the second aspect, an embodiment of the present invention provides a nursing risk intervention decision system based on a knowledge graph, which includes: a construction module, which is used to construct a dynamic medical nursing knowledge graph based on multi-source nursing data and a deep learning model, adopt a bidirectional long short-term memory network and a conditional random field for entity recognition, and realize the dynamic expression of entity relationships and the capture of evolutionary characteristics through a marginal attention mechanism and a temporal graph convolutional network. The medical field knowledge constraints are introduced to optimize the entity semantic similarity calculation, and an efficient knowledge graph structure is constructed based on connectivity constraints and relationship weight thresholds; an optimization module, which is used to dynamically adjust the knowledge structure using a graph optimization mechanism driven by reinforcement learning, optimize the relationship weights between entities in real time by defining the graph state space, action space and reward function, and cooperate with the graph pruning strategy to maintain the simplicity and effectiveness of the knowledge structure; an early warning module, which is used to construct a multi-level risk characterization model based on the optimized knowledge graph, and realize accurate identification and early warning of nursing risks through local risk feature extraction, global risk information aggregation and temporal risk evolution analysis.

[0089] The beneficial effects of the present invention are as follows: the present invention deeply integrates the complex correlation between medical entities with nursing risk decisions, constructs multi-dimensional medical entity features into an inter-entity correlation network, and realizes early risk identification. And through intelligent analysis of the combination changes of medical entity features, the law of temporal evolution and the correlation propagation mode, the corresponding nursing decisions are triggered and targeted intervention plans are pushed. All aspects of nursing work are organically integrated to form a closed-loop management model from risk prediction, plan formulation to measure implementation, which significantly reduces the probability of adverse events while improving the quality of nursing. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0091] Figure 1 This is a flow chart of the nursing risk intervention decision-making method based on knowledge graph. DETAILED DESCRIPTION

[0092] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0093] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0094] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0095] Example 1: Reference Figure 1 , which is the first embodiment of the present invention, and provides a nursing risk intervention decision method based on a knowledge graph, including:

[0096] S1: Based on the multi-source data acquisition interface and the improved graph neural network, the initial medical care knowledge graph is constructed, and the bidirectional long short-term memory network and conditional random field are used for entity recognition. The dynamic expression and evolution characteristics of entity relationships are captured through the marginal attention mechanism and the temporal graph convolutional network, and the medical field knowledge constraints are introduced to optimize the entity semantic similarity calculation, and the knowledge graph structure is optimized based on connectivity constraints and relationship weight thresholds.

[0097] Collect nursing text data, patient physiological indicator data and expert knowledge base data;

[0098] The nursing text data include nursing records, doctor's order execution records, nursing assessment records, nursing handover records, nursing adverse event records, and nursing quality monitoring records;

[0099] The patient's physiological index data include real-time monitoring data of body temperature, blood pressure, heart rate, respiratory rate, blood oxygen saturation, blood sugar, weight, urine output, bowel movement frequency, and consciousness state;

[0100] Expert knowledge base data includes standard nursing specifications, expert experience summary, nursing case library, nursing guidelines, nursing technical operation procedures, nursing quality evaluation standards, and nursing risk warning rules;

[0101] The nursing text data is segmented and medical entities are annotated. Each word is processed by the annotation function, which is expressed as:

[0102] ;

[0103] in, Represents the input word A function that maps to the corresponding annotation category, Represents the input word, Indicates the label category of the first position, Indicates the annotation category of the second position, Indicates The annotation categories of each position include the medical entity types of "disease", "symptom", "treatment", "examination", "drug", "surgery", "instrument", "nursing measures", "vital signs", "abnormal events", "risk level", "nursing outcome", "nursing time", "nursing staff", and "nursing part".

[0104] Then the initial nursing knowledge entity set is generated, which is expressed as:

[0105] ;

[0106] in, Represents the initial nursing knowledge entity set, represents the first nursing knowledge entity, Represents the second nursing knowledge entity, Indicates Nursing knowledge entities, is the total number of entities.

[0107] Each entity They all contain entity name, entity type, entity occurrence location, entity frequency, and entity timestamp;

[0108] The entity name indicates the specific content of the entity, the entity type belongs to one of the 15 medical entity types mentioned above, the entity occurrence position indicates the position index in the original text, the entity frequency indicates the number of times it appears in the corpus, and the entity timestamp indicates the time information when the entity was recognized.

[0109] Constructing feature vector matrix based on initial nursing knowledge entity set , extracting semantic features through context windows and introducing medical domain knowledge constraints to improve The model calculates the entity semantic similarity of each entity, expressed as:

[0110] ;

[0111] in, and Represent entities and The characteristic vector of Represents the vector norm.

[0112] Organize the semantic similarities between all entity pairs into a matrix form and construct a × Dimensional matrix, namely, entity semantic association matrix ,pass Characterizes the strength of semantic association between all entity pairs.

[0113] Improve the objective function by adding medical domain knowledge constraints Model, expressed as:

[0114] ;

[0115] in, For the original Loss function, is a set of medical knowledge constraints, is the balance parameter, Represents the total loss function.

[0116] The entity semantic association matrix Input into the marginal attention network, the network adaptively learns the importance weights between entities through the attention mechanism, and the attention score calculation formula is:

[0117] ;

[0118] in, Representation Node For neighbor nodes The attention coefficient, is the attention weight matrix, is the hidden layer transformation matrix, is the weight matrix of the attention layer, Indicates the current node The characteristic vector of The target neighbor node The characteristic vector of Representation Node Neighbor nodes The characteristic vector of Representation Node The neighborhood set of As the activation function, the initial association weight matrix between nursing knowledge entities is calculated .

[0119] Furthermore, the output feature of the attention mechanism is expressed as:

[0120] ;

[0121] in, is a nonlinear activation function, Representation Node Updated feature vector.

[0122] Based on the initial association weight matrix A temporal graph convolution model is established. The temporal graph convolution model captures the characteristics of knowledge evolution by combining time information and graph structure information. The feature extraction formula of the temporal graph convolution model is expressed as:

[0123] ;

[0124] in, , is the adjacency matrix after adding the self-loop, is the original adjacency matrix, which represents the connection relationship between nodes in the graph; for dimensional identity matrix, Indicates The node feature matrix of the layer, Indicates The node feature matrix of the layer, for The corresponding degree matrix, Indicates The learnable weight matrix of the layer, represents the activation function, is the time coding matrix.

[0125] Furthermore, the time coding matrix The elements in are represented as:

[0126] ;

[0127] in, Represents the time coding matrix No. OK Column elements, For Node The corresponding timestamp, is the frequency parameter, is the phase parameter.

[0128] The time information is encoded into the node features through the time encoding matrix to capture the changes of the nodes in the time dimension, so that the model can distinguish the node states at different times.

[0129] Extract the original features of the entity directly from the nursing documents to form the initial feature matrix , the initial feature matrix Input to the temporal graph convolution model for feature extraction and compare with the temporal encoding matrix Perform temporal convolution to obtain the output feature matrix after being processed by the temporal graph convolution model .

[0130] The output feature matrix Input to the named entity recognition module, which uses a bidirectional long short-term memory network structure to improve the accuracy of entity recognition through forward and backward information flow. The calculation process is expressed as:

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] in, , , are the activation values ​​of the forget gate, input gate, and output gate, respectively. for The cell state at a moment in time, for The hidden layer output at time t, for The input vector at time instant, for The hidden state output at time t, for The cell state at a moment in time, , , , are the weight matrices of each gate, , , , are the bias terms of each gate respectively.

[0137] The final output layer uses conditional random fields Perform sequence labeling, expressed as:

[0138] ;

[0139] in, Indicates that given an input sequence Under the condition, output the label sequence The conditional probability of is a normalization factor to ensure that the sum of the probabilities of all possible labeled sequences is 1, Represents the potential function, calculating the current annotation and the previous annotation The transition probability, represents the sequence length, express Marking of time, express Marking of time, Represents the input sequence.

[0140] pass Sequence annotation can identify new entities from the text. The newly identified entities will be added to the original nursing knowledge entity set to form a new nursing knowledge entity set. , to achieve dynamic expansion and updating of nursing knowledge graph.

[0141] Based on initial nursing knowledge entity set , Added nursing knowledge entity set and the initial association weight matrix Building the initial knowledge graph , where the node set , represents the nursing knowledge entity; the constructed initial knowledge graph satisfies the connectivity constraint:

[0142] ;

[0143] in, represents any two nodes in the initial knowledge graph, Represents a collection of nodes. Indicates that there is a path between nodes. Connectivity constraints ensure that there is a reachable path between any two nodes in the knowledge graph, that is, the knowledge graph is connected.

[0144] At the same time, through the relationship weight threshold Control the density of the knowledge graph, expressed as:

[0145] ;

[0146] in, Representation Node and The edge between represents the relationship weight between any two nodes, Represents the weight threshold.

[0147] Only when the relationship weight between two nodes exceeds the threshold When the edge connection is established, the knowledge graph is controlled to avoid being too dense and to achieve a balance between the effectiveness (connectivity) and efficiency (moderate sparseness) of the knowledge graph.

[0148] S2: Use the reinforcement learning-driven graph optimization mechanism to dynamically adjust the knowledge structure. By defining the graph state space, action space, and reward function, the relationship weights between entities are optimized in real time. At the same time, the graph pruning strategy is used to maintain the simplicity and effectiveness of the knowledge structure, ensuring the timeliness and accuracy of the knowledge graph.

[0149] The number of entity nodes in the knowledge graph , weight of relationship between entities And the graph density is mapped to a state vector, expressed as:

[0150] ;

[0151] in, Represents the state vector, describing the knowledge graph in The state of the moment; represents the total number of initial knowledge graph nodes, is the weight threshold for controlling the density of the graph. This state vector fully characterizes the current structural characteristics of the graph and provides a basis for reinforcement learning optimization.

[0152] Importance scoring of entity nodes based on nursing entities Clinical relevance Constructing the action space , the importance score calculation formula is:

[0153] ;

[0154] in, Representation Node The entity importance score of Represents an entity node The degree characteristics of Representation Node The centrality characteristics of Represents an entity node The frequency of occurrence in clinical data, , and is the weight coefficient of each feature.

[0155] Clinical relevance The value is obtained through medical literature verification and the value range is [0,1].

[0156] Action Space Also includes relationship weight adjustment and node addition and deletion operations, It is expressed as:

[0157] ;

[0158] in, is the weight adjustment step size, Representation Node The entity importance score of , and is the weight coefficient, which is used to balance the contribution of the importance of the two nodes.

[0159] Based on the semantic similarity between the new entities and the existing entities in the nursing documents Establish a reward function, expressed as:

[0160] ;

[0161] in, Indicates that at time step The total reward value, is the indicative function, is the similarity threshold is the positive reward value when the similarity is higher than the threshold, indicating the benefit of knowledge fusion; It is the negative penalty value when the similarity is lower than the threshold, indicating the cost of knowledge redundancy; Indicates a new entity With existing entities The semantic similarity of .

[0162] Furthermore, The cosine similarity is calculated as follows:

[0163] ;

[0164] in, For the entities to be added extracted from the new nursing instrument, The existing entity.

[0165] By introducing importance scoring As a reward weight adjustment factor, new knowledge that is similar to important entities receives higher rewards, thereby giving priority to retaining high-value knowledge updates and evaluating the rationality of graph updates through semantic similarity.

[0166] Based on the calculation results of the reward function, through the depth The network calculates the value of each action, expressed as:

[0167] ;

[0168] in, Indicates the current state Next action The estimated value of and is the weight parameter of temporal difference learning and satisfies + = 1; To perform actions The next state after Represents a set of optional actions for the next state.

[0169] Then define the network loss function as:

[0170] ;

[0171] in, Represents the loss value of network training, represents the expected calculation, Indicates the target value, Indicates the current forecast value.

[0172] Based on the current state choose The action with the largest value executes the selected action and obtains the new state and rewards , according to the loss function renew The network parameters are iteratively optimized until the graph structure is stable.

[0173] choose The action with the largest value updates the relationship weight. and graph structure, and continuously iterate and optimize until the graph structure is stable, and obtain the optimized relationship weight matrix .

[0174] pass Learning combines immediate rewards and long-term benefits to ensure the strategic and continuous nature of knowledge updating, and achieve dynamic optimization and continuous learning of graph knowledge.

[0175] According to the accessibility between nursing entities Prune the graph and define the path score, which is expressed as:

[0176] ;

[0177] in, Represents the entity arrive The path quality score is is the weight of the edge on the path, Represents a connected entity and A complete path of is the path length, Represents a single edge on a path.

[0178] When the path length exceeds the preset threshold If there is an alternative path, delete Minimum redundant edges, preset threshold The calculation is expressed as:

[0179] ;

[0180] in, is the maximum acceptable path length threshold, is the total number of nodes in the current graph, Indicates rounding up.

[0181] By rounding up, the connectivity of the graph is ensured, while the path complexity is controlled and additional redundant space is provided for path optimization.

[0182] By evaluating the path quality and controlling the path length, the efficiency of the graph is improved while maintaining the integrity of knowledge. The dynamic threshold ensures that the pruning strategy can adapt to knowledge graphs of different sizes and maintain the availability and scalability of the graph.

[0183] By constructing a dynamic optimization framework of knowledge graphs based on reinforcement learning, the knowledge in the newly added nursing documents is intelligently integrated with the existing graphs. The entity importance and clinical relevance are used to guide action selection, and a semantic similarity-driven reward mechanism is used to evaluate the rationality of knowledge updating. The simplicity of the graph is maintained through an adaptive pruning strategy, ultimately achieving dynamic optimization and continuous learning of nursing knowledge.

[0184] S3: Based on the optimized knowledge graph, a multi-level risk characterization model is constructed. Through local risk feature extraction, global risk information aggregation and time-series risk evolution analysis, accurate identification and early warning of nursing risks are achieved, and the risk assessment effect of the system is continuously verified and improved through the expert feedback mechanism.

[0185] Based on the optimized knowledge graph, a multi-level risk characterization model is constructed, including attention layer, knowledge fusion layer, graph neural network layer, feature aggregation layer and prediction layer.

[0186] The attention layer is based on the optimized knowledge graph The collection of nursing entity nodes in and the relationship weight matrix In order to accurately capture the risk-related characteristics of each nursing entity node, A multi-layer attention mechanism is applied to consider the local semantic environment and clinical attributes of the node to obtain the local risk feature vector, which is expressed as:

[0187] ;

[0188] in, Representation Node The local risk feature vector contains the risk information of the node and its neighboring nodes. Representation Node The set of neighboring nodes of All directly connected nodes, and Represents the current central node and neighboring nodes The entity importance score of represents the feature transformation matrix, which is used to map the concatenated node representation to the risk feature space. is the attention weight coefficient, indicating the neighborhood node For the central node degree of impact.

[0189] Furthermore, the attention weight coefficient pass Normalized calculation, expressed as:

[0190] ;

[0191] in, is a trainable attention weight matrix used to calculate the importance of different neighborhood nodes. is the activation function.

[0192] In order to more comprehensively characterize the risk propagation model, the optimized knowledge graph Each entity node in Build Hop-Neighborhood Subgraph , this subgraph contains possible risk propagation paths. Combined with the updated relationship weight matrix , the structured risk pattern vector is extracted through a multi-layer graph convolutional network, which is expressed as:

[0193] ;

[0194] in, Representation Node In the The risk pattern vector of the layer, For Node of The set of hop neighbor nodes, is the normalization coefficient, For Node of The set of hop neighbor nodes, For the The learnable weight matrix of the layer, is the activation function, Indicated in In the layer, nodes Risk Pattern Vector.

[0195] go through The final risk pattern vector obtained after layer graph convolution , the final risk pattern vector is concatenated with the local risk feature vector to form an enhanced risk feature vector, which is expressed as:

[0196] ;

[0197] in, represents the enhanced risk feature vector.

[0198] Neighborhood information is aggregated through graph convolution operations, and normalization coefficients are used to balance node influences. Deep risk patterns are obtained through multi-layer feature extraction, and finally local and global features are fused to obtain an enhanced risk feature vector.

[0199] Based on the acquired enhanced risk feature vector A multi-layer risk propagation network is constructed, and a multi-head attention mechanism is used to model the complex correlation between different risk types. Attention heads are used to calculate the correlation strength matrix, which can be expressed as:

[0200] ;

[0201] in, Indicates The output matrix of the attention head is yes The query matrix represents The query matrix of the attention heads; yes The key matrix represents The key matrix of the attention heads; yes The value matrix represents The value matrix of the attention heads; is the scaling factor.

[0202] The outputs of multiple heads are integrated by splicing, expressed as:

[0203] ;

[0204] in, represents the final output matrix, Represents the output matrix of each attention head, Represents the output projection matrix.

[0205] The knowledge fusion layer combines the clinical expert rating set Adjust the strength of the association:

[0206] ;

[0207] in, represents the association strength matrix after expert knowledge adjustment, Represents a collection of clinical expert ratings.

[0208] The multi-angle risk association patterns are captured through a multi-head mechanism, and the association strength is corrected using expert knowledge, thus achieving an effective combination of model automatic learning and expert experience.

[0209] Graph neural network layer based on adjusted correlation strength matrix Construct a risk propagation network and model the risk propagation process as message passing in a graph neural network. The risk transfer of the layer is expressed as:

[0210] ;

[0211] in, For Node In the The risk propagation status of the layer, For the The parameter matrix of the layer is used to learn the transformation relationship in the risk propagation process; Representation Node The set of neighboring nodes of is the association strength after adjustment by expert knowledge, indicating that the node and The intensity of risk transmission between Represents the neighboring nodes The risk propagation status in the previous layer.

[0212] The feature aggregation layer integrates risk information from different layers through skip connections, which can be expressed as:

[0213] ;

[0214] in, is the learnable aggregation matrix, is the number of network layers, Representation Node The final aggregate risk representation.

[0215] Multi-scale information is captured through jump connections, including using low layers to capture local risk features and high layers to capture global risk patterns, avoiding the loss of information during propagation, retaining useful features at different levels, and providing a more comprehensive risk representation for prediction.

[0216] The prediction layer calculates the final score based on the aggregated risk vector, expressed as:

[0217] ;

[0218] in, and is the prediction layer parameter, is a weight matrix used to map the risk vector to the scoring space and is a trainable parameter matrix. is the bias term, is the activation function, Representation Node The final risk score.

[0219] Decision intervention is performed based on the output of the prediction layer, including:

[0220] If the prediction output shows that the correlation between symptom aggravation and treatment response is abnormal, the system will associate the symptom observation records, match the treatment plan execution status, and push the corresponding nursing measure adjustment decision.

[0221] If the prediction output indicates that disease progression is associated with changes in vital signs, the severity of the disease is analyzed, the dynamic changes in vital signs are tracked, and the trend of pathological outcomes is further predicted.

[0222] If the prediction output identifies an abnormal correlation with the test results after medication, the medication records are traced back, the changes in the associated test results are correlated, and an abnormal event warning is generated.

[0223] If the prediction output captures the timing association between surgical operations and instrument use, then the nursing time nodes are planned, the instrument preparation list is matched, and a perioperative nursing plan is developed.

[0224] If the prediction output reflects the association between nursing operations and area risks, nursing staff responsibilities are assigned, key nursing areas are identified, and risk prevention measures are determined.

[0225] In summary, the present invention deeply integrates the complex association relationship between medical entities with nursing risk decisions, constructs multidimensional medical entity features into an inter-entity association network, and realizes early risk identification. And through intelligent analysis of the combination changes of medical entity features, the law of temporal evolution and the associated propagation mode, the corresponding nursing decisions are triggered and targeted intervention plans are pushed. All aspects of nursing work are organically integrated to form a closed-loop management model from risk prediction, program formulation to measure implementation, which significantly reduces the probability of adverse events while improving the quality of nursing. And with the help of medical entity association analysis, the nursing knowledge map is continuously accumulated, the performance of the prediction model is continuously optimized, the scientificity and accuracy of nursing decisions are further improved, and intelligent solutions are provided for clinical nursing safety management.

[0226] This embodiment further provides a nursing risk intervention decision system based on a knowledge graph, including:

[0227] A construction module for building a dynamic medical care knowledge graph based on multi-source nursing data and deep learning models, using a bidirectional long short-term memory network and conditional random fields for entity recognition, and using a marginal attention mechanism and a temporal graph convolutional network to achieve dynamic expression of entity relationships and capture of evolutionary characteristics. It also introduces medical domain knowledge constraints to optimize entity semantic similarity calculations, and builds an efficient knowledge graph structure based on connectivity constraints and relationship weight thresholds.

[0228] The optimization module is used to dynamically adjust the knowledge structure using the graph optimization mechanism driven by reinforcement learning. By defining the graph state space, action space and reward function, the weights of the relationships between entities are optimized in real time, while cooperating with the graph pruning strategy to maintain the simplicity and effectiveness of the knowledge structure.

[0229] The early warning module is used to build a multi-level risk characterization model based on the optimized knowledge graph, and realize the accurate identification and early warning of nursing risks through local risk feature extraction, global risk information aggregation and time-series risk evolution analysis.

[0230] This embodiment also provides a computer device, which is suitable for the case of a nursing risk intervention decision-making method based on a knowledge graph, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the nursing risk intervention decision-making method based on a knowledge graph as proposed in the above embodiment.

[0231] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0232] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for implementing a nursing risk intervention decision-making method based on a knowledge graph as proposed in the above embodiment is implemented.

[0233] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0234] Example 2: This example provides a nursing risk intervention decision-making method based on a knowledge graph. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0235] This experiment conducted a 6-month nursing data analysis in the intensive care unit of a tertiary hospital, and the sample included nursing records of 1,000 patients. The experimental data covers 200,000 nursing record texts, vital signs data (blood pressure, heart rate, body temperature, etc.), 50,000 doctor's order execution records, 20,000 examination results and 10,000 nursing evaluation forms. The bidirectional long short-term memory network combined with the CRF combined model of the present invention was used for entity recognition, achieving an accuracy of 92.3%, a recall rate of 90.1% and an F1 value of 91.2%, which is 6.9% higher than the traditional method. In the relationship extraction link, the marginal attention mechanism achieved a relationship recognition accuracy of 88.7%, and the time series feature capture accuracy reached 85.6%, which is 8.5% higher than the traditional method.

[0236] Specifically, take an intensive care case as an example: Patient Zhang, a 65-year-old male, was admitted to the hospital due to acute myocardial infarction. The symptoms upon admission were severe chest pain, accompanied by sweating and dyspnea. Admission examination showed: BP 165 / 95mmHg, HR 92 times / min, T37.2℃, SpO2 92%, electrocardiogram showed ST segment elevation, and elevated troponin T. The system instantly identifies key entities including basic information, main symptoms, vital signs, and test results through a combined model, and uses the marginal attention mechanism to establish a complex relationship network between entities, including the symptom association "chest pain-association-acute myocardial infarction" with a weight of 0.85, the test confirmation "ST segment elevation-indication-myocardial ischemia" with a weight of 0.92, and the vital sign relationship "abnormal blood pressure-aggravated-cardiac load" with a weight of 0.78.

[0237] During the dynamic monitoring process, the system tracked and recorded in real time the change of the patient's blood pressure from 165 / 95 to 145 / 85 mmHg, the trend of the heart rate from 92 beats / min to 85 beats / min, and the improvement of blood oxygen saturation from 92% to 97%. Through the comprehensive analysis of these parameters, the system successfully warned of possible unstable heart function 3 hours in advance and predicted the potential risk of arrhythmia 2.5 hours in advance.

[0238] Through this typical case, we can clearly see that the dynamic medical care knowledge graph constructed by the present invention not only performs well in technical indicators, but more importantly, it shows significant practical value in actual clinical applications. The system can warn of 85% of potential risks 2 to 4 hours in advance. This intelligent nursing model based on knowledge graph provides medical institutions with an efficient, accurate and reliable nursing management solution, which truly improves the quality and optimizes the efficiency of nursing work.

[0239] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A nursing risk intervention decision-making method based on knowledge graph, characterized in that: include: Build a dynamic medical care knowledge graph based on multi-source nursing data and deep learning models, and use bidirectional long short-term memory networks and conditional random fields for entity recognition; The dynamic expression and evolution characteristics of entity relationships are captured through the marginal attention mechanism and the temporal graph convolutional network. The medical domain knowledge constraints are introduced to optimize the entity semantic similarity calculation, and the knowledge graph structure is optimized based on connectivity constraints and relationship weight thresholds. The knowledge structure is dynamically adjusted using the graph optimization mechanism driven by reinforcement learning. By defining the graph state space, action space and reward function, the weights of the relationships between entities are optimized in real time. At the same time, the graph pruning strategy is used to maintain the simplicity and effectiveness of the knowledge structure. Based on the optimized knowledge graph, a multi-level risk characterization model is constructed to achieve accurate identification and early warning of nursing risks through local risk feature extraction, global risk information aggregation and time-series risk evolution analysis. The multi-source nursing data includes nursing text data, patient physiological index data and expert knowledge base data; The nursing text data includes nursing records, doctor's order execution records, nursing assessment records, nursing handover records, nursing adverse event records and nursing quality monitoring records; The patient's physiological index data include real-time monitoring data of body temperature, blood pressure, heart rate, respiratory rate, blood oxygen saturation, blood sugar, weight, urine output, bowel movement frequency and consciousness state; The expert knowledge base data includes standard nursing specifications, expert experience summary, nursing case library, nursing guidelines, nursing technical operation procedures, nursing quality evaluation standards and nursing risk warning rules; The nursing text data is segmented and medical entity labeled, and each word is processed by the labeling function, which is expressed as: ； in, Represents the input word, Indicates the label category of the first position, Indicates the annotation category of the second position, Indicates The annotation categories of each position include the medical entity types of "disease", "symptom", "treatment", "examination", "drug", "surgery", "instrument", "nursing measures", "vital signs", "abnormal events", "risk level", "nursing outcome", "nursing time", "nursing staff" and "nursing part"; Generate the initial nursing knowledge entity set, expressed as: ； in, Represents the initial nursing knowledge entity set, represents the first nursing knowledge entity, Represents the second nursing knowledge entity, Indicates Nursing knowledge entities, is the total number of entities; Each entity They all contain entity name, entity type, entity occurrence location, entity frequency, and entity timestamp; Constructing a feature vector matrix based on the initial nursing knowledge entity set , extracting semantic features through context windows and introducing medical domain knowledge constraints to improve The model calculates the entity semantic similarity of each entity, expressed as: ； in, and Represent entities and The characteristic vector of represents the vector norm; Organize the semantic similarities between all entity pairs into a matrix form and construct a × Dimensional matrix, namely, entity semantic association matrix ; Improvement of optimization objective function by adding medical knowledge constraints Model, expressed as: ； in, For the original Loss function, is a set of medical knowledge constraint pairs, is the balance parameter, Represents the total loss function; The entity semantic association matrix Input into the marginal attention network, adaptively learn the importance weights between entities through the attention mechanism, and calculate the initial association weight matrix between nursing knowledge entities ; Based on the initial association weight matrix A temporal graph convolution model is established to capture the knowledge evolution characteristics by combining time information and graph structure information. The feature extraction formula of the temporal graph convolution model is expressed as: ； in, , is the adjacency matrix after adding the self-loop, is the original adjacency matrix; for dimensional identity matrix, Indicates The node feature matrix of the layer, Indicates The node feature matrix of the layer, for The corresponding degree matrix, Indicates The learnable weight matrix of the layer, represents the activation function, is the time coding matrix; Furthermore, the time coding matrix The elements in are represented as: ； in, Represents the time coding matrix No. OK Column elements, For Node The corresponding timestamp, is the frequency parameter, is the phase parameter; Extract the original features of the entity directly from the nursing documents to form the initial feature matrix , the initial feature matrix Input to the temporal graph convolution model for feature extraction and compare with the time encoding matrix Perform temporal convolution to obtain the output feature matrix after being processed by the temporal graph convolution model ; The output feature matrix The data is input to the named entity recognition module, which uses a bidirectional long short-term memory network structure to improve the accuracy of entity recognition through forward and backward information flow; The final output layer of the bidirectional long short-term memory network uses a conditional random field Perform sequence labeling, expressed as: ； in, Indicates that given an input sequence Under the condition, output the label sequence The conditional probability of is the normalization factor, Represents the potential function, calculating the current annotation and the previous annotation The transition probability, represents the sequence length, express Marking of time, express Marking of time, represents the input sequence; pass Perform sequence annotation to identify new entities from the text. The newly identified entities will be added to the original nursing knowledge entity set to form a new nursing knowledge entity set. .

2. The nursing risk intervention decision-making method based on knowledge graph according to claim 1, characterized in that: Based on the initial nursing knowledge entity set , Added nursing knowledge entity set and the initial association weight matrix Building the initial knowledge graph , where the node set , represents the nursing knowledge entity; The constructed initial knowledge graph satisfies the connectivity constraints: ； in, represents any two nodes in the initial knowledge graph, Represents a collection of nodes. Indicates that there is a path between nodes; By relationship weight threshold Control the density of the knowledge graph, expressed as: ； in, Representation Node and The edge between represents the relationship weight between any two nodes, Represents the weight threshold.

3. The nursing risk intervention decision-making method based on knowledge graph according to claim 2, characterized in that: The number of entity nodes in the initial knowledge graph , weight of relationship between entities And the graph density is mapped to a state vector, expressed as: ； in, represents the state vector, represents the total number of initial knowledge graph nodes, is the weight threshold for controlling the density of the graph; Importance scoring of entity nodes based on nursing entities Clinical relevance Constructing the action space , the importance score calculation formula is: ； in, Representation Node The entity importance score of Represents an entity node The degree characteristics of Representation Node The centrality characteristics of Represents an entity node The frequency of occurrence in clinical data, , and is the weight coefficient of each feature; The action space Also includes relationship weight adjustment and node addition and deletion operations, It is expressed as: ； in, is the weight adjustment step size, Representation Node The entity importance score of , and is the weight coefficient, Indicates clinical relevance; Based on the semantic similarity between the new entities and the existing entities in the nursing documents Establish a reward function, expressed as: ； in, Indicates that at time step The total reward value, is the indicative function, is the similarity threshold, Indicates a new entity With existing entities The semantic similarity of and are positive and negative reward values, respectively; Based on the calculation results of the reward function, through the depth The network calculates the value of each action, expressed as: ； in, Indicates the current state Next action The estimated value of and is the weight parameter of temporal difference learning and satisfies + = 1; To perform actions The next state after Represents a set of optional actions for the next state; The network loss function is defined as: ； in, represents the loss value of network training, represents the expected calculation, Indicates the target value, Indicates the current forecast value; Based on the current state choose The action with the largest value executes the selected action and obtains the new state and rewards , according to the loss function renew The network parameters are iteratively optimized until the graph structure is stable, and the optimized relationship weight matrix is ​​obtained. ; Based on the accessibility between nursing entities Prune the graph and define the path score, which is expressed as: ； in, Represents the entity arrive The path quality score is is the weight of the edge on the path, Represents a connected entity and A complete path of is the path length, Represents a single edge on a path; When the path length exceeds the preset threshold If there is an alternative path, delete Minimum redundant edges, preset threshold The calculation is expressed as: ； in, is the maximum acceptable path length threshold, is the total number of nodes in the current graph.

4. The nursing risk intervention decision-making method based on knowledge graph according to claim 3, characterized in that: Based on the optimized knowledge graph, a multi-level risk representation model is constructed, including the attention layer, knowledge fusion layer, graph neural network layer, feature aggregation layer, and prediction layer; The attention layer is based on the optimized knowledge graph The collection of nursing entity nodes in and the relationship weight matrix , for each entity node A multi-layer attention mechanism is applied to consider the local semantic environment and clinical attributes of the node to obtain the local risk feature vector, which is expressed as: ； in, Representation Node The local risk feature vector of Representation Node The set of neighboring nodes of and Respectively represent the current central node and neighboring nodes The entity importance score of represents the feature transformation matrix, is the attention weight coefficient; For the optimized knowledge graph Each entity node in Build Hop-Neighborhood Subgraph , which contains possible risk propagation paths, combined with the updated relationship weight matrix , the structured risk pattern vector is extracted through a multi-layer graph convolutional network, which is expressed as: ； in, Representation Node In the The risk pattern vector of the layer, For Node of The set of hop neighbor nodes, is the normalization coefficient, For Node of The set of hop neighbor nodes, For the The learnable weight matrix of the layer, is the activation function, Indicated in In the layer, nodes The risk pattern vector of go through The final risk pattern vector obtained after layer graph convolution , the final risk pattern vector is concatenated with the local risk feature vector to form an enhanced risk feature vector, which is expressed as: ； in, represents the enhanced risk feature vector; Based on the acquisition of enhanced risk feature vector Construct a multi-layer risk propagation network and use a multi-head attention mechanism to model the complex correlation between different risk types. Attention heads are used to calculate the correlation strength matrix, which can be expressed as: ； in, Indicates The output matrix of the attention head is yes The query matrix represents The query matrix of the attention heads; yes The key matrix represents The key matrix of the attention heads; yes The value matrix represents The value matrix of the attention heads; is the scaling factor; The outputs of multiple heads are integrated by splicing, expressed as: ； in, represents the final output matrix, Represents the output matrix of each attention head, Represents the output projection matrix.

5. The nursing risk intervention decision-making method based on knowledge graph according to claim 4, characterized in that: The knowledge fusion layer combines the clinical expert score set Adjust the strength of the association: ； in, represents the association strength matrix after expert knowledge adjustment, represents the set of clinical expert ratings; The graph neural network layer is based on the adjusted correlation strength matrix Construct a risk propagation network and model the risk propagation process as message passing in a graph neural network. The risk transfer of the layer is expressed as: ； in, For Node In the The risk propagation status of the layer, For the The parameter matrix of the layer; Representation Node The set of neighboring nodes of is the association strength after adjustment by expert knowledge, indicating that the node and The intensity of risk transmission between Represents the neighboring nodes The risk propagation status in the previous layer; The feature aggregation layer integrates risk information from different layers through skip connections, expressed as: ； in, is the learnable aggregation matrix, is the number of network layers, Representation Node The final aggregate risk representation of The prediction layer calculates the final score based on the aggregated risk vector, expressed as: ； in, and is the prediction layer parameter, is the weight matrix, is the bias term, is the activation function, Representation Node The final risk score of Decision intervention is performed based on the output of the prediction layer, including: If the prediction output shows that the correlation between symptom aggravation and treatment response is abnormal, the system will associate the symptom observation record, match the treatment plan execution status, and push the corresponding nursing measure adjustment decision; If the prediction output indicates that disease progression is associated with changes in vital signs, the severity of the disease is analyzed, the dynamic changes in vital signs are tracked, and the trend of nursing outcomes is further predicted; If the prediction output identifies abnormal correlations with post-medication test results, the medication records are traced back, the changes in the associated test results are correlated, and an abnormal event warning is generated; If the prediction output captures the timing association between surgical operations and instrument use, then the nursing time nodes are planned, the instrument preparation list is matched, and a perioperative nursing plan is developed; If the prediction output reflects the association between nursing operations and area risks, nursing staff responsibilities are assigned, key nursing areas are identified, and risk prevention measures are determined.

6. A nursing risk intervention decision system based on a knowledge graph, based on the nursing risk intervention decision method based on a knowledge graph according to any one of claims 1 to 5, characterized in that: include: A construction module for building a dynamic medical care knowledge graph based on multi-source nursing data and deep learning models, using a bidirectional long short-term memory network and conditional random fields for entity recognition, and using a marginal attention mechanism and a temporal graph convolutional network to achieve dynamic expression of entity relationships and capture of evolutionary characteristics. It also introduces medical domain knowledge constraints to optimize entity semantic similarity calculations, and builds an efficient knowledge graph structure based on connectivity constraints and relationship weight thresholds. The optimization module is used to dynamically adjust the knowledge structure using the graph optimization mechanism driven by reinforcement learning. By defining the graph state space, action space and reward function, the weights of the relationships between entities are optimized in real time, while cooperating with the graph pruning strategy to maintain the simplicity and effectiveness of the knowledge structure. The early warning module is used to build a multi-level risk characterization model based on the optimized knowledge graph, and realize the accurate identification and early warning of nursing risks through local risk feature extraction, global risk information aggregation and time-series risk evolution analysis.

Images