Hospital diagnosis and treatment quality real-time supervision method based on big data analysis

Abstract

The invention discloses a hospital diagnosis and treatment quality real-time supervision method based on big data analysis, and the method comprises the following steps: collecting and processing internal multi-source diagnosis and treatment behavior data of a hospital, and obtaining structured diagnosis and treatment behavior sequence data; constructing a diagnosis and treatment behavior graph based on the structured diagnosis and treatment behavior sequence data; obtaining a node initial representation vector based on the diagnosis and treatment behavior graph; based on the node initial representation vector and the diagnosis and treatment behavior graph structure, constructing a behavior perception graph attention network, and obtaining a node embedding representation set of the diagnosis and treatment behavior graph; based on the node embedding representation set, comparing the node embedding representation set with a standard diagnosis and treatment path diagram to obtain an abnormal diagnosis and treatment behavior node set and a corresponding abnormal type label; generating an abnormal diagnosis and treatment behavior path diagram based on the abnormal diagnosis and treatment behavior node set; and pushing the abnormal diagnosis and treatment behavior path diagram and the abnormal type label to a hospital quality control system or a medical management system. The behavior perception graph attention network is adopted, and real-time monitoring of hospital diagnosis and treatment quality is achieved.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a method for real-time monitoring of hospital diagnosis and treatment quality based on big data analysis. Background Technology

[0002] With the continuous development of medical information systems, hospitals generate a large amount of structured and unstructured behavioral data during the diagnosis and treatment process, including medical orders, examinations and tests, surgical procedures, nursing interventions, and admission and discharge management. This data is recorded in the hospital information platform according to time sequence, operator, and department affiliation, providing a foundation for the digital management of medical quality. Existing methods for monitoring medical quality largely rely on manual sampling, static rule matching, or post-event traceability, which are insufficient to cover high-frequency and complex medical behavioral paths, and also cannot achieve real-time detection and response to abnormal medical behaviors.

[0003] Some studies have attempted to introduce graph structures to model medical behavior sequences, using graph neural networks to mine potential relationships between behavioral nodes and identify potential risks. However, these methods are typically based on standard graph convolutional networks or general graph attention mechanisms, making it difficult to effectively capture fine-grained feature variations between medical behaviors caused by differences in intent, abnormal risks, or contextual semantics. Furthermore, existing solutions often focus on node classification or edge prediction when identifying abnormal paths, lacking modeling and evaluation of the complete path structure in the medical behavior graph. This leads to insufficient identification of complex abnormal patterns such as logical jumps, missing key nodes, or conflicting behavioral sequences, affecting the real-time performance and accuracy of medical monitoring.

[0004] Therefore, how to provide a real-time monitoring method for hospital treatment quality based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a real-time monitoring method for hospital medical quality based on big data analysis. This invention employs a behavior-aware graph attention network to model the medical behavior graph, integrating behavior type, time information, intent labels, and abnormal risk features to achieve high-precision modeling of the medical behavior path and intelligent identification of abnormal behaviors. By constructing the medical behavior graph, extracting node features, introducing an intent difference control mechanism and an anomaly-sensitive gating structure, this invention generates node embedding representations and compares them with standard medical paths. Ultimately, it achieves accurate identification and real-time feedback of abnormal patterns such as non-standard path jumps, missing key nodes, and conflicting behavior sequences. It possesses advantages such as high granularity of anomaly identification, fine-grained graph structure modeling, and adaptability to the characteristics of multi-source medical data.

[0006] The real-time monitoring method for hospital medical treatment quality based on big data analysis according to an embodiment of the present invention includes the following steps: Collect multi-source diagnosis and treatment behavior data within the hospital and perform data cleaning, missing value imputation, timestamp alignment, and behavior type labeling to obtain structured diagnosis and treatment behavior sequence data; A diagnosis and treatment behavior graph is constructed based on structured diagnosis and treatment behavior sequence data; For each node in the diagnosis and treatment behavior graph, extract node features and perform unified encoding to obtain the initial node representation vector; Based on the initial node representation vector and the diagnosis and treatment behavior graph structure, a behavior-aware graph attention network is constructed. Feature aggregation is performed on each node in the diagnosis and treatment behavior graph to obtain the node embedding representation set of the diagnosis and treatment behavior graph. Based on the node embedding representation set, it is compared with the standard embedding representation set of the corresponding node in the standard diagnosis and treatment path diagram to obtain the node set of abnormal diagnosis and treatment behavior and the corresponding abnormal type label; Based on the set of abnormal diagnosis and treatment behavior nodes, and combined with attention weight information and path structure features, an abnormal diagnosis and treatment behavior path graph is generated. The abnormal diagnosis and treatment behavior path map and abnormal type label are pushed to the hospital quality control system or medical management system to trigger the early warning mechanism and record abnormal event information.

[0007] Optionally, the structured diagnosis and treatment behavior sequence data includes the original diagnosis and treatment behavior data after cleaning, completion, alignment and type labeling; the original diagnosis and treatment behavior data includes medical order record data, examination and test data, surgical record data, nursing record data and admission and discharge record data.

[0008] Optionally, the construction of the diagnosis and treatment behavior graph based on structured diagnosis and treatment behavior sequence data specifically includes: All medical behavior events are extracted from the structured medical behavior sequence data to form a set of medical behavior events; Based on the set of diagnosis and treatment behavior events, a set of nodes for the diagnosis and treatment behavior graph is constructed, each diagnosis and treatment behavior is associated with a graph node, and initial feature information is extracted for each node, combined and encoded to form the initial feature vector of the node; Based on the temporal order of each behavior in the set of diagnosis and treatment behavior events and the dependencies in the diagnosis and treatment process, the connection relationships between nodes are determined, and the set of edges in the diagnosis and treatment behavior graph is generated. For each edge in the edge set, the time interval between actions is calculated based on the time difference between the starting node and the target node connected by the edge. The edge feature information is generated by combining the process dependency relationship type between the corresponding diagnosis and treatment actions of the two nodes. All edge feature information forms the edge feature set. The set of nodes, the set of edges, the set of initial feature vectors of nodes, and the set of feature vectors of edges are integrated to form a diagnostic behavior graph.

[0009] Optionally, the step of extracting node features and performing unified encoding processing on each node in the diagnosis and treatment behavior graph to obtain the initial node representation vector specifically includes: Based on each node in the diagnostic behavior graph, a behavior type vector, a time interval encoding vector, and an intent embedding vector are generated. Based on the graph structure of the diagnosis and treatment behavior graph, the predecessor and successor nodes associated with each node are identified, and the initial feature representations of the adjacent nodes and the edge feature information between them and the current node are extracted to obtain the behavior context combination features of the current node, forming a behavior context feature set. The behavior type vector, time interval encoding vector, intent embedding vector and context combination features in the node feature set are concatenated to generate the fused feature vector of each node, thus obtaining the fused feature set. A unified linear transformation and nonlinear activation process is applied to the fused feature set to obtain the initial representation vector of each node, forming a set of initial representation vectors for the nodes.

[0010] Optionally, the step of constructing a behavior-aware graph attention network based on the initial node representation vector and the diagnosis and treatment behavior graph structure, and performing feature aggregation on each node in the diagnosis and treatment behavior graph to obtain the node embedding representation set of the diagnosis and treatment behavior graph specifically includes: The initial node representation vectors of each target node and its neighboring nodes in the diagnostic behavior graph are input into the behavior-aware graph attention network, which includes an intent difference regulation attention unit, a path decay regulation unit, and an anomaly sensitivity gating unit. In the intent difference-modulated attention unit, based on the degree of difference between the intent embedding vectors of the target node and its neighboring nodes, the attention adjustment factor is calculated, and the attention weights of the neighboring nodes to the target node are adjusted to obtain the intent difference attention weight set. In the path attenuation adjustment unit, the path attenuation factor is calculated based on the path depth information in the graph between each target node and its neighboring nodes, resulting in a set of path attenuation factors. In the anomaly-sensitive gating unit, for each target node, an anomaly risk score is calculated based on the frequency of anomalies identified in historical behavior, and an anomaly-sensitive gating factor is generated. The set of intent difference attention weights, the set of path decay factors, and the set of anomaly sensitivity gating factors are weighted and fused to obtain the embedded representation of the target node; By integrating the embedded representations of each target node in the diagnosis and treatment behavior graph, we obtain the node embedded representation set of the diagnosis and treatment behavior graph.

[0011] Optionally, in the intent difference-controlled attention unit, the attention adjustment factor is calculated based on the degree of difference between the intent embedding vectors of the target node and its neighboring nodes, and the attention weights of the neighboring nodes to the target node are adjusted to obtain the intent difference attention weight set, specifically including: Obtain the behavioral intent embedding vectors of the target node and its neighboring nodes from the treatment behavior graph, and calculate the intent difference between the target node and its neighboring nodes. Obtain the initial node representation vectors of the target node and its neighboring nodes, encode the node representations through a linear transformation function, concatenate them, and input them into the attention scoring function. Combine the intention difference to calculate the adjusted attention score. Softmax normalization is performed on the attention scores of all neighboring nodes to obtain the intent difference attention weights. The attention weights corresponding to all neighboring nodes are combined to form the intent difference attention weight set.

[0012] Optionally, in the anomaly-sensitive gating unit, for each target node, calculating an anomaly risk score based on the frequency of anomalies identified in historical behavior, and generating an anomaly-sensitive gating factor specifically includes: For each target node in the diagnosis and treatment behavior graph, count the number of times the target node participates in abnormal behavior paths in the historical diagnosis and treatment behavior graph, obtain the total number of times the target node appears in all historical diagnosis and treatment behavior graphs, and calculate the abnormal frequency component of the target node. Based on the set of neighboring nodes of the target node, the anomaly prediction probability of the neighboring nodes is extracted, the neighborhood average anomaly probability component is calculated, and the anomaly frequency component is combined with a weighted sum to obtain the anomaly risk score of the target node. The abnormal risk score is input into the abnormality sensitive gating function to generate the abnormality sensitive gating factor.

[0013] Optionally, the comparison of the node embedding representation set with the standard embedding representation set of corresponding nodes in the standard diagnosis and treatment path diagram to obtain the abnormal diagnosis and treatment behavior node set and corresponding abnormality type label specifically includes: A standard diagnosis and treatment path diagram is constructed, and feature extraction and embedding calculation are performed on each standard node to obtain a set of standard embedding representations; Based on the node embedding representation set and the standard embedding representation set, the representation offset between the target node and the corresponding node in the standard diagnosis and treatment path diagram is calculated to obtain the node representation offset set. Based on each offset value in the node representation offset set, a representation offset threshold is set. When the representation offset of the target node is greater than the offset threshold, it is marked as a behavioral representation aberration node, and all marked nodes are aggregated to form a behavioral representation aberration node set. Based on the edge information in the diagnosis and treatment behavior graph, a structural comparison is made with the standard diagnosis and treatment path graph. If the diagnosis and treatment behavior graph is missing a key node from the standard diagnosis and treatment path graph, the missing node is included in the set of missing key nodes. If there is a behavioral path in the diagnosis and treatment behavior graph that jumps to a non-adjacent standard node, then the starting node of the path is included in the set of abnormal path jump nodes. If the execution order of a behavior node conflicts with the standard treatment path diagram, the corresponding node will be included in the set of behavior order conflict nodes. The system integrates the set of abnormal nodes, the set of missing key nodes, the set of abnormal nodes for path jumps, and the set of nodes with conflicting behavior order to construct a set of abnormal diagnosis and treatment behavior nodes, and generates a set of corresponding abnormal type labels based on the source of the abnormality.

[0014] Optionally, the step of generating an abnormal diagnosis and treatment behavior path graph based on the set of abnormal diagnosis and treatment behavior nodes, combined with attention weight information and path structure features, specifically includes: Based on the set of abnormal diagnosis and treatment behavior nodes, each abnormal node is located in the diagnosis and treatment behavior graph, the predecessor and successor nodes and connecting edges of the abnormal node are extracted, the corresponding path context subgraph is constructed, and all path context subgraphs are summarized into a path context subgraph set. For each path context subgraph in the path context subgraph set, extract the attention weights between the target abnormal node and its neighboring nodes in the behavior-aware graph attention network, and calculate the context attention aggregation vector. The anomaly score of the target anomaly node is calculated by combining the initial representation vector of the target anomaly node with its embedding representation in the node embedding representation set. Based on the path context subgraph set and anomaly score, the differences between the current path of the target abnormal node and the standard diagnosis and treatment path are analyzed, and the trigger condition description of each abnormal node is generated. Based on the contextual attention representation features, anomaly scores, and anomaly triggering conditions, an abnormal diagnosis and treatment behavior path diagram is constructed. Optionally, the step of pushing the abnormal diagnosis and treatment behavior path map and abnormal type labels to the hospital quality control system or medical management system to trigger the early warning mechanism and record abnormal event information specifically includes: The abnormal diagnosis and treatment behavior path map set and abnormal type labels are pushed to the hospital quality control system and medical management system respectively. In the hospital quality control system, based on the abnormality type label, the corresponding early warning rule is matched to generate an early warning notification data packet, which is then sent to the designated department administrators. In the medical management system, abnormal diagnosis and treatment behavior path diagrams and abnormal type labels are recorded to the quality control log module to establish a structured abnormal event record; Based on the early warning notification data packet, an intervention task item is generated and assigned to the designated person in charge. At the same time, the task status information is updated and synchronously transmitted back to the abnormal event record. The system continuously monitors changes in the responsible physician's behavior throughout the diagnosis and treatment process. If any abnormal node status is eliminated, the closed-loop process is considered complete, and a completion marker is added.

[0015] The beneficial effects of this invention are: This invention proposes a real-time monitoring method for hospital medical quality based on big data analytics. Addressing the limitations of traditional medical behavior analysis methods, such as coarse data granularity, weak anomaly detection capabilities, and limited behavioral modeling abilities, this method constructs a medical behavior graph with medical behaviors as nodes and logical relationships as edges. This enables structured modeling of complex medical processes within hospitals. By uniformly integrating multi-source data including medical orders, examinations, surgeries, nursing care, and admissions / discharges, it comprehensively and meticulously reflects the actual medical behavior process, providing a clear and semantically rich foundation for subsequent abnormal behavior identification.

[0016] Based on the aforementioned diagnostic behavior graph, this invention designs a graph attention network model with behavior perception capabilities. It introduces an intent difference regulation mechanism and an anomaly-sensitive gating structure, enabling the model to consider both the intent correlation between behaviors and the historical distribution of abnormal risks when calculating attention weights. This improves the accuracy of node feature aggregation and context awareness. Compared to traditional graph neural networks, the method of this invention is more effective in capturing potential behavioral deviations and non-standard process features in the diagnostic path, enhancing the model's sensitivity to abnormal behavior identification.

[0017] After identifying abnormal medical behaviors, this invention further generates an abnormal behavior path diagram. Combining attention distribution and contextual path information, it classifies the abnormality type and determines the triggering mechanism. This information is then systematically output to the hospital's quality control system or medical management platform, achieving real-time early warning, information recording, and a closed-loop regulatory feedback mechanism for abnormal behaviors. This mechanism effectively assists medical administrators in identifying and responding to potential process violations, operational errors, or system risks, improving the standardization and safety of the hospital's overall medical processes. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart of the real-time monitoring method for hospital treatment quality based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of the construction of the treatment behavior graph and the node feature extraction structure in the real-time monitoring method for hospital treatment quality based on big data analysis proposed in this invention; Figure 3 This is a schematic diagram of the internal structure of the behavior-aware graph attention network in the real-time monitoring method for hospital diagnosis and treatment quality based on big data analysis proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A real-time monitoring method for hospital treatment quality based on big data analysis includes the following steps: Collect multi-source diagnosis and treatment behavior data within the hospital and perform data cleaning, missing value imputation, timestamp alignment, and behavior type labeling to obtain structured diagnosis and treatment behavior sequence data; A diagnosis and treatment behavior graph is constructed based on structured diagnosis and treatment behavior sequence data; For each node in the diagnosis and treatment behavior graph, extract node features and perform unified encoding to obtain the initial node representation vector; Based on the initial node representation vector and the diagnosis and treatment behavior graph structure, a behavior-aware graph attention network is constructed. Feature aggregation is performed on each node in the diagnosis and treatment behavior graph to obtain the node embedding representation set of the diagnosis and treatment behavior graph. Based on the node embedding representation set, it is compared with the standard embedding representation set of the corresponding node in the standard diagnosis and treatment path diagram to obtain the node set of abnormal diagnosis and treatment behavior and the corresponding abnormal type label; Based on the set of abnormal diagnosis and treatment behavior nodes, and combined with attention weight information and path structure features, an abnormal diagnosis and treatment behavior path graph is generated. The abnormal diagnosis and treatment behavior path map and abnormal type label are pushed to the hospital quality control system or medical management system to trigger the early warning mechanism and record abnormal event information.

[0022] In this embodiment, the process of collecting multi-source medical behavior data within the hospital and performing data cleaning, missing value imputation, timestamp alignment, and behavior type labeling to obtain structured medical behavior sequence data specifically includes: Raw data of diagnosis and treatment behavior are collected from the hospital information system to construct a raw dataset of diagnosis and treatment behavior. The raw dataset of diagnosis and treatment behavior includes a medical order record dataset, an examination and test dataset, a surgical record dataset, a nursing record dataset, and an admission and discharge record dataset. Data cleaning is performed on the original dataset of medical behavior to remove null values, duplicate records and logically conflicting records, and all time fields are uniformly converted into standard timestamp format to obtain the cleaned dataset of medical behavior. For missing behavioral items in the cleaned medical behavior dataset, linear interpolation is used to fill in the missing values, resulting in a complete medical behavior dataset. All diagnostic and treatment behaviors in the completed diagnostic and treatment behavior dataset are globally sorted according to their timestamps, and the behaviors are aligned based on the time of occurrence to construct a behavior time series. Each behavior item is bound to its corresponding timestamp to form an ordered behavior pair sequence, resulting in a behavior time series data structure. For each diagnosis and treatment behavior in the behavioral time series data structure, it is classified and labeled according to the semantic content, execution purpose, process position and source system attributes of the diagnosis and treatment behavior, and each diagnosis and treatment behavior is assigned a corresponding behavior type label. The behavior type label is selected from a predefined set of behavior categories, which includes examination, diagnosis, treatment, management and follow-up. By integrating behavioral time-series data structures with behavioral type labels, structured diagnostic and treatment behavioral sequence data is formed.

[0023] In this embodiment, constructing a diagnosis and treatment behavior graph based on structured diagnosis and treatment behavior sequence data specifically includes: All diagnosis and treatment behavior events are extracted from structured diagnosis and treatment behavior sequence data. Each diagnosis and treatment behavior event includes a behavior identifier, behavior occurrence time, behavior type label, executing department identifier, and operator identifier, forming a set of diagnosis and treatment behavior events. Based on each medical behavior in the set of medical behavior events, a set of nodes for the medical behavior graph is constructed, with each medical behavior corresponding to a graph node. Initial feature information is extracted for each node. The initial features include behavior identifier, behavior type, behavior occurrence time, executing department identifier, and operator identifier. The above features are combined and encoded to form the initial feature vector of the node. Based on the temporal order of each behavior in the set of diagnosis and treatment behavior events and the dependencies in the diagnosis and treatment process, the connection relationship between nodes is determined. When a diagnosis and treatment behavior is earlier than another behavior in time, and there is a causal or process connection relationship between the two, a directed edge is constructed from the node corresponding to the former to the node corresponding to the latter in the diagnosis and treatment behavior graph. The edge set in the diagnosis and treatment behavior graph is generated in this way. For each edge in the edge set, the time interval between actions is calculated based on the time difference between the starting node and the target node connected by the edge. The edge feature information is generated by combining the process dependency relationship type between the corresponding diagnosis and treatment actions of the two nodes. The edge feature includes the time interval value and the logical association type encoding. All edge feature information constitutes the edge feature set. By integrating the set of nodes, the set of edges, the set of initial feature vectors of nodes, and the set of feature vectors of edges, a directed graph containing node attributes and edge attributes is constructed, forming a diagnosis and treatment behavior graph.

[0024] In this embodiment, the step of extracting node features and performing unified encoding processing on each node in the diagnosis and treatment behavior graph to obtain the initial node representation vector specifically includes: Based on each node in the diagnosis and treatment behavior graph, extract the behavior type label, behavior timestamp, department identifier, operator identifier and behavior intention label corresponding to the node, and generate behavior type vector, time interval encoding vector and intention embedding vector respectively. Among them, the time interval encoding vector is obtained by processing the time difference between the current behavior and the previous behavior through the time encoding function, and finally form the basic feature set of the node. Based on the graph structure of the diagnosis and treatment behavior graph, the predecessor and successor nodes associated with each node are identified, and the initial feature representations of these neighboring nodes and the edge feature information between them and the current node are extracted respectively. The feature and edge feature of the neighboring nodes are combined and calculated using the aggregation function to obtain the behavior context combination feature of the current node, forming a behavior context feature set. The behavior type vector, time interval encoding vector, and intent embedding vector in the node feature set are concatenated with the context combination features in the behavior context feature set to generate the fused feature vector for each node, thus obtaining the fused feature set. The fused feature set is subjected to a unified linear transformation and nonlinear activation process. Specifically, each fused feature vector is input into the encoding function, first linearly mapped through the weight matrix and bias term, and then nonlinearly transformed using the activation function to obtain the initial representation vector of each node, thus forming the initial representation vector set of the nodes.

[0025] In this embodiment, the step of constructing a behavior-aware graph attention network based on the initial node representation vector and the diagnosis and treatment behavior graph structure, and performing feature aggregation on each node in the diagnosis and treatment behavior graph to obtain the node embedding representation set of the diagnosis and treatment behavior graph specifically includes: The initial node representation vectors of each target node and its neighboring nodes in the diagnostic behavior graph are input into the behavior-aware graph attention network, which includes an intent difference regulation attention unit, a path decay regulation unit, and an anomaly sensitivity gating unit. In the intent difference-modulated attention unit, based on the degree of difference between the intent embedding vectors of the target node and its neighboring nodes, the attention adjustment factor is calculated, and the attention weights of the neighboring nodes to the target node are adjusted to obtain the intent difference attention weight set. In the path decay adjustment unit, a path decay factor is calculated for the path depth information in the graph between each target node and its neighboring nodes. The path depth represents the shortest directed path length from the target node to the neighboring node. The path decay factor decreases exponentially according to the set decay coefficient, reflecting the degree of influence decay caused by the increase of distance during information propagation. Finally, a set of path decay factors is obtained, which is used to adjust the contribution weight of each neighboring node to the embedded representation of the target node. In the anomaly-sensitive gating unit, for each target node, an anomaly risk score is calculated based on the frequency of anomalies identified in historical behavior, and an anomaly-sensitive gating factor is generated. The intention difference attention weight set, path decay factor set, and anomaly sensitivity gating factor set are weighted and fused, and applied to the feature representation of the target node's neighbor nodes to obtain the embedded representation of the target node. By integrating the embedded representations of each target node in the diagnosis and treatment behavior graph, we obtain the node embedded representation set of the diagnosis and treatment behavior graph.

[0026] In this embodiment, the intention difference-controlled attention unit calculates an attention adjustment factor based on the degree of difference in the intention embedding vectors between the target node and its neighboring nodes, and adjusts the attention weights of the neighboring nodes towards the target node to obtain the intention difference attention weight set. Specifically, this includes: Obtain the behavioral intent embedding vectors of the target node and its neighboring nodes from the diagnostic behavior graph, and calculate the intent difference between the target node and its neighboring nodes: ; in, This represents the degree of difference in intent between the target node and its neighboring nodes. This represents the intent embedding vector of the target node in the treatment behavior graph. This represents the intent embedding vector of a neighboring node in the treatment behavior graph. The Euclidean norm (L2 norm) is used to measure the distance between two vectors, that is, the degree of difference in behavioral intentions within the embedding space. This represents the index of the current target node in the treatment behavior graph. Indicates the index of the neighboring node adjacent to the target node; Obtain the initial node representation vectors of the target node and its neighboring nodes, encode the node representations using a linear transformation function, concatenate them, and input them into the attention scoring function. Calculate the adjusted attention score based on the intent difference. ; in, This represents the attention score that the target node gives to its neighboring nodes. This represents the transpose of the trainable attention weight vector. Represents the node feature transformation matrix. This represents the initial representation vector of the target node. This represents the initial representation vector of the neighboring nodes. To represent the adjustment coefficient for the intended difference; The above formula is derived from the structural evolution of the attention coefficient calculation formula in graph attention networks. This invention, targeting the "hospital treatment behavior" scenario, introduces behavioral intent embedding vectors and quantifies the intent differences between nodes, proposing a Behavior-aware Graph Attention Network (GAT) that introduces intent difference degree. And intention difference adjustment coefficient Cancel the nonlinear activation function; Softmax normalization is performed on the attention scores of all neighboring nodes to obtain the intent difference attention weights. The attention weights corresponding to all neighboring nodes are combined to form the intent difference attention weight set.

[0027] In this embodiment, the step of calculating anomaly risk scores and generating anomaly sensitive gating factors for each target node in the anomaly sensitive gating unit specifically includes: For each target node in the diagnosis and treatment behavior graph, count the number of times the target node participates in abnormal behavior paths in the historical diagnosis and treatment behavior graph, and record it as the historical abnormal frequency of the target node. Also, obtain the total number of times the target node appears in all historical diagnosis and treatment behavior graphs and calculate the abnormal frequency component of the target node. Based on the set of neighboring nodes of the target node, the anomaly prediction probability of the neighboring nodes is extracted, the neighborhood average anomaly probability component is calculated, and then weighted and summed with the anomaly frequency component to obtain the anomaly risk score of the target node: ; in, This represents the abnormal risk score of the target node in the diagnosis and treatment behavior graph. Indicates the historical frequency influence coefficient. This represents the abnormal frequency components of the target node. This represents the neighbor perception impact coefficient. This represents the set of neighboring nodes of a node in the diagnosis and treatment behavior graph. This represents the probability of anomaly prediction for neighboring nodes; The abnormal risk score is input into the abnormality sensitive gating function. First, it is linearly mapped with the preset weight matrix and a bias term is added. Then, it is nonlinearly transformed by the Sigmoid activation function to generate the abnormality sensitive gating factor corresponding to the target node.

[0028] In this embodiment, the step of comparing the node embedding representation set with the standard embedding representation set of corresponding nodes in the standard diagnosis and treatment path diagram to obtain the abnormal diagnosis and treatment behavior node set and the corresponding abnormal type label specifically includes: A standard diagnosis and treatment path graph is constructed. The standard diagnosis and treatment path graph is a directed graph structure, in which each standard node represents a standardized diagnosis and treatment behavior, and each standard edge represents the sequential relationship between standard behavior processes. Feature extraction and embedding calculation are performed on each standard node to obtain a set of standard node embedding representations. Based on the node embedding representation set and the standard embedding representation set, for each target node in the diagnosis and treatment behavior graph, the representation offset between the target node and the corresponding node in the standard diagnosis and treatment path graph is calculated. Specifically, the node representation offset is obtained by calculating the Euclidean distance between the two embedding representations, and the representation offsets of all nodes are summarized to obtain the node representation offset set. Based on each offset value in the node representation offset set, a representation offset threshold is set. When the representation offset of a target node is greater than the offset threshold, the node is marked as a behavior representation aberration node, and all marked nodes are aggregated to form a behavior representation aberration node set. Based on the edge information in the diagnosis and treatment behavior graph, a structural comparison is performed with the standard diagnosis and treatment path graph to identify whether the following structural anomalies exist: If a key node in the standard treatment path diagram is missing from the treatment behavior diagram, the missing node will be included in the key node missing set. If there is a behavioral path in the diagnosis and treatment behavior graph that jumps to a non-adjacent standard node, then the starting node of the path is included in the set of abnormal path jump nodes. If the execution order of a behavior node conflicts with the standard treatment path diagram, the corresponding node will be included in the set of behavior order conflict nodes. The system integrates the set of abnormal nodes, the set of missing key nodes, the set of abnormal nodes for path jumps, and the set of nodes with conflicting behavior order to construct a set of abnormal diagnosis and treatment behavior nodes, and generates a set of corresponding abnormal type labels based on the source of the abnormality.

[0029] In this embodiment, generating an abnormal diagnosis and treatment behavior path graph based on the set of abnormal diagnosis and treatment behavior nodes, combined with attention weight information and path structure features, specifically includes: Based on the set of abnormal diagnosis and treatment behavior nodes, each abnormal node is located in the diagnosis and treatment behavior graph, the predecessor and successor nodes and their connecting edges of the abnormal node are extracted, the corresponding path context subgraph is constructed, and the path context subgraphs corresponding to all abnormal nodes are summarized into a path context subgraph set. For each path context subgraph in the path context subgraph set, extract the attention weights between the target abnormal node and its neighboring nodes in the behavior-aware graph attention network. Combine the connection edge type and node type to calculate the context attention aggregation vector, which serves as the attention representation feature reflecting the local path dependency relationship. The initial representation vector of the target abnormal node is combined with the embedded representation in the node embedding representation set. The vector distance between the two is calculated as the abnormal score of the target abnormal node. The vector distance adopts the Euclidean distance metric, which reflects the degree of semantic feature shift of the diagnosis behavior before and after the behavior-aware graph attention network modeling. The larger the abnormal score, the more significant the deviation of the node's behavior features from the standard. Based on the path context subgraph set and anomaly score, the differences between the current path of the target abnormal node and the standard diagnosis and treatment path are analyzed, the key path offset position, behavior sequence conflict and key node missing are extracted, and the trigger condition description of each abnormal node is generated. Based on the contextual attention representation features, abnormal scores, and abnormal triggering conditions, an abnormal diagnosis and treatment behavior path diagram is constructed.

[0030] In this embodiment, the step of pushing the abnormal diagnosis and treatment behavior path map and abnormal type labels to the hospital quality control system or medical management system, triggering the early warning mechanism and recording abnormal event information specifically includes: The abnormal diagnosis and treatment behavior path map set and abnormal type labels are used as input data. The information push interface is called to push the data to the hospital quality control system and medical management system respectively. In the hospital quality control system, based on the abnormality type label, the corresponding early warning rule is matched to trigger the abnormality early warning process, generate an early warning notification data package including abnormal node identifier, abnormal score, path structure, trigger logic and recommended treatment suggestions, and send it to the designated department management personnel through the internal message channel; In the medical management system, abnormal diagnosis and treatment behavior path diagrams and abnormal type labels are recorded to the quality control log module, and the occurrence time of abnormal behavior, responsible person information and department to which it belongs are associated to establish a structured abnormal event record; Based on the recommended handling suggestions contained in the early warning notification data packet, subsequent intervention task items are generated and automatically assigned to the designated responsible persons by the system. At the same time, the task status information is updated and synchronously transmitted back to the abnormal event record. In subsequent diagnosis and treatment processes, the changes in the responsible physician's behavioral nodes are continuously monitored, and the new nodes are embedded and compared with the previous abnormal diagnosis and treatment behavior path diagram. If the abnormal node status is eliminated, it is determined that the closed-loop processing is completed, and the processing completion mark is marked in the system.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the information system of a comprehensive hospital for real-time quality monitoring of inpatient treatment processes. Prior to actual deployment, the hospital's quality control department reported that current rule-based quality control methods suffer from significant delays and high false negative rates. This is particularly problematic in multidisciplinary collaborative treatment pathways, where complex processes make it difficult to cover all quality control nodes, often resulting in missed treatment steps or incorrect timing of actions that cannot be identified in a timely manner, thus impacting treatment quality and patient safety.

[0032] In the implementation of this invention, multiple system interfaces of the hospital information system were first accessed, including medical orders, laboratory examinations, surgeries, nursing care, and admission / discharge records. Raw data on the treatment behaviors of over 16,000 inpatients were collected over two consecutive months, totaling over 48 million data entries. After data cleaning, behavior annotation, and time alignment, structured treatment behavior sequence data was constructed, further transformed into a treatment behavior graph, and the behavior type, behavior time, and intent embedding features of each node were extracted. Subsequently, based on the constructed behavior graph, the behavior-aware graph attention network of this invention was applied to deeply aggregate the node features, obtaining a more contextually semantically rich node embedding representation.

[0033] Next, by comparing the deviation between the node embedding representation and the standard path graph, suspected abnormal nodes were identified. Furthermore, attention weights and path features were combined to construct an abnormal behavior path graph. The hospital information system automatically pushes abnormal information to the hospital quality control platform, enabling real-time early warning and in-process intervention. In actual operation, a total of 2871 abnormal paths with significant path deviations were identified, with an average abnormality rate of 17.6% per thousand complete treatment paths. Among these, the highest proportion of abnormalities was due to missed key examinations, reaching 39.2%, followed by behavioral sequence conflicts (28.7%), abnormal path jumps (19.5%), and missing key behaviors (12.6%).

[0034] Compared with the hospital's original rule-based post-event quality control methods, the hospital information system implemented in this invention has advantages in terms of anomaly identification accuracy and response speed. It can achieve early identification and rapid intervention of potential risky behaviors, thereby improving the safety of diagnosis and treatment and process compliance.

[0035] As can be seen from the above applications, this invention not only solves the problems of lag and insufficient coverage in existing hospital quality control processes, but also achieves accurate analysis of complex behavioral relationships through graph neural network modeling, which has high practical value and promotion potential.

[0036] The following are statistics on the types and frequencies of abnormal behavior paths identified during the operation of the hospital information system implemented in this invention. The data comes from the actual operation logs of a certain period, and the statistical dimensions cover multiple types of medical departments and path types: Table 1: Statistical Table for Identifying Abnormal Diagnostic and Treatment Behavior Paths

[0037] As can be seen from the table above, the behavior-aware graph attention network proposed in this invention has good application results in real-time monitoring of hospital diagnosis and treatment quality, especially in identifying complex abnormal paths, where it demonstrates higher accuracy and response efficiency.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of hospital treatment quality based on big data analysis, characterized in that, Includes the following steps: Collect multi-source diagnosis and treatment behavior data within the hospital and perform data cleaning, missing value imputation, timestamp alignment, and behavior type labeling to obtain structured diagnosis and treatment behavior sequence data; A diagnosis and treatment behavior graph is constructed based on structured diagnosis and treatment behavior sequence data; For each node in the diagnosis and treatment behavior graph, extract node features and perform unified encoding to obtain the initial node representation vector; Based on the initial node representation vector and the diagnosis and treatment behavior graph structure, a behavior-aware graph attention network is constructed. Feature aggregation is performed on each node in the diagnosis and treatment behavior graph to obtain the node embedding representation set of the diagnosis and treatment behavior graph. Based on the node embedding representation set, it is compared with the standard embedding representation set of the corresponding node in the standard diagnosis and treatment path diagram to obtain the node set of abnormal diagnosis and treatment behavior and the corresponding abnormal type label; Based on the set of abnormal diagnosis and treatment behavior nodes, and combined with attention weight information and path structure features, an abnormal diagnosis and treatment behavior path graph is generated. The abnormal diagnosis and treatment behavior path map and abnormal type label are pushed to the hospital quality control system or medical management system to trigger the early warning mechanism and record abnormal event information.

2. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The structured diagnosis and treatment behavior sequence data includes the original diagnosis and treatment behavior data after cleaning, completion, alignment and type labeling; the original diagnosis and treatment behavior data includes medical order record data, examination and test data, surgical record data, nursing record data and admission and discharge record data.

3. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The construction of the diagnosis and treatment behavior graph based on structured diagnosis and treatment behavior sequence data specifically includes: All medical behavior events are extracted from the structured medical behavior sequence data to form a set of medical behavior events; Based on the set of diagnosis and treatment behavior events, a set of nodes for the diagnosis and treatment behavior graph is constructed, each diagnosis and treatment behavior is associated with a graph node, and initial feature information is extracted for each node, combined and encoded to form the initial feature vector of the node; Based on the temporal order of each behavior in the set of diagnosis and treatment behavior events and the dependencies in the diagnosis and treatment process, the connection relationships between nodes are determined, and the set of edges in the diagnosis and treatment behavior graph is generated. For each edge in the edge set, the time interval between actions is calculated based on the time difference between the starting node and the target node connected by the edge. The edge feature information is generated by combining the process dependency relationship type between the corresponding diagnosis and treatment actions of the two nodes. All edge feature information forms the edge feature set. The set of nodes, the set of edges, the set of initial feature vectors of nodes, and the set of feature vectors of edges are integrated to form a diagnostic behavior graph.

4. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The extraction of node features and unified encoding processing for each node in the diagnosis and treatment behavior graph to obtain the initial node representation vector specifically includes: Based on each node in the diagnostic behavior graph, a behavior type vector, a time interval encoding vector, and an intent embedding vector are generated. Based on the graph structure of the diagnosis and treatment behavior graph, the predecessor and successor nodes associated with each node are identified, and the initial feature representations of the adjacent nodes and the edge feature information between them and the current node are extracted to obtain the behavior context combination features of the current node, forming a behavior context feature set. The behavior type vector, time interval encoding vector, intent embedding vector and context combination features in the node feature set are concatenated to generate the fused feature vector of each node, thus obtaining the fused feature set. A unified linear transformation and nonlinear activation process is applied to the fused feature set to obtain the initial representation vector of each node, forming a set of initial representation vectors for the nodes.

5. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The process of constructing a behavior-aware graph attention network based on the initial node representation vector and the diagnosis and treatment behavior graph structure, and performing feature aggregation on each node in the diagnosis and treatment behavior graph to obtain the node embedding representation set of the diagnosis and treatment behavior graph specifically includes: The initial node representation vectors of each target node and its neighboring nodes in the diagnostic behavior graph are input into the behavior-aware graph attention network, which includes an intent difference regulation attention unit, a path decay regulation unit, and an anomaly sensitivity gating unit. In the intent difference-modulated attention unit, based on the degree of difference between the intent embedding vectors of the target node and its neighboring nodes, the attention adjustment factor is calculated, and the attention weights of the neighboring nodes to the target node are adjusted to obtain the intent difference attention weight set. In the path attenuation adjustment unit, the path attenuation factor is calculated based on the path depth information in the graph between each target node and its neighboring nodes, resulting in a set of path attenuation factors. In the anomaly-sensitive gating unit, for each target node, an anomaly risk score is calculated based on the frequency of anomalies identified in historical behavior, and an anomaly-sensitive gating factor is generated. The set of intent difference attention weights, the set of path decay factors, and the set of anomaly sensitivity gating factors are weighted and fused to obtain the embedded representation of the target node; By integrating the embedded representations of each target node in the diagnosis and treatment behavior graph, we obtain the node embedded representation set of the diagnosis and treatment behavior graph.

6. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, In the intent difference-controlled attention unit, based on the degree of difference in intent embedding vectors between the target node and its neighboring nodes, an attention adjustment factor is calculated, and the attention weights of neighboring nodes towards the target node are adjusted to obtain the intent difference attention weight set, specifically including: Obtain the behavioral intent embedding vectors of the target node and its neighboring nodes from the treatment behavior graph, and calculate the intent difference between the target node and its neighboring nodes. Obtain the initial node representation vectors of the target node and its neighboring nodes, encode the node representations through a linear transformation function, concatenate them, and input them into the attention scoring function. Combine the intention difference to calculate the adjusted attention score. Softmax normalization is performed on the attention scores of all neighboring nodes to obtain the intent difference attention weights. The attention weights corresponding to all neighboring nodes are combined to form the intent difference attention weight set.

7. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, In the anomaly-sensitive gating unit, for each target node, an anomaly risk score is calculated based on the frequency of anomalies identified in historical behavior, and an anomaly-sensitive gating factor is generated. Specifically, this includes: For each target node in the diagnosis and treatment behavior graph, count the number of times the target node participates in abnormal behavior paths in the historical diagnosis and treatment behavior graph, obtain the total number of times the target node appears in all historical diagnosis and treatment behavior graphs, and calculate the abnormal frequency component of the target node. Based on the set of neighboring nodes of the target node, the anomaly prediction probability of the neighboring nodes is extracted, the neighborhood average anomaly probability component is calculated, and the anomaly frequency component is combined with a weighted sum to obtain the anomaly risk score of the target node. The abnormal risk score is input into the abnormality sensitive gating function to generate the abnormality sensitive gating factor.

8. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The set of node embedding representations is compared with the standard embedding representation set of corresponding nodes in the standard diagnosis and treatment path diagram to obtain the set of abnormal diagnosis and treatment behavior nodes and corresponding abnormality type labels, specifically including: A standard diagnosis and treatment path diagram is constructed, and feature extraction and embedding calculation are performed on each standard node to obtain a set of standard embedding representations; Based on the node embedding representation set and the standard embedding representation set, the representation offset between the target node and the corresponding node in the standard diagnosis and treatment path diagram is calculated to obtain the node representation offset set. Based on each offset value in the node representation offset set, a representation offset threshold is set. When the representation offset of the target node is greater than the offset threshold, it is marked as a behavioral representation aberration node, and all marked nodes are aggregated to form a behavioral representation aberration node set. Based on the edge information in the diagnosis and treatment behavior graph, a structural comparison is made with the standard diagnosis and treatment path graph. If the diagnosis and treatment behavior graph is missing a key node from the standard diagnosis and treatment path graph, the missing node is included in the set of missing key nodes. If there is a behavioral path in the diagnosis and treatment behavior graph that jumps to a non-adjacent standard node, then the starting node of the path is included in the set of abnormal path jump nodes. If the execution order of a behavior node conflicts with the standard treatment path diagram, the corresponding node will be included in the set of behavior order conflict nodes. The system integrates the set of abnormal nodes, the set of missing key nodes, the set of abnormal nodes for path jumps, and the set of nodes with conflicting behavior order to construct a set of abnormal diagnosis and treatment behavior nodes, and generates a set of corresponding abnormal type labels based on the source of the abnormality.

9. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The process of generating an abnormal diagnosis and treatment behavior path graph based on a set of abnormal diagnosis and treatment behavior nodes, combined with attention weight information and path structure features, specifically includes: Based on the set of abnormal diagnosis and treatment behavior nodes, each abnormal node is located in the diagnosis and treatment behavior graph, the predecessor and successor nodes and connecting edges of the abnormal node are extracted, the corresponding path context subgraph is constructed, and all path context subgraphs are summarized into a path context subgraph set. For each path context subgraph in the path context subgraph set, extract the attention weights between the target abnormal node and its neighboring nodes in the behavior-aware graph attention network, and calculate the context attention aggregation vector. The anomaly score of the target anomaly node is calculated by combining the initial representation vector of the target anomaly node with its embedding representation in the node embedding representation set. Based on the path context subgraph set and anomaly score, the differences between the current path of the target abnormal node and the standard diagnosis and treatment path are analyzed, and the trigger condition description of each abnormal node is generated. Based on the contextual attention representation features, abnormal scores, and abnormal triggering conditions, an abnormal diagnosis and treatment behavior path diagram is constructed.

10. The method for real-time monitoring of hospital treatment quality based on big data analysis according to claim 1, characterized in that, The step of pushing the abnormal diagnosis and treatment behavior path map and abnormal type labels to the hospital quality control system or medical management system to trigger the early warning mechanism and record abnormal event information specifically includes: The abnormal diagnosis and treatment behavior path map set and abnormal type labels are pushed to the hospital quality control system and medical management system respectively. In the hospital quality control system, based on the abnormality type label, the corresponding early warning rule is matched to generate an early warning notification data packet, which is then sent to the designated department administrators. In the medical management system, abnormal diagnosis and treatment behavior path diagrams and abnormal type labels are recorded to the quality control log module to establish a structured abnormal event record; Based on the early warning notification data packet, an intervention task item is generated and assigned to the designated person in charge. At the same time, the task status information is updated and synchronously transmitted back to the abnormal event record. The system continuously monitors changes in the responsible physician's behavior throughout the diagnosis and treatment process. If any abnormal node status is eliminated, the closed-loop process is considered complete, and a completion marker is added.

Images