Intelligent supervision and resource scheduling system for patient ward monitoring state
Patent Information
- Application Number
- CN202610765003.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-12-05
- Filing Date
- 2026-05-29
- Publication Date
- 2026-09-29
AI Technical Summary
[0003]现有技术中，病区的资源调度通常基于规则引擎的自动化调度系统，该系统中预定义有一个静态规则库，通过实时接收监护数据，由规则引擎将患者体征等参数与规则库中的条件进行匹配，当触发某条规则时，便自动生成一条资源调度指令，但是该系统在处理请求时，规则之间相互独立，无法模拟多目标之间复杂的权衡与协商过程，例如，当系统中仅有一台呼吸机，却同时有三位病情持续恶化且危重等级相同的患者提出需求时，仅通过配置的“危重等级优先”规则，可能因三者等级相同而陷入决策僵局，或随机分配，导致该系统无法在复杂的资源竞争场景下做出真正最优的全局决策
[0036]1.本发明通过患者的监护数据建立语义知识图谱，并通过构建语义预测模型对语义知识图谱进行深度分析，从患者实体与其关联资源、人员构成的局部子图中，捕捉复杂的、非线性的病情演变规律，实现了对患者体征趋势和资源需求更早期、更准确的预测，为主动干预提供了关键时间窗口，克服了传统系统仅能被动响应阈值告警的滞后性；
Smart Images
Figure CN122840465A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical information technology and intelligent decision-making technology, and more specifically, to an intelligent monitoring and resource scheduling system for the monitoring status of patients in wards. Background Technology
[0002] Real-time monitoring of patient wards and efficient allocation of medical resources are core elements for ensuring medical safety and improving the success rate of treatment. Among them, how to make rapid, fair and overall optimal decisions when multiple critically ill patients are competing for limited key resources has become a technical bottleneck that modern smart medical systems urgently need to overcome.
[0003] In existing technologies, resource scheduling in wards is usually based on automated scheduling systems using rule engines. These systems have a predefined static rule base. By receiving real-time monitoring data, the rule engine matches patient vital signs and other parameters with conditions in the rule base. When a rule is triggered, a resource scheduling instruction is automatically generated. However, when processing requests, the rules in this system are independent of each other and cannot simulate the complex trade-offs and negotiations between multiple objectives. For example, when there is only one ventilator in the system, but three patients with continuously deteriorating conditions and the same level of criticality make requests at the same time, the system may get stuck in a decision-making deadlock due to the three patients having the same level of criticality, or it may be randomly assigned, making it impossible for the system to make a truly optimal global decision in complex resource competition scenarios.
[0004] In view of this, the present invention proposes an intelligent monitoring and resource scheduling system for the monitoring status of patients in wards to solve the above problems. Summary of the Invention
[0005] To overcome the above-mentioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an intelligent monitoring and resource scheduling system for patient ward monitoring status, comprising: a data acquisition module, used to extract entities and the relationships between entities from the acquired patient monitoring data, and to establish a semantic knowledge graph;
[0006] The prediction module is used to analyze the semantic knowledge graph based on the constructed semantic prediction model and generate a list of prediction query responses that includes predictions of vital signs trends and resource demand.
[0007] The game supervision module is used to deploy multiple agents based on the predictive query response list, interact with the semantic knowledge graph to obtain the collaborative scheduling decision vector, and generate a supervision dashboard and a list of executable tasks based on the collaborative scheduling decision vector.
[0008] The scheduling and execution module executes tasks based on the list of executable tasks and updates the semantic knowledge graph by providing feedback on the execution results.
[0009] Furthermore, the specific process of establishing a semantic knowledge graph is as follows:
[0010] The system receives patient monitoring data from the ward, extracts entities from the cleaned and standardized monitoring data, and obtains a list of patient entities, which includes patient entities and various resource entities. Each entity contains textual and numerical attributes. A relationship list is determined based on the relationships between entities in each entity list.
[0011] A semantic knowledge graph is constructed based on an entity list and a relationship list. Each entity in the entity list is treated as a node, the textual and numerical attributes of the entity are treated as the attribute information of the node, and the relationships between entities in the relationship list are treated as edges, thus completing the construction of the semantic knowledge graph.
[0012] Furthermore, the specific process of constructing the semantic prediction model is as follows:
[0013] A semantic prediction model is constructed, comprising sub-layers, a fusion layer, and a task layer. The sub-layers are used to extract all entities and their relationships connected to the target patient entity at the first layer from the semantic knowledge graph, centered on the target patient entity, forming a local semantic subgraph. The fusion layer has an attention aggregation mechanism based on a heterogeneous graph attention network to encode the local semantic subgraph. The task layer has parallel branches including a first branch and a second branch. The first branch is composed of a classifier and is used to output the patient's vital sign trend prediction. The second branch is composed of a linear regressor and is used to output the patient's resource demand prediction.
[0014] Furthermore, the specific process for generating a predictive query response list that includes vital sign trend predictions and resource demand predictions is as follows:
[0015] The sub-layer traverses the semantic knowledge graph to obtain the local semantic sub-graph of the current patient entity; the fusion layer encodes the local semantic sub-graph based on the attention aggregation mechanism to obtain the patient state semantic vector of the current patient entity; the patient state semantic vector is input into the task layer, and the patient state semantic vector is analyzed through the first branch and the second branch respectively to obtain the current patient entity's vital sign trend prediction and resource demand prediction.
[0016] The patient identifier, prediction timestamp, predicted value of vital signs trend, and predicted resource demand corresponding to the current patient entity are combined into a prediction record. The prediction records of all patient entities are then aggregated to generate a prediction query response list.
[0017] Furthermore, the specific process of deploying multi-agent systems is as follows:
[0018] Based on the predicted query response list, identify patients whose probability of deterioration exceeds a preset deterioration threshold in the predicted trend of vital signs. Construct a patient agent for each patient agent and identify resource types whose estimated demand is greater than the current available capacity in the resource demand prediction. Construct a resource agent for each type of resource.
[0019] Extract the predicted value of vital signs or the predicted resource demand for each agent from the predicted query response list, and use it as the initial state vector for each agent; configure the corresponding action space and decision preferences for each agent based on the preset policy rule base, thereby completing the deployment of multiple agents.
[0020] Furthermore, the specific process of configuring corresponding action spaces and decision preferences for each agent based on a preset policy rule base is as follows:
[0021] The preset strategy rule base defines the action space and decision preferences corresponding to various intelligent agents. The decision preferences are used to guide the intelligent agents to select actions from the action space. The action space of the patient intelligent agent includes requesting resources, releasing resources, and waiting for resources. The action space of each resource intelligent agent includes accepting allocation, rejecting allocation, and negotiating scheduling.
[0022] Based on the type of each agent in the multi-agent system, the corresponding action space and decision preferences are loaded from the preset policy rule base and bound to the corresponding agent.
[0023] Furthermore, the specific process for obtaining the collaborative scheduling decision vector is as follows:
[0024] Based on the initial state vector and decision preferences of each agent, the semantic knowledge graph is queried, and the resource nodes that are simultaneously requested by multiple patient agents are identified as competitive resource points; the current available capacity of each competitive resource point is analyzed to obtain the conflict intensity of each competitive resource point.
[0025] The Nash equilibrium is iterated for the action selection of each agent in the action space. Each agent updates its action selection in each iteration based on its decision preferences and the conflict intensity of its associated competing resource nodes. After reaching the equilibrium state, the action selection of each resource agent and the target patient agent are extracted to generate a collaborative scheduling decision vector consisting of a resource identifier, the action to be executed, and the target patient identifier.
[0026] Furthermore, the specific process for analyzing and obtaining the conflict intensity of the corresponding competing resource nodes is as follows:
[0027] Obtain the number of patient agents corresponding to each competing resource point, analyze the number of patient agents with the current available capacity corresponding to the competing resource point, and obtain the expected supply and demand gap of each competing resource node;
[0028] Based on the semantic knowledge graph, the severity of the patient's condition and the waiting time are obtained; the expected supply and demand gap, severity of the condition and waiting time of each competing resource point are weighted and summed to determine the conflict intensity of each competing resource node.
[0029] Furthermore, the specific process of generating the supervision dashboard based on the collaborative scheduling decision vector is as follows:
[0030] The collaborative scheduling decision vector is associated with the corresponding resource entities and patient entities in the semantic knowledge graph to generate a list of executable tasks;
[0031] Based on the real-time status of all resource entities in the executable task list, a resource scheduling sequence diagram and a department load prediction view are generated; a vital sign trend curve is generated based on the predicted vital sign trends of each patient's intelligent agent; the vital sign trend curve, the resource scheduling sequence diagram, and the department load prediction view are integrated to generate a visual monitoring dashboard.
[0032] Furthermore, the specific process by which the execution terminal executes tasks based on the list of executable tasks and updates the semantic knowledge graph by reporting the execution results is as follows:
[0033] The executable task list is converted into an instruction format, and a task identifier and timestamp are added to each instruction to form an executable instruction list.
[0034] The execution terminal executes tasks based on the list of execution instructions and provides feedback on the execution results after the instructions are completed. Based on the task identifier and execution result status in the execution results, the nodes, edges between nodes, and attribute information of the nodes in the semantic knowledge graph are updated.
[0035] The technical effects and advantages of the intelligent monitoring and resource scheduling system for patient ward monitoring status of the present invention are as follows:
[0036] 1. This invention establishes a semantic knowledge graph through patient monitoring data and performs in-depth analysis of the semantic knowledge graph by constructing a semantic prediction model. From the local subgraph of patient entities and their associated resources and personnel, it captures complex and nonlinear disease evolution patterns, enabling earlier and more accurate prediction of patient vital signs trends and resource needs. This provides a key time window for proactive intervention and overcomes the lag of traditional systems that can only passively respond to threshold alarms.
[0037] 2. This invention transforms resource allocation from a static rule-based process into a dynamic, multi-agent collaborative decision-making process by deploying patient and resource agents for Nash equilibrium iteration. It can quantitatively assess the conflict intensity of resource nodes and automatically construct decision-making criteria based on the severity of the patient's condition, waiting time, and resource supply-demand gap, thus solving the "decision deadlock" problem caused by patients with similar conditions. It generates collaborative scheduling decision vectors, reducing decision deadlocks or suboptimal allocations in complex competitive scenarios, and maximizing the overall treatment efficiency and fairness of the ward.
[0038] 3. This invention not only visualizes data by generating a monitoring dashboard, but also transforms scheduling decision vectors into a specific list of executable tasks and issues them to the execution terminal. At the same time, the system can update the feedback data of the execution effect to the semantic knowledge graph in real time, so that the semantic prediction model and the multi-agent game strategy can continuously evolve, forming a closed-loop optimization system, which significantly improves the long-term practical value and adaptability of the system. Attached Figure Description
[0039] Figure 1 This is a block diagram of an intelligent monitoring and resource scheduling system for patient ward monitoring status according to the present invention;
[0040] Figure 2 This is a schematic diagram of the process for deploying multiple agents according to the present invention. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Example 1
[0043] Please see Figure 1 and Figure 2 As shown in this embodiment, the intelligent monitoring and resource scheduling system for patient ward monitoring status mainly includes the following design contents:
[0044] Real-time monitoring of patient wards and efficient allocation of medical resources are core elements for ensuring medical safety and improving the success rate of treatment.
[0045] In existing technologies, resource scheduling in wards is usually based on automated scheduling systems with static rule bases. The core logic is to match patient vital signs parameters with preset rule conditions and trigger corresponding scheduling instructions. This is efficient when dealing with single-dimensional, conditional decisions. However, real medical decisions, especially in monitoring environments with extremely limited resources, require difficult trade-offs between multiple conflicting objectives.
[0046] For example, when three patients are all classified as having the same level of critical illness and simultaneously request the only available ventilator, the system can only rely on the single rule of "priority based on level of critical illness." However, since the conditions are completely matched, it is impossible to further distinguish the priorities, and ultimately it can only fall into a decision-making deadlock or degenerate into random allocation.
[0047] Based on this, an intelligent monitoring and resource scheduling system for patient ward monitoring status is designed, including:
[0048] The acquisition module is used to extract entities and relationships between entities from the acquired patient monitoring data and to build a semantic knowledge graph.
[0049] The specific process of building a semantic knowledge graph is as follows:
[0050] Receive patient monitoring data from the ward; the monitoring data includes time-series data of patient vital signs (continuous sampling values of ECG, heart rate, blood oxygen saturation, non-invasive blood pressure, respiratory rate, body temperature, etc.), patient static and dynamic information (patient identifier, name, age, gender, admission diagnosis, attending physician, nursing and medical order information, etc.), resource status data (bed occupancy status, medical equipment working status and parameters, drug inventory quantity, medical staff scheduling information), and medical event and record data (nursing observation records, condition assessment records, laboratory test reports, and records of performed clinical operations), etc.
[0051] Entities are extracted from the cleaned and standardized monitoring data to obtain a list of patient entities, including patient entities and various resource entities; each entity contains textual and numerical attributes. Cleaning includes missing value handling, outlier handling, formatting error correction, and duplicate record removal; standardization includes numerical data standardization and textual data standardization; entities are divided into two categories, and patient entities and various resource entities (e.g., bed number, ventilator serial number, nurse employee number) are extracted from the monitoring data and their attribute types are labeled. Textual attributes (e.g., condition assessment records, laboratory reports, medical equipment parameters) and numerical attributes (e.g., heart rate, blood oxygen saturation, non-invasive blood pressure) are encapsulated into an attributed entity list.
[0052] The relationship list is determined based on the association between each entity in each entity list; specifically, it includes: extracting the "use" relationship from medical order records, extracting the "use" relationship from equipment allocation records, extracting the "admission" relationship from bed allocation records, and extracting the "responsibility" relationship from medical staff allocation records, etc.
[0053] For example, iterate through each medical order record, extract the patient identifier and drug identifier from the medical order record, extract the patient entity that matches the patient identifier from the entity list, and extract the drug entity that matches the drug identifier. If both the patient entity and the drug entity exist, create a relation record: the relation name is "Use", the starting entity identifier is the entity identifier of the patient entity, the pointing entity identifier is the entity identifier of the drug entity, and add this relation record to the relation list.
[0054] A semantic knowledge graph is constructed based on an entity list and a relationship list. Each entity in the entity list is treated as a node, the textual and numerical attributes of the entity are treated as the attribute information of the node, and the relationships between entities in the relationship list are treated as edges, thus completing the construction of the semantic knowledge graph.
[0055] The prediction module is used to analyze the semantic knowledge graph based on the constructed semantic prediction model and generate a list of prediction query responses that includes predictions of vital signs trends and resource demand.
[0056] The specific process of constructing a semantic prediction model is as follows:
[0057] A semantic prediction model is constructed, comprising sub-layers, a fusion layer, and a task layer. The sub-layers extract all entities and their relationships connected to the target patient entity at the first layer from the semantic knowledge graph, forming a local semantic subgraph. The fusion layer, based on a heterogeneous graph attention network, incorporates an attention aggregation mechanism to encode the local semantic subgraph. The task layer includes parallel branches containing a first branch and a second branch. The first branch, composed of a classifier, outputs a prediction of the patient's vital signs trend; the second branch, composed of a linear regressor, outputs a prediction of the patient's resource needs. The classifier consists of a fully connected layer and a softmax activation function; the linear regressor consists of a fully connected layer and a linear activation function.
[0058] The specific process for generating a list of forecast query responses that includes predictions of vital sign trends and resource demand is as follows:
[0059] The sub-layer traverses the semantic knowledge graph to obtain the local semantic subgraph of the current patient entity. It retrieves the identifiers of all nodes of type patient entity in the semantic knowledge graph and processes each patient entity sequentially according to the order of the patient entity identifiers. For each currently processed patient entity, the following operations are performed: locating the node corresponding to the target patient entity identifier; extracting all neighboring nodes directly connected to the target patient entity node and their attribute information; extracting all edges connecting the target patient entity node and each neighboring node and their relationship types; and combining the target patient entity node (including attribute information), neighboring nodes (including attribute information), connecting edges, and their relationship types to form the local semantic subgraph of the current patient entity.
[0060] The fusion layer encodes the local semantic subgraph based on an attention aggregation mechanism to obtain the patient state semantic vector of the current patient entity. The attention aggregation mechanism takes the target patient entity node as the center, calculates the attention coefficients with each neighbor node for different relation types in its local semantic subgraph, and generates the state semantic vector of the target patient entity by weighted aggregation of the information of all neighbor nodes.
[0061] The patient state semantic vector is input into the task layer. The first and second branches are used to analyze the patient state semantic vector to obtain the current patient entity's vital sign trend prediction and resource demand prediction. The vital sign trend prediction outputs the probability distribution of the current patient's health status deteriorating within the prediction time. The resource demand prediction is used to quantify the current patient's demand for various resources (such as ICU beds, ventilators, and nursing staff) within the prediction time.
[0062] The fully connected layer in the classifier of the first branch performs a nonlinear transformation on the patient state semantic vector, and the softmax activation function converts the transformation result into a probability distribution, outputting the predicted value of the current patient entity's vital signs trend; the fully connected layer in the linear regressor of the second branch performs a linear transformation on the patient state semantic vector, and the linear activation function directly outputs the transformation result, obtaining the predicted value of the current patient entity's resource demand.
[0063] The patient identifier, prediction timestamp, predicted value of vital signs trend, and predicted resource demand corresponding to the current patient entity are combined into a prediction record. The prediction records of all patient entities are then aggregated to generate a prediction query response list.
[0064] The game supervision module is used to deploy multiple agents based on the predictive query response list, interact with the semantic knowledge graph to obtain the collaborative scheduling decision vector, and generate a supervision dashboard and a list of executable tasks based on the collaborative scheduling decision vector.
[0065] The specific process of deploying multi-agent systems is as follows:
[0066] Based on the predicted query response list, patient identifiers whose probability of deterioration exceeds a preset deterioration threshold in the predicted vital signs trend are identified, and a patient agent is constructed for each patient identifier. Each predicted record in the predicted query response list is traversed, and the predicted vital signs trend value in each record is read. All patient identifiers whose probability of deterioration is greater than the preset deterioration threshold are then selected.
[0067] For example, if the preset deterioration probability threshold is 0.7, and the predicted value of the vital signs trend of a certain record in the prediction query response list shows a deterioration probability of 0.85, then the patient identifier "P001" corresponding to that record is recorded.
[0068] It should be explained that the deterioration threshold is set based on historical patient monitoring data and corresponding records of actual deterioration of the condition.
[0069] It identifies resource types whose estimated demand exceeds current available capacity in resource demand forecasting, and constructs a resource agent for each resource type. It reads the resource demand forecast value from each forecast record to obtain the estimated demand quantity for each resource type. It queries the current available capacity attribute of the corresponding resource type node in the semantic knowledge graph. When the estimated demand quantity for a certain resource type exceeds the current available capacity, it records the resource type identifier and constructs a resource agent for each resource type.
[0070] For example, if multiple records in the predictive query response list show that the estimated total demand for ventilators is 5 units, and the current available capacity of ventilator resources in the semantic knowledge graph is 3 units, then the record resource type is identified as "ventilator".
[0071] Extract the predicted vital sign trends or resource demand predictions corresponding to each agent from the prediction query response list, and use them as the initial state vectors for each agent. Extract the predicted vital sign trends of the corresponding patients from the prediction query response list as the initial state vectors for the patient agent, and extract the predicted resource demand predictions for the corresponding resources as the initial state vectors for the resource agent.
[0072] Based on a pre-defined policy rule base, each agent is configured with a corresponding action space and decision preferences, thereby completing the deployment of multiple agents.
[0073] The specific process of configuring corresponding action spaces and decision preferences for each agent based on a preset policy rule base is as follows:
[0074] The preset policy rule base defines the action space and decision preferences corresponding to various intelligent agents. The policy rule base is stored in the form of a structured key-value table. The key field is "intelligent agent type name", and the value field contains "action space list" and "decision preference parameters".
[0075] Decision preferences are used to guide the agent's action selection from the action space. Decision preferences include predefined static preference weights for each action and policy weighting coefficients used to weigh different decision factors.
[0076] The action space of the patient agent includes requesting resources, releasing resources, and waiting for resources; the action space of each resource agent includes accepting allocation, rejecting allocation, and negotiating scheduling.
[0077] Based on the type of each agent in the multi-agent system, the corresponding action space and decision preferences are loaded from the preset policy rule base and bound to the corresponding agent. For each agent identified as a patient agent, the action space definition (requesting resources, releasing resources, and waiting for resources) is read from the patient agent partition of the policy rule base, and the read action space definition is stored in the action space attribute of the patient agent.
[0078] For each agent identified as a resource agent, read the action space definition (accept allocation, reject allocation, and negotiate scheduling) from the resource agent partition of the policy rule base, and store the read action space definition in the action space attribute of the resource agent.
[0079] For example, for the patient agent "P001", the loaded action space is {request resources, release resources, wait for resources}, and its decision preferences are set as follows: its static preference weights are preset as follows: (Resource requested) = 0.8 (Waiting for resources) = 0.2; its strategy weighting coefficient is preset as (personal preference weight is 0.6, resource conflict state weight is 0.3, and social cooperation weight is 0.1), indicating that its decision-making depends more on its own preferences than on resource conflict.
[0080] The specific process of obtaining the collaborative scheduling decision vector is as follows:
[0081] Based on the initial state vectors and decision preferences of each agent, the semantic knowledge graph is queried, and resource nodes that are simultaneously requested by multiple patient agents are identified as competitive resource points. The semantic knowledge graph is queried to retrieve all patient agents that have initiated the "request resource" action and their target resource nodes; the number of times each target resource node is requested is counted, and resource nodes that are simultaneously requested by two or more patient agents are selected and marked as competitive resource points.
[0082] The current available capacity of each competing resource point is analyzed to obtain the conflict intensity of each competing resource point; Nash equilibrium iteration is performed on the action selection of each agent in the action space; each agent updates its action selection in each iteration based on its decision preferences and the conflict intensity of its associated competing resource nodes. A hypothetical game algorithm is used for Nash equilibrium iteration. All agents randomly select actions based on their decision preferences and initial state vectors; in each iteration, each agent evaluates the expected utility of each action in its own action space based on its current state, decision preferences, the conflict intensity of associated competing resource points, and the current action selections of other agents, and each agent selects the action with the highest expected utility as the new action selection for this round.
[0083] The iterative formula for obtaining the expected utility of the agent's choice is:
[0084] ;
[0085] in, Represents intelligent agents For resources Select action above Expected utility; Represents intelligent agents Candidate actions currently being evaluated; Represents intelligent agents Action The decision preference weights; Representing resources The intensity of the conflict; This indicates the maximum conflict intensity among all resources; This is an indicator function, representing if other intelligent agents... The action selected in the previous iteration With current candidate actions If compatible, the indicator function value is 1; otherwise, it is 0. Indicates the number of agents participating in the game; , and These are weighting coefficients, and they satisfy... + + =1; , and The weights are used to balance the three factors: personal preferences, resource conflict, and social cooperation.
[0086] After reaching equilibrium, the action choices of each resource agent and the target patient agent they act upon are extracted, generating a collaborative scheduling decision vector consisting of a triplet of resource identifier, execution action, and target patient identifier. If the action choices of all agents remain unchanged across multiple iterations, a Nash equilibrium is reached. The final action choices of all resource agents and the identifiers of their target patient agents are recorded at this point. The resource identifier, execution action, and target patient identifier are combined into a triplet; the set of all such triplets constitutes the collaborative scheduling decision vector.
[0087] For example, the collaborative scheduling decision vector contains triples ["ventilator 001", "accept allocation", "patient agent P001"] and ["ICU bed 005", "negotiated scheduling", "patient agent P003"].
[0088] The specific process for analyzing and obtaining the conflict intensity of corresponding competing resource nodes is as follows:
[0089] Obtain the number of patient agents corresponding to each competing resource point, analyze the number of patient agents and the current available capacity corresponding to the competing resource point to obtain the expected supply and demand gap for each competing resource node; subtract the current available capacity from the number of patient agents to obtain the expected supply and demand gap value; for example, if a ventilator resource node is requested by 3 patient agents, the current available capacity is 1, and the expected supply and demand gap is 2.
[0090] Based on a semantic knowledge graph, the severity of the patient's condition and its waiting time are obtained for each patient agent. The system then queries the entity nodes corresponding to each patient agent in the semantic knowledge graph, retrieving the severity value from the node attributes. The severity values of all relevant patient entities are compared, and the maximum value is taken as the representative severity value for that competing resource node. Finally, the system retrieves the waiting time attribute for each patient entity. The maximum waiting time value of all relevant patient entities is then taken as the representative waiting time value for that competing resource node.
[0091] The expected supply and demand gap, the severity of the illness, and the waiting time corresponding to each competing resource point are weighted and summed to determine the conflict intensity of each competing resource node.
[0092] The conflict intensity is obtained as follows:
[0093] ;
[0094] in, Indicates competing nodes The intensity of the conflict; Indicates nodes competing for resources The expected supply-demand gap figure; Indicates nodes competing for resources The maximum severity of the condition for all corresponding patient entities; Indicates nodes competing for resources The maximum waiting time for all corresponding patient entities; , and These represent the weighting coefficients for the expected supply-demand gap, the severity of the illness, and the waiting time, respectively. ;
[0095] It needs to be explained that, , and Adjustments can be made based on the core objectives of different periods. For example, when the objective is to maximize resource utilization efficiency, the weight of the supply-demand gap can be increased. This makes the system more sensitive to resource shortages; when the goal is to prioritize medical safety and equity, the weighting of the severity of the illness can be increased. Prioritize resources for the most critically ill patients; when the goal is to reduce patient wait times and improve satisfaction, the weighting of wait time can be increased. This is to avoid delays for patients due to long waiting times.
[0096] The scheduling and execution module executes tasks based on the list of executable tasks and updates the semantic knowledge graph by providing feedback on the execution results.
[0097] The specific process of generating a supervision dashboard based on the collaborative scheduling decision vector is as follows:
[0098] The collaborative scheduling decision vector is associated with the corresponding resource entities and patient entities in the semantic knowledge graph to generate an executable task list. In the semantic knowledge graph, the resource entity corresponding to the resource identifier and the patient entity corresponding to the target patient identifier in each triple of the collaborative scheduling decision vector are found and associated to form an executable task record. The set of all executable task records constitutes the executable task list.
[0099] Based on the real-time status of all resource entities in the executable task list, a resource scheduling sequence diagram and a departmental load prediction view are generated. Based on the generation time of the collaborative scheduling decision vector and the estimated task execution duration, a plan execution timestamp is assigned to each record in the executable task list. Real-time status data of all resource entities, including resource location, usage status, and current load, are obtained from the semantic knowledge graph. With the time axis as the horizontal axis and the resource status as the vertical axis, the status change curve of each resource entity at different time points is plotted to form a resource scheduling sequence diagram.
[0100] Based on the real-time status of all resource entities in the executable task list, the current total load rate of the department is calculated. Combined with the predicted resource demand values in the predicted query response list, an area map is used to display the load formation department load prediction view.
[0101] A vital sign trend curve is generated based on the predicted vital sign trends of each patient agent. The curve of the predicted vital sign trend over time is plotted for each patient entity with the time axis as the horizontal axis and the predicted vital sign trend as the vertical axis. Different colors are used to distinguish the curves of different patient entities to form a vital sign trend curve.
[0102] The vital sign trend curve, resource allocation time series chart, and departmental load prediction view are integrated to generate a visual monitoring dashboard. These three views are arranged in the same layout on the same display interface, sharing the same timeline, allowing users to compare and view the relationship between resource allocation, load prediction, and patient vital sign trends.
[0103] The specific process by which the execution terminal executes tasks based on the list of executable tasks and updates the semantic knowledge graph by reporting the execution results is as follows:
[0104] The executable task list is converted into an instruction format, and a task identifier and timestamp are added to each instruction to form an execution instruction list. Each instruction in the executable task list is converted into a structured data format that can be parsed by the execution terminal. A globally unique task identifier is generated for each converted instruction. The current timestamp is obtained from the operating environment as the instruction generation timestamp. The task identifier, instruction generation timestamp, and converted instruction are combined into a scheduled execution instruction. The set of all scheduled execution instructions constitutes the execution instruction list.
[0105] For example, the triple ["ventilator001", "accept allocation", "patient agent P001"] in the executable task record is converted to JSON format {"action": "accept_allocation", "resource": "ventilator001", "target_patient": "P001"}.
[0106] The execution terminal executes tasks based on the list of execution instructions and provides feedback on the execution results after the instructions are completed. The execution terminal parses the structured data in each scheduled execution instruction, executes the actual operations required by the instruction, and records the actual execution result status of each instruction, including successful execution, failure, or partial execution, as well as the actual completion timestamp for each instruction. The execution terminal combines the task identifier, actual execution result status, and actual completion timestamp into an execution result record. The collection of all execution result records constitutes the execution result feedback list.
[0107] Based on the task identifier and execution result status in the execution effect, the nodes, edges between nodes, and node attribute information in the semantic knowledge graph are updated. According to the task identifier in the execution effect feedback list, the corresponding original instruction is searched in the execution instruction list. For successfully executed instructions, the resource node corresponding to the resource identifier in the instruction is searched in the semantic knowledge graph, the numerical attribute of the resource node is updated, the resource status is changed to "allocated", and the numerical attribute of the patient node corresponding to the target patient identifier is updated, and the allocated resource attribute is added to the resource identifier.
[0108] For a successfully executed "Accept Allocation" or "Negotiation Scheduling" instruction, the corresponding resource node and patient node are searched in the semantic knowledge graph. A new edge is created between the resource node and the patient node, with the edge type set to "Assigned to". The edge's attributes record the allocation timestamp and task identifier. For a successfully executed "Release Resource" instruction, the corresponding "Assigned to" edge is searched in the semantic knowledge graph, and the edge's status attribute is updated to "Released".
[0109] For all instructions involving resource allocation, regardless of whether the execution is successful or not, the last operation timestamp attribute of the corresponding resource node in the semantic knowledge graph is updated.
[0110] In this embodiment, a semantic knowledge graph is established through the patient's monitoring data, and a semantic prediction model is constructed to conduct in-depth analysis of the semantic knowledge graph. From the local subgraph of the patient entity and its associated resources and personnel, the complex and nonlinear disease evolution pattern is captured, which enables earlier and more accurate prediction of the patient's vital signs and resource needs, providing a key time window for proactive intervention and overcoming the lag of traditional systems that can only passively respond to threshold alarms.
[0111] By deploying patient and resource agents for Nash equilibrium iteration, resource allocation is transformed from a static rule-based process into a dynamic, multi-agent collaborative decision-making process. This allows for the quantitative assessment of conflict intensity among resource nodes and the automatic construction of decision-making criteria based on factors such as the severity of the patient's condition, waiting time, and resource supply-demand gap. This solves the "decision-making deadlock" problem caused by patients with similar conditions. Furthermore, the generation of collaborative scheduling decision vectors reduces decision-making deadlocks or suboptimal allocations in complex competitive scenarios, maximizing the overall treatment efficiency and fairness of the ward.
[0112] By generating a monitoring dashboard for data visualization, the system also transforms scheduling decision vectors into a list of specific executable tasks and issues them to the execution terminal. At the same time, the system can update the feedback data of the execution effect to the semantic knowledge graph in real time, enabling the semantic prediction model and the multi-agent game strategy to continuously evolve, forming a closed-loop optimization system, which significantly improves the long-term practical value and adaptability of the system.
[0113] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0114] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0115] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
[0116] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An intelligent monitoring and resource scheduling system for patient ward monitoring status, characterized in that, The intelligent monitoring and resource scheduling system for patient ward monitoring status includes: The acquisition module is used to extract entities and relationships between entities from the acquired patient monitoring data and to build a semantic knowledge graph. The prediction module is used to analyze the semantic knowledge graph based on the constructed semantic prediction model and generate a list of prediction query responses that includes predictions of vital signs trends and resource demand. The game supervision module is used to deploy multiple agents based on the predictive query response list, interact with the semantic knowledge graph to obtain the collaborative scheduling decision vector, and generate a supervision dashboard and a list of executable tasks based on the collaborative scheduling decision vector. The scheduling and execution module executes tasks based on the list of executable tasks and updates the semantic knowledge graph by providing feedback on the execution results.
2. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 1, characterized in that, The specific process for establishing the semantic knowledge graph is as follows: The system receives patient monitoring data from the ward, extracts entities from the cleaned and standardized monitoring data, and obtains a list of patient entities, which includes patient entities and various resource entities; each entity contains text attributes and numerical attributes. Determine the relationship list based on the associations between each entity in each entity list; A semantic knowledge graph is constructed based on an entity list and a relationship list. Each entity in the entity list is treated as a node, the textual and numerical attributes of the entity are treated as the attribute information of the node, and the relationships between entities in the relationship list are treated as edges, thus completing the construction of the semantic knowledge graph.
3. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 2, characterized in that, The specific process for constructing the semantic prediction model is as follows: A semantic prediction model is constructed, comprising a sub-layer, a fusion layer, and a task layer. The sub-layer is used to extract all entities and their relationships connected to the first layer of the target patient entity from the semantic knowledge graph, centered on the target patient entity, to form a local semantic subgraph. The fusion layer is based on a heterogeneous graph attention network and has an attention aggregation mechanism to encode the local semantic subgraph. The task layer has parallel branches containing a first branch and a second branch. The first branch is composed of a classifier and is used to output the patient's vital sign trend prediction. The second branch consists of a linear regressor, used to output a prediction of the patient's resource needs.
4. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 3, characterized in that, The specific process for generating the prediction query response list that includes vital sign trend prediction and resource demand prediction is as follows: The sub-layer traverses the semantic knowledge graph to obtain the local semantic sub-graph of the current patient entity; the fusion layer encodes the local semantic sub-graph based on the attention aggregation mechanism to obtain the patient state semantic vector of the current patient entity; the patient state semantic vector is input into the task layer, and the patient state semantic vector is analyzed through the first branch and the second branch respectively to obtain the current patient entity's vital sign trend prediction and resource demand prediction. The patient identifier, prediction timestamp, predicted value of vital signs trend, and predicted resource demand corresponding to the current patient entity are combined into a prediction record. The prediction records of all patient entities are then aggregated to generate a prediction query response list.
5. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 4, characterized in that, The specific process of deploying multiple agents is as follows: Based on the predicted query response list, identify patients whose probability of deterioration exceeds a preset deterioration threshold in the predicted trend of vital signs. Construct a patient agent for each patient agent and identify resource types whose estimated demand is greater than the current available capacity in the resource demand prediction. Construct a resource agent for each type of resource. Extract the predicted vital signs or resource requirements corresponding to each agent from the predicted query response list, and use them as the initial state vector of each agent. Based on a pre-defined policy rule base, each agent is configured with a corresponding action space and decision preferences, thereby completing the deployment of multiple agents.
6. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 5, characterized in that, The specific process of configuring corresponding action spaces and decision preferences for each agent based on a preset policy rule base is as follows: The preset strategy rule base defines the action space and decision preferences corresponding to various intelligent agents. The decision preferences are used to guide the intelligent agents to select actions from the action space. The action space of the patient intelligent agent includes requesting resources, releasing resources, and waiting for resources. The action space of each resource intelligent agent includes accepting allocation, rejecting allocation, and negotiating scheduling. Based on the type of each agent in the multi-agent system, the corresponding action space and decision preferences are loaded from the preset policy rule base and bound to the corresponding agent.
7. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 5, characterized in that, The specific process for obtaining the collaborative scheduling decision vector is as follows: Based on the initial state vector and decision preferences of each agent, the semantic knowledge graph is queried, and the resource nodes that are simultaneously requested by multiple patient agents are identified as competitive resource points. Analyze the current available capacity of each competing resource point to obtain the conflict intensity of each competing resource point; The Nash equilibrium is iterated for the action selection of each agent in the action space. Each agent updates its action selection in each iteration based on its decision preferences and the conflict intensity of its associated competing resource nodes. After reaching the equilibrium state, the action selection of each resource agent and the target patient agent are extracted to generate a collaborative scheduling decision vector consisting of a resource identifier, the action to be executed, and the target patient identifier.
8. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 7, characterized in that, The specific process for obtaining the conflict intensity of the corresponding competing resource nodes through analysis is as follows: Obtain the number of patient agents corresponding to each competing resource point, analyze the number of patient agents with the current available capacity corresponding to the competing resource point, and obtain the expected supply and demand gap of each competing resource node; Based on semantic knowledge graphs, the severity of the patient's condition and the waiting time are obtained for the corresponding patient's intelligent agent. The expected supply and demand gap, the severity of the illness, and the waiting time corresponding to each competing resource point are weighted and summed to determine the conflict intensity of each competing resource node.
9. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 7, characterized in that, The specific process of generating the supervision dashboard based on the collaborative scheduling decision vector is as follows: The collaborative scheduling decision vector is associated with the corresponding resource entities and patient entities in the semantic knowledge graph to generate a list of executable tasks; Based on the real-time status of all resource entities in the list of executable tasks, a resource scheduling sequence diagram and a departmental load prediction view are generated. Based on the predicted values of each patient's vital signs, a vital signs trend curve is generated; the vital signs trend curve, the resource scheduling time sequence diagram, and the department load prediction view are integrated to generate a visual monitoring dashboard.
10. The intelligent monitoring and resource scheduling system for patient ward monitoring status according to claim 9, characterized in that, The specific process by which the execution terminal executes tasks based on the list of executable tasks and updates the semantic knowledge graph by reporting the execution results is as follows: The executable task list is converted into an instruction format, and a task identifier and timestamp are added to each instruction to form an executable instruction list. The execution terminal executes tasks based on the list of execution instructions and provides feedback on the execution results after the instructions are completed. Based on the task identifier and execution result status in the execution results, the nodes, edges between nodes, and attribute information of the nodes in the semantic knowledge graph are updated.