An agent-generated interactive system and method applied to hospital operation management

Abstract

The application discloses an intelligent agent generated interactive system and method applied to hospital operation management, relates to the technical field of hospital operation management, and captures each interactive operation of a user for an operation entity from a hospital heterogeneous system and records each operation as an atomic interactive event; analyzes the atomic interactive event to generate a user long-term preference vector and stores the user long-term preference vector; when a natural language query is received, a target operation entity identifier and an analysis index are analyzed; the analyzed index is matched with a candidate analysis model, an execution model is generated, and an index analysis result set is output; the data type and the dimension quantity of the index analysis result set are analyzed, and a basic UI component requirement is generated; the UI component requirement and a UI component preference are fused and decided, a target UI component selection is generated, and a final interactive interface is rendered and generated; and the frequency distribution is updated according to the user interaction. Personalized decision is supported, and the use experience of the user is improved.

Description

Technical Field

[0001] This invention relates to the field of hospital operation management technology, specifically to an intelligent agent generative interaction system and method applied to hospital operation management. Background Technology

[0002] Hospital operation and management is an important component of hospital administration. With the continuous reform and improvement of the economic system, the reform of the medical system is also deepening. Strengthening hospital operation and management to improve the overall efficiency of hospitals is an inevitable requirement for their future survival and sustainable development.

[0003] Existing hospital management software suffers from significant rigidity in its interaction model. The user interface typically consists of multi-level menus and fixed controls, resulting in lengthy and indiscriminate operation paths. Managers at different levels and with different responsibilities encounter the exact same interface layout and information presentation, making it impossible for the system to recognize user roles, current task focus, or personal usage habits. This not only increases the learning and operational costs for users but also makes it difficult to efficiently understand and utilize even the intelligent analysis results generated in the background due to the rigid front-end interaction, significantly diminishing the value of intelligent technology in the "last mile" of interaction.

[0004] Therefore, this invention discloses an intelligent agent generative interaction system and method for hospital operation management to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent agent generative interaction system and method for hospital operation and management, so as to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a generative interaction method for intelligent agents applied to hospital operation management, characterized in that the method includes the following steps: S1: Collect time-series operational data from heterogeneous hospital systems, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operational entities; after cleaning and aligning the time-series operational data, generate an operational entity time-series feature matrix; S2: Capture every user interaction with the operational entity and record each interaction as an atomic interaction event; analyze the atomic interaction events to generate and store the user's long-term preference vector; S3: When a user's natural language query for a specific operational entity is received, the target operational entity identifier and analysis metrics are parsed; the parsed metrics are matched with candidate analysis models to generate an execution model and output a set of metric analysis results. S4: Analyze the data type and number of dimensions of the indicator analysis result set to generate basic UI component requirements; integrate UI component requirements and UI component preferences to generate target UI component selection, render and generate the final interactive interface; update the frequency distribution based on user interaction.

[0007] S1 includes the following: S101: Collect time-series operational data from heterogeneous information systems within the hospital, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities; specifically, using the operating room number as an identifier, collect the operating room opening duration and actual occupancy time for each operating room at a given time point from the operating room management system; using the equipment serial number as an identifier, collect the current running duration and fault status code for each large medical device at a given time point from the equipment management system; and using the department code as an identifier, collect the human resource cost for each clinical department at a given time point from the human resources system. S102: Assign a unified, discrete timestamp to all collected raw data records; allocate cumulative data for each time period to the corresponding timestamp according to the data properties, and mark instantaneous event data to the corresponding event occurrence time; the cumulative data for each time period includes, but is not limited to, duration data and cost data; the instantaneous event data includes, but is not limited to, the occurrence of fault codes; If a certain data record is missing for a particular operating entity at a certain unified timestamp, a preset strategy is used to fill it in, so as to ensure that each operating entity has a complete status record at each sampling time; the strategy includes a forward filling method and a zeroing method; The forward filling method is to fill continuous indicators with data from the most recent valid time for the corresponding operating entity if there is no data at the current moment; the continuous indicators include, but are not limited to, utilization rate and labor cost. The zeroing method is to set the data item to zero or a specific default code representing "none" or "normal" for discrete events or count indicators if no event occurs or no record is recorded at the current time. S103: Based on the data of each operating entity at each timestamp, perform feature extraction and normalization calculation to obtain numerical features that characterize the operating status of the operating entity at the corresponding time. The numerical features include at least utilization rate features, cost features, duration features and status coding features. The utilization rate characteristics are as follows: for operating rooms, the ratio of actual occupancy time to operating room opening time is used as the utilization rate characteristic; for equipment, the ratio of current operation time to standard shift time is used as the utilization rate characteristic. The cost feature is the cost per unit time; the duration feature directly uses or logarithmically processes the actual duration occupied and the absolute duration of the current run; the status coding feature maps equipment fault codes to digital status flags. The numerical features are normalized to generate the feature vector V of the operating entity at the corresponding time. X(t) =[f1,f2,…,fn], where X is the operating entity identifier, t represents time; f1 to fn are specific normalized feature values; collect the feature vectors of all operating entities under all timestamps to construct the operating entity time series feature matrix, which includes timestamp index, operating entity identifier index and operating entity feature vector index.

[0008] This invention collects time-series data from specific operational entities (including but not limited to operating rooms, equipment, and departments) from heterogeneous, independent systems within a hospital, and explicitly constructs a time-series feature matrix for these operational entities, incorporating a three-dimensional index of time, entity, and features. By forcibly collecting data from independent systems and processing it uniformly, it achieves substantial convergence of cross-system data at the technical level, laying the foundation for global analysis. Anchoring the analysis target to core hospital resource entities such as "operating rooms, equipment, and departments" ensures that all subsequent intelligent analyses revolve around specific business objects, avoiding analysis generalization. Forward imputation utilizes historical data to maintain the continuity of the time-series sequence, suitable for scenarios where data is temporarily missing but the trend continues (such as equipment status). Zeroing clearly distinguishes between "no activity" and "missing data," preventing noise introduced by erroneous interpolation, especially suitable for cumulative indicators such as cost and duration. It provides deterministic rules for handling missing values, enabling the system to generate stable and interpretable feature matrices even when facing incomplete data sources, improving the system's reliability in the complex data environment of real hospitals.

[0009] S2 includes the following: S201: Capture every interaction operation of user U and record each operation as an atomic interaction event, wherein the atomic interaction event is represented as a triple Event = (Target, Model, Component); Wherein, Target represents the target operational entity identifier for the interactive operation, Model represents the analysis model identifier of the system call, and Component represents the user interface (UI) component identifier enabled by the operation result; S202: Set a statistical time window and count all atomic interaction events that occur to the user within the time window; calculate the frequency distribution of user U’s operation on each target operation entity, the frequency distribution of user U’s call to each analysis model, and the frequency distribution of user U’s use of each UI component, respectively. The operation frequency distribution F_U(X) is equal to the proportion of the number of operation events of user U on the operation entity X to the total number of operation events; the call frequency distribution F_U(M) is equal to the proportion of the number of events of user U triggering model M to the total number of events; and the usage frequency distribution F_U(C) is equal to the proportion of the number of events of user U viewing the results through UI component C to the total number of events. The frequency distributions of the above three dimensions are concatenated into a long-term preference vector in a predetermined order; This invention transforms vague "user habits" into a measurable and storable mathematical vector representing the frequency distribution of entities, models, and UI components, thus providing a data foundation for personalization. By setting a "statistical time window," the system can focus on users' recent stable behavioral patterns, avoiding premature or outdated interactions that could interfere with current preferences, making the preference vector more timely and representative. By concatenating and normalizing the three frequency distributions into a single vector, a compact and comprehensive user profile is created, facilitating efficient vector retrieval and similarity calculation in subsequent steps, supporting personalized decision-making.

[0010] S3 includes the following: S301: Receive the natural language query submitted by the user, parse the natural language query, and extract the target operating entity identifier, the key index to be analyzed, and the time range. Based on the parsed target operational entity identifier, extract the corresponding conditional event sub-vector; calculate the indicator applicability between the indicator keywords and each candidate analysis model in the conditional event sub-vector; The conditional event subvector analysis process includes: for any operational entity, retrieving all atomic interaction events that satisfy the user identifier and the target operational entity identifier within the statistical time window from the interaction event history database, forming a conditional event subset; traversing each event in the conditional event subset and counting the number of times each analysis model appears; calculating the call frequency of each model in the conditional event subset; the call frequency is equal to the ratio of the number of times the model appears in the conditional event subset to the total number of events in the conditional event subset; arranging the call frequencies corresponding to all analysis models in a predefined model index order to form a conditional event subvector. Each candidate analysis model is associated with descriptive metadata, including the indicator tags it excels at; the indicator applicability is calculated by the cosine similarity of word vectors between the indicator keywords and the indicator tags of the candidate analysis model, or by statistically analyzing the success rate or user satisfaction of each candidate analysis model in handling indicator keyword queries with similarity greater than a threshold in the past. S302: Based on the indicator applicability and conditional event sub-vectors, calculate a comprehensive selection index for each candidate analysis model; the comprehensive selection index is equal to the weighted fusion of the call frequency corresponding to the conditional event sub-vector and the indicator applicability; select the candidate analysis model with the highest comprehensive selection score as the execution model; extract data slices corresponding to the target operation entity identifier and time range from the operation entity time series feature matrix, and input the data slices into the execution model for calculation to obtain the calculation result; S303: Identify the output characteristics of the execution model and determine the result type of the calculation result; extract the key dimension information of the calculation result; and encapsulate the calculation result, result type, and key dimension information together to generate an indicator analysis result set; The result types include, but are not limited to, time-series numerical pairs, category labels, ranking lists, and statistical summaries; wherein a time-series numerical pair is a set of timestamps and numerical pairs; a category label is one or more category identifiers; a ranking list is the sorting result of several operating entities; and a statistical summary is an object containing statistics of mean, maximum, and minimum values; the key dimension information includes, for time-series results, the time range and unit of measurement; and for ranking lists, the key dimension information includes the indicators on which the ranking is based and the set of operating entities involved. This invention not only analyzes the metrics of the current query but also incorporates the user's historical model preferences for that entity. Through weighted fusion, it selects the model most likely to align with the user's understanding and task requirements, making the analysis results more accurate and user-friendly. The same query (e.g., "view operating room efficiency") may trigger different analysis models due to different user historical preferences (e.g., some users frequently view "predictive models," while others frequently view "comparative models"), truly achieving personalized intelligent analysis. The conditional event sub-vector does not calculate a user's general preference for all entities but precisely calculates the model preference within the context of a specific target entity. This makes preference information highly relevant, resulting in more granular and precise decision-making. Through a clear "retrieval-statistics-frequency calculation-ranking" process, "conditional preference" is transformed from a concept into a rigorously calculable data vector, enhancing the technical feasibility and repeatability of the solution.

[0011] S4 includes the following: S401: Generate a list of basic UI component requirements that matches the result type and dimension information according to the preset visualization mapping rules; S402: Based on the result type of the indicator analysis result set, retrieve the conditional UI component preference sub-vector that matches the result type from the current user's long-term preference vector; the conditional UI component preference sub-vector represents the frequency distribution of various UI components used by the user when viewing the analysis results of the corresponding result type in history; traverse each candidate UI component type in the basic UI component requirement list, and for each candidate UI component type, find the corresponding frequency distribution from the conditional UI component preference sub-vector; select the UI component type with the highest frequency distribution and mark it as the target UI component selection; S403: The instantiated and data-bound UI component object is placed in a preset interface layout template for assembly, and the final interactive interface is rendered and output to the user; the user's subsequent interaction with the UI component is monitored, and when a preset positive interaction behavior is identified, the corresponding frequency distribution is updated according to the positive interaction behavior.

[0012] The update algorithm uses a sliding window average or an exponentially weighted moving average; based on the updated frequency distribution, the normalized long-term preference vector is re-stored. When generating the interface, this invention considers both the objective attributes of the analysis results to ensure that the chart types accurately convey the information, and the user's subjective historical preferences, selecting the UI components that are most familiar to the user and have the highest operational efficiency. This achieves a balance between scientific rigor and ease of use. By monitoring "positive interactive behaviors" and updating preferences, the system can transform users' immediate feedback (such as drilling down into a particular chart) into long-term memory, allowing the UI generation capability to continuously evolve with use and become increasingly tailored to individual user habits.

[0013] Another aspect of this application provides an intelligent agent generative interaction system for hospital operation management. The system is implemented using the aforementioned intelligent agent generative interaction method for hospital operation management. The system includes a data acquisition and representation module, a user preference analysis module, a real-time query and matching module, and an interaction update module. The data acquisition and characterization module is used to collect time-series operational data from heterogeneous hospital systems, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities; after cleaning and aligning the data, a time-series feature matrix of the operating entities is generated. The user preference analysis module is used to capture every interaction operation of the user with the operating entity and record each operation as an atomic interaction event; analyze the atomic interaction events to generate and store the user's long-term preference vector; The real-time query matching module is used to parse the target operational entity identifier and analysis indicators when it receives a natural language query from a user for a specific operational entity; match the parsed indicators with candidate analysis models, generate an execution model, and output an indicator analysis result set. The interactive update module is used to parse the data type and number of dimensions of the indicator analysis result set, generate basic UI component requirements; integrate UI component requirements and UI component preferences to make decisions, generate target UI component selection, render and generate the final interactive interface; and perform frequency distribution updates based on user interaction.

[0014] The data acquisition and characterization module includes a data acquisition unit, a data preprocessing unit, and a data integration unit; The data acquisition unit is used to collect time-series operational data from independent heterogeneous information systems within the hospital, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities. The data preprocessing unit is used to unify all collected time-series operational data into the same discrete time series and to stamp each data item with a corresponding timestamp; for operational entity data items that are missing records at a certain timestamp, a preset strategy is used to fill them in; The data integration unit performs feature extraction and normalization calculations based on the data of each operating entity at each timestamp, obtaining numerical features representing the operating entity's operational status at the corresponding time. It then combines all normalized numerical features of the same operating entity at the same timestamp to construct a feature vector for the operating entity at the corresponding time. Finally, it aggregates the feature vectors of all operating entities at all timestamps to construct a temporal feature matrix for the operating entities. The user preference analysis module includes an interaction recording unit and an operation analysis unit; The interaction recording unit is used to capture every interaction operation of user U in the system and record each operation as an atomic interaction event; The operation analysis unit is used to set a statistical time window and count all atomic interaction events that occur to the user within the time window. Based on the statistics, it calculates the operation frequency distribution of the user to each target operation entity, the call frequency distribution of each analysis model, and the usage frequency distribution of each UI component. The operation frequency distribution, call frequency distribution, and usage frequency distribution are concatenated, and the concatenated overall vector is normalized to form the user's long-term preference vector.

[0015] The real-time query matching module is used to adapt to the analysis unit, the fusion analysis unit, and the result processing unit; The adaptation analysis unit is used to receive natural language queries submitted by users, parse the natural language queries, extract the target operating entity identifier, the indicator keywords to be analyzed, and the time range; extract the corresponding conditional event sub-vectors based on the parsed target operating entity identifiers; and calculate the indicator applicability between the indicator keywords and each candidate analysis model in the conditional event sub-vectors. The fusion analysis unit is used to calculate a comprehensive selection index for each candidate analysis model based on the indicator applicability and conditional event sub-vectors; select the candidate analysis model with the highest comprehensive selection score as the execution model; extract data slices corresponding to the target operation entity identifier and time range from the time series feature matrix of the operation entity, and input the data slices into the execution model for calculation to obtain the original calculation results; The result processing unit is used to identify the output characteristics of the execution model, determine the result type of the calculation result, extract the key dimension information of the calculation result, and encapsulate the calculation result, result type and key dimension information together to generate an indicator analysis result set.

[0016] The interactive update module includes a UI component matching unit, a target confirmation unit, and a data update unit; The UI component matching unit is used to generate a list of basic UI component requirements that match the result type and dimension information according to preset visualization mapping rules; The target confirmation unit is used to retrieve the conditional UI component preference sub-vector that matches the result type from the current user's long-term preference vector based on the result type of the indicator analysis result set; iterate through each candidate UI component type in the basic UI component requirement list, and for each candidate UI component type, find the corresponding usage frequency distribution from the conditional UI component preference sub-vector; select the UI component type with the highest usage frequency distribution and mark it as the target UI component selection. The data update unit is used to assemble the instantiated and data-bound UI component objects in a preset interface layout template, and render and generate the final interactive interface for the user; monitor the user's subsequent interaction with the UI components, and when a preset positive interaction behavior is identified, update the corresponding frequency distribution according to the positive interaction behavior.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention collects time-series data of specific operational entities from heterogeneous systems that are independent within a hospital, and explicitly constructs a time-series feature matrix of operational entities containing a three-dimensional index of time, entity, and features. By forcibly collecting data from independent systems and processing it uniformly, it achieves substantial convergence of cross-system data at the technical level, laying the foundation for global analysis. Anchoring the analysis target to core hospital resource entities such as "operating rooms, equipment, and departments" ensures that all subsequent intelligent analyses revolve around specific business objects, avoiding analysis generalization. Forward imputation can utilize historical data to maintain the continuity of the time-series sequence, suitable for scenarios where data is temporarily missing but the trend continues. Zeroing processing can clearly distinguish between "no activity" and "missing data," preventing noise introduced by erroneous interpolation, especially suitable for cumulative indicators such as cost and duration. It provides deterministic missing value handling rules, enabling the system to generate a stable and interpretable feature matrix even when facing incomplete data sources, improving the system's reliability in the complex data environment of actual hospitals. This invention transforms vague "user habits" into a measurable and storable mathematical vector representing the frequency distribution of entities, models, and UI components, providing a data foundation for personalization. By setting a "statistical time window," the system can focus on users' recent stable behavioral patterns, avoiding premature or outdated interactions that could interfere with current preferences, making the preference vector more timely and representative. By concatenating and normalizing the three frequency distributions into a single vector, a compact and comprehensive user profile is created, facilitating efficient vector retrieval and similarity calculation in subsequent steps, supporting personalized decision-making. This invention not only analyzes the indicators of the current query but also combines the user's historical model preferences for that entity, using weighted fusion to select the model most likely to match the user's cognition and task needs, making the analysis results more accurate and easier for users to accept. The same query may trigger different analysis models due to different user historical preferences, truly achieving "personalized analysis for each user." When generating the interface, this invention considers both the objective attributes of the analysis results, ensuring that the chart types correctly convey information, and the user's subjective historical preferences, selecting the UI components that the user is most familiar with and that have the highest operational efficiency. This achieves a unity of scientific rigor and ease of use. By monitoring "positive interaction behaviors" and updating preferences, the system can transform users' immediate feedback into long-term memory, enabling the UI generation capabilities to continuously evolve with use and become increasingly tailored to individual user habits. 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: Figure 1 This is a flowchart illustrating an intelligent agent generative interaction method for hospital operation management according to the present invention. Figure 2This is a schematic diagram of the structure of an intelligent agent generative interactive system for hospital operation management according to the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1 The present invention provides the following technical solution: A generative interaction method for intelligent agents applied to hospital operation management, characterized by the following steps: S1: Collect time-series operational data from heterogeneous hospital systems, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operational entities; after cleaning and aligning the time-series operational data, generate a time-series feature matrix of operational entities; S2: Capture every user interaction with the operational entity and record each interaction as an atomic interaction event; analyze the atomic interaction events to generate and store the user's long-term preference vector; S3: When a user's natural language query for a specific operational entity is received, the target operational entity identifier and analysis metrics are parsed; the parsed metrics are matched with candidate analysis models to generate an execution model and output a set of metric analysis results. S4: Analyze the data type and number of dimensions of the indicator analysis result set to generate basic UI component requirements; integrate UI component requirements and UI component preferences to make decisions, generate target UI component selection, render and generate the final interactive interface; update the frequency distribution based on user interaction.

[0021] S1 includes the following: S101: Collect time-series operational data from heterogeneous information systems within the hospital, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities; specifically, using the operating room number as an identifier, collect the operating room opening duration and actual occupancy time for each operating room at a given time point from the operating room management system; using the equipment serial number as an identifier, collect the current running duration and fault status code for each large medical device at a given time point from the equipment management system; and using the department code as an identifier, collect the human resource cost for each clinical department at a given time point from the human resources system. S102: Assign a uniform, discrete timestamp to all collected raw data records; allocate cumulative data for each time period to the corresponding timestamp based on the data nature, and mark instantaneous event data to the corresponding event occurrence time; cumulative data for each time period includes, but is not limited to, duration data and cost data; instantaneous event data includes, but is not limited to, the occurrence of fault codes; If a certain data record is missing for a certain operating entity at a certain unified timestamp, a preset strategy is used to fill it in, so as to ensure that each operating entity has a complete status record at each sampling time; the strategy includes forward filling method and zeroing method; The forward filling method is to fill in continuous indicators by using the data from the most recent valid time for the corresponding operating entity if there is no data at the current moment. Continuous indicators include, but are not limited to, utilization rate and labor cost. The zeroing method is used for discrete events or count-type indicators. If no event occurs or no record is recorded at the current time, the data item is set to zero or a specific default code representing "none" or "normal". S103: Based on the data of each operating entity at each timestamp, perform feature extraction and normalization calculation to obtain numerical features that characterize the operating status of the operating entity at the corresponding time. The numerical features include at least utilization rate features, cost features, duration features and status coding features. The utilization rate characteristics are as follows: for operating rooms, the ratio of actual occupancy time to operating room opening time is used as the utilization rate characteristic; for equipment, the ratio of current operation time to standard shift time is used as the utilization rate characteristic. The cost feature is the cost per unit time; the duration feature directly uses or logarithmically processes the actual duration occupied and the absolute duration of the current run; the status coding feature maps equipment fault codes to digital status flags. Example 1: In this example, the status coding feature is 0 to represent normal, and non-zero to represent different levels or types of abnormalities; including but not limited to 1 representing a warning and 2 representing a fault; The numerical features are normalized to generate the feature vector V of the operating entity at the corresponding time. X(t) =[f1,f2,…,fn], where X is the operating entity identifier, t represents time; f1 to fn are specific normalized feature values; collect the feature vectors of all operating entities under all timestamps to construct the operating entity time series feature matrix, which includes timestamp index, operating entity identifier index and operating entity feature vector index.

[0022] S2 includes the following: S201: Capture every interaction operation of user U and record each operation as an atomic interaction event. An atomic interaction event is represented as a triple Event = (Target, Model, Component). Wherein, Target represents the target operational entity identifier for the interactive operation, Model represents the analysis model identifier for the system call, and Component represents the UI component identifier for the operation result. Example 2: In this example, when a user clicks to view detailed data for operating room R_101, Target is R_101; when a user filters and displays all CT equipment, Target is the equipment category identifier or specific equipment identifier; when displaying the cost trend of a department to the user, the "Monthly Cost Prediction Model" is called, and Model is the monthly cost prediction model; when performing equipment failure risk assessment for the user, the "Failure Warning Model" is called, and Model is the failure warning model. Example 3: In this example, when the user enters the natural language query "equipment time utilization data of the MR room in the last six months", the system displays the equipment time utilization data in a table format, so the Component is the table; when the user continues to enter the natural language query "change to a line graph", the system changes to display the equipment time utilization data as a line graph, so the Component is the line graph; S202: Set a statistical time window and count all atomic interaction events that occur to the user within the time window; calculate the frequency distribution of user U’s operation on each target operation entity, the frequency distribution of user U’s call to each analysis model, and the frequency distribution of user U’s use of each UI component, respectively. The operation frequency distribution F_U(X) is equal to the proportion of the number of operation events of user U on the operation entity X to the total number of operation events; the call frequency distribution F_U(M) is equal to the proportion of the number of events of user U triggering model M to the total number of events; and the usage frequency distribution F_U(C) is equal to the proportion of the number of events of user U viewing the results through UI component C to the total number of events. The frequency distributions of the above three dimensions are concatenated into a long-term preference vector in a predetermined order; Example 4: In this example, all operational entity frequency values, all model frequency values, and all UI component frequency values ​​are concatenated sequentially; then, the concatenated vector is normalized as a whole to generate a long-term preference vector P_U; the long-term preference vector P_U = [F_U(X1), F_U(X2), ..., F_U(M1), F_U(M2), ..., F_U(C1), F_U(C2), ...]; S3 includes the following: S301: Receive natural language queries submitted by users, parse the natural language queries, and extract the target operational entity identifier, the key metrics to be analyzed, and the time range. Example 5: In this example, the natural language query is: What was the utilization rate of operating room R101 last month? Operational entity identification: Identify and extract the target operational entity identifier explicitly mentioned in the query. In this embodiment, the target operational entity identifier is R101. Intent extraction: Identify and extract the core indicator keywords of the query; in this embodiment, the core indicator keyword is utilization rate; intent extraction is achieved through keyword matching, intent classification models, or operational entity linking technology. Time range extraction: Identify the time range in the query; in this embodiment, the time range is the previous month. Based on the parsed target operational entity identifier, extract the corresponding conditional event sub-vectors; calculate the applicability of the indicator keywords to each candidate analysis model in the conditional event sub-vectors; The conditional event subvector analysis process includes: for any operational entity, retrieving all atomic interaction events that satisfy the user identifier and the target operational entity identifier within the statistical time window from the interaction event history database, forming a conditional event subset; traversing each event in the conditional event subset and counting the number of times each analysis model appears; calculating the call frequency of each model in the conditional event subset; the call frequency is equal to the ratio of the number of times the model appears in the conditional event subset to the total number of events in the conditional event subset; and arranging the call frequencies corresponding to all analysis models in a predefined model index order to form a conditional event subvector. Each candidate analysis model is associated with descriptive metadata, including the indicator tags it excels at; indicator applicability is calculated by the cosine similarity of word vectors between indicator keywords and indicator tags of candidate analysis models, or by statistically analyzing the success rate or user satisfaction of each candidate analysis model in handling indicator keyword queries with similarity greater than a threshold in the past. S302: Based on the applicability of the indicator and the sub-vectors of the conditional events, calculate a comprehensive selection indicator for each candidate analysis model; the comprehensive selection indicator is equal to the weighted fusion of the call frequency and the applicability of the indicator in the sub-vectors of the conditional events; select the candidate analysis model with the highest comprehensive selection score as the execution model; extract data slices corresponding to the target operation entity identifier and time range from the time-series feature matrix of the operation entity, and input the data slices into the execution model for calculation to obtain the calculation results; S303: Identify the output characteristics of the execution model and determine the result type of the calculation result; extract the key dimension information of the calculation result; and encapsulate the calculation result, result type, and key dimension information together to generate an indicator analysis result set; Result types include, but are not limited to, time-series numerical pairs, category labels, ranking lists, and statistical summaries; where time-series numerical pairs are a set of timestamps and numerical pairs; category labels are one or more category identifiers; ranking lists are the sorting results of several operational entities; and statistical summaries are objects containing statistics of mean, maximum, and minimum values; key dimension information: for time-series results, key dimension information includes the time range and unit of measurement; for ranking lists, key dimension information includes the indicators on which the ranking is based and the set of operational entities involved. S4 includes the following: S401: Generate a list of basic UI component requirements that match the result type and dimension information according to the preset visualization mapping rules; Example 6: In this example, the visualization mapping rules include, but are not limited to: if the result type is "time series numerical pair" and the dimension information indicates a single indicator sequence, then the UI component requirement list = ["line chart", "area chart", "bar chart (time series)"]; If the result type is "Ranking List", then the UI component requirement list = ["Horizontal Bar Chart", "Table"]; If the result type is "category tag", then the UI component requirement list = ["pie chart", "donut chart", "tag cloud"]; S402: Based on the result type of the indicator analysis result set, retrieve the conditional UI component preference sub-vector that matches the result type from the current user's long-term preference vector; the conditional UI component preference sub-vector represents the frequency distribution of user usage of various UI components when the user has viewed the analysis results of the corresponding result type in history; traverse each candidate UI component type in the basic UI component requirement list, and for each candidate UI component type, find the corresponding usage frequency distribution from the conditional UI component preference sub-vector; select the UI component type with the highest usage frequency distribution and mark it as the target UI component selection; S403: The instantiated and data-bound UI component object is placed in a preset interface layout template for assembly, and the final interactive interface is rendered and output to the user; the user's subsequent interaction with the UI component is monitored, and when a preset positive interaction behavior is identified, the corresponding frequency distribution is updated according to the positive interaction behavior.

[0023] The update algorithm uses a sliding window average or an exponentially weighted moving average; based on the updated frequency distribution, the normalized long-term preference vector is re-stored. Example 7: In this example, the operation records of the MR room over the past 7 days are collected from the medical technology department management system, including: planned start time, actual start time, planned end time, actual end time, operation status, and associated equipment number; the status logs of the fixed equipment in the MR room, including but not limited to X-ray machine E001 and CT scanner E002, are collected from the equipment management system for the same time period, including: power on / off time, operating parameters, and fault codes; the shift records and hourly costs of the nurses and doctors on duty in the MR room are collected from the human resources system; the data is cleaned; feature calculations are performed on each time slice with a 15-minute time granularity; based on the data of each operating entity at each timestamp, feature extraction and normalization calculations are performed to obtain numerical features representing the operating status of the operating entity at the corresponding time; the numerical features are normalized to generate the feature vector V of the operating entity at the corresponding time. X(t) =[f1,f2,…,fn], where X is the operational entity identifier, t represents the time; f1 to fn are specific normalized feature values; a total of 672 feature vector sets are generated over a week (7 days × 96 time slices). The feature vectors of all operational entities under all timestamps are used to construct the operational entity time series feature matrix. The operational entity time series feature matrix includes timestamp index, operational entity identifier index and operational entity feature vector index; The logs are parsed to extract 300 interaction records of Director 1, and each operation is recorded as an atomic interaction event. In this embodiment, Event1 = (Target: MR room, Model: equipment time utilization trend model, Component: line graph); Event2 = (Target: MR room, Model: equipment time utilization anomaly model, Component: warning card). Set a statistical time window and count all atomic interaction events that occur within the time window for Director 1; take Director 1's interaction with the MR room as the operating entity; among which, Director 1's operation with the MR room accounts for 40% of all operations; calling the "Equipment Time Utilization Trend Model" accounts for 60%, the "Equipment Time Utilization Anomaly Model" accounts for 40%, using the "Line Chart" accounts for 50%, the "Early Warning Card" accounts for 30%, and the "Detailed Table" accounts for 20%; concatenate and normalize the three types of frequency values ​​to obtain Director 1's long-term preference vector P_Director 1 = [0.4, 0.6, 0.4, 0.5, 0.3, 0.2]; The system receives a natural language query submitted by the user. In this embodiment, the query text of Director 1 is: "What is the equipment time utilization rate of the MR room in the past 7 days? Are there any abnormalities?"; the parsing results are: target operating entity = MR room, core indicator keywords = equipment time utilization rate, abnormality, time range = the past 7 days; Based on the analysis of conditional event subvectors in the MR chamber, the conditional event subvectors in this embodiment are: (Equipment time utilization trend model: 0.6, Equipment time utilization anomaly model: 0.4). The applicability of each candidate analysis model was analyzed; among them, the applicability of the indicator (equipment time utilization trend model, equipment time utilization) = 0.9, the applicability of the indicator (equipment time utilization trend model, anomaly) = 0.1; the applicability of the indicator (equipment time utilization anomaly model, anomaly) = 0.8; and the applicability of the indicator (equipment time utilization anomaly model, equipment time utilization) = 0.2. In this embodiment, the weight of the call frequency corresponding to the conditional event subvector is 0.7; the weight of the indicator applicability is 0.3. Therefore, for equipment time utilization: Score1 (equipment time utilization trend model) = 0.7 × 0.9 + 0.3 × 0.6 = 0.81; Score2 (equipment time utilization anomaly model) = 0.7 × 0.2 + 0.3 × 0.4 = 0.26; For anomalies: Score3 (equipment time utilization trend model) = 0.7 × 0.1 + 0.3 × 0.6 = 0.25; Score4 (equipment time utilization anomaly model) = 0.7 × 0.8 + 0.3 × 0.4 = 0.78; The candidate analysis model with the highest comprehensive selection score is selected as the execution model; therefore, the equipment time utilization trend model is used for equipment time utilization analysis, and the equipment time utilization anomaly model is used for anomaly analysis. Extract all feature vectors from the MR room over the past 7 days; input the equipment time utilization trend model and the equipment time utilization anomaly model respectively, and output the equipment time utilization sequence for the past 7 days; the equipment time utilization rate was 65% on a certain afternoon (below the threshold of 80%), indicating an anomaly; Based on the indicator analysis result set and the preset visualization mapping rules, for the equipment time utilization sequence of the past 7 days: UI component requirement list 1 = ["line chart", "area chart", "bar chart (time series)"]; for abnormal records: UI component requirement list 2 = ["early warning card", "table"]; in this embodiment, the frequency distribution of Director 1's use of UI component requirement list 1 is line chart = 0.5, area chart = 0.3, bar chart = 0.2; the frequency distribution of Director 1's use of UI component requirement list 2 is early warning card = 0.7, table = 0.3; The UI component type with the highest usage frequency is selected and marked as the target UI component selection. Therefore, in this embodiment, a line graph is used to display the device time utilization curve, and an alert card is used to display the anomaly. The instantiated and data-bound UI component object is placed in a preset interface layout template for assembly, and the final interactive interface is rendered and output to the user. If the supervisor clicks on the alert card, the detailed anomaly record of the device time utilization is viewed. The system records this click as a "positive interactive behavior" and updates the corresponding frequency distribution based on the positive interactive behavior.

[0024] Please see Figure 2 The present invention provides a technical solution: an intelligent agent generative interactive system for hospital operation management, the system comprising a data acquisition and representation module, a user preference analysis module, a real-time query and matching module, and an interactive update module; The data acquisition and characterization module is used to collect time-series operational data from heterogeneous hospital systems, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operational entities; after cleaning and aligning the data, it generates a time-series feature matrix of the operational entities. The user preference analysis module is used to capture every interaction operation of the user with the operational entity and record each operation as an atomic interaction event; analyze the atomic interaction events to generate and store the user's long-term preference vector. The real-time query matching module is used to parse the target operational entity identifier and analysis indicators when it receives a natural language query from a user for a specific operational entity; it matches the parsed indicators with candidate analysis models, generates an execution model, and outputs a set of indicator analysis results. The interactive update module is used to parse the data type and number of dimensions of the indicator analysis result set, generate basic UI component requirements, integrate UI component requirements and UI component preferences to make decisions, generate target UI component selection, render the final interactive interface, and update the frequency distribution based on user interaction.

[0025] The data acquisition and characterization module includes a data acquisition unit, a data preprocessing unit, and a data integration unit; The data acquisition unit is used to collect time-series operational data from independent heterogeneous information systems within the hospital, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities. The data preprocessing unit is used to unify all collected time-series operational data into the same discrete time series and to stamp each data item with a corresponding timestamp; for operational entity data items that are missing records at a certain timestamp, a preset strategy is used to fill them in; The data integration unit performs feature extraction and normalization calculations based on the data of each operating entity at each timestamp, obtaining numerical features representing the operating entity's operational status at the corresponding time. It then combines all normalized numerical features of the same operating entity at the same timestamp to construct a feature vector for that operating entity at the corresponding time. Finally, it aggregates the feature vectors of all operating entities at all timestamps to construct a time-series feature matrix for each operating entity. The user preference analysis module includes an interaction recording unit and an operation analysis unit; The interaction recording unit is used to capture every interaction operation of user U in the system and record each operation as an atomic interaction event; The operation analysis unit is used to set a statistical time window and count all atomic interaction events that occur within the time window. Based on the statistics, it calculates the frequency distribution of user operations on each target operation entity, the frequency distribution of user calls to each analysis model, and the frequency distribution of user usage for each UI component. The operation frequency distribution, call frequency distribution, and usage frequency distribution are concatenated, and the concatenated overall vector is normalized to form the user's long-term preference vector.

[0026] The real-time query and matching module is used to adapt to the analysis unit, the fusion analysis unit, and the result processing unit; The adaptation analysis unit is used to receive natural language queries submitted by users, parse the natural language queries, extract the target operational entity identifier, the indicator keywords to be analyzed, and the time range; extract the corresponding conditional event sub-vectors based on the parsed target operational entity identifiers; and calculate the indicator applicability between the indicator keywords and each candidate analysis model in the conditional event sub-vectors. The fusion analysis unit is used to calculate a comprehensive selection index for each candidate analysis model based on the applicability of the index and the conditional event sub-vectors; the candidate analysis model with the highest comprehensive selection score is selected as the execution model; data slices corresponding to the target operational entity identifier and time range are extracted from the time series feature matrix of the operational entity, and the data slices are input into the execution model for calculation to obtain the original calculation results; The results processing unit is used to identify the output characteristics of the execution model, determine the result type of the calculation results, extract key dimension information of the calculation results, and encapsulate the calculation results, result type, and key dimension information together to generate an indicator analysis result set.

[0027] The interactive update module includes a UI component matching unit, a target confirmation unit, and a data update unit; The UI component matching unit is used to generate a list of basic UI component requirements that match the result type and dimension information according to preset visualization mapping rules. The target confirmation unit is used to retrieve the conditional UI component preference sub-vector that matches the result type from the current user's long-term preference vector based on the result type of the indicator analysis result set; it iterates through each candidate UI component type in the basic UI component requirement list, and for each candidate UI component type, it finds the corresponding usage frequency distribution from the conditional UI component preference sub-vector; it selects the UI component type with the highest usage frequency distribution and marks it as the target UI component selection. The data update unit is used to assemble instantiated and data-bound UI component objects in a preset interface layout template, and render the final interactive interface to the user; it monitors the user's subsequent interaction with the UI components, and when a preset positive interaction behavior is identified (i.e., the user clicks on a UI component recommended by the system), it updates the corresponding frequency distribution according to the positive interaction behavior.

[0028] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one operating entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these operating entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A generative interaction method for intelligent agents applied to hospital operation management, characterized in that, The method includes the following steps: S1: Collect time-series operational data from heterogeneous hospital systems, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operational entities; after cleaning and aligning the time-series operational data, generate an operational entity time-series feature matrix; S2: Capture every user interaction with the operational entity and record each interaction as an atomic interaction event; analyze the atomic interaction events to generate and store the user's long-term preference vector; S3: When a user's natural language query for a specific operational entity is received, the target operational entity identifier and analysis metrics are parsed; the parsed metrics are matched with candidate analysis models to generate an execution model and output a set of metric analysis results. S4: Analyze the data type and number of dimensions of the indicator analysis result set to generate basic UI component requirements; integrate UI component requirements and UI component preferences to generate target UI component selection, render and generate the final interactive interface; update the frequency distribution based on user interaction.

2. The intelligent agent generative interaction method for hospital operation management according to claim 1, characterized in that: In S1, Includes the following: S101: Collect time-series operational data from heterogeneous information systems within the hospital, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities; S102: Unify all collected time-series operational data into the same discrete time series and assign a corresponding timestamp to each data item; for operational entity data items that are missing records at a certain timestamp, fill them in using a preset strategy; S103: Based on the data of each operating entity at each timestamp, perform feature extraction and normalization calculation to obtain numerical features that characterize the operating status of the operating entity at the corresponding time. The numerical features include at least utilization rate features, cost features, duration features and status coding features. Combine all normalized numerical features of the same operating entity at the same timestamp to construct the feature vector of the operating entity at the corresponding time. Collect the feature vectors of all operating entities under all timestamps to construct the time-series feature matrix of operating entities. The time-series feature matrix of operating entities includes timestamp index, operating entity identifier index and feature vector index of operating entities.

3. The intelligent agent generative interaction method for hospital operation management according to claim 2, characterized in that: The filling strategy employs preset methods, including forward filling and zeroing.

4. The intelligent agent generative interaction method for hospital operation management according to claim 2, characterized in that: S2 includes the following: S201: Capture every user interaction in the system and record each interaction as an atomic interaction event; S202: Set a statistical time window to count all atomic interaction events that occur within the user's time window; Based on the statistics, the frequency distribution of user operations on each target operational entity, the frequency distribution of calls to each analysis model, and the frequency distribution of usage of each UI component were calculated respectively. The operation frequency distribution, call frequency distribution, and usage frequency distribution are concatenated, and the concatenated overall vector is normalized to form the user's long-term preference vector.

5. The intelligent agent generative interaction method for hospital operation management according to claim 4, characterized in that: S3 includes the following: S301: Receive a natural language query submitted by a user, parse the natural language query, and extract the target operational entity identifier, the key index to be analyzed, and the time range; extract the corresponding conditional event sub-vector based on the parsed target operational entity identifier; Calculate the applicability of the index keywords to each candidate analysis model in the conditional event subvector; S302: Based on the applicability of the indicator and the sub-vector of the conditional events, calculate a comprehensive selection indicator for each candidate analysis model; the comprehensive selection indicator is equal to the weighted fusion of the call frequency corresponding to the sub-vector of the conditional events and the applicability of the indicator. The candidate analysis model with the highest comprehensive selection score is selected as the execution model; data slices corresponding to the target operational entity identifier and time range are extracted from the time series feature matrix of the operational entity, and the data slices are input into the execution model for calculation to obtain the original calculation results; S303: Identify the output characteristics of the execution model and determine the result type of the calculation result; Extract the key dimension information from the calculation results; The calculation results, result types, and key dimension information are collectively packaged into a set of indicator analysis results.

6. The intelligent agent generative interaction method for hospital operation management according to claim 5, characterized in that: The conditional event subvector analysis process includes: for any operational entity, retrieving all atomic interaction events that satisfy the user identifier and the target operational entity identifier within the statistical time window from the interaction event history database, forming a conditional event subset; traversing each event in the conditional event subset and counting the number of times each analysis model appears; calculating the call frequency of each model in the conditional event subset; the call frequency is equal to the ratio of the number of times the model appears in the conditional event subset to the total number of events in the conditional event subset; and arranging the call frequencies corresponding to all analysis models in a predefined model index order to form a conditional event subvector.

7. The intelligent agent generative interaction method for hospital operation management according to claim 5, characterized in that: S4 includes the following: S401: Generate a list of basic UI component requirements that matches the result type and dimension information according to the preset visualization mapping rules; S402: Based on the result type of the indicator analysis result set, retrieve the conditional UI component preference sub-vector that matches the result type from the current user's long-term preference vector; traverse each candidate UI component type in the basic UI component requirement list, and for each candidate UI component type, find the corresponding usage frequency distribution from the conditional UI component preference sub-vector; select the UI component type with the highest usage frequency distribution and mark it as the target UI component selection. S403: The instantiated and data-bound UI component object is placed in a preset interface layout template for assembly, and the final interactive interface is rendered and output to the user; the user's subsequent interaction with the UI component is monitored, and when a preset positive interaction behavior is identified, the corresponding frequency distribution is updated according to the positive interaction behavior.

8. The intelligent agent generative interaction method for hospital operation management according to claim 7, characterized in that: The update algorithm uses a sliding window average or an exponentially weighted moving average; based on the updated frequency distribution, the normalized long-term preference vector is re-stored.

9. A generative intelligent agent interaction system for hospital operation management, wherein the system is implemented using the generative intelligent agent interaction method for hospital operation management as described in any one of claims 1-8, characterized in that, The system includes a data acquisition and characterization module, a user preference analysis module, a real-time query and matching module, and an interactive update module; The data acquisition and characterization module is used to collect time-series operational data from heterogeneous hospital systems, including but not limited to operating rooms, large medical equipment, clinical departments, and medical technology departments as operating entities; after cleaning and aligning the data, a time-series feature matrix of the operating entities is generated. The user preference analysis module is used to capture every interaction operation of the user with the operating entity and record each operation as an atomic interaction event; Analyze atomic interaction events to generate and store long-term user preference vectors; The real-time query matching module is used to parse the target operational entity identifier and analysis indicators when it receives a natural language query from a user for a specific operational entity; match the parsed indicators with candidate analysis models, generate an execution model, and output an indicator analysis result set. The interactive update module is used to parse the data type and number of dimensions of the indicator analysis result set, generate basic UI component requirements; integrate UI component requirements and UI component preferences to make decisions, generate target UI component selection, render and generate the final interactive interface; and perform frequency distribution updates based on user interaction.

Images