A method for rapid assessment of venous thromboembolism risk based on agent and multi-source clinical data fusion

Abstract

The application provides a kind of based on agent and multi-source clinical data fusion's venous thromboembolism risk rapid assessment method, constructs the data integration agent for VTE risk assessment, can automatically plan data demand according to different scoring system, locate data source and call corresponding MCP tool, complete the cleaning, standardization and semantic alignment of structured and unstructured data, provide stable, unified data basis for risk factor analysis;The application constructs a complete VTE evaluation result traceability chain, so that each risk factor, scoring item and clinical suggestion can be traced back to the original data, analysis process and rules, providing transparent, auditable evidence support for result interpretation, quality management and manual review, significantly enhancing the credibility and clinical usability of the system.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical information processing technology, specifically to a rapid risk assessment method for venous thromboembolism based on intelligent agents and the fusion of multi-source clinical data. Background Technology

[0002] Venous thromboembolism (VTE) is a common and high-risk complication in hospitalized patients. Commonly used VTE risk assessment systems in clinical practice include Caprini, Padua, Wells, and RCOG, which have detailed items and complex logic, imposing strict requirements on the time sequence of the patient's diagnosis and treatment process, disease characteristics, and procedural records. However, these scales are all described in natural language and lack a semantic structure that machines can directly understand, making automated assessment difficult to implement.

[0003] With the rapid development of Hospital Information Systems (HIS), Electronic Medical Records (EMR), Laboratory Information Systems (LIS), and Picture Archiving and Communication Systems (PACS), medical institutions have accumulated a large amount of structured and unstructured clinical data. However, due to the lack of uniformity in data formats, naming conventions, and coding standards among systems, the information required for VTE risk assessment is scattered across multiple systems, making automatic integration and accurate identification difficult.

[0004] Currently, VTE risk assessment in most medical institutions still relies heavily on manual operation. Healthcare staff need to search for information item by item in multiple systems such as HIS, EMR, LIS, and PACS, manually match scale items, and calculate scores. This method is not only labor-intensive and dependent on human experience, but the assessment results are also easily affected by subjective judgment, making it difficult to guarantee consistency. At the same time, manual assessment requires frequent switching between multiple platforms, and repetitive work significantly increases the burden on healthcare staff.

[0005] In current clinical practice, VTE risk assessment generally faces the following problems:

[0006] (1) Risk factors are difficult to locate and integrate quickly in multi-source systems: The information required for assessment is scattered in multiple systems such as HIS, EMR, LIS, and PACS. There is a lack of a unified data access and integration mechanism, which makes manual review time-consuming and data utilization low.

[0007] (2) Manual form filling relies on item-by-item judgment, resulting in low assessment efficiency and insufficient consistency: Clinical staff need to manually judge whether patients meet the various conditions according to the scoring system. The process is cumbersome, easily affected by subjective factors, and it is difficult to guarantee the consistency and accuracy of the assessment conclusions.

[0008] (3) The assessment results lack systematic management, which is not conducive to hospital collaboration and continuous quality control: Risk factors, scoring results and recommended plans are not uniformly recorded and structured, which does not support cross-department sharing, intelligent quality control or follow-up management, and affects the overall management efficiency.

[0009] (4) Lack of traceability mechanism, and the assessment process cannot be audited and reviewed: The current assessment cannot clearly record the source, analysis link and rules of each risk factor, and lacks the ability to trace the entire link, resulting in insufficient transparency and auditability of the assessment process.

[0010] Currently, existing research on VTE risk mainly focuses on several directions, including clinical scoring scale optimization, machine learning prediction models, multi-omics and genetic susceptibility analysis, and time-series-based event prediction. However, there is still a lack of a unified method that can efficiently integrate multi-source data and automatically complete the assessment in a real clinical environment. For example, patent CN112259232B discloses a deep learning-based automatic VTE risk assessment system. This patent includes a first analysis module, which, based on a preset VTE knowledge graph and information related to all risk items, obtains the probability that all unstructured risk indicators of a patient are selected; a second analysis module, which, based on information related to a patient and a certain unstructured risk indicator, obtains the probability that a certain unstructured risk indicator of a patient is selected; a first judgment module, which, based on the results of the first and second analysis modules, determines whether an unstructured risk indicator is selected; a second judgment module, which determines whether a structured risk indicator is selected; and a risk item assessment module, which, based on the judgment results of the first and second judgment modules, determines that a risk item is risky when any risk indicator is selected. This disclosure significantly reduces the workload of nursing staff and improves the accuracy of VTE risk assessment. Another publication, CN120954726A, proposes a genetic risk assessment method and testing kit based on VTE-PRS-15. This method, which detects 15 VTE-related gene loci and combines them with the Caprini scale for comprehensive stratification, improves the accuracy of VTE genetic risk assessment in the Chinese population. However, it relies on specialized gene testing and does not utilize real-time clinical data from existing HIS, EMR, LIS, and PACS systems. It also lacks automated data integration, dynamic scoring, and outcome traceability mechanisms for bedside scenarios. Therefore, it cannot yet achieve rapid, automated, and interpretable integrated assessment of VTE risk in hospitalized patients in a routine clinical setting. Summary of the Invention

[0011] To address the aforementioned shortcomings, this invention proposes a rapid risk assessment method for venous thromboembolism based on the fusion of intelligent agents and multi-source clinical data. The method includes the following steps:

[0012] S1. Based on the Model Context Protocol (MCP), encapsulate the interfaces of multi-source hospital information systems to build a standardized data access tool that can be automatically invoked by intelligent agents;

[0013] S2. Construct a VTE data exploration intelligent agent to automatically complete task decomposition and data demand planning according to the established scoring system, and select and schedule the corresponding MCP tools in a unified context to clean, structure and semantically standardize multi-source hospital information system data.

[0014] S3, entity extraction and risk factor identification, is performed based on the unified data context output after S2 processing. This unified data context includes cleaned and standardized structured data, normalized unstructured text fragments, and a clinical event chain organized in chronological order, serving as the unified input for entity identification and logical judgment. Based on rule knowledge and the semantic capabilities of large models, the system automatically completes clinical entity extraction and risk factor identification from the above unified input, and performs semantic alignment with the scoring items to generate a preliminary list of structured risk factors.

[0015] S4. Perform multi-dimensional verification on the extracted risk factors, including semantic consistency verification, temporal logic verification, cross-data source consistency comparison, automatically correct and mark contradictions, and output the verified risk factor set and verification result record.

[0016] S5. Construct a VTE assessment intelligent agent to automatically generate and fill in the VTE risk assessment form based on risk factors, calculate the total score and risk level according to the scoring rules, and simultaneously generate the corresponding anticoagulation regimen and follow-up recommendations.

[0017] S6. Construct a traceability chain for risk factors and their data sources, analysis paths, and scoring criteria to achieve interpretable, traceable, and verifiable display of assessment results.

[0018] Preferably, the multi-source hospital information system in S1 includes a Hospital Information System (HIS), an Electronic Medical Record System (EMR), a Laboratory Information System (LIS), and an Image Archiving and Communication System (PACS), and is a standardized data access tool that can be automatically invoked by an intelligent agent by the multi-source hospital information system.

[0019] Preferably, a standardized data access tool that can be automatically invoked by an intelligent agent is constructed; specifically, this includes the following steps:

[0020] S1-1, Multi-source medical data capability extraction and unified service construction: For the key data types required for VTE risk assessment, the system first performs structured classification of the hospital's existing multi-source system data.

[0021] After sorting out the data entry points of the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Image Archiving and Communication System (PACS), Medical Record System (PN), and Order Information System (ORD), the system aggregates similar data capabilities based on medical business semantics to form coarse-grained data access services. Through this step, the originally scattered and differently formatted data access interfaces are integrated into a set of services with consistent semantics and clear logic.

[0022] (1)

[0023] In formula (1), each MCP declares its access capabilities using a structured semantic description: In formula (1), MCP represents the multi-source data access tool set constructed in this invention, and each MCP i Each corresponds to a structured data access capability unit, which is used to encapsulate the data access capabilities of Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Image Archiving and Communication System (PACS), Medical Record System (PN), and Order Information System (ORD).

[0024] S1-2, Service Registration and Standardized Interface Definition: After completing the coarse-grained data access service construction in S1-1, the system will register the functional descriptions, input conditions, output content and applicable scope of each service in a unified service directory, making it a standardized capability unit that can be automatically discovered and reused.

[0025] (2)

[0026] In formula (2), S1, S2, ..., S n The structured service tasks constructed by the system in the preceding steps are numbered sequentially according to the execution order. They are used to uniformly represent the semantic definitions of different services. DESC(Si) represents the structured semantic description of the i-th service, which includes the service's functional description, input conditions, output content, and scope of application. R is the aggregation result of all service semantic descriptions, used to form a unified service semantic space at the system level.

[0027] Subsequently, standardized interface definitions are generated for each service, specifying the call parameter format, return field structure, and call method to ensure that the VTE assessment process has a consistent interaction method when calling data from different source systems.

[0028] (3)

[0029] In formula (3), MCP_Services represents the set of MCP data services that can be directly called after encapsulation; MCP_Interfaces represents the set of standardized interface specifications corresponding to the services, used to constrain the input parameter format, return field structure and calling mode; R is the service semantic description aggregation result defined in formula (2), and the three together constitute the data access abstraction layer. Establish a unified mapping relationship between the semantic space and the data access space, so that the VTE assessment process has a consistent access method when calling data across hospital information systems (HIS), electronic medical record systems (EMR), laboratory information systems (LIS), and image archiving and communication systems (PACS).

[0030] Preferably, the key data types required for VTE risk assessment include diagnostic, surgical, laboratory, imaging, medication, medical orders, nursing records, and vital signs data. Multi-source system data are structured and categorized into database tables, electronic medical record texts, laboratory / imaging reports, and various tabular files.

[0031] Preferably, in S2, a VTE data exploration agent is constructed to automatically complete task decomposition and data requirement planning according to a predetermined scoring system. Within a unified context, appropriate MCP tools are selected and scheduled to clean, structure, and semantically standardize multi-source data. Specifically, this includes the following steps:

[0032] S2-1, Construction of VTE Data Integration Agent: Construct an agent framework for automated VTE evaluation. The agent consists of an input parsing module, a task scheduling module, a logic execution module, and a memory module. Each module achieves collaborative operation through shared cache and asynchronous task queue.

[0033] (1) Input parsing module: used to identify the evaluation scenario and applicable scoring system (Caprini, Padua, RCOG, etc.), parse the variable definition, data dependency and verification requirements corresponding to each system; map the evaluation task in natural language form into structured task units, and generate a machine-executable task description containing variable items, constraints and data call requirements;

[0034] (2) Task scheduling module: Automatically constructs task links based on the task description generated by the input parsing module, and completes the inference of the binding relationship between variables and MCP tools; generates execution plans for data access and processing according to the variable dependency order, and schedules relevant MCP tools in a unified context to support the synchronous extraction and preprocessing of multi-source data;

[0035] (3) Logic execution module: According to the execution plan determined by the task scheduling module, the acquired data is processed step by step through a structured process, including logical operations such as data cleaning, semantic normalization, field standardization and dependency verification, to generate standardized data input that can be used for risk factor analysis;

[0036] (4) Memory module: Records input parsing results, the binding relationship between variables and tools, processing intermediate states and context cache information, and stores parsed variables and completed subtasks to support cross-step reuse, reduce repeated calls and maintain context consistency in the evaluation process;

[0037] (4)

[0038] In formula (4), This refers to the data integration agent used for automated VTE assessment; The input parsing module is responsible for parsing the evaluation task and scoring system; This is the task scheduling module, responsible for building task chains and scheduling the MCP tool; As the logic execution module, it is responsible for data cleaning, standardization, and dependency verification; As a memory module, it is responsible for recording context cache and intermediate states. Through the formal representation of the agent structure, it is easy to reference and extend the capabilities of each module in a unified way in subsequent steps.

[0039] S2-2, Risk Factor Identification and Data Source Determination: Based on the item requirements of the selected scoring system, the system first clarifies the range of risk factors to be extracted. Then, based on the triggering conditions of each factor, it locates its original source in a unified data context and selects the corresponding MCP tool that can provide this type of information. Through the automatic mapping of factor-source-tool, a structured data requirement list is formed, laying the foundation for subsequent data capture and processing.

[0040] (5)

[0041] In formula (5), Represents the i-th risk factor; source( ) indicates the original data source system corresponding to this factor; tool( ) indicates the data capability tool that supports access to this factor; map( This is used to establish the binding relationship between factors and data sources and tools, ensuring that risk factors can be automatically located and retrieved by the system;

[0042] S2-3. Data capture and execution based on MCP tools: According to the data requirement list generated in S2-2, the system calls the corresponding MCP tools to extract structured and unstructured data from multiple source systems. After the captured data is input, it is automatically cleaned, formatted, timestamp-calibrated, and text segmented to remove redundant fields and abnormal records. The data from different systems is also aligned to ensure that it can be consistently identified and referenced in subsequent steps.

[0043] S2-4. Data Structuring and Risk Factor Alignment: After data capture, the system transforms multi-source inputs into a structured context, including standardized field formats, event chain sequences, and normalized text fragments. It also cross-verifies multi-source records to generate data usable for risk factor analysis. Simultaneously, the internal memory module is updated to maintain the parsed state of each variable and data dependencies, providing consistent and complete input for the next stage of risk factor identification.

[0044] (6)

[0045] In formula (6), This represents a unified data context used for VTE risk assessment; It is a structured data set that has been cleaned and standardized, and it is an important basis for triggering execution rules, comparing values, and locating events in risk factor judgment; It is a collection of unstructured text fragments that have been segmented and normalized to supplement semantic information that is difficult to cover by structured data; A sequence of clinical events organized chronologically to support temporal logic verification and event chain reconstruction;

[0046] Preferably, entity extraction and risk factor identification in S3 are performed based on the unified data context output in S2. This unified data context includes cleaned and standardized structured data, normalized unstructured text fragments, and a chronologically organized chain of clinical events, serving as unified input for entity identification and logical judgment. Based on rule knowledge and the semantic capabilities of a large model, the system automatically completes clinical entity extraction and risk factor identification from the above inputs, and semantically aligns them with the scoring items to generate a preliminary list of structured risk factors. Specifically, this includes the following steps:

[0047] S3-1, Target Entity Set Construction: Building upon the risk factor range identified in S2-2, this step further constructs a structured target clinical entity set. Based on the selected scoring system and the "factor-source-tool" mapping results completed in S2-2, the system standardizes and organizes the clinical entity types and triggering conditions corresponding to each risk factor. On this basis, the system generates a unified target entity set that can be used for entity extraction, clarifying the semantic definition, value rules, and boundary conditions of each entity type, providing a consistent standardized target set for subsequent automatic extraction, rule matching, and feature generation.

[0048] S3-2, Clinical Entity Extraction: In a unified data context, the system extracts structured data. Complete the extraction of entities that can be directly mapped by the rules, and at the same time utilize the rule templates and the semantic capabilities of the large model. Perform semantic recognition to extract risk factors that are difficult to obtain directly from structured data, thereby forming a basic entity set corresponding to the scoring system;

[0049] (7)

[0050] In formula (7), E represents the set of clinical entities extracted from the patient's medical records; each entity It is defined as a complete semantic unit, where type i Indicates entity type, used to distinguish different clinical concepts, value i Indicates the specific value or content of an entity, time i Indicates the occurrence or recording time corresponding to this entity, source i A data system or document representing the source of an entity, confidence i The confidence score or credibility score of the entity extraction result is used to reflect the reliability of rule matching or semantic recognition. Through the above definition, the system expands the original extraction result from simple field values ​​into structured semantic entities with type, time, source and confidence information, providing a unified and complete basic semantic unit for subsequent risk factor mapping, multi-dimensional verification and traceability chain construction.

[0051] S3-3, Entity and scoring factor mapping: In the structured modeling process, the system aligns the extracted entities with the specific scoring items of each scoring system; the mapping process converts the extracted and verified clinical information into a structured expression, including: factor category, triggering conditions, corresponding scoring system and score, and the data evidence from which it originates, forming a structured preliminary risk factor list;

[0052] (8)

[0053] In formula (8), This represents a structured set of risk factor mappings; each risk factor It is defined as a complete rating factor instance, where This refers to the clinical entities extracted and verified through the aforementioned steps. This indicates the triggering conditions or judgment rules for this factor in the corresponding scoring system. This represents the set of data evidence supporting the validity of the factor, including original fields, text fragments, and time records. This indicates the specific scoring system to which the factor belongs. This indicates the score or weight of the factor in the corresponding scoring system. Through the above definition, the system will uniformly model the extracted clinical entities, scoring items, triggering conditions, and sources of evidence to form a structured list of risk factors that is calculable, interpretable, and traceable, providing standardized input for subsequent multidimensional verification, score calculation, and result interpretation.

[0054] Preferably, step S4 involves performing multi-dimensional verification on the extracted risk factors, including semantic consistency verification, temporal logic verification, and cross-data source consistency comparison. Inconsistencies are automatically corrected and marked, and the verified risk factor set and verification result record are output. Specifically, this includes the following steps:

[0055] S4-1, Rule Verification and Semantic Consistency Verification: The system performs semantic consistency verification on the risk factors output by S3, checking whether the factors meet the triggering conditions, meaning definition, value range and applicable scenarios of the scoring items; It serves as the primary basis for participating in time-series verification and field consistency validation. This is used to supplement missing structured data, resolve cross-source record inconsistencies, and enhance semantic consistency judgment. The verification process compares the guideline definitions with the original clinical records to verify whether the risk factors truly meet the item requirements. This includes, but is not limited to, confirming whether "prolonged bed rest" meets the scale's required number of days, whether "postoperative events" occurred within the effective time after surgery, whether "malignant tumors" are in the active phase or still affect the current risk, and whether "age grouping" falls within the correct score range. For factors with ambiguous text descriptions, insufficient conditions, inconsistent time sequences, or mismatches with the item definitions, the system will automatically identify and remove them or mark them as items to be reviewed, thereby ensuring that the risk factors are consistent with the scoring system rules and providing reliable input for subsequent scoring.

[0056] S4-2, Time Sequence Rationality and Cross-Source Data Consistency Verification: Based on a unified data context, perform time sequence verification on the timestamps, event chains, and triggering conditions of risk factors to verify the consistency between the order of events, the effective time window, and clinical logic; at the same time, use multi-source data for cross-comparison to identify time conflicts, missing content, or inconsistencies in cross-system records, and automatically correct or mark obviously contradictory factors as requiring manual review.

[0057] (9)

[0058] In formula (9), This indicates a check of the semantic rationality of risk factors; Indicates time sequence verification; This indicates cross-source consistency verification; valid(fi) = true when all three conditions are true, and the factor is valid; otherwise, the factor is marked as abnormal or requires further verification.

[0059] S4-3, Multidimensional Integration and Validation of Risk Factors: The system integrates the results of semantic verification, temporal verification, and cross-source comparison to screen all risk factors for validity and generate a final verified risk factor set; automatic correction is performed on factors that have conflicts but can be resolved through rules; explicit annotation is performed on factors that still have uncertainties and verification result records are output.

[0060] (10)

[0061] In formula (10), This represents the final set of valid risk factors retained after multi-dimensional verification; valid(fi) is the comprehensive verification function for semantic, temporal, and cross-source consistency mentioned above. When the verification result is true, it means that the factor meets the requirements in terms of semantic definition, temporal logic, and data source. For risk factors that still have uncertainty after verification, the system will not discard them directly, but will mark them as "pending verification" and include them in a separate set of questionable factors. This set does not participate in automatic scoring, but will be saved together with its source data, parsing path, and verification results for manual review and re-inclusion or exclusion as needed, thereby ensuring the traceability and reusability of the assessment process.

[0062] Preferably, the VTE assessment agent described in step S5 automatically generates and fills out a VTE risk assessment form based on the extracted risk factors, calculates the total score and risk level according to the scoring rules, and simultaneously generates the corresponding anticoagulation regimen and follow-up recommendations. Specifically, this includes the following steps:

[0063] S5-1, Construction of VTE Assessment Agent: Constructing a VTE assessment agent for automatic scoring and clinical suggestion generation; the agent consists of an assessment planning unit, a scale processing toolkit, an assessment memory unit, and a clinical action unit, which work together in a unified context to complete the entire process from risk factor input to risk stratification and suggestion output.

[0064] (1) Assessment planning module: responsible for developing the assessment process after loading and verifying the risk factors, including planning the form generation steps according to the corresponding scoring system, determining the required scoring items, scoring calculation path and subsequent decision triggering conditions, and providing clear execution instructions for subsequent tool modules;

[0065] (2) Scale and scoring processing module: including form generation tool and scoring calculation tool, used to complete the construction of calculable forms and score execution; the form generation tool automatically maps risk factors to scale items according to the scoring system template and fills in missing fields according to the rules; the scoring calculation tool completes the score calculation and risk level determination according to the scale rules and outputs structured scoring results;

[0066] (3) Clinical decision-making module: Based on the scoring results and risk level, the module automatically generates clinical recommendations; the module calls the prevention and treatment rules in the knowledge base to derive individualized anticoagulation regimens, physical prevention measures, monitoring requirements and re-evaluation time; the output results are presented in a structured manner and can be directly used for clinical decision-making or patient management, realizing the closed loop from assessment to action;

[0067] (4) Assessment memory module: It is responsible for storing short-term and long-term information during the assessment process; short-term memory stores risk factors, field mapping results, scoring forms and intermediate calculation states; long-term memory stores guidelines, scoring system knowledge and clinical pathway content; the module ensures data continuity and traceability of execution results during the assessment process, providing a stable basis for subsequent decision-making and traceability;

[0068] S5-2, Assessment Form Generation: After receiving the risk factors output by S4, the system calls the form generation tool, loads the corresponding scale template according to the selected scoring system, and matches the risk factors with the standard items one by one; it automatically fills in the factors that can be clearly identified, and completes the missing items according to the preset rules to form a structured VTE assessment form that can be directly used for calculation.

[0069] S5-3. Scoring and Risk Level Calculation: Based on the generated structured form, the system calls the scoring calculation tool and performs score calculation according to the rules of each scoring system, including item scores, condition combinations and cumulative summation. It also generates risk levels based on preset thresholds. The scoring process is executed using a unified logical framework to ensure that the calculation paths between different scales are clear, the rules are consistent, and the results are traceable. The final output includes item scores, total scores, risk levels and corresponding judgment criteria, and is synchronously written into the assessment context.

[0070] (11)

[0071] In formula (11), This represents the patient's overall VTE risk score; Indicates the first A set of risk factors under each scoring dimension, with each scoring dimension being independent of the others; This indicates the score or weight of the risk factor in the corresponding scoring system; This is a trigger indicator variable for risk factors, taking a value of 1 if and only if its trigger condition is met, otherwise taking a value of 0; constraint conditions. This is used to ensure that the triggering conditions of each risk factor within the same scoring dimension are mutually exclusive, and to avoid the repeated accumulation of scores due to the simultaneous fulfillment of multiple conditions within a single dimension, thereby ensuring that the scoring calculation logic is consistent and the results are interpretable.

[0072] (12)

[0073] In formula (12), , The preset risk threshold is denoted as Score, which is the total score calculated by formula (7). Risk is the final risk level, which enables automated stratification and classification of VTE risk in patients.

[0074] S5-4. Clinical Recommendation Generation: After completing the assessment form and risk level determination, the system automatically generates individualized prevention and treatment recommendations based on the combination of scoring results and key risk factors. The recommendations cover drug or non-drug preventive measures, anticoagulant medication selection and monitoring requirements, physical prevention methods, and recommended reassessment times. The derivation of the recommendations is based on long-term memory knowledge to complete the judgment.

[0075] (13)

[0076] In formula (13), Plan represents the set of individualized clinical recommendations generated by the system, including anticoagulation regimen, physical precautions, monitoring requirements, and reassessment time; Score is the aforementioned total VTE score; This is a subset of key factors selected from all valid risk factors; the function g(·) represents the decision mapping process based on guideline rules and knowledge base, mapping "score level + key factor pattern" to specific clinical action plans, thereby achieving an automatic closed loop from quantitative assessment to actionable recommendations.

[0077] Preferably, the construction of the traceability chain of risk factors and their data sources, analysis paths, and scoring criteria, as described in S6, to achieve interpretable, traceable, and verifiable display of the assessment results, specifically includes the following steps:

[0078] S6-1, Source Tracing Element Collection: Automatically collect key elements related to risk factors from the entire assessment process. The system records various evidence information in a unified context to form a basic evidence set that can be used for subsequent source tracing.

[0079] S6-2, Source tracing chain construction: Based on the content of the evidence set, risk factors, data sources, analysis steps, reasoning logic and scoring criteria are organized into a chain structure according to time order and semantic dependencies. The system automatically generates a full path mapping of "factor - data - analysis - rule" to realize a continuous and structured source tracing chain from the original record to the final scoring result.

[0080] (14)

[0081] In formula (14), Trace represents the traceability chain constructed for a single VTE risk assessment; each As a node in the traceability chain, it contains key information such as risk factors, original data sources, analysis steps, scoring items used and corresponding scores; by organizing nodes into a sequence structure according to execution order and dependencies, the system can completely record the entire path "from original data to final score and recommendation", providing a visual and verifiable chain of evidence for post-audit, result interpretation and manual review;

[0082] S6-3, Interpretable Output and Verification Display, transforms the completed traceability chain into auditable and interpretable output. The output results support clinical staff review, quality control checks, and manual auditing, ensuring that the evaluation results are interpretable, traceable, and verifiable.

[0083] The beneficial effects of this invention are:

[0084] This invention constructs a data integration intelligent agent for VTE risk assessment, which can automatically plan data requirements, locate data sources and call corresponding MCP tools according to different scoring systems, and complete the cleaning, standardization and semantic alignment of structured and unstructured data, providing a stable and unified data foundation for risk factor analysis.

[0085] This invention constructs a complete traceability chain for VTE assessment results, enabling each risk factor, scoring item, and clinical recommendation to be traced back to the original data, analysis process, and rules followed. This provides transparent and auditable evidence support for result interpretation, quality management, and manual review, significantly enhancing the credibility and clinical usability of the system. Attached Figure Description

[0086] Figure 1 This is a flowchart of the overall process of a rapid VTE risk assessment method based on an intelligent agent according to the present invention.

[0087] Figure 2 This is a flowchart illustrating the abstract description of the multi-source data access interface and the construction of the MCP tool in this invention;

[0088] Figure 3 This is a flowchart of the VTE data integration intelligent agent processing of the present invention;

[0089] Figure 4 This is a flowchart illustrating the risk factor identification, entity extraction, and multi-dimensional verification process of this invention.

[0090] Figure 5 This is a flowchart of the VTE assessment agent of the present invention performing form generation, scoring reasoning, and clinical suggestion generation based on risk factors;

[0091] Figure 6 This is a flowchart of the risk factor tracing chain construction and interpretability output method of the present invention. Detailed Implementation

[0092] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0093] This invention proposes a rapid VTE risk assessment method based on intelligent agents, the method comprising the following steps:

[0094] S1. Based on the MCP (Model Context Protocol), encapsulate the interfaces of multiple systems such as HIS, EMR, LIS, and PACS to build a standardized data access tool that can be automatically invoked by intelligent agents.

[0095] S2. Construct a VTE data exploration intelligent agent to automatically complete task decomposition and data demand planning according to the established scoring system, and select and schedule the corresponding MCP tools in a unified context to clean, structure and semantically standardize multi-source data.

[0096] S3, entity extraction and risk factor identification, is performed based on the unified data context output by S2. This context includes cleaned and standardized structured data, normalized unstructured text fragments, and a chronologically organized chain of clinical events, serving as unified input for entity identification and logical judgment. Based on rule knowledge and the semantic capabilities of a large model, the system automatically performs clinical entity extraction and risk factor identification from the above inputs, semantically aligns them with scoring items, and generates a preliminary list of structured risk factors.

[0097] S4. Perform multi-dimensional verification on the extracted risk factors, including semantic consistency verification, temporal logic verification, cross-data source consistency comparison, automatically correct and mark contradictions, and output the verified risk factor set and verification result record.

[0098] S5. Construct a VTE assessment intelligent agent to automatically generate and fill in the VTE risk assessment form based on risk factors, calculate the total score and risk level according to the scoring rules, and simultaneously generate the corresponding anticoagulation regimen and follow-up recommendations.

[0099] S6. Construct a traceability chain for risk factors and their data sources, analysis paths, and scoring criteria to achieve interpretable, traceable, and verifiable display of assessment results.

[0100] The above-described S1, which encapsulates multi-source system interfaces such as HIS, EMR, LIS, and PACS based on the MCP protocol to construct a standardized data access tool that can be automatically invoked by an intelligent agent, specifically includes the following steps:

[0101] S1-1, Multi-source medical data capability extraction and unified service construction: For the key data types required for VTE risk assessment (diagnosis, surgery, laboratory, imaging, medication, medical orders, nursing records, vital signs, etc.), the system first performs structured classification of the hospital's existing multi-source system data, including database tables, electronic medical record texts, laboratory / imaging reports, and various tabular files.

[0102] After analyzing the data entry points of business systems such as HIS, EMR, LIS, PACS, Progress Notes (PN), and Physician Orders Management (ORD), the system aggregates similar data capabilities based on medical business semantics to form coarse-grained data access services. Through this step, the originally scattered and inconsistent data access interfaces are integrated into a set of services with consistent semantics and clear logic.

[0103] (1)

[0104] In formula (1), each MCP declares its access capabilities using a structured semantic description: In formula (1), MCP represents the multi-source data access tool set constructed in this invention, and each MCP i Each corresponds to a structured data access capability unit, used to encapsulate the data access capabilities of systems such as HIS, EMR, LIS, PACS, PN, and ORD.

[0105] S1-2, Service Registration and Standardized Interface Definition: After completing the coarse-grained service construction in S1-1, the system registers the functional descriptions, input conditions, output content, and applicable scope of each service in a unified service directory, making them standardized capability units that can be automatically discovered and reused.

[0106] (2)

[0107] In formula (2), S1, S2, ..., S n This represents the structured service tasks that the system has constructed in the preceding steps, numbered sequentially according to the execution order, used to uniformly represent the semantic definitions of different services. DESC(Si) represents the structured semantic description of the i-th service, including the service's functional description, input conditions, output content, and scope of application; R is the aggregation result of all service semantic descriptions, used to form a unified service semantic space at the system level.

[0108] Subsequently, standardized interface definitions are generated for each service, specifying the call parameter format, return field structure, and call method to ensure that the VTE assessment process has a consistent interaction method when calling data from different source systems.

[0109] (3)

[0110] In formula (3), MCP_Services represents the set of MCP data services that can be directly called after encapsulation; MCP_Interfaces represents the set of standardized interface specifications corresponding to the services, used to constrain the input parameter format, return field structure, and calling mode; R is the aggregated result of the service semantic description defined in formula (2). The three together constitute the data access abstraction layer. A unified mapping relationship is established between the semantic space and the data access space, so that the VTE evaluation process has a consistent access method when calling data across systems such as HIS, EMR, LIS, and PACS.

[0111] Through this step, the multi-source data upon which VTE evaluation relies can be accessed by the agent in a unified manner.

[0112] The VTE data integration intelligent agent described in S2 above automatically completes task decomposition and data requirement planning according to a predetermined scoring system, and selects and schedules the corresponding MCP tools in a unified context to clean, structure, and semantically standardize multi-source data. Specifically, this includes the following steps:

[0113] S2-1, Construction of VTE Data Integration Agent: Construct an agent framework for automated VTE evaluation. The agent consists of an input parsing module, a task scheduling module, a logic execution module, and a memory module. Each module operates collaboratively through a shared cache and an asynchronous task queue.

[0114] (1) Input parsing module: used to identify the evaluation scenario and applicable scoring system (Caprini, Padua, RCOG, etc.), parse the variable definition, data dependency and verification requirements corresponding to each system; map the evaluation task in natural language form into structured task units, and generate a machine-executable task description containing variable items, constraints and data call requirements.

[0115] (2) Task scheduling module: Automatically constructs task links based on the task description generated by the input parsing module, and completes the inference of the binding relationship between variables and MCP tools; generates execution plans for data access and processing according to the variable dependency order, and schedules relevant MCP tools in a unified context to support the synchronous extraction and preprocessing of multi-source data.

[0116] (3) Logic execution module: According to the execution plan determined by the task scheduling module, the acquired data is processed step by step through a structured process, including logical operations such as data cleaning, semantic normalization, field standardization and dependency verification, to generate standardized data input that can be used for risk factor analysis.

[0117] (4) Memory module: Records input parsing results, the binding relationship between variables and tools, processing intermediate states and context cache information, and stores parsed variables and completed subtasks to support cross-step reuse, reduce repeated calls and maintain context consistency in the evaluation process.

[0118] (4)

[0119] In formula (4), This refers to the data integration agent used for automated VTE assessment; The input parsing module is responsible for parsing the evaluation task and scoring system; This is the task scheduling module, responsible for building task chains and scheduling the MCP tool; As the logic execution module, it is responsible for data cleaning, standardization, and dependency verification; This is a memory module responsible for recording context caches and intermediate states. A formal representation of the agent's structure facilitates the unified referencing and extension of the capabilities of each module in subsequent steps.

[0120] S2-2. Risk Factor Identification and Data Source Determination: Based on the requirements of the selected scoring system (Caprini, Padua, RCOG, etc.), the system first clarifies the range of risk factors to be extracted, such as age, type of surgery, duration of bed rest, malignancy, pregnancy-related risks, and history of VTE. Then, based on the triggering conditions of each factor, its original source is located within a unified data context, such as diagnosis, surgery, laboratory tests, medication, nursing records, progress notes, or imaging reports, and the corresponding MCP tool that can provide this type of information is selected. Through automatic factor-source-tool mapping, a structured data requirement list is formed, laying the foundation for subsequent data capture and processing.

[0121] (5)

[0122] In formula (5), Represents the i-th risk factor; source( ) indicates the original data source system corresponding to this factor (such as EMR, LIS, PACS, etc.); tool( The ) indicates the data capability tool that supports access to this factor (such as MCP_EMR, MCP_LIS). map( This is used to establish the binding relationship between factors and data sources and tools, ensuring that risk factors can be automatically located and retrieved by the system.

[0123] S2-3. Data Acquisition and Execution Based on MCP Tools: Based on the data requirement list generated in 2.2, the system calls the corresponding MCP tools to extract structured and unstructured data from multiple sources, including diagnostic codes, surgical records, test results, medication information, vital signs, and medical records and nursing texts. After input, the acquired data is automatically cleaned, formatted, timestamped, and segmented to remove redundant fields and abnormal records. Basic alignment is performed on data from different systems to ensure consistent identification and referencing in subsequent steps.

[0124] S2-4. Data Structuring and Risk Factor Alignment: After data capture, the system transforms multi-source inputs into a structured context, including standardized field formats, event chain sequences, and normalized text fragments. It also cross-verifies multi-source records to generate data usable for risk factor analysis. Simultaneously, the internal memory module is updated to maintain the parsed state of each variable and data dependencies, providing consistent and complete input for the next stage of risk factor identification.

[0125] (6)

[0126] In formula (6), This represents the unified data context used for VTE risk assessment; The cleaned and standardized structured data set (such as fields for diagnosis, surgery, testing, and medication) serves as an important basis for triggering execution rules, comparing values, and locating events in risk factor judgment. It is a collection of unstructured text fragments (such as medical records, nursing records, and imaging reports) that have been segmented and normalized to supplement semantic information that is difficult to cover by structured data; This is a chronologically organized sequence of clinical events, used to support temporal logic verification and event chain reconstruction. Through this contextual structure, subsequent risk factor extraction and verification can be performed under a unified view. Structured Events All events extracted from the text will be projected onto a unified event timeline. It is used to perform the timing logic verification of S4, the event chain reconstruction, and the "time judgment of scoring trigger conditions" in S5.

[0127] Provides highly deterministic basic data. It undertakes the tasks of complex semantic supplementation and state recognition. Both work together within a unified event time structure to support the complete risk factor extraction and verification process, enabling the system to simultaneously possess structured rule inference capabilities and unstructured understanding capabilities.

[0128] The entity extraction and risk factor identification described in S3 above are performed based on the unified data context output in S2. This context includes cleaned and standardized structured data, normalized unstructured text fragments, and a chronologically organized chain of clinical events, serving as the unified input for entity identification and logical judgment. Based on rule knowledge and the semantic capabilities of a large model, the system automatically performs clinical entity extraction and risk factor identification from the above input, semantically aligns it with the scoring items, and generates a preliminary structured risk factor list, specifically including the following steps:

[0129] S3-1, Target Entity Set Construction: Building upon the risk factors identified in S2-2, this step further constructs a structured set of target clinical entities. Based on the selected scoring system (Caprini, Padua, RCOG, etc.) and the "factor-source-tool" mapping results completed in S2-2, the system standardizes and organizes the clinical entity types and triggering conditions corresponding to each risk factor. Examples include age, previous VTE, surgical type and duration, activity limitation, bedridden time, malignancy, pregnancy and postpartum period, central venous access, infection, heart failure, and respiratory failure. Based on this, the system generates a unified target entity set that can be used for entity extraction, clearly defining the semantic definition, value rules, and boundary conditions for each type of entity, providing a consistent standardized target set for subsequent automatic extraction, rule matching, and feature generation.

[0130] S3-2, Clinical Entity Extraction: In a unified data context, the system extracts structured data. Complete the extraction of entities that can be directly mapped by the rules (such as age, test value, whether surgery has occurred, etc.), and at the same time utilize the rule templates and the semantic capabilities of the large model to... Semantic recognition is performed to extract risk factors that are difficult to obtain directly from structured data, such as "limited activity", "infection status" and "postoperative condition", thereby forming a basic entity set corresponding to the scoring system.

[0131] (7)

[0132] In formula (7), E represents the set of clinical entities extracted from the patient's medical records; each entity It is defined as a complete semantic unit, where type i Indicates entity type, used to distinguish different clinical concepts such as diagnosis name, symptom presentation, laboratory indicators, medication information, surgical events, etc. i Indicates the specific value or content of an entity, time i Indicates the occurrence or recording time corresponding to this entity, source iThe data system or document (including EMR, LIS, PACS, medical records, etc.) from which the entity originates, confidence i This represents the credibility or confidence score of the entity extraction results, used to reflect the reliability of rule matching or semantic recognition. Through the above definition, the system expands the original extraction results from simple field values ​​into structured semantic entities with type, time, source, and credibility information, providing a unified and complete basic semantic unit for subsequent risk factor mapping, multi-dimensional verification, and traceability chain construction.

[0133] S3-3, Entity-Rating Factor Mapping: In the structured modeling process, the system aligns the extracted entities with the specific scoring items of each scoring system. The mapping process converts the extracted and verified clinical information into a structured expression, including: factor category (age, past medical history, surgery-related factors, activity restriction, etc.), triggering conditions ("≥72 hours of bed rest", "events within 48 hours postoperatively"), corresponding scoring system (Caprini, Padua, RCOG, etc.), score, and the data evidence from which it originates (raw fields, text fragments, time records, etc.), forming a structured preliminary list of risk factors.

[0134] (8)

[0135] In formula (8), This represents a structured set of risk factor mappings; each risk factor It is defined as a complete rating factor instance, where This refers to the clinical entities extracted and verified through the aforementioned steps. This indicates the triggering conditions or judgment rules for the factor in the corresponding scoring system. This represents the set of data evidence supporting the validity of the factor, including original fields, text fragments, and time records, etc. This indicates the specific scoring system to which the factor belongs. This indicates the score or weight of the factor in the corresponding scoring system. Through the above definition, the system uniformly models the extracted clinical entities, scoring items, triggering conditions, and sources of evidence, forming a structured list of risk factors that is calculable, interpretable, and traceable, providing standardized input for subsequent multidimensional verification, score calculation, and result interpretation.

[0136] The above-described S4 describes performing multi-dimensional verification on the extracted risk factors, including semantic consistency verification, temporal logic verification, and cross-data source consistency comparison. It automatically corrects and marks inconsistencies, and outputs the verified risk factor set and verification result record. Specifically, this includes the following steps:

[0137] S4-1, Rule Verification and Semantic Consistency Validation. The system performs semantic consistency verification on the risk factors output by S3, checking whether the factors meet the triggering conditions, meaning definition, value range, and applicable scenarios of the scoring items. It serves as the primary basis for participating in time sequence verification and field consistency validation. This is used to supplement missing structured data, resolve inconsistencies in cross-source records, and enhance semantic consistency. The verification process compares the risk factors with the guideline definitions and original clinical records (such as diagnosis names, time descriptions, nursing record statements, etc.) to verify whether the risk factors truly meet the item requirements. This includes, but is not limited to, confirming whether "prolonged bed rest" meets the scale's required number of days, whether "postoperative events" occurred within the effective timeframe after surgery, whether "malignant tumors" are in the active phase or still affect current risk, and whether "age grouping" falls within the correct score range. For factors with ambiguous text descriptions, insufficient conditions, inconsistent time sequences, or mismatches with the item definitions, the system will automatically identify and remove them or mark them as items to be reviewed, thereby ensuring that the risk factors are consistent with the scoring system rules and providing reliable input for subsequent scoring.

[0138] S4-2, Timing Reasonableness and Cross-Source Data Consistency Verification. Based on a unified data context, timing verification is performed on the timestamps, event chains, and triggering conditions of risk factors to verify the consistency between the order of events, effective time windows, and clinical logic. Simultaneously, cross-referencing is performed using multi-source data such as diagnostics, medical orders, laboratory tests, imaging, and nursing to identify issues such as time conflicts, missing content, or inconsistencies across systems. Factors with obvious contradictions are automatically corrected or marked as requiring manual review.

[0139] (9)

[0140] In formula (9), This indicates a check of the semantic rationality of risk factors; Indicates time sequence verification; This indicates cross-source consistency verification. If all three conditions are true, valid(fi) = true, and the factor is valid; otherwise, the factor is marked as abnormal or requires further verification.

[0141] S4-3. Multidimensional Integration and Risk Factor Validity Confirmation. The system integrates the results of semantic verification, temporal validation, and cross-source comparison to screen all risk factors for validity, generating a final verified risk factor set. Automatic correction is performed on factors with conflicts that can be resolved through rules; factors with remaining uncertainties are explicitly labeled, and verification result records are output.

[0142] (10)

[0143] In formula (10), This represents the final set of valid risk factors retained after multi-dimensional verification. `valid(fi)` is the comprehensive verification function for semantic, temporal, and cross-source consistency mentioned above. When the verification result is true, it indicates that the factor meets the requirements in terms of semantic definition, temporal logic, and data source. For risk factors that still have uncertainty after verification, the system will not discard them directly, but will mark them as "pending review" and include them in a separate set of questionable factors. This set does not participate in automatic scoring, but will be saved along with its source data, parsing path, and verification results for manual review and re-inclusion or exclusion as needed, thereby ensuring the traceability and reusability of the assessment process.

[0144] The VTE assessment agent described in S5 above, which automatically generates and fills out a VTE risk assessment form based on extracted risk factors, calculates the total score and risk level according to scoring rules, and simultaneously generates corresponding anticoagulation regimens and follow-up recommendations, is characterized by including the following steps:

[0145] S5-1, Construction of VTE Assessment Agent: This section describes the construction of a VTE assessment agent for automated scoring and clinical suggestion generation. The agent consists of an assessment planning unit, a scale processing toolkit, an assessment memory unit, and a clinical action unit. Through coordinated operation within a unified context, it completes the entire process from risk factor input to risk stratification and suggestion output.

[0146] (1) Assessment planning module: responsible for developing the assessment process after loading and verifying the risk factors, including planning the form generation steps according to the corresponding scoring system, determining the required scoring items, scoring calculation path and subsequent decision triggering conditions, and providing clear execution instructions for subsequent tool modules.

[0147] (2) Scale and scoring processing module: including form generation tool and scoring calculation tool, used to complete the construction of calculable forms and score execution. The form generation tool automatically maps risk factors to scale items according to the scoring system template and fills in missing fields according to the rules; the scoring calculation tool completes the score calculation and risk level determination according to the scale rules and outputs structured scoring results.

[0148] (3) Clinical Decision Module: Based on the scoring results and risk level, the module automatically generates clinical recommendations. The module calls on the prevention and treatment rules in the knowledge base to derive individualized anticoagulation regimens, physical prevention measures, monitoring requirements, and reassessment times. The output results are presented in a structured manner and can be directly used for clinical decision-making or patient management, realizing a closed loop from assessment to action.

[0149] (4) Assessment Memory Module: This module is responsible for storing short-term and long-term information during the assessment process. Short-term memory stores risk factors, field mapping results, scoring forms, and intermediate calculation states; long-term memory stores guidelines, scoring system knowledge, and clinical pathway content. This module ensures data continuity and traceability of results during the assessment process, providing a stable basis for subsequent decision-making and traceability.

[0150] S5-2, Assessment Form Generation. After receiving the risk factors output from S4, the system calls the form generation tool, loads the corresponding scale template according to the selected scoring system, and matches the risk factors with the standard items one by one. Automatic filling is performed on factors with clear corresponding values, and missing items are completed according to preset rules, forming a structured VTE assessment form that can be directly used for calculation.

[0151] S5-3. Scoring and Risk Level Calculation: Based on the generated structured form, the system calls the scoring calculation tool to perform score calculation according to the rules of each scoring system, including item scores, condition combinations, and cumulative summation, and generates a risk level (low / medium / high risk) based on preset thresholds. The scoring process is executed using a unified logical framework to ensure clear calculation paths, consistent rules, and traceable results across different scales. The final output includes item scores, total scores, risk levels, and corresponding judgment criteria, and is synchronously written into the assessment context.

[0152] (11)

[0153] In formula (11), This represents the patient's overall VTE risk score; Indicates the first A set of risk factors under each scoring dimension, with each scoring dimension being independent of the others; This indicates the score or weight of the risk factor in the corresponding scoring system; This is a trigger indicator variable for risk factors, taking a value of 1 if and only if its trigger condition is met, otherwise taking a value of 0; constraint conditions. This is used to ensure that the triggering conditions of each risk factor within the same scoring dimension are mutually exclusive, avoiding the repeated accumulation of scores due to the simultaneous fulfillment of multiple conditions within a single dimension, thereby ensuring that the scoring calculation logic is consistent and the results are interpretable.

[0154] (12)

[0155] In formula (12), , The preset risk threshold is denoted as ; Score is the total score calculated by formula (7); and Risk is the final risk level. This formula enables automated stratification and classification of patients' VTE risk.

[0156] S5-4. Clinical Recommendation Generation: After completing the assessment form and risk level determination, the system automatically generates individualized prevention and treatment recommendations based on the combination of scoring results and key risk factors. The recommendations cover pharmacological or non-pharmacological preventative measures, anticoagulant medication selection and monitoring requirements, physical prevention methods, and recommended follow-up assessment times. The derivation of these recommendations is based on long-term memory knowledge.

[0157] (13)

[0158] In formula (13), Plan represents the set of individualized clinical recommendations generated by the system, including anticoagulation regimens, physical precautions, monitoring requirements, and reassessment times; Score is the aforementioned total VTE score; The key factor subset selected from all effective risk factors (such as malignant tumors, previous VTE history, long surgery, etc.) is represented by the function g(·), which represents the decision mapping process based on guideline rules and knowledge base, mapping "score level + key factor pattern" to specific clinical action plan, thereby realizing an automatic closed loop from quantitative assessment to actionable recommendations.

[0159] The above-described S6 section describes constructing a traceability chain for risk factors, their data sources, analytical paths, and scoring criteria to achieve interpretable, traceable, and verifiable presentation of assessment results. This specifically includes the following steps:

[0160] S6-1, Source Tracing Element Collection: This feature automatically collects key elements related to risk factors from the entire assessment process, including original data sources, structured fields, semantically extracted fragments, event timestamps, dependency rules, and processing paths. The system records all evidentiary information within a unified context, forming a basic evidence set that can be used for subsequent source tracing.

[0161] S6-2. Source tracing chain construction: Based on the evidence set content, risk factors, data sources, analysis steps, reasoning logic, and scoring criteria are organized into a chain structure according to chronological order and semantic dependencies. The system automatically generates a full-path mapping of "factor—data—analysis—rule," realizing a continuous and structured source tracing chain from the original record to the final scoring result.

[0162] (14)

[0163] In formula (14), Trace represents the traceability chain constructed for a single VTE risk assessment; each As a node in the traceability chain, it contains key information such as risk factors, original data sources (e.g., specific diagnostic records, test results, or text fragments), parsing steps (entity extraction, rule matching, or large model inference), scoring items used, and corresponding scores. By organizing nodes into a sequence structure according to execution order and dependencies, the system can completely record the entire path "from original data to final score and recommendation," providing a visual and verifiable chain of evidence for post-audit, result interpretation, and manual review.

[0164] S6-3, Interpretable Output and Verification Display, transforms the completed traceability chain into auditable and interpretable output, including factor source comparison, trigger logic explanation, scoring basis analysis, and decision path display. The output supports clinical staff review, quality control checks, and manual auditing, ensuring the assessment results are interpretable, traceable, and verifiable.

[0165] The above are preferred embodiments of the present invention. Those skilled in the art should understand that the embodiments of the present invention are not limited to the above content. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A rapid risk assessment method for venous thromboembolism based on intelligent agents and multi-source clinical data fusion, characterized in that: Includes the following steps: S1. Based on the Model Context Protocol (MCP), encapsulate the interfaces of multi-source hospital information systems to build a standardized data access tool that can be automatically invoked by intelligent agents; S2. Construct a VTE data exploration intelligent agent to automatically complete task decomposition and data demand planning according to the established scoring system, and select and schedule the corresponding MCP tools in a unified context to clean, structure and semantically standardize multi-source hospital information system data. S3, entity extraction and risk factor identification, is performed based on the unified data context output after S2 processing. This unified data context includes cleaned and standardized structured data, normalized unstructured text fragments, and a clinical event chain organized in chronological order, serving as the unified input for entity identification and logical judgment. Based on rule knowledge and the semantic capabilities of large models, the system automatically completes clinical entity extraction and risk factor identification from the above unified input, and performs semantic alignment with the scoring items to generate a preliminary list of structured risk factors. S4. Perform multi-dimensional verification on the extracted risk factors, including semantic consistency verification, temporal logic verification, cross-data source consistency comparison, automatically correct and mark contradictions, and output the verified risk factor set and verification result record. S5. Construct a VTE assessment intelligent agent to automatically generate and fill in the VTE risk assessment form based on risk factors, calculate the total score and risk level according to the scoring rules, and simultaneously generate the corresponding anticoagulation regimen and follow-up recommendations. S6. Construct a traceability chain for risk factors and their data sources, analysis paths, and scoring criteria to achieve interpretable, traceable, and verifiable display of assessment results.

2. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 1, characterized in that: The multi-source hospital information system in S1 includes the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), and Picture Archiving and Communication System (PACS). It is a standardized data access tool built by the multi-source hospital information system that can be automatically invoked by intelligent agents.

3. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 2, characterized in that: Build standardized data access tools that can be automatically invoked by intelligent agents; Specifically, the following steps are included: S1-1, Multi-source medical data capability extraction and unified service construction: For the key data types required for VTE risk assessment, the system first performs structured classification of the hospital's existing multi-source system data. After sorting out the data entry points of the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Image Archiving and Communication System (PACS), Medical Record System (PN), and Order Information System (ORD), the system aggregates similar data capabilities based on medical business semantics to form coarse-grained data access services. Through this step, the originally scattered and differently formatted data access interfaces are integrated into a set of services with consistent semantics and clear logic. （1） In formula (1), each MCP declares its access capabilities using a structured semantic description: In formula (1), MCP represents the multi-source data access tool set constructed in this invention, and each MCP i Each corresponds to a structured data access capability unit, which is used to encapsulate the data access capabilities of Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Image Archiving and Communication System (PACS), Medical Record System (PN), and Order Information System (ORD). S1-2, Service Registration and Standardized Interface Definition: After completing the coarse-grained data access service construction in S1-1, the system will register the functional descriptions, input conditions, output content and applicable scope of each service in a unified service directory, making it a standardized capability unit that can be automatically discovered and reused. （2） In formula (2), S1, S2, ..., S n This represents the structured service tasks constructed by the system in the preceding steps, numbered sequentially according to execution order. These tasks are used to uniformly represent the semantic definitions of different services. DESC(Si) represents the structured semantic description of the i-th service, including its functional description, input conditions, output content, and scope of application. R is the aggregated result of all service semantic descriptions, used to form a unified service semantic space at the system level. Subsequently, standardized interface definitions are generated for each service, specifying the call parameter format, return field structure, and calling method, ensuring that the VTE evaluation process has a consistent interaction method when calling data from different source systems. （3） In formula (3), MCP_Services represents the set of MCP data services that can be directly called after encapsulation; MCP_Interfaces represents the set of standardized interface specifications corresponding to the services, used to constrain the input parameter format, return field structure and calling mode; R is the service semantic description aggregation result defined in formula (2), and the three together constitute the data access abstraction layer. Establish a unified mapping relationship between the semantic space and the data access space, so that the VTE assessment process has a consistent access method when calling data across hospital information systems (HIS), electronic medical record systems (EMR), laboratory information systems (LIS), and image archiving and communication systems (PACS).

4. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 3, characterized in that: Key data types required for VTE risk assessment include diagnostic, surgical, laboratory, imaging, medication, medical orders, nursing records, and vital signs data. Multi-source system data are structured and categorized, including database tables, electronic medical record texts, laboratory / imaging reports, and various tabular files.

5. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 1, characterized in that: In S2, a VTE data exploration agent is built to automatically complete task decomposition and data requirement planning based on a predetermined scoring system. Within a unified context, appropriate MCP tools are selected and scheduled to clean, structure, and semantically standardize multi-source data. Specifically, this includes the following steps: S2-1, Construction of VTE Data Integration Agent: Construct an agent framework for automated VTE evaluation. The agent consists of an input parsing module, a task scheduling module, a logic execution module, and a memory module. Each module achieves collaborative operation through shared cache and asynchronous task queue. (1) Input parsing module: used to identify the evaluation scenario and applicable scoring system (Caprini, Padua, RCOG, etc.), parse the variable definition, data dependency and verification requirements corresponding to each system; map the evaluation task in natural language form into structured task units, and generate a machine-executable task description containing variable items, constraints and data call requirements; (2) Task scheduling module: Automatically constructs task links based on the task description generated by the input parsing module, and completes the inference of the binding relationship between variables and MCP tools; generates execution plans for data access and processing according to the variable dependency order, and schedules relevant MCP tools in a unified context to support the synchronous extraction and preprocessing of multi-source data; (3) Logic execution module: According to the execution plan determined by the task scheduling module, the acquired data is processed step by step through a structured process, including logical operations such as data cleaning, semantic normalization, field standardization and dependency verification, to generate standardized data input that can be used for risk factor analysis; (4) Memory module: Records input parsing results, the binding relationship between variables and tools, processing intermediate states and context cache information, and stores parsed variables and completed subtasks to support cross-step reuse, reduce repeated calls and maintain context consistency in the evaluation process; (4) In formula (4), This refers to the data integration agent used for automated VTE assessment; The input parsing module is responsible for parsing the evaluation task and scoring system; This is the task scheduling module, responsible for building task chains and scheduling the MCP tool; As the logic execution module, it is responsible for data cleaning, standardization, and dependency verification; As a memory module, it is responsible for recording the context cache and intermediate states. Through the formal representation of the agent structure, it is easy to reference and extend the capabilities of each module in a unified way in subsequent steps. S2-2, Risk Factor Identification and Data Source Determination: Based on the item requirements of the selected scoring system, the system first clarifies the range of risk factors to be extracted. Then, based on the triggering conditions of each factor, it locates its original source in a unified data context and selects the corresponding MCP tool that can provide this type of information. Through the automatic mapping of factor-source-tool, a structured data requirement list is formed, laying the foundation for subsequent data capture and processing. (5) In formula (5), Represents the i-th risk factor; source( ) indicates the original data source system corresponding to this factor; tool( ) indicates the data capability tool that supports access to this factor; map( This is used to establish the binding relationship between factors and data sources and tools, ensuring that risk factors can be automatically located and retrieved by the system; S2-3. Data capture and execution based on MCP tools: According to the data requirement list generated in S2-2, the system calls the corresponding MCP tools to extract structured and unstructured data from multiple source systems. After the captured data is input, it is automatically cleaned, formatted, timestamp-calibrated, and text segmented to remove redundant fields and abnormal records. The data from different systems is also aligned to ensure that it can be consistently identified and referenced in subsequent steps. S2-4. Data Structuring and Risk Factor Alignment: After data capture, the system transforms multi-source inputs into a structured context, including standardized field formats, event chain sequences, and normalized text fragments. It also cross-verifies multi-source records to generate data usable for risk factor analysis. Simultaneously, the internal memory module is updated to maintain the parsed state of each variable and data dependencies, providing consistent and complete input for the next stage of risk factor identification. (6) In formula (6), 𝐶 represents the unified data context used for VTE risk assessment; It is a structured data set that has been cleaned and standardized, and it is an important basis for triggering execution rules, comparing values, and locating events in risk factor judgment; It is a collection of unstructured text fragments that have been segmented and normalized to supplement semantic information that is difficult to cover by structured data; It is a sequence of clinical events organized in chronological order to support temporal logic verification and event chain reconstruction.

6. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 1, characterized in that: Entity extraction and risk factor identification in S3 are performed based on the unified data context output by S2. This unified data context includes cleaned and standardized structured data, normalized unstructured text fragments, and a chronologically organized chain of clinical events, serving as unified input for entity identification and logical judgment. Based on rule knowledge and the semantic capabilities of a large model, the system automatically performs clinical entity extraction and risk factor identification from the above inputs, semantically aligns them with scoring items, and generates a preliminary list of structured risk factors. Specifically, this includes the following steps: S3-1, Target Entity Set Construction: Based on the risk factor range identified in S2-2, this step further constructs a structured target clinical entity set. The system, according to the selected scoring system and the "factor-source-tool" mapping results completed in S2-2, standardizes and organizes the clinical entity types and triggering conditions corresponding to each risk factor. Based on this, the system generates a unified target entity set that can be used for entity extraction, clarifying the semantic definition, value rules, and boundary conditions of each type of entity, providing a consistent standardized target set for subsequent automatic extraction, rule matching, and feature generation. S3-2, Clinical Entity Extraction: In a unified data context, the system extracts structured data. Complete the extraction of entities that can be directly mapped by the rules, and at the same time utilize the rule templates and the semantic capabilities of the large model. Perform semantic recognition to extract risk factors that are difficult to obtain directly from structured data, thereby forming a basic entity set corresponding to the scoring system; (7) In formula (7), E represents the set of clinical entities extracted from the patient's medical records; each entity 𝑒 𝑖 It is defined as a complete semantic unit, where type i Indicates entity type, used to distinguish different clinical concepts, value i Indicates the specific value or content of an entity, time i Indicates the occurrence or recording time corresponding to this entity, source i A data system or document representing the source of an entity, confidence i The confidence score or credibility score of the entity extraction result is used to reflect the reliability of rule matching or semantic recognition. Through the above definition, the system expands the original extraction result from simple field values ​​into structured semantic entities with type, time, source and confidence information, providing a unified and complete basic semantic unit for subsequent risk factor mapping, multi-dimensional verification and traceability chain construction. S3-3, Entity and scoring factor mapping: In the structured modeling process, the system aligns the extracted entities with the specific scoring items of each scoring system; the mapping process converts the extracted and verified clinical information into a structured expression, including: factor category, triggering conditions, corresponding scoring system and score, and the data evidence from which it originates, forming a structured preliminary risk factor list; (8) In formula (8), This represents a structured set of risk factor mappings; each risk factor It is defined as a complete rating factor instance, where This refers to the clinical entities extracted and verified through the aforementioned steps. This indicates the triggering conditions or judgment rules for this factor in the corresponding scoring system. This represents the set of data evidence supporting the validity of the factor, including original fields, text fragments, and time records. This indicates the specific scoring system to which the factor belongs. This indicates the score or weight of the factor in the corresponding scoring system. Through the above definition, the system will uniformly model the extracted clinical entities, scoring items, triggering conditions, and sources of evidence to form a structured list of risk factors that is calculable, interpretable, and traceable, providing standardized input for subsequent multidimensional verification, score calculation, and result interpretation.

7. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 1, characterized in that: S4 describes performing multi-dimensional verification on the extracted risk factors, including semantic consistency verification, temporal logic verification, and cross-data source consistency comparison. It automatically corrects and marks inconsistencies, and outputs the verified risk factor set and verification result record. Specifically, this includes the following steps: S4-1, Rule Verification and Semantic Consistency Verification: The system performs semantic consistency verification on the risk factors output by S3, checking whether the factors meet the triggering conditions, meaning definition, value range and applicable scenarios of the scoring items; It serves as the primary basis for participating in time-series verification and field consistency validation. This is used to supplement missing structured data, resolve inconsistencies in cross-source records, and enhance semantic consistency. The verification process compares the risk factors with the guideline definitions and original clinical records to verify whether the risk factors truly meet the requirements of the item. This includes, but is not limited to, confirming whether "prolonged bed rest" meets the required number of days, whether "postoperative events" occurred within the effective timeframe after surgery, whether "malignant tumors" are in an active phase or still affect the current risk, and whether "age grouping" falls within the correct score range. For factors with ambiguous text descriptions, insufficient conditions, inconsistent timelines, or mismatches with the item definitions, the system will automatically identify and remove them or mark them as items to be reviewed, thereby ensuring that the risk factors are consistent with the scoring system rules and providing reliable input for subsequent scoring. S4-2, Time Sequence Rationality and Cross-Source Data Consistency Verification: Based on a unified data context, perform time sequence verification on the timestamps, event chains, and triggering conditions of risk factors to verify the consistency between the order of events, the effective time window, and clinical logic; at the same time, use multi-source data for cross-comparison to identify time conflicts, missing content, or inconsistencies in cross-system records, and automatically correct or mark obviously contradictory factors as requiring manual review. (9) In formula (9), This indicates a check of the semantic rationality of risk factors; Indicates time sequence verification; This indicates cross-source consistency verification; valid(fi) = true when all three conditions are true, and the factor is valid; otherwise, the factor is marked as abnormal or requires further verification. S4-3, Multidimensional Integration and Validation of Risk Factors: The system integrates the results of semantic verification, temporal verification, and cross-source comparison to screen all risk factors for validity and generate a final verified risk factor set; automatic correction is performed on factors that have conflicts but can be resolved through rules; explicit annotation is performed on factors that still have uncertainties and verification result records are output. (10) In formula (10), This represents the final set of valid risk factors retained after multi-dimensional verification; valid(fi) is the comprehensive verification function for semantic, temporal, and cross-source consistency mentioned above. When the verification result is true, it means that the factor meets the requirements in terms of semantic definition, temporal logic, and data source. For risk factors that still have uncertainty after verification, the system will not discard them directly, but will mark them as "pending verification" and include them in a separate set of questionable factors. This set does not participate in automatic scoring, but will be saved together with its source data, parsing path, and verification results for manual review and re-inclusion or exclusion as needed, thereby ensuring the traceability and reusability of the assessment process.

8. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 1, characterized in that: S5 describes the construction of a VTE assessment agent, which automatically generates and fills out a VTE risk assessment form based on the extracted risk factors, calculates the total score and risk level according to the scoring rules, and simultaneously generates corresponding anticoagulation protocols and follow-up recommendations. Specifically, this includes the following steps: S5-1, Construction of VTE Assessment Agent: Constructing a VTE assessment agent for automatic scoring and clinical suggestion generation; the agent consists of an assessment planning unit, a scale processing toolkit, an assessment memory unit, and a clinical action unit, which work together in a unified context to complete the entire process from risk factor input to risk stratification and suggestion output. (1) Assessment planning module: responsible for developing the assessment process after loading and verifying the risk factors, including planning the form generation steps according to the corresponding scoring system, determining the required scoring items, scoring calculation path and subsequent decision triggering conditions, and providing clear execution instructions for subsequent tool modules; (2) Scale and scoring processing module: including form generation tool and scoring calculation tool, used to complete the construction of calculable forms and score execution; the form generation tool automatically maps risk factors to scale items according to the scoring system template and fills in missing fields according to the rules; the scoring calculation tool completes the score calculation and risk level determination according to the scale rules and outputs structured scoring results; (3) Clinical decision-making module: Based on the scoring results and risk level, the module automatically generates clinical recommendations; the module calls the prevention and treatment rules in the knowledge base to derive individualized anticoagulation regimens, physical prevention measures, monitoring requirements and re-evaluation time; the output results are presented in a structured manner and can be directly used for clinical decision-making or patient management, realizing the closed loop from assessment to action; (4) Assessment memory module: It is responsible for storing short-term and long-term information during the assessment process; short-term memory stores risk factors, field mapping results, scoring forms and intermediate calculation states; long-term memory stores guidelines, scoring system knowledge and clinical pathway content; the module ensures data continuity and traceability of execution results during the assessment process, providing a stable basis for subsequent decision-making and traceability; S5-2, Assessment Form Generation: After receiving the risk factors output by S4, the system calls the form generation tool, loads the corresponding scale template according to the selected scoring system, and matches the risk factors with the standard items one by one; it automatically fills in the factors that can be clearly identified, and completes the missing items according to the preset rules to form a structured VTE assessment form that can be directly used for calculation. S5-3. Scoring and Risk Level Calculation: Based on the generated structured form, the system calls the scoring calculation tool and performs score calculation according to the rules of each scoring system, including item scores, condition combinations and cumulative summation. It also generates risk levels based on preset thresholds. The scoring process is executed using a unified logical framework to ensure that the calculation paths between different scales are clear, the rules are consistent, and the results are traceable. The final output includes item scores, total scores, risk levels and corresponding judgment criteria, and is synchronously written into the assessment context. (11) In formula (11), This represents the patient's overall VTE risk score; This represents the set of risk factors under the d-th rating dimension, where different rating dimensions are independent of each other; This indicates the score or weight of the risk factor in the corresponding scoring system; This is a trigger indicator variable for risk factors, taking a value of 1 if and only if its trigger condition is met, otherwise taking a value of 0; constraint conditions. This is used to ensure that the triggering conditions of each risk factor within the same scoring dimension are mutually exclusive, and to avoid the repeated accumulation of scores due to the simultaneous fulfillment of multiple conditions within a single dimension, thereby ensuring that the scoring calculation logic is consistent and the results are interpretable. (12) In formula (12), , The preset risk threshold is denoted as ; Score is the total score calculated by formula (7); Risk is the final risk level determined by the formula, which realizes the automated stratification and classification of VTE risk in patients. S5-4. Clinical Recommendation Generation: After completing the assessment form and risk level determination, the system automatically generates individualized prevention and treatment recommendations based on the combination of scoring results and key risk factors. The recommendations cover drug or non-drug preventive measures, anticoagulant medication selection and monitoring requirements, physical prevention methods, and recommended reassessment times. The derivation of the recommendations is based on long-term memory knowledge to complete the judgment. (13) In formula (13), Plan represents the set of individualized clinical recommendations generated by the system, including anticoagulation regimen, physical precautions, monitoring requirements, and reassessment time; Score is the aforementioned total VTE score; This is a subset of key factors selected from all valid risk factors; the function g(·) represents the decision mapping process based on guideline rules and knowledge base, which maps "score level + key factor pattern" into specific clinical action plans, thereby achieving an automatic closed loop from quantitative assessment to actionable recommendations.

9. The rapid risk assessment method for venous thromboembolism based on intelligent agent and multi-source clinical data fusion according to claim 1, characterized in that: S6 describes constructing a traceability chain for risk factors, their data sources, analysis paths, and scoring criteria to achieve interpretable, traceable, and verifiable presentation of assessment results. This specifically includes the following steps: S6-1, Source Tracing Element Collection: Automatically collect key elements related to risk factors from the entire assessment process. The system records various evidence information in a unified context to form a basic evidence set that can be used for subsequent source tracing. S6-2, Source tracing chain construction: Based on the content of the evidence set, risk factors, data sources, analysis steps, reasoning logic and scoring criteria are organized into a chain structure according to time order and semantic dependencies. The system automatically generates a full path mapping of "factor - data - analysis - rule" to realize a continuous and structured source tracing chain from the original record to the final scoring result. (14) In formula (14), Trace represents the traceability chain constructed for a single VTE risk assessment; each As a node in the traceability chain, it contains key information such as risk factors, original data sources, analysis steps, scoring items used and corresponding scores; by organizing nodes into a sequence structure according to execution order and dependencies, the system can completely record the entire path "from original data to final score and recommendation", providing a visual and verifiable chain of evidence for post-audit, result interpretation and manual review; S6-3, Interpretable Output and Verification Display, transforms the completed traceability chain into auditable and interpretable output. The output results support clinical staff review, quality control checks, and manual auditing, ensuring that the evaluation results are interpretable, traceable, and verifiable.

Images