Multi-source heterogeneous clinical data virtual integration query method and system

Abstract

The application discloses a multi-source heterogeneous clinical data virtual integration query method and system, which performs diagnosis and treatment semantic coding and value range normalization processing on original clinical data to generate standardized clinical data, constructs a virtual patient journey data chain, establishes a clinical natural language understanding model to analyze a user natural language query request to generate a structured query instruction, positions a candidate data node in the virtual patient journey data chain and performs time sequence clipping, determines a cross-source data recall strategy according to an intention category, issues a localized query statement to a data source to obtain an original sub-result set, performs credibility weighting fusion on synonymous data and adds a grade identifier, assembles query response content for clinical decision-making, and iteratively optimizes the model and parameters according to user feedback behavior. The application realizes natural language clinical query and cross-source time sequence association tracing, and meets the demand of clinical decision support.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method and system for virtual integration and querying of multi-source heterogeneous clinical data. Background Technology

[0002] With the deepening of hospital informatization, clinical data is scattered across multiple heterogeneous business systems, including hospital information systems, laboratory information systems, image archiving and communication systems, electronic medical record systems, and off-site health monitoring equipment. These systems use different data formats, encoding standards, and communication protocols, resulting in a high degree of heterogeneity in clinical data and making it difficult to achieve unified querying and correlation analysis.

[0003] In existing technologies, multi-source heterogeneous data queries mainly employ two methods: centralized storage in data warehouses or association using intermediate indexes. These methods suffer from poor data timeliness, complex index maintenance, and difficulty in supporting natural language queries and clinical semantic understanding.

[0004] Therefore, there is an urgent need for a virtual integration query method and system for multi-source heterogeneous clinical data that can integrate multi-source heterogeneous clinical data and support natural language queries. Summary of the Invention

[0005] The main objective of this invention is to provide a virtual integrated query method and system for multi-source heterogeneous clinical data. By constructing a virtual patient journey data chain, it achieves cross-source data temporal correlation, uses a clinical natural language understanding model to analyze query intent, drives differentiated data recall strategies based on intent categories, and introduces a credibility-weighted fusion mechanism to ensure data quality. Ultimately, it forms query response content oriented towards clinical decision-making, and achieves system self-evolution based on user feedback behavior.

[0006] The virtual integration and query method and system for multi-source heterogeneous clinical data proposed in this invention employs the following technologies: A method for virtual integration and querying of multi-source heterogeneous clinical data, the method comprising: S1. Collect raw clinical data, perform diagnostic semantic encoding and value range normalization on the raw clinical data, and generate standardized clinical data. S2. Arrange and reorganize the standardized clinical data to construct a virtual patient journey data chain covering the entire life cycle of the patient; S3. Establish a clinical natural language understanding model, and based on the clinical natural language understanding model, extract diagnostic entities, determine query intent, and identify time window limits for clinical query requests submitted by users in natural language form, and generate structured query instructions including target diagnostic entities, intent categories, and time constraints. S4. Based on the target diagnosis and treatment entity, locate the matching set of candidate data nodes in the virtual patient journey data chain, and perform time window pruning on the set of candidate data nodes according to the time constraint to obtain a time-aligned target node sequence. S5. Determine a cross-source data retrieval strategy based on the intent category and the target node sequence; S6. According to the cross-source data retrieval strategy, send a localized query statement after diagnosis and treatment semantic conversion to the corresponding data source to obtain the original sub-result set returned by the data source; S7. Perform confidence-weighted fusion of synonym clinical data in the original sub-result set to generate fused clinical data items, and attach a confidence level identifier to the fused clinical data items; S8. Assemble the query response content based on the fused clinical data items and the intent category; S9. Return the query response to the user and iteratively optimize the clinical natural language understanding model based on user feedback.

[0007] Optionally, the diagnostic semantic encoding and range normalization processing includes: Extract business semantic information and data format features from the original clinical data, and identify the clinical diagnosis and treatment business scenario corresponding to the original clinical data; The data items of the original clinical data are matched with the standard diagnosis and treatment codes in the standard clinical terminology set to establish a correspondence database between the original data fields and the standard diagnosis and treatment codes; Based on the corresponding relational database, the original data value range is converted in units, the encoding is replaced, and the format is unified to generate standardized clinical data.

[0008] Optionally, the construction of a virtual patient journey data chain covering the entire patient lifecycle includes: Establish a globally unique patient identity index for the same patient distributed across different clinical business systems; Using the patient identity index as the root node, the standardized clinical data are sorted in ascending order according to the clinical event timestamps to form a panoramic time-series data stream of patients; Within each clinical event, establish clinical diagnosis and treatment association edges between data nodes, and record the association type, association strength, and bidirectional traceability path.

[0009] Optionally, the establishment of the clinical natural language understanding model includes: Collect clinical diagnosis and treatment corpus; The clinical diagnosis and treatment corpus is labeled with diagnosis and treatment entities, query intent categories, and time descriptive words to form a training sample set with labeled information. The pre-trained language model is fine-tuned based on the training sample set to construct a clinical natural language understanding model.

[0010] Optionally, the generation of structured query instructions, including the target diagnostic entity, intent category, and time constraints, includes: The clinical natural language understanding model is used to extract diagnostic entities and their entity types from the clinical query request; The intent category of the clinical query request is identified based on the clinical natural language understanding model. Based on the diagnosis entity and the intent category, the time window slots in the pre-query instruction template are filled to generate a structured query instruction containing the target diagnosis entity, intent category, and time constraints.

[0011] Optionally, the set of candidate data nodes for location matching includes: The target diagnostic entity is matched with the standard diagnostic codes of each data node in the virtual patient journey data chain, and data nodes with consistent codes or similar semantics are selected to form a preliminary set of nodes. The initial screening node set is prioritized according to the intent category; The sorted set of initially screened nodes is output as the set of candidate data nodes.

[0012] Optionally, determining the cross-source data retrieval strategy includes: When the intent category is medical history tracing, starting from the earliest data node in the target node sequence, the virtual patient journey data chain is traversed in a forward direction to recall all historical data nodes related to the target diagnosis and treatment entity, up to the current time node. When the intent category is efficacy assessment, the latest data node in the target node sequence is taken as the endpoint, and the data chain of the virtual patient journey is traced backward to recall the baseline data node before the intervention and the follow-up data node after the intervention. When the intent category is association analysis, starting from each data node in the target node sequence, multi-hop expansion is performed along the clinical diagnosis and treatment association edge to recall cross-modal data nodes that are directly or indirectly associated with the target diagnosis and treatment entity. The cross-modal data nodes include structured test data, semi-structured medical record text, and unstructured medical image data.

[0013] Optionally, the issuance of localized query statements after diagnostic semantic conversion includes: The traversal direction identifier and associated hop count parameter in the cross-source data retrieval strategy are parsed, and a query sequence is generated based on the traversal direction identifier and associated hop count parameter. Based on the real-time access handles reserved by each data node in the target node sequence, determine the set of target data source addresses to be queried; The standard diagnosis and treatment codes in the query sequence are converted into the local encoding format corresponding to each target data source, and the standard time expression is converted into the time query syntax corresponding to each target data source, and then assembled into a local query statement. The localized query statements are sent in parallel to the corresponding target data sources.

[0014] Optionally, the generation of fused clinical data items includes: Identify data items in the original sub-results set that represent the same clinical diagnosis and treatment concept, and form synonym data groups; Assign basic weights to the data items according to the data source authority level table; Calculate the time decay coefficient and consistency coefficient based on the data items; The base weight, the time decay coefficient, and the consistency coefficient are calculated to obtain the fusion weight of each data item. The values ​​of each data item are then weighted and averaged according to the fusion weight to generate fused clinical data items.

[0015] Optionally, a virtual integration and query system for multi-source heterogeneous clinical data, the system comprising: The standardization processing module is used to collect raw clinical data and perform diagnostic semantic encoding and value range normalization on the raw clinical data to generate standardized clinical data. The journey building module is used to arrange and reorganize the standardized clinical data and build a virtual patient journey data chain covering the entire life cycle of the patient; The instruction generation module includes a model building unit, a request parsing unit, an intent discrimination unit, a time recognition unit, and an instruction generation unit. It is used to establish a clinical natural language understanding model and, based on the clinical natural language understanding model, to discriminate the query intent and time window limitation of the clinical query request submitted by the user in natural language form, so as to generate a structured query instruction. The node localization module is used to locate a matching set of candidate data nodes in the virtual patient journey data chain according to the structured query instruction, and to perform time window pruning on the set of candidate data nodes to obtain a time-aligned target node sequence. The strategy determination module is used to determine a cross-source data retrieval strategy based on the structured query instruction and starting from the target node sequence. The query module is used to send localized query statements that have undergone diagnostic semantic transformation to the corresponding data source according to the cross-source data retrieval strategy, and obtain the original sub-result set returned by the data source; The credibility fusion module is used to perform credibility-weighted fusion of synonym clinical data in the original sub-result set, generate fused clinical data items, and attach credibility level identifiers to the fused clinical data items. An assembly module is used to assemble query response content based on the fused clinical data items and the intent category; The feedback module is used to return the query response content to the user and iteratively optimize the clinical natural language understanding model based on the user's feedback behavior.

[0016] The beneficial effects of this invention are as follows: This invention provides a virtual integration and query method and system for multi-source heterogeneous clinical data. It weaves heterogeneous data scattered in hospital information systems, laboratory information systems, image archiving and communication systems, electronic medical record systems, and outpatient health monitoring equipment into unified data covering the entire patient life cycle, constructs a virtual patient journey data chain, and realizes the logical integration of cross-source heterogeneous clinical data. Based on a multi-source clinical diagnosis and treatment corpus constructed from electronic medical record texts, clinical guidelines, medical textbook chapters, and real clinical query logs, a clinical natural language understanding model with the ability to recognize diagnostic and treatment entities, judge query intent, and extract time windows is formed through multi-task fine-tuning training. This model supports medical staff to directly initiate queries in natural language, realizing accurate parsing and structured conversion of clinical query intent. Through a differentiated recall strategy driven by intent categories, precise adaptation of query scenarios and data traversal paths is achieved. Through an iterative optimization closed loop driven by user feedback behavior, continuous evolution of system capabilities and scenario adaptation are realized. Attached Figure Description

[0017] Figure 1 A flowchart of a virtual integration and query method for multi-source heterogeneous clinical data provided in an embodiment of the present invention; Figure 2 A flowchart of a clinical natural language understanding model processing method provided in an embodiment of the present invention; Figure 3 A flowchart of the decision-making process for locating and pruning a set of candidate data nodes provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the localized query statement delivery process according to an embodiment of the present invention. Figure 5 This is a structural diagram of a virtual integrated query system for multi-source heterogeneous clinical data provided in an embodiment of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The main solution of this invention is: Virtual integration and query methods for multi-source heterogeneous clinical data include: S1. Collect raw clinical data, perform diagnostic semantic coding and value range normalization on the raw clinical data, and generate standardized clinical data. S2. Arrange and reorganize standardized clinical data to construct a virtual patient journey data chain covering the entire patient life cycle; S3. Establish a clinical natural language understanding model, and based on the clinical natural language understanding model, extract diagnostic entities, determine query intent, and identify time window limits for clinical query requests submitted by users in natural language form, and generate structured query instructions including target diagnostic entities, intent categories, and time constraints. S4. Based on the target diagnosis and treatment entity, locate the matching candidate data node set in the virtual patient journey data chain, and perform time window pruning on the candidate data node set according to time constraints to obtain the time-aligned target node sequence. S5. Determine a cross-source data retrieval strategy based on the intent category and the target node sequence; S6. Based on the cross-source data retrieval strategy, send localized query statements that have undergone diagnostic semantic transformation to the corresponding data source to obtain the original sub-result set returned by the data source; S7. Perform confidence-weighted fusion of synonymous clinical data in the original sub-result set to generate fused clinical data items, and attach a confidence level label to the fused clinical data items; S8. Assemble the query response content based on the integrated clinical data items and intent categories; S9. Return the query response to the user and iteratively optimize the clinical natural language understanding model based on user feedback.

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can understand it.

[0021] refer to Figure 1 An embodiment of the present invention provides a virtual integration and query method and system for multi-source heterogeneous clinical data, the implementation steps of which are as follows: S1. Collect raw clinical data, perform diagnostic semantic coding and value range normalization on the raw clinical data, and generate standardized clinical data.

[0022] S1.1 Extract business semantic information and data format features from the original clinical data, and identify the clinical diagnosis and treatment business scenarios corresponding to the original clinical data.

[0023] In the specific implementation process, a method combining metadata parsing and regular expression pattern matching is adopted to extract business semantic information and data format characteristics of raw clinical data from various heterogeneous clinical business systems. Metadata parsing is used to obtain database table structure definitions, field type definitions, and field annotation information; regular expression pattern matching is used to identify keywords of clinical diagnosis and treatment business scenarios from field annotations, table names, and data dictionaries.

[0024] Let the set of original clinical data sources be... ,in For the first Each clinical business system data source. For each data source... Perform metadata retrieval operation: Extract database connection metadata: Extracting metadata from the data table structure: Extract field definition metadata: Based on the metadata defined by the fields The field annotation information is identified using regular expression matching based on a pre-defined keyword library for clinical diagnosis and treatment scenarios. The matching score between the field annotation and the keywords of the clinical diagnosis and treatment scenarios is calculated. : in, A set of keywords for the corresponding scenario; This is a function for calculating the inverse document frequency weight of term frequencies.

[0025] when Exceeding the preset threshold When determining the clinical diagnosis and treatment business scenario to which this field belongs. and mark the data table. Business scenario attributes are Record the recognition results: in, .

[0026] S1.2 Match the data items of the original clinical data with the standard diagnosis and treatment codes in the standard clinical terminology set to establish a correspondence database between the original data fields and the standard diagnosis and treatment codes.

[0027] In the specific implementation process, a hierarchical semantic matching algorithm is used to match the data items of the original clinical data with the standard clinical terminology set. Let the original data field set be... for: in, For the first One original data field, including field name Field comments and sample value set .

[0028] The standard clinical terminology set is as follows: in, For the first A standard diagnostic and treatment coding item, including the code ,name Synonyms With the category .

[0029] Perform hierarchical semantic matching, with the first layer being exact matching: for each original data field Determine whether a standard medical code exists. satisfy: If it exists, then establish an exact correspondence. The matching confidence level is 1.0.

[0030] The second layer is fuzzy matching: for raw data fields that do not pass an exact match. Calculation and standard medical codes Semantic similarity: in, This is a set of word segmentation results; The Jaccard similarity coefficient; For pre-trained medical word vectors; for; This is a category matching function. It takes a value of 1 when the business scenario to which the field belongs matches the coding category, and 0 otherwise. , , Let be the weighting coefficient, satisfying .

[0031] when Exceeding the preset threshold Establish fuzzy correspondences: Match confidence level is .

[0032] The third layer is manual verification and matching: for raw data fields that fail the fuzzy matching. Mark the fields as awaiting manual verification and output a list of candidate matches. The corresponding relationships are established after manual confirmation by the clinical data administrator.

[0033] By summarizing the three-level matching results, a mapping database between the original data fields and standard medical codes is constructed: in, ; To match confidence levels.

[0034] S1.3 Based on the corresponding relational database, the original data value range is converted in units, the encoding is replaced, and the format is unified to generate standardized clinical data.

[0035] In the specific implementation process, for numerical raw data fields Its corresponding standard diagnosis and treatment code Has standard units of measurement Determine the current unit of measurement for the original data field. Whether it is consistent with the standard unit of measurement: If Unit conversion: in, This is a pre-set table of unit conversion factors.

[0036] For encoded raw data fields According to the standard diagnosis and treatment codes in the corresponding relational database Perform encoding replacement: in, The original encoded value; This is the standard encoding for retrieving matches from the corresponding relation database.

[0037] For date and time type raw data fields Convert them to the ISO 8601 standard format: Generate standardized clinical data: in, For patient identification; Standard medical treatment code; The value is the normalized value; Standard units of measurement; A standard format timestamp; Identify the data source; Used as a data table identifier; Used as a field identifier; The original value; Original unit of measurement; This is the original timestamp; To match confidence levels; This is the matching type.

[0038] The generated standardized clinical data units are written into a standardized clinical database, and a composite index structure is established with patient_id as the first index, std_code as the second index, and normalized_time as the third index to support subsequent fast time-series retrieval.

[0039] S2. Arrange and reorganize standardized clinical data to construct a virtual patient journey data chain covering the entire patient life cycle.

[0040] S2.1 Establish a globally unique patient identity index for the same patient distributed across different clinical business systems.

[0041] In the specific implementation process, the original patient identity identifier set in each clinical business system is assumed to be: in, This refers to the outpatient or inpatient number in the hospital information system. The patient identifier associated with the test invention number in the laboratory information system; Patient identifiers associated with image examination numbers in image archiving and communication systems; For example, the medical record number in the electronic medical record system; and the device user identifier uploaded by the off-site health monitoring device.

[0042] Extract the patient identity attribute set: Two patient identifiers from different data sources and Calculate the identity matching score: in, For exact matching, the value is 1 when the match is found and 0 when the match is not found. This is a fuzzy matching function, calculated using edit distance or Jaccard similarity, with a value range of... ; This is a date difference function; it takes the value 1 when the date difference is 0, and the larger the difference, the closer the value is to 0. , , , and All are weighting coefficients, satisfying ,and This reflects the highest matching priority of the ID number.

[0043] when Exceeding the preset threshold At that time, determine the identity identifier and To establish an identity association by referring to the same patient. .

[0044] Construct a patient identity connectivity graph based on identity associations. , where the vertex set Identify all patients and aggregate them. This establishes identity associations. A connected component algorithm is used to identify the complete set of identity identifiers for the same patient. Assign a globally unique patient identity index to each connected component: Establish a patient identity mapping table: in, The confidence score for identity association.

[0045] S2.2 Using the patient identity index as the root node, the standardized clinical data are sorted in ascending order according to the timestamp of the clinical event to form a panoramic time-series data of the patient.

[0046] In the specific implementation process, the patient identity index is used as the root node to retrieve the complete set of standardized clinical data corresponding to the patient: right Extract the timestamp field of clinical events from each standardized clinical data unit. Sort in ascending order: The sorted data is divided into event segments according to clinical diagnosis and treatment business scenarios. The scenario segmentation rules are as follows: If the time difference between adjacent data units and and If they belong to the same diagnostic event coding group, they are grouped into the same event segment. Among them, Maximum interval threshold for the scenario: Medical visit scenario Hourly inspection and testing scenarios Hours, medication scenarios Heavens, surgical scene sky.

[0047] Forming a panoramic temporal data stream of the patient: in, This serves as a unique identifier for the event segment; ; This is the start time of the event segment; This is the deadline for the event segment; This is the set of data nodes contained in this event segment; This refers to scene context information.

[0048] S2.3 Within each clinical event, establish clinical diagnosis and treatment association edges between data nodes, and record the association type, association strength, and bidirectional traceability path.

[0049] In the specific implementation process, in each event segment Internally, for the set of data nodes Establish clinical diagnosis and treatment association edges for each node in the data. Let the data nodes be... Three types of clinical diagnosis and treatment association edges are defined.

[0050] The first type is direct diagnosis-related edges, used to connect data nodes within the same event segment that have a direct causal relationship. Establishment rules: like and ,and exist After the time of diagnosis Within the window, a diagnostic test association edge is established: in, The effect of temporal proximity on association strength. The calculation formula is: like and ,and exist After the time of diagnosis Within the window, a diagnosis-medication association edge is created: The second type is time-series diagnostic association edges, used to connect consecutive data nodes within different event segments that target the same diagnostic goal. Establishment rules: like , and , and , And the inspection items If the project falls under the routine postoperative monitoring group, then a postoperative follow-up association edge should be established: in, This is a key node in the surgical event segment.

[0051] The third type is cross-source diagnosis and treatment association edges, used to connect data nodes distributed across different clinical business systems but belonging to the same clinical event. Establishment rules: like , and , and and If they belong to the same single-code group of the invention being inspected, then a cross-source association edge is established: in, This reflects the impact of data source synchronization reliability on association strength. The calculation formula is: Two-way traceability path record: By aggregating all data nodes and clinical diagnosis and treatment related edges, a virtual patient journey data chain is constructed: in, A real-time access handle is reserved for each data node, pointing to the physical storage location of the original clinical data.

[0052] S3. Establish a clinical natural language understanding model, and based on the clinical natural language understanding model, extract diagnostic entities, determine query intent, and identify time window limits for clinical query requests submitted by users in natural language form, and generate structured query instructions including target diagnostic entities, intent categories, and time constraints.

[0053] S3.1 Establish a clinical natural language understanding model.

[0054] In the specific implementation process, refer to Figure 2 A clinical diagnosis and treatment corpus was constructed by collecting data from multiple clinical information systems. This corpus includes electronic medical record corpora, clinical guideline literature corpora, textbook corpora, and query log corpora. Free text content was extracted from outpatient medical records, inpatient medical records, progress notes, and discharge summaries in the electronic medical record system to establish the electronic medical record corpus. Standardized texts from clinical practice guidelines, expert consensus, and drug instructions are collected from medical knowledge bases to establish a clinical guideline literature corpus. Extracting chapters from textbooks on internal medicine, surgery, and diagnostics from medical education databases to establish a textbook corpus. Extract clinical data retrieval requests actually initiated by medical staff from the hospital information system query logs, and establish a query log corpus. A corpus of clinical diagnosis and treatment statements was compiled and constructed. : The clinical diagnosis and treatment corpus is annotated with diagnosis and treatment entities, query intent categories, and time descriptive terms to form a training sample set with annotation information. The set of diagnosis and treatment entity types is defined as follows: Sequence labeling is performed on the corpus. The query intent category labeling process is as follows: Define the query intent category set as: right Each query log in Intent category is determined based on user click behavior and query result usage: If If the keywords include "effect," "change," or "comparison," and the result includes both before-and-after test results, it is labeled as a therapeutic efficacy assessment; if If the results contain keywords such as "relationship," "related," or "cause," and include data from multiple categories, then it is labeled as an association analysis; if... If the results contain keywords such as "contraindications," "interactions," and "dosage," and include data from multiple categories, then it is labeled as a drug use review; if... If the results contain keywords such as "prognosis," "survival," and "relapse" and include data from multiple categories, they are marked as medication review.

[0055] The process of annotating time descriptors is as follows: Define a set of time descriptor types: Fine-grained annotation is performed on time descriptors in the corpus, including absolute time (e.g., January 15, 2024), relative time (e.g., within the last three months), and treatment stage (e.g., upon admission). After annotation, a training sample set with annotation information is formed. The pre-trained language model is fine-tuned based on the training sample set to construct a clinical natural language understanding model.

[0056] We selected the pre-trained language model BaseModel to construct a multi-task training framework, which includes three parallel prediction heads: a diagnostic entity recognition network, an intent classification network, and a temporal extraction network. The diagnostic entity recognition network uses a Conditional Random Field (CRF) as its decoding layer, and its input is the sequence of context representation vectors output by BaseModel. The output is a sequence of entity labels. Define the CRF transition matrix A, where Let be the score for transitioning from label i to label j. The entity recognition loss function is: in, To score the launch; sum the results by iterating through all possible tag sequences. .

[0057] The intent classification network uses a pooling layer plus a fully connected layer structure, with the input being the [CLS] position vector output by BaseModel. The output is the probability distribution of intent categories. The intent classification loss function is: The temporal extraction network employs a pointer network structure to predict the start and end positions of temporal descriptors. (Start position probability) for: The probability of the termination position is: The time decimation loss function is: The total loss function is a weighted sum of the three losses: in, , and All of these are task weight coefficients.

[0058] set up , , With a learning rate of 2e-5, 32 training epochs, and 32 training rounds, a clinical natural language understanding model was obtained through joint training. : S3.2 Generate a structured query instruction that includes the target diagnosis entity, intent category, and time constraints.

[0059] In practice, it receives clinical query requests submitted by users in natural language. ,Will enter The diagnostic entity recognition network outputs entity annotation sequences; enter An intent classification network outputs an intent category probability distribution, and takes the maximum probability as the intent category; enter The time period is extracted, identifying time descriptors for the past six months and converting them into standardized time expressions. Based on the extraction results, a pre-query instruction template is populated to generate a structured query instruction.

[0060] S4. Based on the target diagnosis and treatment entity, locate the matching set of candidate data nodes in the virtual patient journey data chain, and perform time window pruning on the set of candidate data nodes according to time constraints to obtain the time-aligned target node sequence.

[0061] S4.1 Match the target diagnostic entity with the standard diagnostic codes of each data node in the virtual patient journey data chain, and filter data nodes with consistent codes or similar semantics to form a preliminary set of nodes. In the specific implementation process, refer to Figure 3 Extract the target set of diagnostic entities from the structured query command: Each target diagnostic and treatment entity include: Index by patient identity Using this as the entry point, the virtual patient journey data chain is traversed. All data nodes in: Each data node include: The coding matching is performed in three layers. The first layer is exact coding matching: for each target diagnostic entity... Retrieve data nodes with completely identical standard medical codes: The second layer is semantic layer extended matching. When the exact match results are empty or insufficient, semantic layer extension is initiated. Based on the hierarchical structure of the standard clinical terminology set, it retrieves matches... Encoding that involves hierarchical or synonym relationships: like Then, based on the ICD-10 coding hierarchy, retrieve relevant codes for the same chapter, category, or subcategory: in, Return and The set of codes whose distance in the hierarchical structure does not exceed level.

[0062] like Then, based on the six-dimensional structure of LOINC encoding—component, attribute, time, system, scale, and method—retrieve relevant codes for the same component or attribute: Expand the set of matching nodes: The third layer is text similarity compensation matching. When semantic level expansion still cannot meet the minimum number of results requirement, the similarity between the data node scene context and the target diagnosis entity text is calculated. in, This is a pre-trained medical word vector encoding function.

[0063] Set similarity threshold ,filter: The results of the three-level matching are summarized to form a preliminary set of nodes: Attach a matching source tag to each initial screening node: S4.2 Sort the initial screening node set according to the intent category, and output the sorted initial screening node set as the candidate data node set.

[0064] In the specific implementation process, the intent category is extracted from the structured query command: Determine the sorting strategy based on intent category: when In this case, data nodes with earlier timestamps are prioritized for retention to reflect the entire disease progression. Time priority score is calculated as follows: in, These are the maximum and minimum timestamps in the initial screening node set, respectively.

[0065] At the same time, data nodes from medical visit and diagnosis scenarios are prioritized for retention, while the weight of testing and examination nodes is reduced. : like , ; like , ; like , .

[0066] Overall Priority Score : according to Arrange the initial set of nodes in descending order.

[0067] when Time: Prioritize retaining data nodes with more recent timestamps to reflect the latest treatment status. Calculate time priority score. : At the same time, data nodes for inspection and testing scenarios should be retained with priority.

[0068] like and It is a numerical type. ; like and It is a non-numerical type. ; like , ; like , . Overall priority score: according to Arrange the initial set of nodes in descending order.

[0069] when hour: Data nodes with a higher number of associated edges are prioritized for retention to support multi-hop extended analysis. Calculate the association priority score: in, This represents the maximum node degree in the virtual patient journey data chain.

[0070] At the same time, cross-source data nodes are prioritized for retention to support multi-source heterogeneous associations: like Unlike most nodes in the initial screening node set, like Unlike some nodes in the initial screening node set, like Similar to most nodes in the initial screening node set, Overall priority score: according to Arrange the initial set of nodes in descending order.

[0071] After sorting, output the set of candidate data nodes. : Extracting time constraints from structured query commands: Parse time constraint types: when At that time, extract the start timestamp. and end timestamp : Pruning the candidate data node set: when At that time, based on the current query time Using this as a baseline, calculate the offset time range: Pruning the candidate data node set: when At that time, according to the preset treatment stage time definition table, the treatment stage identifier is converted into the corresponding time interval: Retrieve key time points in the virtual patient journey data chain to determine the current patient's stage time interval. .

[0072] Pruning the candidate data node set: Timing alignment processing: After cutting Sort in ascending order by the timestamp of the clinical event: Perform a time sequence continuity check to determine whether the time interval between adjacent nodes exceeds the maximum clinical event interval threshold: like and , ; like and , ; like , ; like , . For more than The interval, insert time breakpoint markers: Finally, the target node sequence with time alignment is obtained. : Each target node is appended with a timing alignment tag: =in The serial number in; =Identifier of the subsequence of the treatment stage.

[0073] S5. Based on the intent category, determine the cross-source data recall strategy starting from the target node sequence.

[0074] S5.1 When the intent category is medical history tracing, start from the earliest data node in the target node sequence, traverse forward along the virtual patient journey data chain, recall all historical data nodes related to the target diagnosis and treatment entity, up to the current time node.

[0075] In the specific implementation process, the earliest data node is extracted from the time-aligned target node sequence. Starting from the current query time, set the timestamp of the forward traversal to the end of the forward traversal. Along the virtual patient journey data chain Traverse the timeline in forward order and recall data nodes that meet the following conditions: in, A set of semantic extension codes for the target diagnostic entity.

[0076] Calculate the time-series coverage of the recall nodes : when At the same time, the recall scope is expanded to adjacent treatment stages to include related nodes that have clinical treatment-related edges with the target treatment entity: Output medical history tracing and recall strategy parameters . S5.1 When the intent category is efficacy assessment, the latest data node in the target node sequence is used as the endpoint, and the data is traced back along the virtual patient journey data chain to recall the baseline data node before the intervention and the follow-up data node after the intervention.

[0077] In the specific implementation process, the latest data node is extracted from the time-aligned target node sequence. Identify the time points for implementing intervention measures If the target node sequence contains a surgery or medication initiation node, then the timestamp of that node is taken as... Otherwise, take the timestamp of the median node in the target node sequence. Retrace back to retrieve the baseline data nodes. Positive follow-up recall of follow-up data nodes Calculate the baseline follow-up matching rate: when In such cases, the recall can be extended to adjacent time windows or related test groups: Output efficacy assessment recall strategy parameters Strategy_effect.

[0078] S5.3 When the intent category is association analysis, starting from each data node in the target node sequence, perform multi-hop expansion along the clinical diagnosis and treatment association edge to recall cross-modal data that are directly or indirectly related to the target diagnosis and treatment entity.

[0079] In the specific implementation process, all data nodes are extracted from the time-aligned target node sequence as the expansion starting point, and the maximum expansion hop count is set. The number of target diagnostic entities in the structured query command is dynamically adjusted: single entity Multi-entity .

[0080] Perform multi-hop extended recall and aggregate multi-hop recall nodes. Filter cross-modal data nodes, classify them according to data type, and calculate cross-modal coverage. : when At that time, from Secondary expansion is performed by adding nodes with high correlation strength: Output association analysis recall strategy parameters S .

[0081] S5.4 Summarize cross-source data retrieval strategies and generate a list of data source distributions.

[0082] In the specific implementation process, the set of nodes to be recalled is extracted according to the determined recall strategy. Statistics are grouped by data source, and a list of grouped data sources is generated for distribution. Output cross-source data retrieval strategy .

[0083] S6. Based on the cross-source data retrieval strategy, send localized query statements that have undergone diagnostic semantic transformation to the corresponding data source to obtain the original sub-result set returned by the data source.

[0084] S6.1. Parse the traversal direction identifier and associated hop count parameter in the cross-source data retrieval strategy, and generate a query sequence based on the traversal direction identifier and associated hop count parameter.

[0085] In the specific implementation process, refer to Figure 4 Extract traversal direction identifiers from cross-source data retrieval strategies With associated hop count parameter .

[0086] when At that time, from the target node sequence Extract all nodes and sort them in ascending order by the clinical event timestamp; when When, sorted in descending order by timestamp; when At that time, expand the node set Based on this, layer by number of associated hops: in For the first Jump to expand the set of nodes, and arrange the nodes at the same level in descending order of association strength.

[0087] for Execution sequence number assigned to each node: | S6.2 Determine the target data source address set to be queried based on the real-time access handles reserved by each data node in the target node sequence.

[0088] In the specific implementation process, from Each node extracts the real-time access handle Parse the network address, communication protocol, and connection status of each handle to filter out unreachable data sources: For unreachable nodes Triggering standby switchover: Summarize the target data source address set : S6.3. Convert the standard diagnosis and treatment codes in the query sequence into the local encoding format corresponding to each target data source, convert the standard time expression into the time query syntax corresponding to each target data source, and assemble them into a localized query statement.

[0089] In the specific implementation process, Each node in the process performs encoding conversion.

[0090] Standard medical code conversion: based on the corresponding relational database Reverse search: when At that time, the condition for generating the IN set is: Standard time expression conversion: Assume the time constraint is The target data source is hour: The target data source is hour: Assemble localized query statements This allows localized query statements to be sent in parallel to the corresponding target data source.

[0091] S7. Perform confidence-weighted fusion of synonymous clinical data in the original sub-result set to generate fused clinical data items, and attach a confidence level label to the fused clinical data items.

[0092] S7.1 Identify data items that represent the same clinical diagnosis and treatment concept in the original sub-results and form synonym data groups.

[0093] In the specific implementation process, from the original sub-result set Extract all data records: Grouped by patient identification and standard medical code; synonymous data groups are those with the same... The record set, for each synonym data group Further verify the consistency of numerical units: if If this is the case, a dimension conflict flag will be triggered, and the unit will be retained first. Records with high confidence levels. Among them, For the original sub-result set The records in the document.

[0094] S7.2 Assign basic weights to data items based on the data source authority level table, and calculate the time decay coefficient and consistency coefficient based on the data items.

[0095] In practice, the AuthorityTable for data sources is determined based on the generation stage and quality control level in the clinical diagnosis and treatment process: Laboratory information system data sources have an authority level of 1; image archiving systems have an authority level of 0.95; electronic medical record systems and hospital information systems have an authority level of 0.85; and external health monitoring equipment has an authority level of 0.6. Basic weight allocation: Let the current query time be... The data collection time is Time difference : The time decay coefficient adopts an exponential decay model: in The attenuation coefficient is set according to the data type: Test results / hour (diagnostic records) / hour, medication record / Hour.

[0096] For synonym data groups Calculate the within-group mean from the numerical data. with standard deviation Calculate the Z-score for each record. : Calculate the consistency coefficient : when At that time, all records have the same value. .

[0097] S7.3 Calculate the base weight, time decay coefficient and consistency coefficient to obtain the fusion weight of each data item, and calculate the weighted average of the values ​​of each data item according to the fusion weight to generate the fused clinical data item.

[0098] In the specific implementation process, the fusion weight is calculated. : ] Weighted average generates fusion clinical data items : S7.4 Add a confidence level label to the integrated clinical data items. In the specific implementation process, the effective fusion weight entropy H(g) of the synonym data group is calculated, and the credibility score is comprehensively evaluated: in, These are the weighting coefficients. .

[0099] The credibility levels are divided as follows: credibility score [0.90, 1.00], level A, high credibility; credibility score [0.75, 0.90), level B, relatively high credibility; credibility score [0.60, 0.75), level C, generally reliable; credibility score [0.40, 0.60), level D, relatively low credibility; credibility score [0.00, 0.40), level E, low credibility.

[0100] Additional credibility level label: Output a set of fused clinical data items , .

[0101] S8. Assemble the query response content based on the integrated clinical data items and intent categories.

[0102] In practice, when the intent category is medical history tracing, the data is derived from the set of integrated clinical data items. Extract all records Sort by clinical event occurrence time in ascending order: Based on the event segmentation rules in the virtual patient journey data chain, Patients are categorized into corresponding treatment stages, and timeline nodes are assembled hierarchically according to these stages to form a comprehensive patient timeline. .

[0103] When the intent category is efficacy assessment, from the fusion set of clinical data items Extract quantifiable indicator records related to the target diagnostic entity. Quantifiable Records; Identify intervention time points The baseline period and follow-up period are divided; baseline and follow-up statistics are calculated according to standard diagnosis and treatment codes, the rate of change and direction of change are calculated, and a comparison view of indicator change trends is generated. When the intent category is association analysis, from the fused clinical data item set Extract cross-modal data distribution from seed nodes. Starting from the clinical diagnosis and treatment-related edge, construct the associated path. High-value paths are selected based on correlation strength and cross-modal markers, ultimately generating cross-modal diagnostic and treatment data association navigation paths.

[0104] According to intent category Select the corresponding response assembly method.

[0105] S9. Return the query response to the user and iteratively optimize the clinical natural language understanding model based on user feedback.

[0106] In practice, when the user terminal is a visual data analysis system, complete structured data is returned; when the user terminal is a mobile application, concise summary data is returned; and a log of the returned behavior is recorded. A set of feedback behavior types is defined. Each behavior is assigned a feedback weight, and behavior sequences are collected. Calculate the comprehensive feedback score : in, This refers to the first return time; This is the time decay coefficient; The timestamp of the event; .

[0107] Identifying diagnostic entities and extracting error signals Deviation signal from query intent Generate incremental training data ,when At that time, it triggers fine-tuning of the clinical natural language understanding model: The updated model will be deployed to replace the original model.

[0108] Reference Figure 5 Corresponding to the virtual integration and query method for multi-source heterogeneous clinical data provided in the embodiments of the present invention, the embodiments of the present invention also provide a virtual integration and query system for multi-source heterogeneous clinical data, the system comprising: The standardization processing module is used to collect raw clinical data and perform diagnostic semantic encoding and value range normalization on the raw clinical data to generate standardized clinical data. The journey building module is used to arrange and reorganize standardized clinical data and build a virtual patient journey data chain covering the entire patient life cycle; The instruction generation module includes a model building unit, a request parsing unit, an intent discrimination unit, a time recognition unit, and an instruction generation unit. The model building unit is used to establish a clinical natural language understanding model. The request parsing unit is used to extract diagnostic entities from clinical query requests submitted by users in natural language form based on the clinical natural language understanding model. The intent discrimination unit is used to discriminate the query intent of clinical query requests submitted by users in natural language form based on the clinical natural language understanding model. The time recognition unit is used to identify the time window limitation of clinical query requests submitted by users in natural language form based on the clinical natural language understanding model. The instruction generation unit is used to assemble the diagnostic entities extracted by the request parsing subunit, the intent categories identified by the intent discrimination subunit, and the time constraints determined by the time recognition subunit to generate structured query instructions. The node localization module is used to locate a set of matching candidate data nodes in the virtual patient journey data chain based on the target diagnosis and treatment entity, and to perform time window pruning on the set of candidate data nodes according to time constraints to obtain a time-aligned target node sequence. The strategy determination module is used to determine the cross-source data retrieval strategy based on the intent category and the sequence of target nodes. The query module is used to send localized query statements, which have undergone diagnostic semantic transformation, to the corresponding data source according to the cross-source data retrieval strategy, and obtain the original sub-result set returned by the data source; The credibility fusion module is used to perform credibility-weighted fusion of synonym clinical data in the original sub-result set, generate fused clinical data items, and attach credibility level labels to the fused clinical data items; An assembly module is used to assemble query response content based on the fused clinical data items and the stated intent category; The feedback module is used to return the query response to the user and iteratively optimize the clinical natural language understanding model based on user feedback.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0110] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0111] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0112] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0113] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0115] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0116] Furthermore, although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

Claims

1. A method for virtual integration and querying of multi-source heterogeneous clinical data, characterized in that, The method includes: S1. Collect raw clinical data, perform diagnostic semantic encoding and value range normalization on the raw clinical data, and generate standardized clinical data. S2. Arrange and reorganize the standardized clinical data to construct a virtual patient journey data chain covering the entire life cycle of the patient; S3. Establish a clinical natural language understanding model, and based on the clinical natural language understanding model, extract diagnostic entities, determine query intent, and identify time window limits for clinical query requests submitted by users in natural language form, and generate structured query instructions including target diagnostic entities, intent categories, and time constraints. S4. Based on the target diagnosis and treatment entity, locate the matching set of candidate data nodes in the virtual patient journey data chain, and perform time window pruning on the set of candidate data nodes according to the time constraint to obtain a time-aligned target node sequence. S5. Determine a cross-source data retrieval strategy based on the intent category and the target node sequence; S6. According to the cross-source data retrieval strategy, send a localized query statement after diagnosis and treatment semantic conversion to the corresponding data source to obtain the original sub-result set returned by the data source; S7. Perform confidence-weighted fusion of synonym clinical data in the original sub-result set to generate fused clinical data items, and attach a confidence level identifier to the fused clinical data items; S8. Assemble the query response content based on the fused clinical data items and the intent category; S9. Return the query response to the user and iteratively optimize the clinical natural language understanding model based on user feedback.

2. The method for virtual integration and querying of multi-source heterogeneous clinical data according to claim 1, characterized in that, The diagnostic semantic encoding and value range normalization processing include: Extract business semantic information and data format features from the original clinical data, and identify the clinical diagnosis and treatment business scenario corresponding to the original clinical data; The data items of the original clinical data are matched with the standard diagnosis and treatment codes in the standard clinical terminology set to establish a correspondence database between the original data fields and the standard diagnosis and treatment codes; Based on the corresponding relational database, the original data value range is converted in units, the encoding is replaced, and the format is unified to generate standardized clinical data.

3. The method for virtual integration and querying of multi-source heterogeneous clinical data according to claim 2, characterized in that, The construction of the virtual patient journey data chain covering the entire patient lifecycle includes: Establish a globally unique patient identity index for the same patient distributed across different clinical business systems; Using the patient identity index as the root node, the standardized clinical data are sorted in ascending order according to the clinical event timestamps to form a panoramic time-series data stream of patients; Within each clinical event, establish clinical diagnosis and treatment association edges between data nodes, and record the association type, association strength, and bidirectional traceability path.

4. The virtual integration and query method for multi-source heterogeneous clinical data according to claim 3, characterized in that, The establishment of the clinical natural language understanding model includes: Collect clinical diagnosis and treatment corpus; The clinical diagnosis and treatment corpus is labeled with diagnosis and treatment entities, query intent categories, and time descriptive words to form a training sample set with labeled information. The pre-trained language model is fine-tuned based on the training sample set to construct a clinical natural language understanding model.

5. The method for virtual integration and querying of multi-source heterogeneous clinical data according to claim 4, characterized in that, The generated structured query instructions, including the target diagnostic entity, intent category, and time constraints, include: The clinical natural language understanding model is used to extract diagnostic entities and their entity types from the clinical query request; The intent category of the clinical query request is identified based on the clinical natural language understanding model. Based on the diagnosis entity and the intent category, the time window slots in the pre-query instruction template are filled to generate a structured query instruction containing the target diagnosis entity, intent category, and time constraints.

6. The method for virtual integration and querying of multi-source heterogeneous clinical data according to claim 5, characterized in that, The set of candidate data nodes for location matching includes: The target diagnostic entity is matched with the standard diagnostic codes of each data node in the virtual patient journey data chain, and data nodes with consistent codes or similar semantics are selected to form a preliminary set of nodes. The initial screening node set is prioritized according to the intent category; The sorted set of initially screened nodes is output as the set of candidate data nodes.

7. The method for virtual integration and querying of multi-source heterogeneous clinical data according to claim 6, characterized in that, The determination of the cross-source data retrieval strategy includes: When the intent category is medical history tracing, starting from the earliest data node in the target node sequence, the virtual patient journey data chain is traversed in a forward direction to recall all historical data nodes related to the target diagnosis and treatment entity, up to the current time node. When the intent category is efficacy assessment, the latest data node in the target node sequence is taken as the endpoint, and the data chain of the virtual patient journey is traced backward to recall the baseline data node before the intervention and the follow-up data node after the intervention. When the intent category is association analysis, starting from each data node in the target node sequence, multi-hop expansion is performed along the clinical diagnosis and treatment association edge to recall cross-modal data nodes that are directly or indirectly associated with the target diagnosis and treatment entity. The cross-modal data nodes include structured test data, semi-structured medical record text, and unstructured medical image data.

8. The method for virtual integration and querying of multi-source heterogeneous clinical data according to claim 7, characterized in that, The localized query statements that have undergone diagnostic semantic conversion include: The traversal direction identifier and associated hop count parameter in the cross-source data retrieval strategy are parsed, and a query sequence is generated based on the traversal direction identifier and associated hop count parameter. Based on the real-time access handles reserved by each data node in the target node sequence, determine the set of target data source addresses to be queried; The standard diagnosis and treatment codes in the query sequence are converted into the local encoding format corresponding to each target data source, and the standard time expression is converted into the time query syntax corresponding to each target data source, and then assembled into a local query statement. The localized query statements are sent in parallel to the corresponding target data sources.

9. The virtual integration and query method for multi-source heterogeneous clinical data according to claim 8, characterized in that, The generation of fusion clinical data items includes: Identify data items in the original sub-results set that represent the same clinical diagnosis and treatment concept, and form synonym data groups; Assign basic weights to the data items according to the data source authority level table; Calculate the time decay coefficient and consistency coefficient based on the data items; The base weight, the time decay coefficient, and the consistency coefficient are calculated to obtain the fusion weight of each data item. The values ​​of each data item are then weighted and averaged according to the fusion weight to generate fused clinical data items.

10. A virtual integrated query system for multi-source heterogeneous clinical data, characterized in that, The system, used to implement the virtual integration and query method for multi-source heterogeneous clinical data as described in claim 1, comprises: The standardization processing module is used to collect raw clinical data and perform diagnostic semantic encoding and value range normalization on the raw clinical data to generate standardized clinical data. The journey building module is used to arrange and reorganize the standardized clinical data and build a virtual patient journey data chain covering the entire life cycle of the patient; The instruction generation module includes a model building unit, a request parsing unit, an intent discrimination unit, a time recognition unit, and an instruction generation unit. It is used to establish a clinical natural language understanding model and, based on the clinical natural language understanding model, to discriminate the query intent and time window limitation of the clinical query request submitted by the user in natural language form, so as to generate a structured query instruction. The node localization module is used to locate a matching set of candidate data nodes in the virtual patient journey data chain according to the structured query instruction, and to perform time window pruning on the set of candidate data nodes to obtain a time-aligned target node sequence. The strategy determination module is used to determine a cross-source data retrieval strategy based on the structured query instruction and starting from the target node sequence. The query module is used to send localized query statements that have undergone diagnostic semantic transformation to the corresponding data source according to the cross-source data retrieval strategy, and obtain the original sub-result set returned by the data source; The credibility fusion module is used to perform credibility-weighted fusion of synonym clinical data in the original sub-result set, generate fused clinical data items, and attach credibility level identifiers to the fused clinical data items. An assembly module is used to assemble query response content based on the fused clinical data items and the intent category; The feedback module is used to return the query response content to the user and iteratively optimize the clinical natural language understanding model based on the user's feedback behavior.

Images