Heterogeneous mixed mode-based disease quality control labeling system

Abstract

This invention relates to the field of medical data processing technology, specifically to a disease quality control annotation system based on a heterogeneous hybrid model. The system includes a data access and preprocessing module, a unified feature space quality control module, a hybrid annotation guidance module, an annotation execution and arbitration module, and an integrated output module. Specifically: the data access and preprocessing module is used to access raw data of the target disease from heterogeneous data sources; the unified feature space quality control module is used to generate corresponding preliminary quality control opinions; the hybrid annotation guidance module is used to generate hybrid annotation guidance information; and the annotation execution and arbitration module generates the arbitrated annotation results according to preset arbitration rules. This invention, by constructing a unified feature expression, guided annotation execution, and a structured arbitration decision-making mechanism, achieves standardization, controllability, and traceability of the disease data annotation process, effectively improving the consistency and quality control accuracy of the annotation results.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a disease quality control labeling system based on a heterogeneous hybrid model. Background Technology

[0002] In the field of medical artificial intelligence, disease annotation and quality control processing of structured and unstructured medical data is an important prerequisite for building high-quality disease knowledge graphs, clinical decision support systems, and disease risk prediction models. With the widespread deployment of various medical information systems, imaging systems, and experimental testing platforms, massive amounts of heterogeneous disease data are continuously collected and stored in hospital information systems. In order to fully explore the disease expression information contained in these data, it is necessary to rely on a data quality control and annotation system with feature fusion, annotation guidance, and conflict verification capabilities to achieve a complete processing loop from raw disease data to structured, highly reliable annotation results.

[0003] However, most existing systems either only support structured annotation of single-source data or rely solely on model inference results, lacking multi-terminal collaborative annotation and arbitration capabilities. This leads to poor data consistency, difficulty in detecting conflicts, and a lack of traceability for annotation quality, failing to support the needs of large-scale, high-quality disease data management. Therefore, there is an urgent need for a disease quality control annotation system based on a heterogeneous hybrid model to address the shortcomings of existing technologies in data standardization, unified application of quality control rules, conflict detection, and consistency assessment. Summary of the Invention

[0004] To achieve the above objectives, this invention provides a disease quality control labeling system based on a heterogeneous hybrid model.

[0005] A disease quality control annotation system based on a heterogeneous hybrid model includes a data access and preprocessing module, a unified feature space quality control module, a hybrid annotation guidance module, an annotation execution and arbitration module, and an integrated output module; among which:

[0006] Data access and preprocessing module: used to access raw data of the target disease from heterogeneous data sources, and to clean and standardize the raw data to generate standardized disease data;

[0007] Unified Feature Space Quality Control Module: This module maps standardized disease data to a preset unified feature space to form a unified feature vector. It then performs conflict detection on the unified feature vector based on a predefined set of disease quality control rules and generates corresponding preliminary quality control opinions.

[0008] Hybrid annotation guidance module: It is used to integrate the unified feature vector and the preliminary quality control opinions into the preset annotation guidance model, output annotation suggestions and annotation confidence for the corresponding data segment, and generate annotation task prompts and key annotation area indicators in combination with predefined annotation rules to form hybrid annotation guidance information;

[0009] The annotation execution and arbitration module organizes annotation terminals to perform manual annotation based on hybrid annotation guidance information, and generates an arbitrated annotation result according to preset arbitration rules when there are discrepancies in the annotation results for the same data segment.

[0010] The integrated output module is used to associate and integrate the arbitrated annotation results with the corresponding unified feature vectors and standardized disease data, and generate a comprehensive quality control score based on the preliminary quality control opinions and annotation consistency information, and output structured disease quality control annotation results.

[0011] Optionally, the data access and preprocessing module includes a data source access unit, a data structure parsing unit, a data cleaning unit, and a data standardization unit; wherein:

[0012] Data source access unit: used to obtain original data corresponding to the target disease from multiple heterogeneous data sources according to a preset data source access configuration table. The heterogeneous data sources include disease-related data with different data formats, different acquisition systems or different storage structures, and uniformly add disease identifiers and data source identifiers to each original data during the access process.

[0013] Data structure parsing unit: used to perform structure parsing processing on the incoming raw data, mapping non-uniform structured data fields to a preset set of data fields to form an intermediate data representation;

[0014] The data cleaning unit performs abnormal data removal, missing field marking, and redundant record deduplication on intermediate data representation to generate cleaned disease data.

[0015] Data standardization unit: used to unify the numerical scale and standardize the format of the cleaned disease data to generate standardized disease data.

[0016] Optionally, the unified feature space quality control module includes a feature mapping unit, a feature vector construction unit, a quality control rule matching unit, and a conflict detection unit; wherein:

[0017] Feature mapping unit: Used to receive standardized disease data and, according to a preset feature mapping relationship table, map each data field in the standardized disease data to the corresponding feature dimension in a unified feature space, forming a structure-aligned feature mapping result;

[0018] Feature vector construction unit: used to vectorize the feature mapping results according to the arrangement order of each feature dimension in the unified feature space, and construct a unified feature vector for subsequent quality control analysis;

[0019] Quality control rule matching unit: used to match the unified feature vector with the predefined disease quality control rule set item by item, and determine the rule constraints and allowed value ranges corresponding to each feature dimension;

[0020] Conflict detection unit: Based on the quality control rule matching results, it detects feature dimensions in the unified feature vector that violate rule constraints or have logical inconsistencies, identifies the corresponding data conflict items, and generates preliminary quality control opinions for the unified feature vector accordingly.

[0021] Optionally, the disease quality control rule set includes:

[0022] Feature value range consistency rule: used to limit the allowed value range of each numerical feature dimension in a unified feature vector;

[0023] Feature type matching rules: used to verify whether the actual data type of each feature dimension in the unified feature vector is consistent with the predefined data type of the corresponding dimension;

[0024] Required Feature Integrity Rule: This rule is used to constrain the requirement that feature dimensions marked as required in a unified feature vector must have valid values.

[0025] Logical consistency rules between features: used to constrain the predefined logical relationships between multiple feature dimensions;

[0026] Time sequence rationality rule: used to verify that the feature dimensions involving time attributes in a unified feature vector meet the preset time sequence relationship;

[0027] Disease-specific constraint rules: used to confirm that some feature dimensions in a unified feature vector are only allowed to appear under the conditions of the corresponding disease.

[0028] Optionally, the hybrid annotation guidance module includes a feature-quality control fusion unit, an annotation guidance model inference unit, an annotation rule parsing unit, and a guidance information generation unit; wherein:

[0029] The feature-quality control fusion unit is used to receive a unified feature vector and the corresponding preliminary quality control opinions, and to combine the preliminary quality control opinions as auxiliary constraint information with the unified feature vector to form a fusion feature representation for annotation guidance.

[0030] The annotation-guided model inference unit is used to input the fused feature representation into the preset annotation-guided model to perform inference processing. Based on the confirmed annotation patterns and feature associations in the historical annotation knowledge base, it outputs annotation suggestions for the current data segment and the annotation confidence corresponding to the annotation suggestions.

[0031] Annotation rule parsing unit: used to parse the preliminary quality control opinions and annotation suggestions to determine the annotation task types and annotation constraints that need to be considered in the annotation process of the current data segment;

[0032] Guidance information generation unit: Based on the annotation rule parsing results, it generates annotation task prompts for annotation personnel, and combines them with the data conflict items identified in the preliminary quality control opinions to determine the annotation areas that need to be checked in detail, thus forming a mixed annotation guidance information that includes annotation suggestions, annotation confidence levels, annotation task prompts, and key annotation area indications.

[0033] Optionally, the annotation-guided model inference unit includes:

[0034] Pattern retrieval subunit: It receives the fused feature representation output by the feature-quality control fusion unit, and calls the confirmed set of annotation patterns in the historical annotation knowledge base; it performs similarity calculation based on the fused feature representation and the feature representation of each annotation pattern, and filters out a set of candidate annotation patterns that match the current data segment;

[0035] Association weight calculation subunit: Calculate the corresponding association weight based on the similarity results of each labeled pattern in the candidate labeled pattern set;

[0036] The suggestion generation sub-unit is used to read the set of label categories associated with each label pattern in the candidate label pattern set, and to perform weighted summation of each label category based on the pattern association weight to obtain the set of label suggestion scores for the current data segment.

[0037] Confidence calculation subunit: used to convert the set of annotation suggestion item scores into the annotation confidence level corresponding to each annotation suggestion item. It outputs annotation suggestions and the corresponding annotation confidence levels.

[0038] Optionally, the annotation rule parsing unit includes:

[0039] Rule location subunit: Used to receive preliminary quality control opinions and annotation suggestions, and locate the set of rule entries corresponding to the current data segment in the predefined annotation rules based on the type of data conflict identified by the preliminary quality control opinions and the annotation category to which the annotation suggestions belong, forming a set of rules to be parsed;

[0040] Task type determination subunit: used to read the task definition items bound to the annotation category in the set of rules to be parsed, determine the annotation task type corresponding to the current data fragment, and output the annotation task type identifier;

[0041] Constraint Extraction Subunit: Used to read the constraint definition items bound to the data conflict items in the set of rules to be parsed, and extract the labeled constraints corresponding to the current data segment, including the set of mandatory feature dimensions, the set of prohibited label conflicts, and the set of consistency verification items;

[0042] Structured sub-unit of parsing results: used to encapsulate the annotation task type identifier and annotation constraints in a structured way to generate rule parsing results.

[0043] Optionally, the annotation execution and arbitration module includes an annotation task distribution unit, an annotation result collection unit, a disagreement detection unit, and an arbitration decision unit; wherein:

[0044] The annotation task distribution unit is used to receive mixed annotation guidance information, bind the mixed annotation guidance information with the corresponding standardized disease data and unified feature vectors, and push it to at least one annotation terminal according to the preset task distribution strategy, so as to organize the annotation personnel to perform manual annotation operations based on annotation suggestions, annotation task prompts and key annotation area instructions;

[0045] Annotation result acquisition unit: used to receive manual annotation results generated from different annotation terminals for the same data segment, and encapsulate the annotation results in a unified format to form a set of comparable annotation results;

[0046] Disagreement detection unit: Used to compare the consistency of each annotation result in the annotation result set. When the annotation results for the same data segment are found to be inconsistent in annotation category, annotation range or annotation attribute, it is determined that there is an annotation disagreement and the arbitration process is triggered.

[0047] Arbitration Decision Unit: After the arbitration process is triggered, it is used to make a comprehensive judgment on the disputed labeling results according to the preset arbitration rules and generate the post-arbitration labeling results.

[0048] Optionally, the arbitration decision-making unit includes:

[0049] Candidate Result Summarization Subunit: This unit receives all annotation results of data segments marked as conflicting by the divergence detection unit, and summarizes and organizes the results according to the annotation terminal number, annotation category, and annotation attribute field to form a structured candidate annotation result set.

[0050] Arbitration index calculation subunit: Based on the candidate annotation result set, and combining annotation confidence, historical consistency score, and conflict level, calculate the comprehensive arbitration score for each annotation category. ;

[0051] Arbitration rule determination subunit: used to sort and filter the comprehensive arbitration scores of each candidate label category according to a preset arbitration rule set, which includes: highest priority rules, threshold filtering rules and task type constraint rules;

[0052] Arbitration Result Generation Sub-unit: Used to determine the sorting results output by the sub-unit according to the arbitration rules, determine the final arbitration label category, extract the corresponding label attributes and positions, and generate the post-arbitration labeling results.

[0053] Optionally, the integrated output module includes a result correlation unit, a consistency analysis unit, a score generation unit, and a result encapsulation and output unit; wherein:

[0054] Result association unit: Used to receive the post-arbitration annotation results output by the arbitration decision unit, and bind and integrate them with the unified feature vector and standardized disease data corresponding to the same data segment to form a structured data annotation unit entity;

[0055] Consistency Analysis Unit: Used to compare the annotation results after arbitration with the candidate suggestions and actual annotation terminal results during the annotation process, identify whether there are low consistency situations during the annotation process, extract consistency indicators and cross-validate them with the abnormal items in the preliminary quality control opinions to form a set of consistency evaluation factors;

[0056] Scoring Generation Unit: Based on the number, type, and severity of conflicting items in the preliminary quality control comments, and the set of consistency assessment factors, a comprehensive quality control score is generated. ;

[0057] The result encapsulation output unit is used to encapsulate the arbitration-post-annotation results, unified feature vectors, standardized disease data, and comprehensive quality control scores in a structured manner to form a structured disease quality control annotation result file.

[0058] The beneficial effects of this invention are:

[0059] This invention constructs a closed-loop annotation process that includes data access, feature mapping, quality control detection, model guidance, manual annotation, disagreement arbitration, and output scoring. It supports the extraction of unified feature vectors from multi-source heterogeneous disease data and integrates preliminary quality control opinions to guide the execution of annotation tasks, so that the annotation results have clear semantic basis and a manually controllable path.

[0060] This invention, through hierarchical analysis of quality control rules and annotation rules, combined with dual constraints of task type and conflict type, makes annotation suggestions context-adaptable, significantly improving the consistency and relevance of annotation tasks.

[0061] This invention achieves transparency, scoring, and traceability of the entire annotation process by introducing a structured arbitration mechanism in the annotation disagreement handling stage and superimposing consistency analysis and quality control scoring generation in the result output stage. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of the disease quality control labeling system according to an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of the hybrid annotation guidance module according to an embodiment of the present invention. Detailed Implementation

[0065] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0066] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0067] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0068] like Figures 1-2As shown, the disease quality control annotation system based on a heterogeneous hybrid model includes a data access and preprocessing module, a unified feature space quality control module, a hybrid annotation guidance module, an annotation execution and arbitration module, and an integrated output module; wherein:

[0069] Data access and preprocessing module: used to access raw data of the target disease from heterogeneous data sources, and to clean and standardize the raw data to generate standardized disease data;

[0070] Unified Feature Space Quality Control Module: This module maps standardized disease data to a preset unified feature space to form a unified feature vector. It then performs conflict detection on the unified feature vector based on a predefined set of disease quality control rules and generates corresponding preliminary quality control opinions.

[0071] Hybrid annotation guidance module: It is used to integrate the unified feature vector and the preliminary quality control opinions into the preset annotation guidance model, output annotation suggestions and annotation confidence for the corresponding data segment, and generate annotation task prompts and key annotation area indicators in combination with predefined annotation rules to form hybrid annotation guidance information;

[0072] The annotation execution and arbitration module organizes annotation terminals to perform manual annotation based on hybrid annotation guidance information, and generates an arbitrated annotation result according to preset arbitration rules when there are discrepancies in the annotation results for the same data segment.

[0073] The integrated output module is used to associate and integrate the arbitrated annotation results with the corresponding unified feature vectors and standardized disease data, and generate a comprehensive quality control score based on the preliminary quality control opinions and annotation consistency information, and output structured disease quality control annotation results.

[0074] The data access and preprocessing module includes a data source access unit, a data structure parsing unit, a data cleaning unit, and a data standardization unit; among which:

[0075] Data source access unit: It is used to obtain the original data corresponding to the target disease from multiple heterogeneous data sources according to the preset data source access configuration table. The heterogeneous data sources include disease-related data with different data formats, different acquisition systems or different storage structures, and uniformly add disease identifiers and data source identifiers to each original data during the access process.

[0076] Table 1 Data Source Access Configuration Table

[0077] Configuration item number Data source identifier Data source type Data format type Disease Identifier Field Data field mapping rules Collection time field Access frequency C1 DS_001 Medical Information System Structured Tables disease_code Map_Rule_01 record_time real time C2 DS_002 Medical Imaging System binary file disease_id Map_Rule_02 scan_time timing C3 DS_003 Laboratory testing system semi-structured data disease_type Map_Rule_03 test_time timing C4 DS_004 Manual data entry system Text data disease_label Map_Rule_04 input_time Event triggered

[0078] In Table 1 above, the configuration item number is used to uniquely identify a data source access configuration record, serving as an index identifier for internal calls and maintenance within the data access and preprocessing module; the data source identifier is used to uniquely identify the specific data source system, serving as the basis for access channel selection and data source labeling in the data source access unit; the data source type indicates the system category to which the data source belongs, distinguishing different acquisition mechanisms and data generation methods, and supporting the selection of subsequent differentiated parsing strategies; the data format type identifies the organization form of the original data, serving as the basis for the data structure parsing unit to select the corresponding parsing strategy; the disease identifier field specifies the field name in the original data used to identify the target disease, and the data source access unit extracts the original data records related to the target disease accordingly; the data field mapping rule defines the correspondence between the original data fields and the preset data field set, and the data structure parsing unit performs field semantic alignment and structural unification processing according to this rule; the acquisition time field specifies the field in the original data that represents the data generation or acquisition time, serving as the basis field for subsequent time alignment and data consistency checks; and the access frequency specifies the triggering method for the data source access unit to perform data acquisition operations on the corresponding data source, ensuring controllable access to data from different sources in the time dimension.

[0079] Data structure parsing unit: used to perform structure parsing processing on the incoming raw data, mapping non-uniform structured data fields to a preset set of data fields, forming intermediate data representations with clear field meanings and comparability;

[0080] The data cleaning unit performs abnormal data removal, missing field marking, and redundant record deduplication on intermediate data representation to generate cleaned disease data.

[0081] The data standardization unit is used to unify the numerical scale and standardize the format of the cleaned disease data, generating standardized disease data, and outputting the standardized disease data to the unified feature space quality control module. By setting up the data source access unit, data structure parsing unit, data cleaning unit and data standardization unit in sequence in the data access and preprocessing module, the orderly access, structural unification and quality constraint processing of heterogeneous disease raw data are realized. This ensures that the output disease data is consistent in terms of field meaning, value range and data quality, and provides a stable and comparable data foundation for subsequent unified feature space mapping and quality control analysis.

[0082] The unified feature space quality control module includes a feature mapping unit, a feature vector construction unit, a quality control rule matching unit, and a conflict detection unit; among which:

[0083] Feature mapping unit: Used to receive standardized disease data and, according to a preset feature mapping relationship table, map each data field in the standardized disease data to the corresponding feature dimension in a unified feature space, forming a structure-aligned feature mapping result;

[0084] Table 2 Feature Mapping Relationship Table

[0085] Original field name Standard field name Mapping feature dimension number Data types Standardization strategy Missing data handling methods Logical dependency fields age patient_age F01 Numerical Min-Max Normalization Mean fill none gender_code gender F02 Enumeration type One-Hot encoding Outlier removal none diag_code diagnosis_id F03 Category coding type Encoding Mapping Missing markers none record_time timestamp F04 Time type Timestamp Conversion Discard non-conforming values none height_cm height F05 Numerical Z-score standardization Null value removal weight_kg weight_kg weight F06 Numerical Z-score standardization Null value removal height_cm bmi bmi F07 Numerical Direct value Dependency Recalculation height_cm, weight_kg

[0086] In Table 2 above, the original field name refers to the field name of the original field in the standardized disease data, used to identify the data source field; the standard field name refers to the unified feature field name defined internally by the system, used to form the standard feature vector dimension; the mapped feature dimension number indicates the position index of the field in the unified feature vector, ensuring that the dimensions are ordered and do not overlap when constructing the vector; the data type refers to the data type of the original field, used to select the corresponding standardization strategy; the standardization strategy refers to the rules for normalizing, encoding, or normalizing the original field, used for unified feature processing; the missing value handling method refers to the handling method when the field has missing or unqualified values, used to ensure the integrity of feature construction; the logical dependency field indicates that if the standardization or calculation of a certain field depends on other fields, then its dependent fields are listed, used for field scheduling judgment during execution.

[0087] Feature vector construction unit: used to vectorize the feature mapping results according to the arrangement order of each feature dimension in the unified feature space, and construct a unified feature vector for subsequent quality control analysis;

[0088] Quality control rule matching unit: used to match the unified feature vector with the predefined disease quality control rule set item by item, and determine the rule constraints and allowed value ranges corresponding to each feature dimension;

[0089] Conflict detection unit: Based on the quality control rule matching results, it detects feature dimensions in the unified feature vector that violate rule constraints or have logical inconsistencies, identifies the corresponding data conflict items, and generates preliminary quality control opinions for the unified feature vector accordingly.

[0090] In the conflict detection unit, the rule compliance of the i-th feature dimension in the unified feature vector is determined as follows:

[0091] ;in, Represents the first eigenvector in the unified eigenvector. The values ​​of each feature dimension; This indicates that the quality control rules for the disease types are concentrated in the first category. Each feature dimension has a preset allowed value range; Indicates the first The conflict identification results for each feature dimension, where This indicates a conflict. The above steps, by setting up a feature mapping unit, a feature vector construction unit, a quality control rule matching unit, and a conflict detection unit in the unified feature space quality control module, realize the consistent expression of standardized disease data in the unified feature space and the systematic conflict detection under rule constraints. This enables the disease data to complete structured quality control analysis before entering the annotation guidance stage, providing a clear and referable preliminary quality control basis for subsequent annotation guidance and arbitration processing.

[0092] The disease quality control rule set includes:

[0093] Feature value range consistency rule: used to limit the allowed value range of each numerical feature dimension in a unified feature vector, to prevent feature values ​​from exceeding the reasonable range corresponding to the disease;

[0094] Feature type matching rules: used to verify whether the actual data type of each feature dimension in the unified feature vector is consistent with the predefined data type of the corresponding dimension, to prevent quality control deviations caused by type mismatch;

[0095] Required feature integrity rule: This rule is used to constrain the feature dimensions marked as required in the unified feature vector to have valid values, in order to prevent the loss of key disease information.

[0096] Logical consistency rules between features: These rules are used to constrain multiple feature dimensions to meet a predefined logical relationship, preventing semantic conflicts in feature combinations.

[0097] Time sequence rationality rules: used to verify that the feature dimensions involving time attributes in the unified feature vector meet the preset time sequence relationship, and to prevent time logic abnormalities;

[0098] Disease-specific constraint rules: These rules are used to confirm that certain feature dimensions in a unified feature vector are only allowed to appear under the conditions of the corresponding disease, thus preventing cross-disease features from being mixed in.

[0099] After receiving the unified feature vector, the quality control rule matching unit performs item-by-item matching on the disease-specific quality control rule set according to the rule number order. The specific process includes the following steps:

[0100] Rule location steps: The quality control rule matching unit retrieves each quality control rule from the disease quality control rule set in sequence according to the rule number, and parses the target feature dimension or feature dimension combination corresponding to the rule;

[0101] Feature extraction steps: For the feature dimensions associated with the current rule, extract the corresponding feature values ​​from the unified feature vector to form a feature subset for rule validation;

[0102] Rule condition determination steps: Compare the feature subset with the predefined constraints in the current rule to determine whether the rule conditions such as value range, type requirements, logical relationship or time relationship are met;

[0103] Matching result labeling steps: When the judgment result does not meet the current rule constraint, generate a rule violation label corresponding to the rule number, and record the feature dimension number and rule number of the conflict;

[0104] The matching result summary step is as follows: After completing the item-by-item matching of all rules, the matching results corresponding to each rule are summarized to form a set of rule matching results for subsequent conflict detection unit calls. The above steps split the disease quality control rule set into multiple rules with clear constraints and judgment conditions, and the quality control rule matching unit performs item-by-item matching on the unified feature vector according to the rule number. This makes the quality control process of disease data have a clear rule source and a traceable matching path, and can systematically identify abnormal, missing or logically conflicting data at the feature level, providing a stable and interpretable rule basis for subsequent conflict detection and the generation of preliminary quality control opinions.

[0105] The hybrid annotation guidance module includes a feature-quality control fusion unit, an annotation guidance model inference unit, an annotation rule parsing unit, and a guidance information generation unit; among which:

[0106] The feature-quality control fusion unit is used to receive a unified feature vector and the corresponding preliminary quality control opinions, and to combine the preliminary quality control opinions as auxiliary constraint information with the unified feature vector to form a fusion feature representation for annotation guidance.

[0107] The annotation-guided model inference unit is used to input the fused feature representation into the preset annotation-guided model to perform inference processing. Based on the confirmed annotation patterns and feature associations in the historical annotation knowledge base, it outputs annotation suggestions for the current data segment and the annotation confidence corresponding to the annotation suggestions.

[0108] Annotation rule parsing unit: used to parse the preliminary quality control opinions and annotation suggestions to determine the annotation task types and annotation constraints that need to be considered in the annotation process of the current data segment;

[0109] The guidance information generation unit generates annotation task prompts for annotators based on the annotation rule parsing results. It also identifies annotation areas requiring focused verification by combining data conflict indicators from the preliminary quality control feedback. This results in a hybrid annotation guidance information system containing annotation suggestions, annotation confidence levels, annotation task prompts, and key annotation area indications. By introducing a feature-quality control fusion unit, an annotation guidance model inference unit, an annotation rule parsing unit, and a guidance information generation unit into the hybrid annotation guidance module, the unified feature vector and preliminary quality control feedback can be systematically fused before entering the manual annotation stage. Under rule constraints, it generates clearly structured annotation guidance information, providing clear and referable guidance for the annotation execution stage and ensuring the annotation process unfolds systematically with consistent task prompts and focus areas.

[0110] The labeled guided model inference unit includes:

[0111] The pattern retrieval subunit receives the fused feature representation output by the feature-quality control fusion unit and retrieves the confirmed set of annotation patterns from the historical annotation knowledge base. It performs similarity calculations based on the fused feature representation and the feature representations of each annotation pattern, filtering to obtain a set of candidate annotation patterns that match the current data segment. The similarity calculation formula is: , where z represents the fusion feature representation vector of the current data segment; Indicates the first in the historical annotation knowledge base Feature representation vectors of a confirmed labeled pattern; The current fusion feature represents the relationship with the first... Similarity between annotation patterns; Represents the vector norm;

[0112] The association weight calculation subunit calculates the corresponding association weights based on the similarity results of each labeled pattern in the candidate labeled pattern set, forming a weight set for subsequent suggestion generation; the expression for the association weight is: ,in, Indicates the first The association weight of each candidate annotation pattern; This represents the weight sensitivity coefficient; The set of indices representing the set of candidate annotation patterns; Indicates the first The similarity of candidate annotation patterns;

[0113] The suggestion generation sub-unit reads the set of label categories associated with each label pattern in the candidate label pattern set, and performs a weighted summation of each label category based on the pattern association weight to obtain the set of label suggestion scores for the current data segment, from which the label suggestions are determined; the weighting formula is: ,in, Indicates the label category The corresponding suggestion item score; Indicates the first Each candidate annotation pattern corresponds to the annotation category. The strength of the association; Indicates the first The association weight of each candidate annotation pattern; The set of indices representing the set of candidate annotation patterns;

[0114] Confidence calculation subunit: used to convert the set of annotation suggestion item scores into the annotation confidence level corresponding to each annotation suggestion item. It outputs annotation suggestions and their corresponding annotation confidence scores; the expression for the confidence score is: ,in: Indicates the label category The corresponding label confidence level; Indicates the label category The corresponding suggestion item score; This represents the set of candidate annotation categories. By sequentially setting up a pattern retrieval subunit, an association weight calculation subunit, a suggestion item generation subunit, and a confidence calculation subunit in the annotation guidance model inference unit, the fused feature representation can complete traceable retrieval, weighted summarization, and confidence output under the constraints of annotation patterns and feature association relationships in the historical annotation knowledge base. This provides a structured and referable annotation suggestion basis for subsequent annotation rule parsing and guidance information generation.

[0115] The annotation rule parsing unit includes:

[0116] Rule location subunit: Used to receive preliminary quality control opinions and annotation suggestions, and locate the set of rule entries corresponding to the current data segment in the predefined annotation rules based on the type of data conflict identified by the preliminary quality control opinions and the annotation category to which the annotation suggestions belong, forming a set of rules to be parsed;

[0117] Task type determination subunit: used to read the task definition items bound to the annotation category in the set of rules to be parsed, determine the annotation task type corresponding to the current data fragment, and output the annotation task type identifier;

[0118] Constraint Extraction Subunit: Used to read the constraint definition items bound to the data conflict items in the set of rules to be parsed, and extract the labeled constraints corresponding to the current data segment, including the set of mandatory feature dimensions, the set of prohibited label conflicts, and the set of consistency verification items;

[0119] Structured sub-unit of parsing results: used to encapsulate the annotation task type identifier and annotation constraints in a structured way to generate rule parsing results.

[0120] The above predefined annotation rules include:

[0121] Labeling rule 1: This rule is used to ensure that the final labeling result is consistent with the set of labeling categories corresponding to the labeling suggestions, and prohibits the output of labeling categories that do not appear in the labeling suggestions.

[0122] Annotation Rule 2: This rule specifies that when there are conflict markers in the preliminary quality control opinions, the corresponding feature dimension must be included in the key annotation area and used as a mandatory verification object.

[0123] Labeling rule 3: This rule specifies that when a labeling suggestion contains multiple candidate labeling categories, the labeling process must compare the differences between each candidate category to prevent missing valid labeling paths.

[0124] Annotation rule 4: This rule is used to constrain the values ​​or judgments of the annotation results on the key feature dimensions to be consistent with the corresponding features in the unified feature vector, and semantic conflicts are prohibited.

[0125] Annotation Rule 5: This rule specifies that the anomaly types and conflict levels identified in the preliminary quality control comments must be inherited as constraints during the annotation stage and will affect the determination of the annotation task type.

[0126] Annotation rule 6: Used to limit the allowed annotation granularity levels under different annotation task types, to prevent the annotation result granularity from being mismatched with the current task type;

[0127] Labeling rule 7: This rule is used to constrain the completion of all mandatory labeling items associated with the current data segment during the labeling process, and prohibits the generation of incomplete labeling results.

[0128] The annotation execution and arbitration module includes an annotation task distribution unit, an annotation result collection unit, a disagreement detection unit, and an arbitration decision unit; among which:

[0129] The annotation task distribution unit is used to receive mixed annotation guidance information, bind the mixed annotation guidance information with the corresponding standardized disease data and unified feature vectors, and push it to at least one annotation terminal according to the preset task distribution strategy, so as to organize the annotation personnel to perform manual annotation operations based on annotation suggestions, annotation task prompts and key annotation area instructions;

[0130] The annotation result acquisition unit is used to receive manual annotation results generated by different annotation terminals for the same data segment, and to encapsulate the annotation results in a unified format to form a set of comparable annotation results.

[0131] Disagreement detection unit: Used to compare the consistency of each annotation result in the annotation result set. When the annotation results for the same data segment are found to be inconsistent in annotation category, annotation range or annotation attribute, it is determined that there is an annotation disagreement and the arbitration process is triggered.

[0132] Arbitration Decision Unit: After the arbitration process is triggered, it comprehensively judges the disputed annotation results according to the preset arbitration rules and generates the post-arbitration annotation results. By setting up an annotation task distribution unit, annotation result collection unit, dispute detection unit and arbitration decision unit in the annotation execution and arbitration module, the mixed annotation guidance information can effectively drive the manual annotation process, and complete the result convergence through unified arbitration rules when annotation disputes occur. This ensures that the disease annotation results still have consistency and referenceability under the condition of multi-terminal participation, and provides a stable annotation foundation for subsequent quality control scoring and result integration.

[0133] The arbitration decision-making unit includes:

[0134] Candidate Result Summarization Subunit: This unit receives all annotation results of data segments marked as conflicting by the divergence detection unit, and summarizes and organizes the results according to the annotation terminal number, annotation category, and annotation attribute field to form a structured candidate annotation result set.

[0135] Arbitration index calculation subunit: Based on the candidate annotation result set, and combining annotation confidence, historical consistency score, and conflict level, calculate the comprehensive arbitration score for each annotation category. The specific calculation is as follows: ,in, Indicates the label category Comprehensive arbitration score; The average label confidence level corresponding to this category; Historical consistency score represents the consistency rate of annotations on similar data in the past; The conflict level coefficient associated with the current data segment; These are the preset weighting coefficients, satisfying... ;

[0136] Arbitration rule determination subunit: used to sort and filter the comprehensive arbitration scores of each candidate label category according to the preset arbitration rule set. The arbitration rule set includes: the highest priority rule, the threshold filtering rule, and the task type constraint rule, which are used to ensure that the arbitration results meet the requirements of accuracy and contextual rationality.

[0137] The highest score priority rule is defined as follows: when there is disagreement among multiple candidate label categories, the label category with the highest comprehensive arbitration score shall be selected as the arbitration result.

[0138] Applicable scenarios: When all candidate label categories meet the minimum confidence requirement; when there are no other mandatory priority restrictions;

[0139] Judgment logic: Sort by comprehensive arbitration score of each labeled category from high to low; take the category corresponding to the first place as the arbitration output; if there is a tie, make a decision according to the preset priority order.

[0140] The threshold filtering rule is defined as follows: a minimum confidence threshold is set for the comprehensive arbitration score of each candidate label category, and only those scores higher than the threshold are included in the arbitration candidate set;

[0141] Applicable scenarios: Excluding low-consistency results with abnormally high confidence but from a single source; controlling the risk of misjudgment and avoiding misjudgment due to scoring bias;

[0142] Decision logic: Set threshold For each labeled category ,when If the threshold is not met, the candidate is directly eliminated; if none of the candidates meet the threshold, the output requires manual confirmation or is delayed.

[0143] The task type constraint rule is defined as follows: based on the annotation task type corresponding to the current data fragment, the arbitration result must belong to the set of annotation categories allowed by that task type;

[0144] Applicable scenarios: Multi-task mixed annotation scenarios, such as disease identification + severity annotation; certain tasks only allow specific annotation granularity;

[0145] Judgment logic: Parse the current task type identifier; retrieve the set of allowed labeling categories bound to this task type; if the category with the highest score is not in the set, search downwards for matching categories; if there is no match, output a violation of task constraints flag that needs to be manually processed.

[0146] The combined use of these three types of arbitration rules can ensure reasonable screening, quality control, and semantic consistency of conflict labeling results. Among them, the highest score priority ensures efficiency and priority judgment; the threshold filtering ensures confidence boundaries; and the task type constraint ensures semantic boundaries. Together, they constitute a hierarchical decision path mechanism for the arbitration rule's judgment logic.

[0147] The arbitration result generation subunit is used to determine the final arbitration label category based on the ranking results output by the arbitration rule determination subunit, extract the corresponding label attributes and positions, and generate the post-arbitration label results. By setting up subunits such as candidate result aggregation, arbitration score calculation, rule determination and result generation in the arbitration decision unit, the quantitative evaluation of the disputed label results and the comprehensive judgment under the rule constraints are realized, so that the final arbitration result has a clear calculation basis and decision path, effectively improving the stability and consistency of the label results in conflict situations.

[0148] The integrated output module includes a result correlation unit, a consistency analysis unit, a score generation unit, and a result encapsulation and output unit; among which:

[0149] Result association unit: Used to receive the post-arbitration annotation results output by the arbitration decision unit, and bind and integrate them with the unified feature vector and standardized disease data corresponding to the same data segment to form a structured data annotation unit entity;

[0150] Consistency Analysis Unit: Used to compare the annotation results after arbitration with the candidate suggestions and actual annotation terminal results during the annotation process, identify whether there are low consistency situations during the annotation process, extract consistency indicators and cross-validate them with the abnormal items in the preliminary quality control opinions to form a set of consistency evaluation factors;

[0151] Scoring Generation Unit: Based on the number, type, and severity of conflicting items in the preliminary quality control comments, and the set of consistency assessment factors, a comprehensive quality control score is generated. The overall quality control score is calculated using the following formula:

[0152] ,in, This indicates the overall quality control score; Indicates the number of conflicting items identified in the preliminary quality control comments; This represents the total number of feature dimensions in the unified feature vector; This represents the consistency analysis score, ranging from [value missing]. ; Let be the scoring weight coefficient, satisfying ;

[0153] The results encapsulation and output unit is used to encapsulate the post-arbitration annotation results, unified feature vectors, standardized disease data, and comprehensive quality control scores in a structured manner, forming a structured disease quality control annotation result file. By setting up sub-units such as result association, consistency analysis, score generation, and output encapsulation, the post-arbitration annotation results can achieve high consistency integration with the original structured data. Based on the integration of preliminary quality control opinions and annotation consistency, a quantifiable quality control score is output, thereby providing complete and structured data support for quality control traceability, result review, and system evaluation.

[0154] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0155] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A disease quality control labeling system based on a heterogeneous hybrid model, characterized in that, It includes a data access and preprocessing module, a unified feature space quality control module, a hybrid annotation guidance module, an annotation execution and arbitration module, and an integrated output module; among which: Data access and preprocessing module: used to access raw data of the target disease from heterogeneous data sources, and to clean and standardize the raw data to generate standardized disease data; Unified Feature Space Quality Control Module: This module maps standardized disease data to a preset unified feature space to form a unified feature vector. It then performs conflict detection on the unified feature vector based on a predefined set of disease quality control rules and generates corresponding preliminary quality control opinions. Hybrid annotation guidance module: This module integrates a unified feature vector with preliminary quality control opinions, inputting them into a pre-defined annotation guidance model. It outputs annotation suggestions and confidence levels for the corresponding data segment, and combines predefined annotation rules to generate annotation task prompts and key annotation area indicators, forming hybrid annotation guidance information. The hybrid annotation guidance module includes a feature-quality control fusion unit, an annotation guidance model inference unit, an annotation rule parsing unit, and a guidance information generation unit. The feature-quality control fusion unit is used to receive a unified feature vector and the corresponding preliminary quality control opinions, and to combine the preliminary quality control opinions as auxiliary constraint information with the unified feature vector to form a fusion feature representation for annotation guidance. The annotation-guided model inference unit is used to input the fused feature representation into the preset annotation-guided model to perform inference processing. Based on the confirmed annotation patterns and feature associations in the historical annotation knowledge base, it outputs annotation suggestions for the current data segment and the annotation confidence corresponding to the annotation suggestions. Annotation rule parsing unit: used to parse the preliminary quality control opinions and annotation suggestions to determine the annotation task types and annotation constraints that need to be considered in the annotation process of the current data segment; Guidance information generation unit: Based on the annotation rule parsing results, it generates annotation task prompts for annotation personnel, and combines them with the data conflict items identified in the preliminary quality control opinions to determine the annotation areas that need to be checked in detail, thus forming a mixed annotation guidance information that includes annotation suggestions, annotation confidence, annotation task prompts, and key annotation area indications; The annotation-guided model inference unit includes: Pattern retrieval subunit: It receives the fused feature representation output by the feature-quality control fusion unit, and calls the confirmed set of annotation patterns in the historical annotation knowledge base; it performs similarity calculation based on the fused feature representation and the feature representation of each annotation pattern, and filters out a set of candidate annotation patterns that match the current data segment; Association weight calculation subunit: Calculate the corresponding association weight based on the similarity results of each labeled pattern in the candidate labeled pattern set; The suggestion generation sub-unit is used to read the set of label categories associated with each label pattern in the candidate label pattern set, and to perform weighted summation of each label category based on the pattern association weight to obtain the set of label suggestion scores for the current data segment. Confidence calculation subunit: used to convert the set of annotation suggestion item scores into the annotation confidence level corresponding to each annotation suggestion item. It also outputs annotation suggestions and the corresponding annotation confidence scores; The annotation execution and arbitration module organizes annotation terminals to perform manual annotation based on hybrid annotation guidance information, and generates an arbitrated annotation result according to preset arbitration rules when there are discrepancies in the annotation results for the same data segment. The integrated output module is used to associate and integrate the arbitrated annotation results with the corresponding unified feature vectors and standardized disease data, and generate a comprehensive quality control score based on the preliminary quality control opinions and annotation consistency information, and output structured disease quality control annotation results.

2. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 1, characterized in that, The data access and preprocessing module includes a data source access unit, a data structure parsing unit, a data cleaning unit, and a data standardization unit; wherein: Data source access unit: used to obtain original data corresponding to the target disease from multiple heterogeneous data sources according to a preset data source access configuration table. The heterogeneous data sources include disease-related data with different data formats, different acquisition systems or different storage structures, and uniformly add disease identifiers and data source identifiers to each original data during the access process. Data structure parsing unit: used to perform structure parsing processing on the incoming raw data, mapping non-uniform structured data fields to a preset set of data fields to form an intermediate data representation; Data cleaning unit: used to perform abnormal data removal, missing field marking and redundant record deduplication on intermediate data representation, and generate cleaned disease data; Data standardization unit: used to unify the numerical scale and standardize the format of the cleaned disease data to generate standardized disease data.

3. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 1, characterized in that, The unified feature space quality control module includes a feature mapping unit, a feature vector construction unit, a quality control rule matching unit, and a conflict detection unit; wherein: Feature mapping unit: Used to receive standardized disease data and, according to a preset feature mapping relationship table, map each data field in the standardized disease data to the corresponding feature dimension in a unified feature space, forming a structure-aligned feature mapping result; Feature vector construction unit: used to vectorize the feature mapping results according to the arrangement order of each feature dimension in the unified feature space, and construct a unified feature vector for subsequent quality control analysis; Quality control rule matching unit: used to match the unified feature vector with the predefined disease quality control rule set item by item, and determine the rule constraints and allowed value ranges corresponding to each feature dimension; Conflict detection unit: Based on the quality control rule matching results, it detects feature dimensions in the unified feature vector that violate rule constraints or have logical inconsistencies, identifies the corresponding data conflict items, and generates preliminary quality control opinions for the unified feature vector accordingly.

4. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 3, characterized in that, The disease quality control rule set includes: Feature value range consistency rule: used to limit the allowed value range of each numerical feature dimension in a unified feature vector; Feature type matching rules: used to verify whether the actual data type of each feature dimension in the unified feature vector is consistent with the predefined data type of the corresponding dimension; Required Feature Integrity Rule: This rule is used to constrain the requirement that feature dimensions marked as required in a unified feature vector must have valid values. Logical consistency rules between features: used to constrain the predefined logical relationships between multiple feature dimensions; Time sequence rationality rule: used to verify that the feature dimensions involving time attributes in a unified feature vector meet the preset time sequence relationship; Disease-specific constraint rules: used to confirm that some feature dimensions in a unified feature vector are only allowed to appear under the conditions of the corresponding disease.

5. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 1, characterized in that, The annotation rule parsing unit includes: Rule location subunit: Used to receive preliminary quality control opinions and annotation suggestions, and locate the set of rule entries corresponding to the current data segment in the predefined annotation rules based on the type of data conflict identified by the preliminary quality control opinions and the annotation category to which the annotation suggestions belong, forming a set of rules to be parsed; Task type determination subunit: used to read the task definition items bound to the annotation category in the set of rules to be parsed, determine the annotation task type corresponding to the current data fragment, and output the annotation task type identifier; Constraint Extraction Subunit: Used to read the constraint definition items bound to the data conflict items in the set of rules to be parsed, and extract the labeled constraints corresponding to the current data segment, including the set of mandatory feature dimensions, the set of prohibited label conflicts, and the set of consistency verification items; Structured sub-unit of parsing results: used to encapsulate the annotation task type identifier and annotation constraints in a structured way to generate rule parsing results.

6. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 1, characterized in that, The annotation execution and arbitration module includes an annotation task distribution unit, an annotation result collection unit, a disagreement detection unit, and an arbitration decision unit; wherein: The annotation task distribution unit is used to receive mixed annotation guidance information, bind the mixed annotation guidance information with the corresponding standardized disease data and unified feature vectors, and push it to at least one annotation terminal according to the preset task distribution strategy, so as to organize the annotation personnel to perform manual annotation operations based on annotation suggestions, annotation task prompts and key annotation area instructions; Annotation result acquisition unit: used to receive manual annotation results generated from different annotation terminals for the same data segment, and encapsulate the annotation results in a unified format to form a set of comparable annotation results; Disagreement detection unit: Used to compare the consistency of each annotation result in the annotation result set. When the annotation results for the same data segment are found to be inconsistent in annotation category, annotation range or annotation attribute, it is determined that there is an annotation disagreement and the arbitration process is triggered. Arbitration Decision Unit: After the arbitration process is triggered, it is used to make a comprehensive judgment on the disputed labeling results according to the preset arbitration rules and generate the post-arbitration labeling results.

7. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 6, characterized in that, The arbitration decision-making unit includes: Candidate Result Summarization Subunit: This unit receives all annotation results of data segments marked as conflicting by the divergence detection unit, and summarizes and organizes the results according to the annotation terminal number, annotation category, and annotation attribute field to form a structured candidate annotation result set. Arbitration index calculation subunit: Based on the candidate annotation result set, and combining annotation confidence, historical consistency score, and conflict level, calculate the comprehensive arbitration score for each annotation category. ; Arbitration rule determination subunit: used to sort and filter the comprehensive arbitration scores of each candidate label category according to a preset arbitration rule set, which includes: highest priority rules, threshold filtering rules and task type constraint rules; Arbitration Result Generation Sub-unit: Used to determine the sorting results output by the sub-unit according to the arbitration rules, determine the final arbitration label category, extract the corresponding label attributes and positions, and generate the post-arbitration labeling results.

8. The disease quality control labeling system based on heterogeneous hybrid mode according to claim 1, characterized in that, The integrated output module includes a result correlation unit, a consistency analysis unit, a scoring generation unit, and a result encapsulation and output unit; wherein: Result association unit: Used to receive the post-arbitration annotation results output by the arbitration decision unit, and bind and integrate them with the unified feature vector and standardized disease data corresponding to the same data segment to form a structured data annotation unit entity; Consistency Analysis Unit: Used to compare the annotation results after arbitration with the candidate suggestions and actual annotation terminal results during the annotation process, identify whether there are low consistency situations during the annotation process, extract consistency indicators and cross-validate them with the abnormal items in the preliminary quality control opinions to form a set of consistency evaluation factors; Scoring Generation Unit: Based on the number, type, and severity of conflicting items in the preliminary quality control comments, and the set of consistency assessment factors, a comprehensive quality control score is generated. The overall quality control score is calculated using the following formula: ,in, This indicates the overall quality control score; Indicates the number of conflicting items identified in the preliminary quality control comments; This represents the total number of feature dimensions in the unified feature vector; The consistency analysis score represents the range of... ; Let be the scoring weight coefficient, satisfying ; The result encapsulation output unit is used to encapsulate the arbitration-post-annotation results, unified feature vectors, standardized disease data, and comprehensive quality control scores in a structured manner to form a structured disease quality control annotation result file.

Images