Cross-institution medical project mapping method, program product, equipment and storage medium

Abstract

The invention provides a cross-institution medical project mapping method, a program product, equipment and a storage medium, and the method comprises the steps: representing a source project in a source project set and a standard project in a standard project set as multi-dimensional attribute vectors, and carrying out the value of the multi-dimensional attribute vector of the source project through a standardized mapping function, mapping to a standard value space of a corresponding dimension, and generating a standardized attribute vector; for the source project, determining a candidate standard project set from the standard project set through a multi-retrieval path strategy; determining an effective observation dimension set, and calculating a matching score of each dimension and a comprehensive matching degree; and selecting a target standard item corresponding to the source item according to the comprehensive matching degree, and mapping the source item to the target standard item. High-precision and high-coverage project automatic docking in a highly heterogeneous medical data environment is realized, and the continuity of cross-mechanism diagnosis and treatment information is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a cross-institutional medical project mapping method, program product, device, and storage medium. Background Technology

[0002] With the development of internet-based healthcare, it has become common for patients to seek medical treatment at different medical institutions. However, there are significant differences in the names, codes, testing methods, and reference ranges of laboratory tests and examinations across different institutions, making it difficult to effectively correlate and compare data for the same patient across different institutions. For example, the same test item may be called "complete blood count" or "full blood cell count" in different hospitals, or may use different coding systems such as ICD-10, LOINC, or hospital-defined coding. Even if the test item is the same, it may not be directly matched due to different codes. In addition, key metadata such as testing methods, instrument models, and reference ranges are often missing or inconsistently described, further increasing the difficulty of data integration.

[0003] Currently, cross-institutional medical project coordination mainly relies on manual mapping or rule-based methods based on simple text matching. Manual mapping is inefficient and costly, and struggles to handle large amounts of data and frequent updates; simple text matching is inadequate in handling semantic issues such as naming differences, abbreviations, and synonyms, and cannot effectively utilize structured information in the project (such as sample types, detection methods, etc.), resulting in unsatisfactory mapping accuracy and recall. Summary of the Invention

[0004] The purpose of this application is to provide a cross-institutional medical project mapping method, program product, electronic device and storage medium to improve the above-mentioned technical problems.

[0005] In a first aspect, embodiments of this application provide a cross-institutional medical item mapping method, comprising: representing source items in a source item set and standard items in a standard item set as multi-dimensional attribute vectors, and mapping the values ​​of the multi-dimensional attribute vectors of the source items to the standard value space of the corresponding dimension using a standardized mapping function to generate standardized attribute vectors; wherein, one dimension of the multi-dimensional attribute vector represents a data type; for source items, determining a set of candidate standard items from the set of standard items through a multi-retrieval path strategy; the multi-retrieval path strategy includes at least one of core attribute strong constraint retrieval, auxiliary attribute supplementary retrieval, and text semantic fallback retrieval; determining the effective observation dimension set between the source items and the candidate standard items in the candidate standard item set, calculating the matching score of each dimension in the effective observation dimension set, and calculating the comprehensive matching degree between the source items and the corresponding candidate standard items based on the matching score; the dimensions of the effective observation dimension set include dimensions in the standardized attribute vectors of the source items and the candidate standard items where all values ​​are non-empty; selecting the target standard item corresponding to the source item from the candidate standard item set according to the comprehensive matching degree, and mapping the source item to the target standard item if the comprehensive matching degree meets the requirements.

[0006] In the above implementation process, by representing source items and standard items as multi-dimensional attribute vectors and using a standardized mapping function to map the values ​​of each dimension of the source items to the standard value space, the problem of inconsistent descriptions of item names, codes, and units among different medical institutions is effectively solved, providing a semantically consistent data foundation for subsequent matching. Through a multi-retrieval path strategy, comprehensively utilizing core attribute-based strong constraint retrieval, auxiliary attribute-based supplementary retrieval, and textual semantic fallback retrieval, the reliability of high-precision matching is ensured, while semantic fallback guarantees recall when information is missing, improving adaptability in complex scenarios with varying quality of real-world medical data. Simultaneously, weights are dynamically adjusted based on dimensional coverage, ensuring that comparisons with more comprehensive information contribute more to decision-making, thus improving the fairness and robustness of matching. Finally, hierarchical decision-making is performed based on the overall matching degree, and a dynamic threshold adjustment mechanism based on confidence gaps is introduced, achieving a better balance between mapping accuracy and automated coverage.

[0007] Optionally, in this embodiment, the values ​​of the multidimensional attribute vector of the source project are mapped to the standard value space of the corresponding dimension using a normalization mapping function to generate a normalized attribute vector. This includes: extracting the original values ​​of the multidimensional attribute vector of the source project in each dimension from the original text and / or structured fields of the source project using a predefined extraction function; obtaining the normalization mapping function for each dimension; and mapping the original values ​​to a standard value in a predefined standard value space using the normalization mapping function to generate a normalized attribute vector. In the standard value space, multiple semantically equivalent different original values ​​are mapped to the same standard value.

[0008] In the above implementation process, through predefined extraction functions, the original values ​​of each dimension can be automatically identified and extracted from the original text and structured fields of the source project, transforming the unstructured hospital project description into structured attribute data, providing a clear input object for subsequent standardization processing. Secondly, by configuring an independent standardization mapping function for each dimension, various different original values ​​under the same dimension are uniformly mapped to the same standard value in the standard value space, reducing semantic heterogeneity problems caused by different naming issues. Since multiple semantically equivalent original values ​​are mapped to the same standard value, the subsequent matching process no longer relies on complete consistency of name text, but can identify the correspondence between terms that are "different in expression but the same in meaning," thereby improving the adaptability and generalization ability of the mapping method to the differences in data expression among different medical institutions.

[0009] Optionally, in this embodiment of the application, before mapping the original value to a standard value in a predefined standard value space through the standardization mapping function, the method further includes: obtaining a predefined initial standard value set; updating the initial standard value set through a human-machine collaborative feedback mechanism to generate a standard value space; and adjusting the prior weights in the standardization mapping function that map the original value to the standard value.

[0010] In the above implementation process, by obtaining a predefined set of initial standard values, the standard value space is constructed based on authoritative industry standards and medical knowledge systems, ensuring that the standardized mapping function possesses high clinical rationality and professional accuracy from the initial stage. By introducing a human-machine collaborative feedback mechanism, the correct mapping results confirmed by manual review are used as feedback data to continuously update and expand the initial set of standard values, enabling the standard value space to cover new terms and expressions appearing in real medical institution data. The feedback mechanism can also dynamically adjust the prior weights of polysemous terms in the standardized mapping function, ensuring that the mapping results more closely match the actual distribution when encountering the same terms subsequently.

[0011] Optionally, in the embodiments of this application, the conditions for determining the execution order of different strategies in the multi-retrieval path strategy include at least one of the following: the completeness of information of the core attributes or auxiliary attributes of the source project, whether the number of candidate results returned by the strong constraint of the core attributes meets the preset threshold, and the emphasis requirements of the current business scenario on mapping accuracy or recall rate.

[0012] In the aforementioned implementation process, by introducing three types of dynamic configuration conditions—information integrity, the number of candidate results, and business scenario requirements—the execution order and activation decision of the multi-retrieval path strategy are no longer rigidly fixed, but can be adaptively adjusted according to the actual data quality and application goals of the source projects. This dynamically configurable strategy execution mechanism enables the mapping method to achieve a reasonable balance between computational efficiency, matching accuracy, and recall coverage when facing complex scenarios with different medical institutions, different data qualities, and different application goals, thereby improving the overall adaptability and practical value of cross-institutional project mapping.

[0013] Optionally, in this embodiment, the multi-retrieval path strategy includes core attribute strong constraint retrieval; determining a candidate standard item set from the standard item set through the multi-retrieval path strategy includes: if at least one auxiliary attribute in the core attribute set of the source item has a non-empty value in the standardized attribute vector, enabling core attribute strong constraint retrieval; extracting a first standardized value belonging to the core attribute set from the standardized attribute vector of the source item; retrieving a first standard item in the standard item set whose standardized attribute vector value in the same core attribute set is consistent with the first standardized value; merging all first standard items into a subset corresponding to the core attribute strong constraint retrieval path in the candidate standard item set.

[0014] In the above implementation process, the clearly defined enabling and disabling logic for core attributes ensures that the core attribute-strongly constrained retrieval path is activated only when the source item possesses valid core information. This reduces invalid searches or unreliable candidate results when core information is completely missing. The core attribute-strongly constrained retrieval employs a strict exact matching strategy, requiring that the standardized values ​​of the source item and the standard item be completely identical on the same set of core attributes. This guarantees a high degree of certainty between the recalled first standard item and the source item at the core semantic level. By independently merging the search results into subsets corresponding to the core attribute-strongly constrained retrieval path, the system can clearly trace the source evidence for each candidate item, providing a more reliable basic candidate set for subsequent comprehensive matching degree calculations.

[0015] Optionally, in this embodiment, the multi-retrieval path strategy includes auxiliary attribute supplementary retrieval; determining a candidate standard item set from the standard item set through the multi-retrieval path strategy includes: if the candidate results returned by the core attribute strong constraint retrieval are empty or the number is lower than a first preset threshold, and at least one auxiliary attribute in the auxiliary attribute set of the source item has a non-empty value in the standardized attribute vector, then auxiliary attribute supplementary retrieval is enabled: extracting a second standardized value belonging to the auxiliary attribute set from the standardized attribute vector of the source item; for the identified auxiliary attribute, retrieving a second standard item in the standard item set whose standardized attribute vector has the same value in the same auxiliary attribute set as the second standardized value; merging all second standard items into a subset corresponding to the auxiliary attribute supplementary retrieval path in the candidate standard item set; the method further includes: disabling auxiliary attribute supplementary retrieval when all auxiliary attribute values ​​in the standardized attribute vector are missing.

[0016] In the above implementation process, by setting explicit activation conditions and disabling logic, auxiliary attribute supplementary retrieval is only initiated when the core path recall is insufficient and auxiliary information is available, reducing computational overhead. When the core retrieval cannot provide enough candidate items, auxiliary attribute supplementary retrieval uses highly specific auxiliary identifiers such as English abbreviations for precise matching, which can quickly recall standard items with the same auxiliary identifiers as the source item. Since auxiliary identifiers usually have high uniqueness, they can directly locate unique or a few candidate items, thereby effectively compensating for the missed recall problem caused by missing core path information or insufficient retrieval. By independently merging the retrieval results into subsets corresponding to the auxiliary attribute supplementary retrieval path, the system can clearly distinguish the evidence sources of different paths, providing a reliable supplementary candidate set for subsequent comprehensive matching degree calculation. This design significantly improves the adaptability of the mapping method to incomplete data and the overall recall rate while ensuring high-precision retrieval priority.

[0017] Optionally, in this embodiment, the multi-retrieval path strategy includes text semantic fallback retrieval; the multi-retrieval path strategy determines a candidate standard item set from the standard item set, including: if the candidate standard item set determined by strong constraint retrieval of core attributes and supplementary retrieval of auxiliary attributes is empty, the source item contains preset characters, or the source item belongs to a preset business scenario, then the text semantic fallback retrieval is initiated: the first complete text name of the source item, the second complete text name of the standard item, and the synonym set of the second complete text name are obtained; based on the text similarity calculation model, the semantic similarity between the first complete text name and the second complete text name, and between the first complete text name and the elements in the synonym set are calculated; a third standard item with a semantic similarity exceeding a preset threshold, and / or a fourth standard item whose synonym set of the first complete text name satisfies the keyword overlap condition are selected, forming a subset corresponding to the text semantic fallback retrieval path in the candidate standard item set.

[0018] In the above implementation process, by using textual semantic fallback retrieval, generalized recall can be achieved solely based on the item name text when neither the core nor auxiliary paths can provide effective candidates. This path does not rely on the completeness of any structured attributes and can cover situations where core information is completely missing, naming is highly irregular, or new terms have emerged, significantly improving the mapping method's adaptability to low-quality data. By introducing a synonym set and semantic similarity calculation, this path can identify item correspondences that differ in literal expression but have the same clinical semantics, effectively alleviating the missed recall problem caused by terminology diversity and abbreviation ambiguity.

[0019] Optionally, in this embodiment of the application, calculating the matching score of each dimension in the effective observation dimension set includes: for each dimension in the effective observation dimension set: if the standardized value of the source item is the same as the standardized value of the candidate standard item, then the score of the dimension is determined as a first value; if the standardized value of the source item and the standardized value of the candidate standard item have a predefined semantic compatibility relationship, then the score of the dimension is determined as a constant between the second value and the first value; if the standardized value of the source item and the standardized value of the candidate standard item have a predefined conflict relationship, then the score of the dimension is a third value.

[0020] In the above implementation process, by setting fine-grained matching scoring rules for each dimension, the relationship between the source item and the candidate standard item in terms of dimension values ​​is distinguished into three states: completely identical, semantically compatible, and obviously conflicting. These are assigned a first value, a constant between the first and second values, and a third value, respectively. This design ensures that the dimension matching score is no longer limited to simple binary agreement or inconsistency, but accurately reflects the degree of clinical semantic compatibility and conflict. The fine-grained scoring mechanism improves the discriminative power of the dimension matching score, enhancing the accuracy of the overall matching degree calculation and the reliability of the decision.

[0021] Optionally, in this embodiment of the application, if the dimension is the English abbreviation dimension, the process of calculating the matching score of the English abbreviation dimension includes: determining whether the English abbreviation in the source project is consistent with the English abbreviation in the candidate standard project; if consistent, determining the score of the English abbreviation dimension as the first value; if inconsistent, calculating the string similarity between the English abbreviation in the source project and the English abbreviation in the candidate standard project, and using the string similarity as the score of the English abbreviation dimension.

[0022] In the aforementioned implementation process, English abbreviations are relatively stable and highly unique in medical institution information systems, serving as important semantic clues for cross-institutional project mapping. However, different institutions may have subtle differences in the English abbreviations of the same project, such as capitalization, abbreviation format, and separators. This step prioritizes determining whether the standard abbreviations are consistent, assigning a first value when they are consistent, thus fully preserving the value of this dimension as high-confidence evidence. This allows the English abbreviation dimension to play a decisive role when the expressions are completely consistent, while also preserving positive matching evidence when there are differences in writing habits. By increasing the tolerance for the diversity of abbreviation expressions, the mapping method's adaptability to differences in real medical institution data can be improved without sacrificing accuracy, thereby increasing the overall mapping accuracy.

[0023] Optionally, in this embodiment, calculating the comprehensive matching degree between the source item and the corresponding candidate standard item based on the matching score includes: taking a weighted average of the matching scores of all dimensions in the effective observation dimension set according to the preset basic weights of the dimensions to obtain the dimension matching score; obtaining the dynamic effective weight of the dimension matching score, the dynamic effective weight being determined based on the preset basic weights and the dimension coverage rate; the dimension coverage rate being the ratio of the number of dimensions in the effective observation dimension set to the total number of dimensions; weightedly fusing the dimension matching score and its dynamic effective weight with the matching score of at least one other dimension to obtain the comprehensive matching degree; the text similarity score is calculated based on the residual text, the residual text being determined by removing identified terms belonging to the predefined clinical dimensions from the normalized names of the source item and the candidate standard item.

[0024] In the above implementation process, the dynamic effective weights based on dimensional coverage adaptively correlate the contribution of the dimensional matching module to the overall matching degree with the information completeness between the source project and the candidate standard project. When the number of dimensions in the effective observation dimensional set is large and the dimensional coverage is high, it indicates that there is sufficient structured information comparable between the two parties, the evidentiary strength of the dimensional matching score is high, and its dynamic effective weight is correspondingly increased. Conversely, when there are few comparable dimensions, the weight of the dimensional matching module automatically decreases, preventing its score from unduly affecting the overall matching degree due to incomplete information. This improves the fairness, robustness, and discriminative power of the overall matching degree under different data quality conditions, thereby improving the overall accuracy of cross-institutional project mapping.

[0025] Optionally, in this embodiment of the application, selecting the target standard project corresponding to the source project from the candidate standard project set based on the comprehensive matching degree includes: obtaining the candidate standard project with the highest comprehensive matching degree with the source project from the candidate standard project set as the first-choice candidate; if the difference between the comprehensive matching degree of the first-choice candidate and the comprehensive matching degree of the second-best candidate is less than a preset conflict threshold, then arbitrating the first-choice candidate and the second-best candidate according to a predefined clinical priority rule base to determine the target standard project.

[0026] In the above implementation process, by comparing the difference in the overall matching degree between the preferred candidate and the second-best candidate, arbitration is initiated from the clinical priority rule base when their scores are close, avoiding the decision-making difficulties when scores alone cannot distinguish between them. This design transforms clinical experience such as methodological priority and name specificity into executable judgment logic, determining the target standard item based on domain knowledge when score evidence is insufficient, significantly improving the clinical rationality and accuracy of the mapping results.

[0027] Optionally, in this embodiment, the conditions under which the overall matching degree meets the requirements include: the overall matching degree is greater than or equal to a preset reliability threshold; or, the overall matching degree is less than a preset reliability threshold and greater than or equal to a dynamic decision threshold; wherein, the dynamic decision threshold is determined based on the difference between the overall matching degree of the preferred candidate and the overall matching degree of the second-best candidate.

[0028] In the above implementation process, by introducing a dynamic decision threshold based on the difference between the scores of the preferred candidate and the second-best candidate, the judgment of whether the overall matching degree meets the requirements no longer relies on a single fixed threshold. When the preferred candidate has a significant relative advantage, it can still meet the automatic mapping conditions even if its absolute score is slightly lower. This design effectively avoids missing correct matches due to slightly lower absolute scores while ensuring high confidence, thereby improving the coverage of automatic mapping while maintaining high accuracy.

[0029] Optionally, in this embodiment of the application, if the overall matching degree of the source project does not meet the requirements, an audit task is generated based on the matching information between the source project and the candidate standard project set; using the correct mapping pairs confirmed by the audit task, the weight parameters in the calculation of the overall matching degree are optimized by a ranking learning algorithm.

[0030] In the above implementation process, the correctly mapped pairs confirmed by manual review are used as training data. The weight parameters in the comprehensive matching degree calculation are optimized through a ranking learning algorithm, so that the weights of each dimension and module are more in line with the actual data distribution. The human-machine collaborative closed loop enables the system to continuously learn from the results of manual review, constantly improving the accuracy of the comprehensive matching degree calculation, thereby effectively improving the overall mapping accuracy of cross-institutional projects in subsequent mappings.

[0031] Secondly, embodiments of this application also provide a computer program product, including computer program instructions, which are executed by a processor to perform the method provided in the first aspect or any implementation thereof.

[0032] Thirdly, embodiments of this application also provide an electronic device, including: a processor and a memory, the memory storing computer program instructions, which are executed by the processor to perform the method provided in the first aspect or any implementation thereof.

[0033] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, perform the method provided in the first aspect or any implementation thereof.

[0034] This application employs a cross-institutional medical item mapping method, program product, electronic device, and storage medium. By representing source items and standard items as multi-dimensional attribute vectors and utilizing a standardized mapping function to map the values ​​of each dimension of the source items to the standard value space, it effectively solves the problem of inconsistent descriptions of item names, codes, and units among different medical institutions, providing a semantically consistent data foundation for subsequent matching. Through a multi-retrieval path strategy, it comprehensively utilizes core attribute-based strong constraint retrieval, auxiliary attribute-based supplementary retrieval, and textual semantic fallback retrieval. This ensures both the reliability of high-precision matching and the recall rate through semantic fallback when information is missing, improving adaptability in complex scenarios with varying real-world medical data quality. By introducing an effective observation dimension set and calculating matching scores only on dimensions where all values ​​are non-empty, the scores of item pairs with different levels of information completeness are comparable. Simultaneously, by dynamically adjusting weights based on dimension coverage, the more complete the information in the comparison, the higher its contribution to decision-making, improving the fairness and robustness of the matching. Finally, a hierarchical decision is made based on the overall matching degree, and a dynamic threshold adjustment mechanism based on confidence gap is introduced, thereby achieving a better balance between mapping accuracy and automated coverage. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating a cross-institutional healthcare project mapping method provided in this application embodiment; Figure 2This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0037] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this application.

[0039] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0040] Please see Figure 1 The illustrated diagram shows a flowchart of a cross-institutional healthcare project mapping method provided in an embodiment of this application. The cross-institutional healthcare project mapping method provided in this application embodiment can be applied to electronic devices, which may include physical devices such as servers, PCs, tablets, or smartphones, or virtual devices such as virtual machines or containers. The electronic device can be a single device, a combination of multiple devices, or a cluster of a large number of devices. The cross-institutional healthcare project mapping method may include: Step S110: Represent the source items in the source item set and the standard items in the standard item set as multi-dimensional attribute vectors, and use a normalization mapping function to map the values ​​of the multi-dimensional attribute vectors of the source items to the standard value space of the corresponding dimension to generate a normalized attribute vector; wherein, one dimension of the multi-dimensional attribute vector represents a data type.

[0041] Step S120: For the source items, determine the candidate standard item set from the standard item set through a multi-retrieval path strategy; the multi-retrieval path strategy includes at least one of the following: core attribute strong constraint retrieval, auxiliary attribute supplementary retrieval, and text semantic fallback retrieval.

[0042] Step S130: Determine the effective observation dimension set between the source project and the candidate standard project in the candidate standard project set, calculate the matching score of each dimension in the effective observation dimension set, and calculate the comprehensive matching degree between the source project and the corresponding candidate standard project based on the matching score; the dimensions of the effective observation dimension set include the dimensions in the standardized attribute vectors of the source project and the candidate standard project, where all values ​​are non-empty.

[0043] Step S140: Based on the overall matching degree, select the target standard project corresponding to the source project from the candidate standard project set, and if the overall matching degree meets the requirements, map the source project to the target standard project.

[0044] In step S110: The source item set refers to the set of test and examination items from different medical institutions to be mapped. The standard item set refers to a pre-constructed reference item library with standardized descriptions and authoritative codes, serving as the target benchmark for mapping. A multidimensional attribute vector is a vector form representing a medical item as values ​​across multiple attribute dimensions. Each dimension represents a data type, such as categorical (test components, sample type), numerical / interval (reference range), text (full item name), or identifier / code (national standard code). The standardized mapping function is a function pre-defined for each attribute dimension, used to map the original values ​​to a standard value in a predefined standard value space. The standard value space is the set of all standard values ​​under that dimension, where multiple semantically equivalent original values ​​are mapped to the same standard value. The standardized attribute vector is an attribute vector whose values ​​have been normalized after processing by the mapping function.

[0045] For each source item in the source item set, a predefined extraction function is first used to extract its original values ​​for each attribute dimension from the item's original text (such as item name and remarks) and other structured fields (such as independent English abbreviation fields and sample type fields). Then, for each attribute dimension, the corresponding normalization mapping function is called to map the original values ​​of that dimension to standard values ​​in the standard value space of that dimension. For example, for the "detected component" dimension, the original values ​​"blood sugar" and "Glu" are both mapped to the standard value "glucose". After the above processing, the values ​​of each dimension in the multidimensional attribute vector of the source item are replaced with standard values, forming a normalized attribute vector. For standard items in the standard item set, which are already standardized representations, they can be directly represented as multidimensional attribute vectors or mapped in the same way to a unified format.

[0046] In step S120: The multi-search path strategy refers to a search strategy that flexibly combines search methods with different logics, selecting one or more paths to execute based on data quality and business needs. The candidate standard item set is a collection of all potentially matching standard items recalled from the standard item set through the above search paths for a single source item. Strong constraint retrieval based on core attributes refers to a search method that performs precise matching based on core attributes, ensuring high accuracy. Supplementary retrieval based on auxiliary attributes refers to a search method that performs precise matching based on auxiliary attributes, used to quickly narrow down the candidate range. Textual semantic fallback retrieval refers to a search method that performs generalized recall based on text similarity, used to ensure a high recall rate.

[0047] First, the standardized attribute vectors of the source items are obtained. In the core attribute strong constraint retrieval, the standardized values ​​of the source items on the core attribute set are extracted, and all standard items with identical standardized values ​​on the core attribute are retrieved from the standard item set. These items are then included in the candidate standard item set. In the auxiliary attribute supplementary retrieval, the non-empty standardized values ​​of the source items on the auxiliary attribute set are identified. For each identified auxiliary attribute, standard items with completely identical values ​​are independently retrieved from the standard item set, and all retrieval results are merged and included in the candidate set. In the text semantic fallback retrieval, the system obtains the complete text name of the source item, calculates its semantic similarity with the complete text names of each standard item and their synonym sets, and filters out standard items with similarity exceeding a preset threshold or containing overlapping core keywords in their names, including them in the candidate set. These three paths can be selected and activated according to preset strategies or dynamic conditions, and after merging and deduplication, the final candidate standard item set is formed.

[0048] In step S130: The effective observation dimension set is a subset selected for each pair of source items and candidate standard items, where all dimensions in the standardized attribute vectors of both sides are not empty, i.e., all have valid values. The matching score is a numerical value used to quantify the consistency strength between the source items and candidate standard items on a single dimension. The overall matching degree is a single numerical value used to quantify the overall similarity between the source items and candidate standard items after comprehensively considering multiple dimensions and other matching evidence.

[0049] For each candidate standard item in the candidate standard item set, the system first determines a pre-defined set of extended dimensions. This set contains core clinical dimensions used to calculate the dimension matching score, such as detection components, sample type, sampling time point, and English abbreviations. For each dimension in the extended dimension set, the system calculates the single-dimensional matching score δd(q,s) based on its type (text, numerical / range, categorical, or identifier / code). For the English abbreviation dimension, a strategy of "standard matching first, string similarity as a fallback" is adopted: first, both abbreviations are mapped to a unified standard abbreviation space; if the standard abbreviations match, the score is 1; if they do not match but both values ​​are not empty, the similarity of the original strings is calculated as the score for that dimension, thus avoiding excessive penalties due to differences in abbreviation spelling.

[0050] After obtaining the individual dimension scores for each dimension in the extended dimension set, the system further filters out dimensions from this set where both the source item and the candidate standard item have non-empty values, forming the effective observation dimension set (Dobs). The system only performs a weighted average of the corresponding individual dimension scores within the effective observation dimension set, using the preset base weights of each dimension as coefficients to calculate the dimension matching score (Sdim). This design ensures that missing dimensions are excluded from the calculation, avoiding an unreasonable dilution effect on the score due to missing information, while also making scores under different levels of information completeness comparable.

[0051] The system assigns a dynamic effective weight Wdim to the dimension matching module. This weight is determined by the product of the module's preset base weight ωdim and the dimension coverage rate ρ, where ρ is the ratio of the number of dimensions in the effective observation dimension set to the total number of dimensions in the expanded dimension set. A higher dimension coverage rate indicates more available structured information, and the contribution weight of the dimension matching module in the final decision-making process increases accordingly.

[0052] Finally, the system dynamically weights and merges the dimensional matching scores and their dynamic effective weights with the matching scores of at least one other dimension through multi-module dynamic weighting to obtain a comprehensive matching degree. These other dimensions include detection methods, instruments, reference ranges, and project names. The matching scores for these dimensions are calculated according to their respective types using the methods described in the preceding sections, and each has its own independent base weight, effectiveness indicator function, and dynamic weight factor. The fusion process employs a weighted average formula based on the effectiveness indicator function and dynamic weights, ensuring that only modules with available information participate in the calculation, and that modules with more comprehensive information have higher weights, thereby obtaining a robust and fair comprehensive matching degree.

[0053] In step S140: the target standard item is the standard item selected from the set of candidate standard items and determined to be the best match for the source item. Mapping refers to establishing a correspondence between the source item and the target standard item, completing cross-organizational terminology alignment.

[0054] For a source project, its overall matching degree is compared with that of all candidate projects in the candidate standard project set. The candidate standard project with the highest overall matching degree is selected as the first-choice candidate and identified as a potential target standard project. Subsequently, the system determines whether the overall matching degree meets the automatic mapping requirements. The judgment criteria include two scenarios: the first scenario is that the overall matching degree is greater than or equal to a preset high-confidence threshold, in which case it is directly judged as meeting the requirements; the second scenario is that although the overall matching degree is lower than the high-confidence threshold, the difference between the overall matching degree of the first-choice candidate and the second-best candidate is significant. Based on this difference, a dynamic decision threshold lower than the preset threshold is dynamically calculated. If the overall matching degree of the first-choice candidate is still greater than or equal to the dynamic decision threshold, it is also judged as meeting the requirements. This dynamic threshold mechanism allows for an appropriate reduction in the absolute score threshold when the first-choice candidate has a significant relative advantage. If the requirements are met, the system automatically maps the source project to the selected target standard project; if the requirements are not met, it enters the manual review process or is temporarily left unprocessed.

[0055] In the implementation of the above embodiments: by representing source items and standard items as multi-dimensional attribute vectors, and using a standardized mapping function to map the values ​​of each dimension of the source items to the standard value space, the problem of inconsistent descriptions of item names, codes, and units among different medical institutions is effectively solved, providing a semantically consistent data foundation for subsequent matching. Through a multi-retrieval path strategy, comprehensively utilizing core attribute-based strong constraint retrieval, auxiliary attribute-based supplementary retrieval, and textual semantic fallback retrieval, the reliability of high-precision matching is ensured, and recall is guaranteed through semantic fallback when information is missing, improving adaptability in complex scenarios where the quality of real medical data varies greatly. By introducing an effective observation dimension set and calculating matching scores only on dimensions where all values ​​are non-empty, the scores of item pairs with different information completeness are comparable; simultaneously, weights are dynamically adjusted based on dimension coverage, so that comparisons with more complete information contribute more to decision-making, improving the fairness and robustness of matching. Finally, hierarchical decision-making is performed based on the comprehensive matching degree, and a dynamic threshold adjustment mechanism based on confidence gaps is introduced, thereby achieving a better balance between mapping accuracy and automated coverage.

[0056] To address the semantic heterogeneity issue of cross-institutional projects, this application first presents a structured, multi-dimensional representation of the projects and standardizes the attributes of each dimension, laying the foundation for subsequent semantic similarity calculation and automatic mapping.

[0057] Based on the data type of the attributes, this application divides all dimensions into four categories: categorical attributes (such as detected components, sample type, detection method, and instrument), numerical / range attributes (such as reference range), textual attributes (such as the complete Chinese name of the project), and identifier / code attributes (such as national standard code). Different methods can be used to calculate the matching scores for different types of attributes.

[0058] Let Q be the set of source projects and S be the set of standard projects. For any project... This can be represented as an n-dimensional attribute vector:

[0059] in, Indicates project The value of the d-th attribute dimension has a range of values. Let the set of dimension indices for all attributes be . This means that the number of our dimensions is an arbitrary value of n. To facilitate the handling of different types of data, the attribute dimensions in the multidimensional attribute vector can be divided according to their data type, which can be divided into the following data types: (1) Set of categorical attributes Such as the components to be detected, the sample type, and the detection method.

[0060] (2) Numerical / range type attribute set Examples include male reference range (ref_range_male) and female reference range (ref_range_female).

[0061] (3) Text-type attribute set : For example, the complete Chinese name of the project (cn_name).

[0062] (4) Identifier / code type attribute set Examples include national standard codes (nat_code) and internal standard IDs (std_id).

[0063] After generating the multidimensional attribute vectors corresponding to the source and standard items, attribute standardization is performed based on extraction and mapping. Optionally, in this embodiment, the values ​​of the multidimensional attribute vectors of the source items are mapped to the standard value space of the corresponding dimension using a standardization mapping function to generate standardized attribute vectors, including: Using a predefined extraction function, the original values ​​of the multidimensional attribute vector of the source item are extracted from the original text and / or structured fields of the source item.

[0064] Extracting structured attribute values ​​from raw data is a crucial step. For example, a set of extraction functions can be defined. , ,in, NULL indicates that the value is missing.

[0065] For any project By using methods such as rule and dictionary matching, the original values ​​of each dimension are extracted from the original text (such as the project name) and other structured auxiliary fields. This process enables the fusion and extraction of information from multiple sources.

[0066] Obtain the standardized mapping function for each dimension, and use the standardized mapping function to map the original values ​​to a standard value in a predefined standard value space to generate a standardized attribute vector; wherein, in the standard value space, multiple different original values ​​that are semantically equivalent are mapped to the same standard value.

[0067] Because different organizations may use synonyms, abbreviations, or alternative names to describe the same attribute, it is necessary to normalize the original attribute values. Therefore, for each attribute dimension... Define a normalized mapping function The format can be:

[0068] in, It is a project The range of values ​​for the d-th attribute dimension This is a predefined or learned standard value space for this dimension, within which semantically equivalent values ​​are mapped to a unified standard value. For any item x, its standardized attribute vector is denoted as... The calculation formula can be:

[0069] in, Through the above steps, projects from different institutions with varying descriptions are transformed into multi-dimensional feature vectors with consistent structure and unified semantics, providing a standardized data foundation for the next step of cross-institutional project semantic similarity calculation and accurate matching.

[0070] In the implementation of the above embodiments: Through predefined extraction functions, the original values ​​of each dimension can be automatically identified and extracted from the original text and structured fields of the source project, transforming the unstructured hospital project description into structured attribute data, providing a clear input object for subsequent standardization processing. Secondly, by configuring an independent standardization mapping function for each dimension, various different original values ​​under the same dimension are uniformly mapped to the same standard value in the standard value space, reducing semantic heterogeneity problems caused by different naming issues. Since multiple semantically equivalent original values ​​are mapped to the same standard value, the subsequent matching process no longer relies on complete consistency of name text, but can identify the correspondence between terms that are "different in expression but the same in meaning," thereby improving the adaptability and generalization ability of the mapping method to the differences in data expression among different medical institutions.

[0071] Optionally, in this embodiment, before mapping the original value to a standard value within a predefined standard value space using a normalization mapping function, it is necessary to first construct the standard value space. The method includes the following steps: Obtain a predefined set of initial standard values; update the set of initial standard values ​​through a human-machine collaborative feedback mechanism to generate a standard value space, and adjust the prior weights in the standardization mapping function that map the original values ​​to the standard values.

[0072] This application defines a standard value space for each attribute dimension d (such as detected component, sample type). This is not static but follows a dynamic construction principle of "based on authoritative predefined definitions and continuously optimized by data." In addition to the dimensional standard value space, to ensure stable association and traceable referencing of standard items across versions and organizations, this study further constructs a project-level persistent identifier (Std_PID) for each standard item in the standard library. This identifier uses grouped structured encoding and includes check bits, giving it uniqueness, readability, and automatic verification capabilities. This identifier can be decoupled from the project name and external encoding, and after allocation, it can be permanently stored, remaining unchanged when the standard item's attributes are revised, thus ensuring the continuity of cross-version referencing. The standard value space construction and evolution mechanism is as follows: 1. Authoritative predefined initial space: The seed set (i.e., initial standard values) of the standard value space is derived from authoritative industry standards and medical knowledge systems to ensure its clinical rationality and broad acceptance. Specifically: Core source: Based on the provincial and national medical service price item coding standards, standardized descriptions of test item components, sample types, etc., are extracted. (For example, when reproducing the code, a standard item library with multiple data entries was built based on the "Sichuan Province Medical Service Price Item Standard" PDF.) Knowledge Supplement: The above standards are supplemented and improved by referencing authoritative medical textbooks and industry guidelines to ensure the accuracy and completeness of the standard values ​​in a medical sense. This is because a single benchmark may not provide reference standard values ​​for all dimensions; therefore, other authoritative medical textbooks and industry guidelines, such as the "National Clinical Laboratory Operation Procedures," can be used to supplement the standard library.

[0073] By integrating these authoritative sources, a comprehensive and clearly defined initial set of standard values ​​is established for each dimension. For example, the standard value space for the component dimension. Initial values ​​include "hemoglobin", "glucose", "white blood cell count", etc.; initial values ​​for the sample dimension include "serum", "plasma", "venous whole blood", etc.

[0074] 2. Continuous learning and optimization based on human-machine collaborative feedback: The initial standard space may be insufficient to handle the complexity and diversity of real-world data. Therefore, a standard value space is needed. It is a scalable, open set whose optimization process can be implemented within a human-machine collaborative closed loop: Introduction of new standard values: In the manual review process, when inspection experts confirm a recurring value that cannot be mapped to the existing standard value space... When a hospital's description (e.g., a new test component or sample type) truly represents a new clinical concept, the value can be formally added as a standard value after expert review, thereby expanding the standard value range. .

[0075] Optimization of mapping relationships: For terms with ambiguity, map them to a standard value space. The prior weights of multiple candidate standard values ​​are not fixed but can be dynamically adjusted based on historical data from manual verification. For example, if historical data shows that "sugar" is identified as "glucose" rather than "urine sugar" in 90% of cases, then its weight in the mapping function... The prior weights will be adjusted accordingly to make the matching results more consistent with the probability distribution in actual applications.

[0076] The standardized mapping function in this application It is not a static lookup table, but rather a core knowledge base (standard value space). This involves an intelligent module (with mapping rules) that can continuously iterate through a human-machine collaborative feedback mechanism. This design enables the system to possess both initial accuracy based on authoritative standards and strong adaptability to cope with the dynamic evolution of real-world data.

[0077] 3. Standard Project Persistent ID (Std_PID) Construction Mechanism: To support stable referencing of the standard library during long-term maintenance, cross-version evolution, and cross-system association, this study provides a reference for each standard item in the standard project library. An additional project-level persistent identifier, Std_PID, is allocated as a "primary key reference identifier" for the standard project. This identifier does not depend on the project name, organization code, or external specification version, thereby reducing the risk of association breakage due to standard changes.

[0078] The core design principle of persistent identifiers lies in: employing structured numbering and achieving verifiability through a last-digit checksum, thereby reducing the error rate during manual entry / transmission; while emphasizing the decoupling of identifiers from names and their long-term stable referenceability. The Std_PID constructed for the standard project in this application possesses the characteristics of clear structure, high readability, checksum, and stable referenceability.

[0079] (2) Encoding structure of Std_PID: To balance readability and consistency, Std_PID is displayed in a grouped format (e.g., 4-4-4-4 grouping) and includes a check bit. This check bit is calculated based on a preset check algorithm and can automatically verify the overall validity of the identifier, ensuring accuracy in input, import, and interface transmission. For example, the check bit can be "0-9" or "X (representing 10)". The check bit of Std_PID in this study also adopts a similar check mechanism to achieve automatic verification capability.

[0080] (3) Std_PID generation rules (mechanism description): For each standard item The Std_PID generation process can be described as follows: ① Subject number allocation: Assign a unique subject sequence to the standard project (which can be generated by the standard library auto-incrementing sequence number or other internally unique generators) and perform fixed-length processing (such as padding with zeros on the left to uniform length).

[0081] ② Check digit calculation: Calculate the last check character based on the main number, so that Std_PID has consistency verification capability.

[0082] ③ Solidification and Persistence: Once assigned, Std_PID is permanently saved and remains stable during the maintenance and release of the standard library; if the standard project undergoes attribute revisions or name minor adjustments, Std_PID will not change accordingly to ensure the continuity of cross-version references.

[0083] (4) Compatibility with existing provincial standards / internal numbers: If the existing standard library already has a number derived from the provincial standard specification (such as the original std_id), it can be retained as a historical traceability field (such as legacy_std_id), while Std_PID is used as the main reference identifier; the two can establish a corresponding relationship through a mapping table to support historical data backtracking and smooth migration.

[0084] (5) Advantages of introducing Std_PID: Stability and Decoupling: Decoupling of identifiers from names / organization codes / external specification versions reduces the risk of "association breakage" during standard library iterations and supports long-term stable references.

[0085] Verifiability and low error rate: Through the last check bit, consistency verification can be performed during data entry, import, and interface transmission, reducing errors caused by manual copying and system transmission.

[0086] Readability and engineering implementation friendliness: Structured grouping and numbering facilitate manual verification, log auditing and cross-system investigation; at the same time, the retention of historical numbering allows for gradual migration without damaging existing engineering and data assets.

[0087] Governability and Traceability: As an "anchor point" for standard projects, Std_PID is beneficial for recording version changes, mapping evidence and review history in the subsequent human-machine collaborative closed loop, thereby improving the interpretability and maintainability of standard library governance.

[0088] In the implementation of the above embodiments: by obtaining a predefined initial set of standard values, the standard value space is constructed based on authoritative industry standards and medical knowledge systems, ensuring that the standardized mapping function possesses high clinical rationality and professional accuracy from the initial stage. By introducing a human-machine collaborative feedback mechanism, the correct mapping results confirmed by manual review are used as feedback data to continuously update and expand the initial set of standard values, enabling the standard value space to cover new terms and expressions appearing in real medical institution data. The feedback mechanism can also dynamically adjust the prior weights of polysemous terms in the standardized mapping function, making the mapping results more consistent with the actual distribution when encountering the same terms subsequently.

[0089] Optionally, in this embodiment of the application, the conditions for determining the execution order of different strategies in the multi-retrieval path strategy include: The requirements include at least one of the following: the completeness of information on the core or auxiliary attributes of the source project; whether the number of candidate results returned by the strong constraint retrieval of the core attributes meets the preset threshold; and the emphasis of the current business scenario on mapping accuracy or recall rate.

[0090] The core attributes of a source project refer to the set of dimensions used to determine the core semantics of the project, such as the detected components and sample type. Their information completeness is reflected in whether these attributes have non-empty values ​​in the standardized attribute vector. The auxiliary attributes of a source project refer to highly unique identifying dimensions such as English abbreviations; their information completeness is also reflected in whether they have non-empty values.

[0091] The number of candidate results returned by strong constraints on core attributes refers to the number of standard items recalled after precise matching through core dimensions such as component detection. Preset thresholds are pre-defined limits set by the system to determine the sufficiency of candidate results, such as a minimum effective result threshold or a sufficient result threshold. The emphasis on mapping accuracy or recall in different business scenarios refers to the varying requirements for error tolerance and coverage of matching results for different application objectives; for example, clinical decision-making scenarios require high accuracy, while data inventory scenarios require high recall.

[0092] In the implementation of the above embodiments, by introducing three types of dynamic configuration conditions—information integrity, number of candidate results, and business scenario requirements—the execution order and activation decision of the multi-retrieval path strategy are no longer rigidly fixed, but can be adaptively adjusted according to the actual data quality and application goals of the source projects. This dynamically configurable strategy execution mechanism enables the mapping method to achieve a reasonable balance between computational efficiency, matching accuracy, and recall coverage when facing complex scenarios with different medical institutions, different data qualities, and different application goals, thereby improving the overall adaptability and practical value of cross-institutional project mapping.

[0093] The following examples illustrate the triggering conditions for dynamically enabling, disabling, or adjusting core attribute-based strong constraint retrieval, auxiliary attribute-based supplementary retrieval, and text semantic fallback retrieval, as well as the specific implementation methods of the retrieval.

[0094] Optionally, in this embodiment of the application, the retrieval path strategy includes core attribute strong constraint retrieval. The conditions for enabling and disabling core attribute strong constraint retrieval will be introduced first.

[0095] If at least one core attribute in the source project's core attribute set has a non-empty value in the standardized attribute vector, then strong constraint retrieval of core attributes is enabled; if the values ​​of all core attributes in the standardized attribute vector are missing, then strong constraint retrieval of core attributes is disabled.

[0096] Example: Enabling conditions for strong constraint retrieval of core attributes: Enabled by default. When the core attribute set of the source project q... In the text, at least one attribute d has a standardized value. This path is activated when it is not empty.

[0097] Disallowed conditions for strong constraint retrieval based on core attributes: when the core attribute set When all attribute values ​​are missing, the core attribute strong constraint retrieval is automatically skipped, and the subsequent path is triggered directly.

[0098] The following describes the implementation method of strong constraint retrieval based on core attributes.

[0099] From the standardized attribute vectors of the source projects, extract the first standardized value belonging to the core attribute set; in the standard project set, retrieve the first standard project whose standardized attribute vector value on the same core attribute set is consistent with the first standardized value; merge all first standard projects into a subset corresponding to the core attribute strongly constrained retrieval path in the candidate standard project set.

[0100] The first standardized value refers to the specific standard value extracted from the standardized attribute vector of the source project on those dimensions belonging to the core attribute set. The standard project set is a pre-constructed reference project library with standardized descriptions and authoritative codes, serving as the target benchmark for mapping. The first standard project refers to those standard projects in the standard project set whose standardized attribute vector values ​​on the same core attribute set are completely consistent with the first standardized values ​​of the source project.

[0101] When performing a core attribute strong constraint retrieval, the system first determines whether there exists at least one core attribute in the source project's core attribute set, and whether that core attribute has a non-empty value in the source project's standardized attribute vector. If this condition is met, the core attribute strong constraint retrieval path is enabled. Once enabled, the system extracts all dimensions belonging to the core attribute set and their standardized values ​​from the source project's standardized attribute vector, and records these values ​​as the first standardized value.

[0102] The system identifies standard items whose standardized values ​​on the same set of core attributes are completely identical to the first standardized value. Here, "same set of core attributes" means comparisons using the same set of core attribute dimensions. For example, if the source item uses two core attributes—detection component and sample type—for retrieval, the standard items must also have completely identical values ​​on both detection component and sample type. All retrieved standard items are marked as first standard items. The system merges these first standard items to form a subset of the candidate standard item set corresponding to the core attribute strongly constrained retrieval path. On the other hand, if all core attribute values ​​in the standardized attribute vector of the source item are empty (i.e., no core attribute has a non-empty value), the system directly disables the core attribute strongly constrained retrieval path and does not perform any of the above retrieval operations.

[0103] The following example illustrates this. Strongly constrained core attribute retrieval aims to achieve high-precision anchoring using attributes that define the core semantics of a project (such as detected components). Strongly constrained core attribute retrieval can serve as the "main path" of the framework, prioritizing the reliability of matching results.

[0104] (1) Standardized representation of core attributes: for projects from any source Its core attribute set The values ​​on form a standardized vector:

[0105] in, Normalized mapping for attribute d The standard value obtained later.

[0106] (2) Configurable strict matching conditions: Retrieves items from the standard library S that meet the consistency conditions. This framework allows for parameter-based strict matching. The stringency of the matching can be flexibly controlled. The search criteria are:

[0107] like This requires all core attributes to strictly match; if If the subset only requires strict matching of attributes, the remaining core attributes can be used as the basis for subsequent similarity calculations.

[0108] (3) Candidate set generation: Items that meet the conditions constitute a subset corresponding to the core attribute strongly constrained retrieval path (denoted as ). ):

[0109] Pluggable implementation: This path typically relies on inverted indexes built on core attributes or database joins to achieve millisecond-level response times.

[0110] In the implementation of the above embodiments: Clearly defined core attribute enabling conditions and disabling logic ensure that the core attribute-strongly constrained retrieval path is activated only when the source item possesses valid core information, reducing invalid searches or unreliable candidate results when core information is completely missing. The core attribute-strongly constrained retrieval employs a strict exact matching strategy, requiring that the standardized values ​​of the source item and the standard item be completely consistent across the same core attribute set, thus guaranteeing a high degree of certainty between the recalled first standard item and the source item at the core semantic level. By independently merging the search results into subsets corresponding to the core attribute-strongly constrained retrieval path, the system can clearly trace the source evidence of each candidate item, providing a more reliable basic candidate set for subsequent comprehensive matching degree calculations.

[0111] Optionally, in this embodiment, the multi-search path strategy includes auxiliary attribute supplementary retrieval; the conditions for enabling and disabling auxiliary attribute supplementary retrieval will be introduced first.

[0112] If the core attribute strong constraint retrieval returns empty candidate results or the number is less than the first preset threshold, and at least one auxiliary attribute in the auxiliary attribute set of the source project has a non-empty value in the standardized attribute vector, auxiliary attribute supplementary retrieval is enabled; and if the values ​​of all auxiliary attributes in the standardized attribute vector are missing, auxiliary attribute supplementary retrieval is disabled.

[0113] The auxiliary attribute supplementary retrieval path aims to achieve fast and accurate filtering using highly unique auxiliary identifiers. This can be viewed as a high-specificity filtering process. The following example illustrates this, with dynamically enabled triggering conditions including: ① Insufficient results from the core path: When the candidate set returned by path one (strong constraint retrieval based on core attributes) is insufficient. Empty, or quantity | |Below the preset minimum valid result threshold (For example Or 3).

[0114] ② High-confidence availability of auxiliary information: The set of auxiliary attributes of the source item q In the above, at least one attribute has a non-empty value and has been verified to have a high confidence level (e.g., the abbreviation has a success rate of >80% in the organization's historical mapping).

[0115] ③ Business scenarios require high precision: Under the "high precision priority" strategy template, this path is enabled in parallel with path one by default to utilize all available high-reliability information for cross-validation.

[0116] Dynamically disabling / skipping trigger conditions include: Auxiliary information is completely missing: All auxiliary attribute values ​​of the source item q are empty.

[0117] Efficiency-first strategy: Under the "ladder-fast" sub-strategy, if the core attribute strong constraint retrieval has returned sufficient results, auxiliary attribute supplementary retrieval can be skipped to save computing resources.

[0118] The following describes the implementation method for supplementary retrieval of auxiliary attributes.

[0119] From the standardized attribute vector of the source project, extract the second standardized value belonging to the auxiliary attribute set; for the identified auxiliary attribute, in the standard project set, retrieve the second standard project whose standardized attribute vector value on the same auxiliary attribute set is consistent with the second standardized value; merge all second standard projects into a subset corresponding to the auxiliary attribute supplementary retrieval path in the candidate standard project set.

[0120] The second standardized value refers to the specific standardized value extracted from the standardized attribute vector of the source item on those dimensions belonging to the auxiliary attribute set. The second standard item refers to those standard items in the standard item set whose standardized attribute vector values ​​on the same auxiliary attribute set are completely identical to the second standardized values ​​of the source item. The subset corresponding to the auxiliary attribute supplementary retrieval path refers to a subset formed by merging multiple second standard items recalled specifically through this auxiliary attribute supplementary retrieval path; it is part of the candidate standard item set.

[0121] First, it determines whether the candidate results returned by the core attribute strong constraint retrieval are empty, or whether their number is below a first preset threshold. Simultaneously, it determines whether there exists at least one auxiliary attribute in the auxiliary attribute set of the source project whose value in the standardized attribute vector of the source project is not empty. If both conditions are met, the auxiliary attribute supplementary retrieval path is enabled. Once enabled, all dimensions belonging to the auxiliary attribute set and having non-empty values, along with their standardized values, are extracted from the standardized attribute vector of the source project and recorded as second standardized values. Then, for each identified auxiliary attribute, the system independently searches the standard item set to find standard items whose standardized values ​​on the same auxiliary attribute are completely identical to the second standardized values. All retrieved standard items are marked as second standard items, and the system merges these second standard items to form a subset of the candidate standard item set corresponding to the auxiliary attribute supplementary retrieval path.

[0122] On the other hand, if all auxiliary attributes in the standardized attribute vector of the source project have empty values, that is, if no auxiliary attribute has a non-empty value, the system directly disables the auxiliary attributes to supplement the search path and does not perform any of the above search operations.

[0123] The following example illustrates this. To address situations where core attributes are missing or ambiguous, the attribute supplementary retrieval path utilizes highly unique or organization-specific auxiliary identifiers (such as English abbreviations like ABBR) for rapid filtering, serving as an effective supplement to the core path. This retrieval path aims to quickly and accurately narrow down the candidate space using highly reliable auxiliary identifiers (such as project abbreviations), effectively supplementing the core attribute retrieval. Compared to the core attribute retrieval path, auxiliary attribute retrieval has less stringent requirements for semantic consistency but typically possesses higher identifier uniqueness, thus efficiently filtering out a large number of irrelevant items when information is available.

[0124] (1) Auxiliary attribute set: definition This is a set of auxiliary attributes.

[0125] (2) Independent exact match: for each non-empty auxiliary attribute Retrieve items with completely identical attribute values ​​from the standard library to generate a subset of candidates. .

[0126] (3) Candidate set generation: Merge the search results of all auxiliary attributes to form a subset corresponding to the auxiliary attribute supplementary search path (denoted as ). ):

[0127] Pluggable implementation: This path also applies to indexes based on exact value matching. In the configuration, you can specify that this path should only be triggered when the core path returns insufficient results, or it can always be executed in parallel to maximize recall.

[0128] In the implementation of the above embodiments: by setting explicit enable conditions and disable logic, auxiliary attribute supplementary retrieval is only initiated when the core path recall is insufficient and auxiliary information is available, reducing computational overhead. When the core retrieval cannot provide enough candidate items, the auxiliary attribute supplementary retrieval uses highly specific auxiliary identifiers such as English abbreviations for precise matching, which can quickly recall standard items with the same auxiliary identifiers as the source items. Since auxiliary identifiers usually have high uniqueness, they can directly locate unique or a few candidate items, thereby effectively compensating for the missed recall problem caused by missing core path information or insufficient retrieval. By independently merging the retrieval results into subsets corresponding to the auxiliary attribute supplementary retrieval path, the system can clearly distinguish the evidence sources of different paths, providing a reliable supplementary candidate set for subsequent comprehensive matching degree calculation. This design significantly improves the adaptability of the mapping method to incomplete data and the overall recall rate while ensuring high-precision retrieval priority.

[0129] Optionally, in this embodiment, the multi-retrieval path strategy includes text semantic fallback retrieval; the conditions for enabling and disabling the multi-retrieval path strategy are introduced first.

[0130] If the candidate standard item set determined by strong constraint retrieval of core attributes and supplementary retrieval of auxiliary attributes is empty, the source item contains preset characters, or the source item belongs to a preset business scenario, then the text semantic fallback retrieval is initiated.

[0131] For example, the dynamic enabling triggering conditions for text semantic fallback detection may include: ① Pre-path recall failure: The combined candidate set returned by the core attribute strong constraint retrieval and the auxiliary attribute supplementary retrieval (if enabled) is empty, or the number is still lower than the minimum requirement. .

[0132] ② Abnormal source project text description: The source project name T(q) is extremely irregular, contains a large number of special characters or organization-specific codes, which causes the structured attribute extraction to fail. In this case, it is necessary to directly rely on the original text.

[0133] ③ Clear high-recall business scenarios: Under the "high-recall priority" or "data inventory" strategy template, this path is enabled in parallel with the previous path by default to ensure maximum coverage.

[0134] The dynamic disable / skip trigger conditions for text semantic fallback checks include: ① The results of the preceding paths are sufficient: The quantity and quality of the candidate sets returned by the core attribute strong constraint retrieval and / or auxiliary attribute supplementary retrieval meet the requirements, and there is no need to start the computationally expensive semantic retrieval.

[0135] ② Real-time performance constraints: When the system load is too high or the real-time requirements are extremely stringent, this computationally intensive path can be configured to be disabled.

[0136] The following describes the implementation method of text semantic catch-all inspection.

[0137] This process retrieves the first complete text name of the source project, the second complete text name of the standard project, and a set of synonyms for the second complete text name. The first complete text name refers to the original name of the source project after text normalization. The second complete text name refers to the original name of the standard project after text normalization. The synonym set is a pre-constructed set of terms for each standard project, including its common aliases, abbreviations, English acronyms, and historical names.

[0138] Based on a text similarity calculation model, this algorithm calculates the semantic similarity between a first complete text name and a second complete text name, as well as between the first complete text name and elements in the synonym set. A text similarity calculation model is an algorithmic model used to quantify the semantic closeness between two texts.

[0139] The process involves selecting third-criteria items whose semantic similarity exceeds a preset threshold, and / or fourth-criteria items whose first complete text name synonyms meet the keyword overlap condition. These form the subset corresponding to the text semantic fallback retrieval path within the candidate standard item set. The keyword overlap condition refers to a Boolean condition determining whether the source item text name shares specific core vocabulary with the standard item's synonyms. The fourth-criteria item refers to the standard item recalled by meeting the keyword overlap condition. The subset corresponding to the text semantic fallback retrieval path is a subset formed by merging the third-criteria items and fourth-criteria items specifically recalled through this path.

[0140] The following example illustrates this. In cases where core attributes are missing, abbreviations or auxiliary information are incomplete, or there is severe semantic ambiguity, this path uses textual similarity of the complete project name as a fallback for further improving the recall rate of the mapping. This method directly calculates similarity at the textual level of the original project name, aiming to recall terms semantically relevant to the source project and ensuring the framework's coverage of long-tail and non-standard terms. This path does not rely on the completeness of structured attributes; it only relies on the semantic similarity of the name text for recall, effectively mitigating missed recall issues caused by differences in expression, word usage habits, or missing information, thus supplementing and providing a fallback for preceding strongly constrained searches.

[0141] 1. Preliminary Recall Based on Synonym Expansion: A broad recall based on keyword overlap is performed using the Syn(s) thesaurus of standard project names to generate a set. The detailed explanation is as follows: To mitigate differences in terminology, we first construct an extended set of synonyms for the standard project's name. Let the standard project be... The original standard name is T(s), and its synonym set is Syn(s). As one implementation method, the construction, maintenance, and updating of the synonym lexicon follow a semi-automated human-machine collaborative mechanism to ensure its professionalism and adaptability: During the initialization and construction phase: Seed source: The initial vocabulary is derived from laboratory medicine-related terms in authoritative medical thesaurus (such as UMLS, CHV), as well as high-frequency, high-discrimination nouns and abbreviations extracted from all item names T(s) of the standard item library S using statistical methods such as TF-IDF. Expert review: The automatically extracted seed vocabulary needs to be reviewed, categorized (e.g., marked as "ingredients" or "methods") and supplemented by domain experts to form the initial version of the thesaurus.

[0142] For dynamic maintenance and expansion: Learning based on matching logs: The system records every time the KeywordOverlap function is triggered and ultimately leads to a successful mapping. For words that co-occur but are not in the existing vocabulary, the system marks them as candidate new synonyms. Mining based on failed cases: During manual review, for cases missed due to insufficient word overlap, manually marked key new synonyms are prioritized for inclusion in the candidate list. Regular review and inclusion: The system regularly (e.g., monthly) submits the list of candidate new synonyms to experts for review. After expert confirmation, the new words are officially added to the synonym dictionary and take effect immediately in subsequent processes.

[0143] In summary, the thesaurus is a living, evolving knowledge component whose lifecycle is closely integrated with the system's main loop of "human-machine collaborative feedback." It can be continuously optimized as new terms and expressions emerge, thereby ensuring the long-term effectiveness of the "synonym expansion initial recall" strategy.

[0144] For each source project Its original name is T(q). The initial recall condition is defined as: if T(q) is a synonym of any s... If a standard item shares specific core terms (such as "sugar," "SG," or other keywords indicating the detected component), or meets specific word overlap rules, then the standard item is initially recalled. Let the resulting preliminary candidate set be denoted as:

[0145] Among them, the function This is used to determine whether two texts contain co-occurring core keywords.

[0146] The keyword overlap judgment function KeywordOverlap(T(q),t) in the initial retrieval conditions here is specifically defined as follows, aiming to achieve broad recall through computable rules: Core Vocabulary Driven: The system maintains a predefined core vocabulary highly relevant to the field of laboratory medicine. This vocabulary includes two categories of terms: ① Component keywords: such as "sugar", "protein", "bilirubin", "WBC", "RBC", "SG", etc., which are terms and common abbreviations for the substances or cells being detected.

[0147] ② Action / Method Keywords: Terms such as “measurement,” “detection,” “counting,” “experiment,” and “method” that indicate detection actions or methods.

[0148] Word segmentation and normalization: The source item name T(q) and the standard item synonym t are segmented, and the segmentation results are subjected to medical stem extraction and abbreviation normalization (for example, "measurement" and "detection" are normalized to the same stem; "GLU" and "Glu" are normalized to "glucose").

[0149] Overlap detection rule: A function returns True if and only if any of the following conditions are met: ①Condition A (core word co-occurrence): In the normalized word segmentation results of T(q) and t, there is at least one common word that belongs to the above core vocabulary list.

[0150] ② Condition B (Consecutive Character Substring Matching): Any consecutive character substring of length not less than 3 in T(q) appears in t; or vice versa. This rule is used to capture partial matches such as "total protein" and "albumin".

[0151] ③ Condition C (Edit Distance Approximate Match): When T(q) or t is short text (e.g., length ≤ 4), if the edit distance between the normalized strings of the two is no more than 1, they are considered to overlap. This rule is used to tolerate common spelling errors or abbreviations.

[0152] Through the above-mentioned explicit and programmable rule combination, the KeywordOverlap function can maintain a high recall rate while possessing determinism and repeatability.

[0153] 2. Global text similarity calculation: Based on the initial recall, the textual similarity between the source item and the standard item at the level of complete names was further calculated. .

[0154] This framework supports integration with different similarity calculation "engines": (1) Traditional statistical engine: based on cosine similarity of TF-IDF vectors, which is efficient in calculation.

[0155] (2) Semantic Vector Engine: Based on pre-trained models such as Sentence-BERT, sentence vectors are obtained and similarity is calculated, resulting in deeper semantic understanding.

[0156] Set similarity threshold Define a similarity-driven supplementary candidate set:

[0157] 3. Candidate set generation: merging Similarity higher than the threshold The project forms a subset corresponding to the text semantic fallback retrieval path (denoted as...). ):

[0158] Pluggable implementation: The core of text semantic fallback retrieval lies in a replaceable "similarity calculation engine" and an adjustable recall threshold. For example, it can be configured to only be enabled when neither the core nor the secondary paths return results, in order to control computational overhead.

[0159] In an optional embodiment, after obtaining the subsets corresponding to the core attribute strongly constrained retrieval paths, the subsets corresponding to the auxiliary attribute supplementary retrieval paths, and the subsets corresponding to the text semantic fallback retrieval paths, the final global candidate set, i.e., the candidate standard item set C(q), is the union of the results of the above paths:

[0160] After generation, deduplication and size control are performed (e.g., retaining the top K with the highest comprehensive scores) before sending them to the next stage.

[0161] The framework's configurability is reflected in the fact that users can define strategy templates according to actual business needs: (1) High-precision priority strategy: Only enable path one (strong constraint retrieval of core attributes) and path two (supplementary retrieval path of auxiliary attributes), and set strict parameters. and higher (If enabled). This strategy is suitable for scenarios with zero tolerance for errors, such as clinical decision support.

[0162] (2) High recall priority strategy: Activate all three paths in parallel and relax the restrictions. The conditions, and lowered Threshold. This strategy is suitable for scenarios requiring maximum coverage, such as data inventory and scientific research analysis.

[0163] (3) Step-by-step strategy: Execute the path in sequence. If the previous path returns an empty result or the number is insufficient, the next path will be automatically triggered, achieving a balance between efficiency and effectiveness.

[0164] By abstracting the candidate generation process into a pluggable and configurable multi-path strategy layer, this framework not only provides a specific retrieval method, but also forms a flexible generation mechanism that can adapt to different data quality and business objectives, laying a solid foundation for subsequent accurate matching.

[0165] In the implementation of the above embodiments: through textual semantic fallback retrieval, when neither the core nor auxiliary paths can provide effective candidates, generalized recall can be achieved solely based on the text of the item name. This path does not rely on the completeness of any structured attributes and can cover situations where core information is completely missing, naming is highly irregular, or new terms have emerged, significantly improving the mapping method's adaptability to low-quality data. By introducing a synonym set and semantic similarity calculation, this path can identify the correspondence between items with different literal expressions but the same clinical semantics, effectively alleviating the missed recall problem caused by terminology diversity and abbreviation ambiguity.

[0166] Optionally, in this embodiment of the application, calculating the matching score for each dimension in the effective observation dimension set includes: For each dimension in the effective observation dimension set: if the standardized value of the source item is the same as the standardized value of the candidate standard item, then the dimension score is determined as the first value. If the standardized value of the source item and the standardized value of the candidate standard item have a predefined semantic compatibility relationship, then the dimension score is determined as a constant between the second value and the first value; if the standardized value of the source item and the standardized value of the candidate standard item have a predefined conflict relationship, then the dimension score is the third value.

[0167] For example, to quantify the consistency of source item q and candidate standard item s in terms of structured dimensions, this work proposes a missing robust multidimensional matching scoring mechanism. This mechanism aims to address the systematically low scores and incomparability issues caused by missing information in traditional fixed-weight averaging methods. Its core is a two-layer design that ensures fair and comparable matching scores through a normalized weighted average within the observation dimensions and a dynamic modular weighting based on information content perception, while also adapting to comparison pairs with varying degrees of information completeness. Firstly, to comprehensively capture clinical semantics while considering the usability of actual data, this work defines an extended set of dimensions. This set contains two types of dimensions: (1) Core clinical dimensions: such as the components to be detected, sample type, sampling time point, etc., are used to anchor the core semantics of the project.

[0168] (2) Auxiliary identification dimensions: such as English abbreviations, detection methods, etc., can serve as highly specific supplementary evidence. This constitutes the complete set for all subsequent dimensional calculations.

[0169] belong The dimension is determined solely by the aggregated score in this section. The overall matching score is then integrated; these dimensions will no longer be re-integrated as "independent dimensions / independent modules" to avoid double scoring.

[0170] Dimension set and "single-dimensional scoring interface" definition: Let D be the complete set of all available dimensions. This framework selects a subset of dimensions for "core structured alignment". .

[0171] For any dimension Define its single-dimensional matching score interface as .

[0172] (1) If Scoring is achieved by matching textual attribute dimensions. and order = ; (2) If Matching scores using numerical / range attributes and order ; (3) If The score is obtained directly by matching the categorical attribute dimension. ; (4) If Binary values ​​are obtained by matching scores using identifier / code attribute dimensions. .

[0173] Special Handling: Abbreviation (abbr) Dimension Matching: Given that "English abbreviations" are a common but easily confused dimension in clinical practice, this method performs special handling on them, adopting a strategy of "standard matching first, string similarity as a fallback": ① Standard Mapping Matching: First, map the abbreviations of the source and candidate items to a unified standard abbreviation space. If the standard abbreviations are consistent, then... .

[0174] ② String similarity calculation: If the standard mappings are inconsistent but both abbreviations are not empty, then calculate the similarity of their original strings. And this value is used as the dimension matching score, that is:

[0175] generally, The value range is [0,1], to avoid directly imposing negative scores due to inconsistent abbreviations. In the "string similarity calculation" step, various string similarity or distance algorithms can be used, but for clinical abbreviation matching, the Jaro-Winkler algorithm is usually the preferred choice because it often involves prefix consistency and a small number of character differences.

[0176] Jaro-Winkler Distance: This algorithm places particular emphasis on the similarity of string prefixes. It calculates the number of character matches and transpositions between two strings and rewards strings with common prefixes using a scaling factor. It is well-suited for handling abbreviations, names, and medical terms because prefixes (such as the first few letters) often contain key information.

[0177] Furthermore, since the Jaro-Winkler distance is itself a similarity metric algorithm, its output value is usually within the range of [0,1], so no further normalization is needed.

[0178] generally, The value range is [0,1], which avoids directly imposing negative scores due to inconsistent abbreviations, thus better reflecting clinical scenarios.

[0179] In the implementation of the above embodiments: by setting fine-grained matching scoring rules for each dimension, the relationship between the source item and the candidate standard item in terms of dimension values ​​is distinguished into three states: completely identical, semantically compatible, and obviously conflicting. These are then assigned a first value, a constant between the first and second values, and a third value, respectively. This design ensures that the dimension matching score is no longer limited to simple binary agreement or inconsistency, but accurately reflects the degree of clinical semantic compatibility and conflict. The fine-grained scoring mechanism enhances the discriminative power of the dimension matching score, improving the accuracy of the overall matching degree calculation and the reliability of the decision.

[0180] To avoid the unreasonable "dilution" effect of missing values ​​on the matching score calculation, this application introduces an effective observation dimension set. Defined as in the comparison of pairs A subset of dimensions where both sides have non-empty values:

[0181] In this embodiment, the construction of the effective observation dimension set is a dynamic data preprocessing step performed before calculating the matching score for each pair of "hospital item - candidate standard item" (i.e., comparison pair (q, s)). Its purpose is to filter out the available dimensions in the current comparison pair from the predefined set of all dimensions D, thereby avoiding interference from missing values. The construction process follows a clear, programmable logical decision-making flow.

[0182] Dimensional matching total score The above calculation is defined as the weighted average of the scores for each observation dimension:

[0183] in, The preset base weights for dimension d are used to reflect its relative clinical importance (e.g., usually set as follows). (where component is the core dimension of semantic matching). The advantage of this design is: (1) Resistance to missing data and fairness: Missing dimensions are completely excluded from the calculation, avoiding invalid information from "diluting" the overall score and ensuring that the score is based entirely on comparable and valid information; (2) Score Comparability: By normalizing the weights within the observed dimension set, the scores remain comparable regardless of how many dimensions are missing between the comparison pairs. Always normalized within the same meaningful interval (theoretically [ [1,1], after processing, can be mapped to [0,1]), making the matching degree between item pairs with different information completeness directly comparable.

[0184] In subsequent multi-module integrated matching, to reflect the "dimensional matching module" itself in different comparison pairs To address the differences in the strength of evidence, this mechanism introduces a dynamic effective weight. This weight is determined by dimensional coverage. Adjustments are needed: (1) Number of observation dimensions:

[0185] (2) Dimensional coverage:

[0186] Here It is a predefined constant, meaning the total number of all extended dimensions involved in the project matching calculation. Specifically, This represents the number of core, structured attributes predefined in the method to describe a test item. For example, in one embodiment of this method, four core dimensions are defined for matching: d1: detection component, d2: sample type, d3: sampling time point, and d4: English abbreviation. In this case, the total set of dimensions D_extend = {d1, d2, d3, d4}, and its size... =4. It is set during system initialization and remains fixed during a single matching calculation.

[0187] (3) Effective weights:

[0188] in, These are the preset base weights for this module. The contribution of this design is: (1) Information perception: It quantifies the proportion of structured information available in the current comparison pair. The more available dimensions ( The larger the value, the more sufficient the evidence for dimensional matching, and the higher the module's influence (weight) in the final decision.

[0189] (2) Dual robustness: This design, combined with the first layer of this mechanism, forms a dual guarantee: ① Within the module (calculation) (At that time), the absence does not directly lead to a deduction of points.

[0190] ② When merging modules into the overall score, the overall contribution of modules with incomplete information will be reduced proportionally. This encourages data completeness while avoiding disproportionate penalties for incomplete information in the comprehensive matching decision.

[0191] In summary, the "information-incomplete robust multidimensional matching scoring mechanism" proposed in this section systematically addresses the common problems of information incompleteness and incomparability in heterogeneous data fusion through two core designs: "intra-observation dimension normalization" and "information-aware dynamic module weights." Compared to simple global averaging or fixed-weight fusion methods, the matching scores generated by this mechanism are more robust, fair, and interpretable, laying a solid foundation for subsequent high-confidence automatic mapping decisions.

[0192] Overall matching degree of this application The aim is to unify and integrate evidence from different matching modules. To ensure the fairness and robustness of the integration, this framework employs a weighted averaging strategy based on a validity indicator function and dynamic weights. The core of this strategy is that a module is only integrated if it is within the current comparison pair. Only when a matching module possesses valid information will it participate in the final score calculation, and its contribution is determined by its base weight and a specific validity factor. Assuming there are M independent matching modules, the formula for calculating the overall matching degree is as follows: Suppose there are M independent matching modules (e.g., dimension matching module, text attribute dimension module, numeric / range attribute dimension module, category attribute dimension module, identifier / code attribute dimension module, and other modules that can be expanded in the future), then the formula for calculating the overall matching degree is as follows:

[0193] in: (1) It is the matching score calculated by module m, and its value is in the range [0,1].

[0194] (2) It is the validity indicator function of module m, which is a binary function:

[0195] This function is a "switch" that determines whether a module participates in the calculation.

[0196] These are the dynamic effective weights of module m, within the effective (i.e., Under the premise of ), its basic weight And a factor reflecting the completeness or reliability of its information. Joint decision:

[0197] This factor is used to fine-tune the relative importance within effective modules.

[0198] Module definition example: (1) Dimension matching module: Indicator functions: If and only if the effective observation dimension set Not empty.

[0199] Dynamic weights: their weighting factors This refers to dimensional coverage. Therefore .

[0200] (2) Text similarity module: Indicator functions: As a basic fallback module, their information is usually always available, therefore they are always set. .

[0201] Dynamic weights: The weighting factors can be set. Therefore Alternatively, its weight can be appropriately reduced when there is sufficient structured information.

[0202] (3) Specific method / instrument module (if designed as a standalone module): Indicator functions: If and only if the method information of both parties is not empty and can be compared.

[0203] Dynamic weights: can be set Or, it may be adjusted according to the level of detail in the method description.

[0204] The advantages of this fusion mechanism: Explicit validity control: through Clearly defining the eligibility of each module for participation avoids interference from invalid information with the computing framework, resulting in a clear and logical structure.

[0205] Dynamic weight refinement: Based on the effectiveness of the module, through... It enables dynamic adjustment of weights and can sense the amount of information (such as dimensional coverage) or the quality of information.

[0206] High extensibility: Adding a new matching module only requires explicitly defining its... , , The computational logic can be seamlessly integrated.

[0207] In summary, this fusion mechanism achieves robust, refined, and scalable fusion of multi-source matching evidence through the synergistic effect of the "validity indicator function" (I) and the "dynamic validity weight" (W). I is responsible for the Boolean decision of "whether to participate," while W is responsible for the continuous quantification of "how much to participate," jointly ensuring the comprehensive matching score. The accuracy and reliability of [the system / mechanism].

[0208] Confidence-based hierarchical decision-making and human-machine collaborative optimization: Calculate the overall matching degree Then, for each source item q, select the standard item with the highest score in its candidate set. The final decision is made. This framework employs a hierarchical decision-making strategy to balance automation efficiency with mapping accuracy.

[0209] (1) Hierarchical decision-making: Automatic mapping: If ≥ ( If the high confidence threshold is used, the system will automatically map q to... This area corresponds to a high-quality match with complete information and solid evidence.

[0210] Manual review: If ≤ < ( If the threshold is low confidence, the system generates a manual review task, where domain experts confirm or correct the mapping results. This area is a key data source for continuous system optimization.

[0211] Delay mapping: If < This indicates that the current information is insufficient to make a reliable mapping, and the project is marked as "pending" for further in-depth analysis or data quality governance.

[0212] (2) Human-machine collaboration optimization: The results of manual review, as high-quality labeled data, are fed back to the system, forming a closed loop of continuous optimization. ① Knowledge base enhancement: The verified mapping pairs can be used to expand and revise the thesaurus and standardization rules, improving the accuracy of attribute extraction and standardization.

[0213] ② Model parameter tuning: The review results can be used as training data to optimize the weight parameters of each matching module (e.g., ) and decision threshold ( , This makes the system's decisions more aligned with the actual data distribution.

[0214] ③ Strategy Evaluation: By analyzing the distribution of manually reviewed cases, the effectiveness of the candidate generation strategy can be evaluated and optimized.

[0215] Through the aforementioned dynamic weighted fusion and hierarchical decision-making mechanism, this framework not only achieves highly robust automatic mapping, but also constructs a terminology mapping system that can continuously evolve and become more intelligent with use through human-machine collaboration.

[0216] Optionally, in this embodiment of the application, if the dimension is the English abbreviation dimension, the process of calculating the matching score of the English abbreviation dimension includes: determining whether the English abbreviation in the source project is consistent with the English abbreviation in the candidate standard project; if consistent, determining the score of the English abbreviation dimension as the first value; if inconsistent, calculating the string similarity between the English abbreviation in the source project and the English abbreviation in the candidate standard project, and using the string similarity as the score of the English abbreviation dimension.

[0217] For example, based on the multi-dimensional attribute classification of this framework, for any text-type attribute dimension... (e.g., the Chinese name of the project, cn_name, clinical indication testing notes, etc.) This application quantifies the degree of matching between the source project q and the candidate standard project s on this specific dimension by calculating the semantic similarity of their attribute values.

[0218] 1. General matching paradigm: This vector-based similarity measurement framework adopts a unified perspective for handling textual attributes: regardless of the underlying technology used, the core is to map textual attribute values ​​into real-valued vectors, and then calculate the matching score through normalized similarity between vectors. Let the normalized text values ​​of the source item and candidate item on dimension d be... and The matching score is calculated following this paradigm:

[0219] in, It is a text-to-vector encoding function and is the core of the paradigm. The framework supports multiple encoder implementations. It is a normalized similarity function whose output is limited to the interval [0,1] to ensure comparability and fairness of fusion with scores of other types of attributes in the framework. Scoring is assigned to text-based attribute dimensions.

[0220] 2. Core encoding function Two implementation schemes: This framework provides two optional encoding schemes for implementing the above paradigm, focusing on computational efficiency and deep semantic accuracy respectively: Option A: TF-IDF-based statistical coding (efficient baseline) This scheme is based on the Bag-of-Words model, which focuses on statistical information about word co-occurrence and is computationally efficient.

[0221] ① Vocabulary construction: For dimension d, from all items ( Extract text from the attribute values ​​of this dimension of the dataset, perform word segmentation and cleaning, and construct a dataset containing... A unique term-specific vocabulary .

[0222] ② Vectorization: For standardized text values and Calculate its TF-IDF vector representation , The value of the i-th dimension in the vector. The calculation is as follows:

[0223] in, The total number of items with this attribute value. For including entries The number of projects.

[0224] Option B: Semantic encoding (enhancement) based on pre-trained language models.

[0225] This approach utilizes deep pre-trained language models (such as BERT and Sentence-BERT) to generate semantic vectors, which can better capture deep semantic relationships such as synonym substitution, word order changes, and semantic implications.

[0226] ①Text encoding: Introduce a pre-trained Chinese language model As an encoder, the model maps the input text into an h-dimensional dense semantic vector.

[0227] ② Vector generation: Normalizing text values and The semantic vector representations are obtained by inputting each input into the encoder: ,

[0228] 3. Unified normalized similarity calculation and integration: After obtaining the vector representation, cosine similarity is used as the basic metric, and normalization is used to ensure that the score is in the [0,1] interval: , )= ,

[0229] in, It is a normalization function. For scheme A, the cosine similarity result is already in the interval [0,1], therefore... For scheme B, the theoretical range of the cosine similarity between the output vectors of the model is [ [1,1] requires a linear transformation: .

[0230] Matching score of text attributes This will be used as a single-dimensional matching score. A special case, integrated into the "missing robust multidimensional matching mechanism," serves as a special case of single-dimensional fine-grained matching scores: when At that time, its single-dimensional matching score .

[0231] calculate At the same time, the rules regarding missing values ​​in single-dimensional matching score calculation are also followed: if any attribute value is missing, then... .

[0232] In the implementation of the above embodiments: English abbreviations are relatively stable and highly unique in medical institution information systems, serving as important semantic clues for cross-institutional project mapping. However, different institutions may have subtle differences in the English abbreviations of the same project, such as capitalization, abbreviation format, and separators. This step prioritizes determining whether the standard abbreviations are consistent, assigning a first value when they are consistent, thus fully preserving the value of this dimension as high-confidence evidence. This allows the English abbreviation dimension to play a decisive role when the expressions are completely consistent, while also preserving positive matching evidence when there are differences in writing habits. By increasing the tolerance for the diversity of abbreviation expressions, the mapping method's adaptability to differences in real medical institution data can be improved without sacrificing accuracy, thereby improving the overall mapping accuracy.

[0233] The following section describes how to calculate the matching score for numerical / range attribute dimensions.

[0234] The set of numerical or range-type attribute dimensions is denoted as For any such dimension (e.g., reference range, critical value, standard concentration, etc.) This method employs a general scoring strategy based on interval geometric similarity to quantify the degree of matching between the source item q and the candidate standard item s in that numerical dimension. Each dimension is modeled as a numerical interval. , where L and U represent the lower and upper limits of the interval, respectively.

[0235] 1. General interval similarity measure: For dimension d, given the range of values ​​for the source project and the range of values ​​for candidate standard items This method calculates the matching score by integrating information on interval overlap and center point consistency.

[0236] (1) Interval overlap score: This score is measured by calculating the proportion of the overlap between two intervals relative to their union, reflecting the direct consistency of the numerical range.

[0237]

[0238]

[0239]

[0240] (2) Center point consistency score: This score measures the relative deviation of the center points of two intervals through an exponential decay function. It is more sensitive to the position of the interval (rather than just the range) and can effectively identify situations where the center points are aligned but the ranges are different.

[0241] Calculate the center point: ,

[0242] Calculate the width: ,

[0243] Calculate the normalized center point deviation: ,

[0244] The centroid consistency score is defined as:

[0245] When the difference between the center points is small, the score is close to 1; the larger the difference, the score smoothly decays to 0.

[0246] (3) Interval similarity fusion: The final interval matching score is a convex combination of the two scores mentioned above:

[0247] Among them, the fusion coefficient It can be used to adjust the relative importance of overlap and center point consistency in the final score, and can be empirically set or optimized according to dimensional characteristics.

[0248] 2. Integration and processing within the dimensional matching system: Matching score for numerical / range attributes This will be used as a single-dimensional matching score. Integrate into a weighted averaging framework based on the effective observation set: when At that time, its single-dimensional matching score ; Consistent with all dimensions, if any value in that dimension is invalid or missing, then .

[0249] The criterion for determining missing data is that, and only if, after the system performs data parsing and standardization on item q or item s, the corresponding standardized value v_d(q) or v_d(s) on dimension d is NULL or a null value (e.g., an empty string "", a NULL value in the database, etc.). This is usually caused by one or more of the following situations: ① Original data not provided: The original project information does not contain any description of this dimension.

[0250] ② Information extraction failed: The information corresponding to this dimension could not be successfully extracted from the unstructured text (such as the name) of the project.

[0251] The criterion for determining invalidity is: A value for item q or item s, after data parsing and standardization by the system, belongs to a predefined set of invalid values ​​on dimension d. This set typically includes: ① Meaningless placeholders: such as “unknown”, “not provided”, “ / ”, “-”, etc., which clearly indicate invalid information.

[0252] ②Contradictory or unresolvable values: For example, text such as "normal-abnormal" which is a non-numerical, non-qualitative standard description in the "reference range" dimension; or strings that cannot be converted into numbers in the "numerical" dimension.

[0253] ③ Abnormal values ​​that exceed the reasonable range: For example, values ​​such as "999" or "0" appearing in the "red blood cell count" dimension, which are clearly unreasonable in clinical testing.

[0254] 3. Compatibility with special numeric types (such as single numeric values): The above model is inherently compatible with single numerical (point value) cases, which can be regarded as a special interval with equal upper and lower limits (i.e., L=U). In this case, the interval overlap score degenerates into a point value equality judgment (1 for equality, 0 for inequality), while the center point consistency score directly reflects the relative deviation between point values. This method achieves universal processing of continuous numerical attributes through a unified interval representation and calculation framework.

[0255] The following section describes how to calculate the matching score for categorical attribute dimensions.

[0256] The set of categorical attribute dimensions is denoted as These dimensions (such as component, sample, method, instrument, department, etc.) are typically represented by a discrete, standardized classification code. This framework defines a unified matching score calculation template for all categorical attributes.

[0257] 1. General category matching template: For any categorical attribute dimension The standardized codes for the source project and the candidate standard project in this dimension are respectively... and Its single-dimensional matching score The definition is as follows:

[0258] in: This means that the two are semantically identical in this dimension, thus receiving full marks; This indicates that although the codes for the two methods are not completely identical, they belong to a predefined, acceptable compatibility relationship. This relationship needs to be predefined for each dimension d (for example, for the method dimension, "chemiluminescence" and "immunoturbidimetry" can both belong to the category of "immunological methods"; for the sample dimension, "serum" and "plasma" can be considered partially compatible). It is a pre-defined partial compatibility score constant for this dimension; The “conflict” relationship also needs to be predefined, indicating that there is a fundamental contradiction between the two values ​​in clinical or logical terms (for example, one of the detected components is “glucose” and the other is “hemoglobin”).

[0259] 2. Integration in the dimensional matching system: Matching score of categorical attributes This will be directly used as its single-dimensional score and integrated into a weighted average framework based on the effective observation set: when At that time, its Calculate using the general template described above.

[0260] Consistent with all dimensions, if any value in that dimension is invalid or missing, then .

[0261] The following section describes how to calculate the matching score for identifier / code-type attribute dimensions.

[0262] The set of identifier / code attribute dimensions is denoted as These dimensions (such as national standard codes (nat_code), internal standard IDs (std_id), LOINC codes, SNOMED CT codes, etc.) have explicit, standardized, and typically unique identifier characteristics within a specific context. In matching calculations, the processing logic for these dimensions is the most straightforward, yet they offer the highest confidence level.

[0263] 1. Exact matching logic: For any identifier / code type attribute dimension Its value is an encoding defined by authoritative standards. This framework uses precise binary matching logic, and its single-dimensional matching score... The definition is as follows:

[0264] 2. Core role and weight setting in the matching system: Despite its simple computation, the identifier / code dimension plays a decisive role in matching decisions, which manifests in two aspects: (1) High base weight: In a weighted average based on a valid observation set, it should be Assign a significantly higher base weight than other types of dimensions. For example, it can be set This is because a match based on an authoritative standard encoding has far stronger evidentiary value than a similarity based on a textual description.

[0265] (2) Candidate generation and fast filtering: In the generation of candidate standard itemsets based on multi-path retrieval, such attributes are ideal data sources for the "auxiliary attribute supplementary retrieval" path. A successful coded exact match can efficiently locate unique or very few candidate standard items, greatly reducing the search space, and can be directly used as a high-confidence matching result.

[0266] 3. Coordination with the overall framework: Collaboration with candidate generation: In During generation, a successful nat_code match can directly produce a very small set of high-quality candidates, or even a single candidate.

[0267] In the comprehensive matching process, a single successful match (score of 1) in the final comprehensive matching score calculation is sufficient to elevate the candidate item's overall score, even if other dimensions show slight differences. Conversely, a core identifier conflict will significantly reduce the overall score.

[0268] Identifier / code attributes serve as the most deterministic and computationally efficient bridge in cross-agency term mapping. This framework assigns them "precise binary matching" logic and high weight, enabling them to act as anchors and arbitrators in candidate generation and final decision-making. This significantly improves mapping accuracy and system efficiency while ensuring high recall.

[0269] After calculating the matching scores for each dimension and module, this framework calculates the final comprehensive matching degree through a dynamic weighted fusion mechanism, and makes hierarchical decisions based on this, forming an optimization closed loop of human-machine collaboration.

[0270] Optionally, in this embodiment of the application, calculating the comprehensive matching degree between the source item and the corresponding candidate standard item based on the matching score includes: An extended dimension set can be predefined, which contains core clinical dimensions used to calculate dimension matching scores, such as test components, sample type, sampling time point, English abbreviations, etc.

[0271] The matching scores of all dimensions in the effective observation dimension set are weighted and averaged according to the preset basic weights of the dimensions to obtain the dimension matching score. It should be noted that the dimensions belonging to the extended dimension set are only included in the comprehensive matching degree in the form of dimension matching score, and will not participate in the fusion as independent dimensions in the future.

[0272] The dynamic effective weights of the dimension matching score are obtained. The dynamic effective weights are determined based on the preset base weights and the dimension coverage rate. The dimension coverage rate is the ratio of the number of dimensions in the effective observation dimension set to the total number of dimensions. The dimensional matching score and its dynamic effective weight are weighted and fused with the matching score of at least one other dimension to obtain the comprehensive matching degree; the text similarity score is calculated based on the residual text, which is determined by removing identified terms belonging to the predefined clinical dimension from the normalized names of the source item and the candidate standard item.

[0273] Example: Overall matching degree The aim is to unify and integrate evidence from different matching modules. To ensure the fairness and robustness of the integration, this framework employs a weighted averaging strategy based on a validity indicator function and dynamic weights. The core of this strategy is that a module is only integrated if it is within the current comparison pair. Only when a matching module possesses valid information will it participate in the final score calculation, and its contribution is determined by its base weight and a specific validity factor. Assuming there are M independent matching modules, the formula for calculating the overall matching degree is as follows: Suppose there are M independent matching modules (e.g., dimension matching module, text attribute dimension module, numeric / range attribute dimension module, category attribute dimension module, identifier / code attribute dimension module, and other modules that can be expanded in the future), then the formula for calculating the overall matching degree is as follows:

[0274] in: (1) It is the matching score calculated by module m, and its value is in the range [0,1].

[0275] (2) It is the validity indicator function of module m, which is a binary function:

[0276] This function is a "switch" that determines whether a module participates in the calculation.

[0277] These are the dynamic effective weights of module m, within the effective (i.e., Under the premise of ), its basic weight And a factor reflecting the completeness or reliability of its information. Joint decision:

[0278] This factor is used to fine-tune the relative importance within effective modules.

[0279] Module definition example: (1) Dimension matching module: Indicator functions: If and only if the effective observation dimension set Not empty.

[0280] Dynamic weights: their weighting factors This refers to dimensional coverage. Therefore The dimension matching module aggregates the matching scores for each dimension in the extended dimension set.

[0281] (2) Text similarity module: Indicator functions: As a basic fallback module, their information is usually always available, therefore they are always set. .

[0282] Dynamic weights: The weighting factors can be set. Therefore Alternatively, its weight can be appropriately reduced when there is sufficient structured information. The text similarity module can correspond to the matching score of the project name dimension.

[0283] (3) Other dimension modules (such as reference range matching module, method matching module, instrument matching module, etc.): Indicator functions: If and only if the method information of both parties is not empty and can be compared.

[0284] Dynamic weights: can be set Or, it may be adjusted according to the level of detail in the method description.

[0285] The advantages of this fusion mechanism: Explicit validity control: through Clearly defining the eligibility of each module for participation avoids interference from invalid information with the computing framework, resulting in a clear and logical structure.

[0286] Dynamic weight refinement: Based on the effectiveness of the module, through... It enables dynamic adjustment of weights and can sense the amount of information (such as dimensional coverage) or the quality of information.

[0287] High extensibility: Adding a new matching module only requires explicitly defining its... , , The computational logic can be seamlessly integrated.

[0288] In summary, this fusion mechanism achieves robust, refined, and scalable fusion of multi-source matching evidence through the synergistic effect of the "validity indicator function" (I) and the "dynamic validity weight" (W). I is responsible for the Boolean decision of "whether to participate," while W is responsible for the continuous quantification of "how much to participate," jointly ensuring the comprehensive matching score. The accuracy and reliability of [the system / mechanism].

[0289] In the implementation of the above embodiments: based on the dynamic effective weight of dimensional coverage, the contribution of the dimensional matching module to the overall matching degree is adaptively correlated with the information completeness between the source project and the candidate standard project. When the number of dimensions in the effective observation dimensional set is large and the dimensional coverage is high, it indicates that the structured information comparable between the two parties is sufficient, the evidentiary strength of the dimensional matching score is high, and its dynamic effective weight is correspondingly increased. Conversely, when there are few comparable dimensions, the weight of the dimensional matching module is automatically reduced, so that its score will not have an undue impact on the overall matching degree due to incomplete information. This improves the fairness, robustness, and discriminative ability of the overall matching degree under different data quality conditions, thereby improving the overall accuracy of cross-institutional project mapping.

[0290] Optionally, in this embodiment of the application, the target standard project corresponding to the source project is selected from the candidate standard project set based on the comprehensive matching degree, including: From the candidate standard item set, the candidate standard item with the highest overall matching degree with the source item is selected as the first-choice candidate. If the difference between the overall matching degree of the first-choice candidate and the second-best candidate is less than a preset conflict threshold, the first-choice candidate and the second-best candidate are arbitrated according to a predefined clinical priority rule base to determine the target standard item.

[0291] For example, the overall matching degree is calculated. Then, for each source item q, select the standard item with the highest score in its candidate set. The final decision is made. This framework employs a hierarchical decision-making strategy to balance automation efficiency with mapping accuracy.

[0292] The conditions under which the overall matching degree meets the requirements include: the overall matching degree is greater than or equal to the preset reliability threshold; or, the overall matching degree is less than the preset reliability threshold, but greater than or equal to the dynamic decision threshold; wherein, the dynamic decision threshold is determined based on the difference between the overall matching degree of the preferred candidate and the overall matching degree of the second-best candidate.

[0293] (1) Hierarchical decision-making: Automatic mapping: If ≥ ( If the high confidence threshold is used, the system will automatically map q to... This area corresponds to a high-quality match with complete information and solid evidence.

[0294] Manual review: If ≤ < ( If the threshold is low confidence, the system generates a manual review task, where domain experts confirm or correct the mapping results. This area is a key data source for continuous system optimization.

[0295] Delay mapping: If < This indicates that the current information is insufficient to make a reliable mapping, and the project is marked as "pending" for further in-depth analysis or data quality governance.

[0296] (2) Human-machine collaboration optimization: The results of manual review, as high-quality labeled data, are fed back to the system, forming a closed loop of continuous optimization. ① Knowledge base enhancement: The verified mapping pairs can be used to expand and revise the thesaurus and standardization rules, improving the accuracy of attribute extraction and standardization.

[0297] ② Model parameter tuning: The review results can be used as training data to optimize the weight parameters of each matching module (e.g., ) and decision threshold ( , This makes the system's decisions more aligned with the actual data distribution.

[0298] ③ Strategy Evaluation: By analyzing the distribution of manually reviewed cases, the effectiveness of the candidate generation strategy can be evaluated and optimized.

[0299] Through the aforementioned dynamic weighted fusion and hierarchical decision-making mechanism, this framework not only achieves highly robust automatic mapping, but also constructs a terminology mapping system that can continuously evolve and become more intelligent with use through human-machine collaboration.

[0300] In the implementation of the above embodiments: by comparing the difference in the comprehensive matching degree between the preferred candidate and the second-best candidate, arbitration is initiated from the clinical priority rule base when their scores are close, avoiding the decision-making difficulties when scores alone cannot distinguish between them. This design transforms clinical experience such as methodological priority and name specificity into executable judgment logic, and determines the target standard item based on domain knowledge when score evidence is insufficient, significantly improving the clinical rationality and accuracy of the mapping results.

[0301] Optionally, in this embodiment, if the overall matching degree of the source project does not meet the requirements, an audit task is generated based on the matching information between the source project and the candidate standard project set; using the correct mapping pairs confirmed by the audit task, the weight parameters in the overall matching degree calculation are optimized through a ranking learning algorithm.

[0302] Another core advantage of this application's embodiments lies in the learnability of its parameters. Initial parameter settings (such as the basic weights of each module) Search threshold (e.g.,) may rely on domain experience or grid search. To enable the framework to adapt to the characteristics of data from different institutions, we propose a supervised learning-based automatic parameter optimization mechanism, which enables it to learn the optimal configuration from labeled data, thereby improving mapping performance and reducing reliance on manual parameter tuning.

[0303] We formalize the parameter learning problem as a supervised learning task. Training data: Assume we have a set of mapping pairs annotated by experts. in It is the source project The corresponding correct standard item. Also, for each We construct a set of negative samples This includes standard items that were not marked as matching.

[0304] Learning objective: Our goal is to learn a set of global parameters. (include , , ), such that for every positive sample pair in the training data ( ), its overall matching score It should be significantly higher than it compared to any negative sample. The score.

[0305] Specifically, Θ mainly includes the following types of parameters: the basic weights of each matching module, retrieval and decision thresholds, and other learnable model parameters.

[0306] The basic weights of each matching module are as follows: ① Weights of the dimensional matching module: set {ω_d}_{d∈D}, which is the basic weight ω_d of each dimension (such as detected components, sample types, etc.) mentioned in this method. ② Weights of other modules: set {ω_m}_{m∈M}, which includes the basic weights of other modules besides dimensional matching (e.g., possibly including text semantic matching module, reference range matching module, etc.).

[0307] Retrieval and decision thresholds: ① Text similarity threshold: τ_text, which is explicitly mentioned in this method, is used to control the strictness of text semantic retrieval (path three).

[0308] Other learnable model parameters: ① The parameter set Θ mentioned in this method may also include other relevant model parameters that affect the calculation of the comprehensive matching score. Specifically, this may include: when calculating text similarity, if a neural network model (such as Sentence-BERT) is used, its model weights can also be optimized; when calculating string similarity (such as for abbreviation matching), if a learnable similarity metric is used, its parameters can also be optimized; other adjustable hyperparameters or model parameters introduced in the matching score calculation process can also be optimized.

[0309] We use the pairwise method from Learning to Rank (LTR) for optimization, which aims to minimize the loss due to ranking errors caused by positive and negative samples.

[0310] Score difference and probabilization: For a training triplet ( We calculate the score difference:

[0311] The score difference is mapped using the logistic function. Sort by The probability of:

[0312] Loss function: The widely used cross-entropy loss is employed as the optimization objective. For the entire training set, the loss function is defined as follows:

[0313] in It is all possible triples ( The training set consists of .

[0314] Optimization process: Minimize the loss function using gradient descent (e.g., Adam). .parameter The gradient can be calculated through backpropagation, due to the overall matching degree. It's about parameters. The function is a differentiable function (provided that the underlying text similarity calculations are also differentiable, such as using a neural network model), so the entire learning process is end-to-end.

[0315] This parameter learning module integrates seamlessly with the aforementioned framework, forming a complete "training-inference" closed loop: Offline training: In the early stages of system deployment or periodically (e.g., quarterly), it utilizes accumulated human review results. The learned parameters are used as training data to run a parameter learning algorithm, resulting in an optimized parameter set Θ. Online inference: During the automatic mapping phase, the system uses the learned optimal parameters. The system performs candidate generation, matching degree calculation, and decision-making. Newly generated manually reviewed data can be continuously added to the training set, enabling the system to adapt to the addition or change of terms and achieve continuous self-optimization.

[0316] The beneficial effects of this application include, but are not limited to: Data-driven and adaptive: The parameters are learned through data rather than being fixed by humans, enabling the framework to adapt to differences in terminology and data distribution among different medical institutions, thus improving its generalization ability.

[0317] Reduced reliance on manual intervention: Significantly reduces the subjectivity and workload caused by manual parameter tuning, making the framework easier to promote and apply in different scenarios.

[0318] Verifiable innovation: By designing comparative experiments (such as “learned weights” vs. “manually set weights”), the performance improvement brought about by data-driven learning can be empirically verified, which is a key contribution of the paper.

[0319] In the implementation of the above embodiments: the correctly mapped pairs confirmed by manual review are used as training data. The weight parameters in the comprehensive matching degree calculation are optimized through a ranking learning algorithm, so that the weights of each dimension and module are more in line with the real data distribution. The human-machine collaborative closed loop enables the system to continuously learn from the results of manual review, constantly improving the accuracy of the comprehensive matching degree calculation, thereby effectively improving the overall mapping accuracy of cross-institutional projects in subsequent mappings.

[0320] Please see Figure 2 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. An electronic device 300 provided in this application includes a processor 310 and a memory 320. The memory 320 stores machine-readable instructions executable by the processor 310. When the machine-readable instructions are executed by the processor 310, the method described above is performed.

[0321] Figure 2 The components shown can be implemented using hardware, software, or a combination thereof. Electronic device 300 may be a physical device, such as a server or PC, or a virtual device, such as a virtual machine or virtualization container. Furthermore, electronic device 300 is not limited to a single device; it can be a combination of multiple devices or a cluster of numerous devices.

[0322] This application also provides a storage medium storing a computer program, which is executed by a processor to perform the above-described method.

[0323] The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0324] This application also provides a computer program product, including computer program instructions, which are executed by a processor to perform the method described above.

[0325] It should be understood that the disclosed apparatus and methods can also be implemented in other ways, given the several embodiments provided in this application. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0326] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0327] The above description is only an optional implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application.

Claims

1. A method for mapping cross-institutional healthcare projects, characterized in that, include: The source items in the source item set and the standard items in the standard item set are represented as multi-dimensional attribute vectors, and the values ​​of the multi-dimensional attribute vectors of the source items are mapped to the standard value space of the corresponding dimension using a normalization mapping function to generate a normalized attribute vector; wherein, one dimension of the multi-dimensional attribute vector represents a data type; For the source items, a candidate standard item set is determined from the standard item set using a multi-search path strategy; the multi-search path strategy includes at least one of core attribute strong constraint retrieval, auxiliary attribute supplementary retrieval, and text semantic fallback retrieval. Determine the effective observation dimension set between the source item and the candidate standard item in the candidate standard item set, calculate the matching score of each dimension in the effective observation dimension set, and calculate the comprehensive matching degree between the source item and the corresponding candidate standard item based on the matching score; the dimensions of the effective observation dimension set include dimensions in the standardized attribute vectors of the source item and the candidate standard item where all values ​​are non-empty. Based on the overall matching degree, a target standard project corresponding to the source project is selected from the candidate standard project set, and if the overall matching degree meets the requirements, the source project is mapped to the target standard project.

2. The method according to claim 1, characterized in that, The values ​​of the multidimensional attribute vectors of the source project are mapped to the standard value space of the corresponding dimension using a normalization mapping function to generate standardized attribute vectors, including: Using a predefined extraction function, the original values ​​of the multidimensional attribute vector of the source item in each dimension are extracted from the original text and / or structured fields of the source item. Obtain the standardized mapping function for each dimension, and map the original values ​​to a standard value in a predefined standard value space using the standardized mapping function to generate a standardized attribute vector; wherein, in the standard value space, multiple different original values ​​that are semantically equivalent are mapped to the same standard value.

3. The method according to claim 2, characterized in that, Before mapping the original value to a standard value within a predefined standard value space using the normalization mapping function, the method further includes: Obtain a predefined set of initial standard values; The initial set of standard values ​​is updated through a human-machine collaborative feedback mechanism to generate the standard value space, and the prior weights in the standardization mapping function that map the original values ​​to the standard values ​​are adjusted.

4. The method according to claim 1, characterized in that, The conditions for determining the execution order of different strategies in the multi-retrieval path strategy include: The information completeness of the core or auxiliary attributes of the source project, whether the number of candidate results returned by the strong constraint retrieval of the core attributes meets the preset threshold, and the emphasis requirements of the current business scenario on mapping accuracy or recall rate, are all considered.

5. The method according to claim 1, characterized in that, Multi-search path strategies include retrieval with strong constraints based on core attributes; A candidate set of standard items is determined from the set of standard items using a multi-retrieval path strategy, including: If at least one core attribute in the core attribute set of the source project has a non-empty value in the standardized attribute vector, then the core attribute strong constraint retrieval is enabled: Extract the first standardized value belonging to the core attribute set from the standardized attribute vector of the source project; In the set of standard items, retrieve the first standard item whose value of the standardized attribute vector on the same core attribute set is consistent with the first standardized value; Merge all the first standard items into a subset corresponding to the core attribute strongly constrained retrieval path in the candidate standard item set; The method further includes disabling strong constraint retrieval of the core attributes when all core attribute values ​​are missing in the standardized attribute vector.

6. The method according to claim 1, characterized in that, The multi-retrieval path strategy includes supplementary retrieval using auxiliary attributes; determining a candidate standard item set from the standard item set using the multi-retrieval path strategy includes: If the core attribute strong constraint retrieval returns empty candidate results or the number is less than a first preset threshold, and at least one auxiliary attribute in the auxiliary attribute set of the source project has a non-empty value in the standardized attribute vector, then the auxiliary attribute supplementary retrieval is enabled: Extract the second standardized value belonging to the auxiliary attribute set from the standardized attribute vector of the source project; For the identified auxiliary attribute, in the standard item set, a second standard item whose value of the standardized attribute vector on the same auxiliary attribute set is consistent with the second standardized value is retrieved; All second standard items are merged into a subset corresponding to the auxiliary attribute supplementary retrieval path in the candidate standard item set; The method further includes disabling the supplementary retrieval of auxiliary attributes when all auxiliary attribute values ​​are missing in the standardized attribute vector.

7. The method according to claim 1, characterized in that, Multi-path retrieval strategies include textual semantic fallback retrieval; A candidate set of standard items is determined from the set of standard items using a multi-retrieval path strategy, including: If the candidate standard item set determined by the core attribute strong constraint retrieval and the auxiliary attribute supplementary retrieval is empty, the source item contains preset characters, or the source item belongs to a preset business scenario, then the text semantic fallback retrieval is initiated: Obtain the first complete text name of the source project, the second complete text name of the standard project, and a set of synonyms for the second complete text name; Based on the text similarity calculation model, the semantic similarity between the first complete text name and the second complete text name, as well as between the first complete text name and the elements in the synonym set, is calculated. The third standard items whose semantic similarity exceeds a preset threshold, and / or the fourth standard items whose first complete text name synonym set satisfies the keyword overlap condition, are selected to form a subset of the text semantic fallback retrieval path in the candidate standard item set.

8. The method according to claim 1, characterized in that, Calculate the matching score for each dimension in the effective observation dimension set, including: For each dimension in the effective observation dimension set, if the standardized value of the source item is the same as the standardized value of the candidate standard item, then the score of the dimension is determined as the first value. If the standardized value of the source item and the standardized value of the candidate standard item have a predefined semantic compatibility relationship, then the score of the dimension is determined to be a constant between the second value and the first value; If there is a predefined conflict between the standardized value of the source item and the standardized value of the candidate standard item, then the score of the dimension is the third value.

9. The method according to claim 8, characterized in that, If the dimension is an English abbreviation dimension, the process of calculating the matching score for the English abbreviation dimension includes: Determine whether the English abbreviation in the source project is consistent with the English abbreviation in the candidate standard project; If they match, the score for the English abbreviation dimension will be determined as the first value; If they are inconsistent, the string similarity between the English abbreviation in the source project and the English abbreviation in the candidate standard project is calculated, and the string similarity is used as the score for the English abbreviation dimension.

10. The method according to claim 1, characterized in that, Calculating the overall matching degree between the source item and the corresponding candidate standard item based on the matching score includes: The matching scores of all dimensions in the effective observation dimension set are weighted and averaged according to the preset basic weights of the dimensions to obtain the dimension matching score. The dynamic effective weights of the dimension matching scores are obtained, and the dynamic effective weights are determined based on the preset basic weights and the dimension coverage rate; the dimension coverage rate is the ratio of the number of dimensions in the effective observation dimension set to the total number of dimensions. The dimension matching score and its dynamic effective weight are weighted and fused with the text similarity score to obtain the comprehensive matching degree; the text similarity score is obtained based on residual text, which is determined by removing identified terms belonging to the predefined clinical dimension from the normalized names of the source item and the candidate standard item.

11. The method according to claim 1, characterized in that, Based on the comprehensive matching degree, the target standard project corresponding to the source project is selected from the candidate standard project set, including: From the set of candidate standard projects, select the candidate standard project with the highest overall matching degree with the source project as the preferred candidate; If the difference between the overall matching degree of the preferred candidate and the second-best candidate is less than a preset conflict threshold, then the preferred candidate and the second-best candidate are arbitrated according to a predefined clinical priority rule base to determine the target standard item.

12. The method according to claim 11, characterized in that, The conditions under which the overall matching degree meets the requirements include: The overall matching degree is greater than or equal to the preset confidence threshold; Alternatively, the overall matching degree is less than the preset confidence threshold and greater than or equal to the dynamic decision threshold; wherein the dynamic decision threshold is determined based on the difference between the overall matching degree of the preferred candidate and the overall matching degree of the second-best candidate.

13. The method according to claim 1, characterized in that, If the overall matching degree of the source project does not meet the requirements, an audit task is generated based on the matching information between the source project and the candidate standard project set; using the correct mapping pairs confirmed by the audit task, the weight parameters in the overall matching degree calculation are optimized through a ranking learning algorithm.

14. A computer program product, characterized in that, It includes computer program instructions that are executed by a processor at runtime as the method described in any one of claims 1 to 13.

15. An electronic device, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, perform the method as described in any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, perform the method as described in any one of claims 1 to 13.

Images