A medical terminology standardization quality intelligent review method and device, electronic equipment and storage medium

Abstract

The present application relates to the medical information technology field, disclose a kind of medical terminology standardization quality intelligent examination method, device, electronic equipment and storage medium. Obtain the multi-source standardization data containing original terminology, standardization result and source identification, construct conflict term matrix;Original terminology in the matrix is normalized pretreatment, extract terminology core and associated clinical modification information, generate the conflict pair list after sorting;All medical terms are encoded into semantic vectors, and the four-way cross-validation is performed on the candidate confusion pairs, and the risk classification is combined with the clinical modification information;Fusion conflict pair list and risk classification results generate quality review report.The present application can effectively improve the efficiency and accuracy of medical terminology standardization quality review.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method, device, electronic device, and storage medium for intelligent review of the standardization quality of medical terminology. Background Technology

[0002] With the accelerated advancement of medical informatization, medical terminology standardization has become a crucial foundation for clinical data governance. Mapping unstructured medical terms from free clinical texts to a unified standard terminology system is a necessary prerequisite for achieving medical data interoperability and supporting subsequent clinical decision analysis and research applications. However, against the backdrop of the continuously expanding scale of medical big data, existing technological research mainly focuses on optimizing the mapping accuracy of the terminology standardization process itself, neglecting to implement efficient and reliable quality control mechanisms for the standardized results. This deficiency leads to potential quality risks in standardized medical terminology resources, making it difficult to meet the needs of high-reliability applications of medical data. Currently, the quality verification of medical terminology standardization results generally relies on manual sampling. However, due to the massive and continuously growing volume of data to be processed, manual review not only consumes a large amount of human resources but also faces the problem of highly subjective review standards: different reviewers have individual differences in their understanding of the medical standard system, and the same person may have fluctuating judgment standards at different working times, resulting in unstable quality assessment results. Meanwhile, existing automated verification technologies lack a deep understanding of medical semantics, failing to effectively identify legitimate terminology variations in clinical scenarios (such as reasonable modifications based on anatomical location or pathological features), and easily misjudging such legitimate differences as standardization errors. Conversely, existing methods cannot accurately capture hidden mapping errors that are difficult to detect in isolation (such as standardization bias caused by contextual semantic conflicts) through multi-dimensional cross-validation. These limitations result in significant loopholes in the quality assurance system of standardized terminology resources, severely restricting the application value of medical data in key areas such as clinical research and public health management. Therefore, an innovative mechanism that integrates medical semantic analysis and automated quality control is urgently needed to address the core issues of low efficiency and insufficient accuracy in the quality inspection of standardized results. Summary of the Invention

[0003] The purpose of this invention is to provide a method, device, electronic device and storage medium for intelligent review of the standardization quality of medical terminology, which can improve the efficiency and accuracy of the review of the standardization quality of medical terminology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent review of the standardization quality of medical terminology, comprising: Acquire multi-source standardized data, which contains standardized records from different sources. Each standardized record contains the original terminology, the standardized result, and the source identifier. A conflict term matrix is ​​constructed based on multi-source standardized data. The conflict term matrix contains the original terms with multi-standardization results and their corresponding quantitative analysis results. The original terms in the conflict terminology matrix are preprocessed for standardization to extract the standardized terminology core and associated clinical modification information; Generate a sorted list of conflict pairs based on the conflict term matrix; Semantic vector encoding is performed on all medical terms to obtain normalized sparse semantic vectors corresponding to each term. Four-way cross-validation is performed on candidate confusion pairs based on semantic vectors to obtain cross-validation results; Risk stratification was performed based on cross-validation results and clinical modification information to obtain risk stratification results; A quality review report is generated by merging the conflict pair list with the risk classification results.

[0005] Furthermore, a conflict terminology matrix is ​​constructed based on multi-source standardized data, including: Batch parsing of multi-source standardized data is performed to extract the original terms, standardized results and source file identifiers from each standardized record, and an initial dataset is constructed. The standardized results are subjected to one-to-many mapping standardization processing, which splits the composite standardized results containing delimiters into independent standardized terms, and then reassembles them into standardized strings after deduplication and lexicographical sorting. Aggregate the standardized data using the original terms as keys, and calculate the number of unique standardized results, the distribution entropy of the standardized results, the main standardized results, and the list of source documents for each original term; Records with a unique number of standardized results greater than 1 are selected to construct a conflict term matrix.

[0006] Furthermore, the original terms in the conflict terminology matrix undergo normalization preprocessing, including: The original terms are truncated by the position of the first question mark in the term, retaining the core description part before the question mark; Recursive noise removal is performed on the truncated terms, removing the numerical codes, punctuation marks, and non-semantic suffixes at the beginning and end of the terms, while separating and extracting clinical modifiers that conform to the preset medical dimension pattern. If the terminology core is empty after preprocessing, then revert to the original complete terminology to obtain the standardized terminology core and associated clinical modification information.

[0007] Furthermore, a sorted list of conflict pairs is generated based on the conflict term matrix, including: For each original term with multiple standardized results, all its different standardized results are combined pairwise to generate conflict pairs; For each conflict pair, a priority score is calculated. The parameters for calculating the priority score include the entropy value of the original term, the frequency of occurrence of the original term, the frequency difference of the conflict pair standardization results, and the distribution of the source documents of the conflict results. All conflict pairs are sorted in descending order based on priority scores to obtain a sorted list of conflict pairs.

[0008] Furthermore, semantic vector encoding is performed on all medical terms to obtain the normalized sparse semantic vectors corresponding to each term, including: Initialize a zero vector of a specified dimension and perform validity checks on the input terms; The input terms are subjected to multi-level feature hashing and weighting. Character-level features, N-gram-level features, and prefix-suffix structure-level features are hashed and mapped and their corresponding weights are accumulated. The weighted vector is subjected to Top-K sparse compression, retaining only the K dimensions with the largest weights and setting the remaining dimensions to zero. The sparse compressed vector is normalized using the L2 norm to obtain a normalized sparse semantic vector.

[0009] Furthermore, semantic vector encoding can also be implemented using any of the following methods: pre-trained medical domain word vector models, TF-IDF combined with N-gram features, or character-level neural network encoders.

[0010] Furthermore, four-way cross-validation is performed on the candidate confusion pairs based on semantic vectors, including: The original terms are initially screened based on character-level similarity, and the original term pairs with character similarity exceeding a preset threshold are listed as candidate confusion pairs. For each candidate confusion pair, a four-way cross similarity is calculated based on the corresponding sparse semantic vector. The four-way cross similarity includes the first similarity between the original term A and its own standardized result A', the second similarity between the original term A and the standardized result B' of the original term B, the third similarity between the original term B and its own standardized result B', and the fourth similarity between the original term B and the standardized result A' of the original term A. Calculate the fusion similarity of candidate confusion pairs. The fusion similarity is the weighted fusion result of character similarity and semantic similarity.

[0011] Furthermore, risk stratification is performed based on cross-validation results and clinical modification information, including: Call the preset knowledge base of legal differences in medical terminology. The knowledge base contains clinically recognized legal variation dimensions and their value sets. For the original terms and standardized results in the candidate confusion pairs, the corresponding dimensional vocabulary is extracted, dimensional consistency is checked, and the dimensional consistency, dimensional conflict, or lack of dimensional information is marked.

[0012] Furthermore, risk stratification based on cross-validation results and clinical modification information also includes: Based on the four-way cross-similarity and dimensionality consistency verification results, a risk score is calculated, and the risk score satisfies: (, dimensional consistency flag), where The first similarity score, For the second similarity, The third similarity The fourth similarity score; Based on the risk score range, candidate confusion pairs are labeled as high-risk, medium-risk, or low-risk.

[0013] Furthermore, a quality review report is generated by merging the conflict pair list and risk classification results, including: Deduplication is performed on the intra-term conflict list and the inter-term confusion list respectively. The conflict list records are merged using the original terms as the key, and the confusion list records are merged using the normalized original term pairs as the key. The issue records are sorted based on preset priority rules, which include intra-term conflicts taking precedence over inter-term confusions, larger issue clusters having higher priority, and issues containing multiple types of evidence taking precedence over issues containing a single type of evidence. A structured quality review report is generated based on the aggregated issue records. The report includes issue cluster identifiers, core issue descriptions, issue details, and terminology statistics.

[0014] This invention also proposes a medical terminology standardization quality intelligent review device, comprising: The data acquisition module is used to acquire multi-source standardized data, which includes standardized records from different sources. Each standardized record includes the original terminology, the standardized result, and the source identifier. The conflict construction module is used to construct a conflict term matrix based on multi-source standardized data. The conflict term matrix contains the original terms with multi-standardization results and their corresponding quantitative analysis results. The preprocessing module is used to perform normalization preprocessing on the original terms in the conflict terminology matrix, and extract the standardized terminology core and associated clinical modification information. The conflict sorting module is used to generate a sorted list of conflict pairs based on the conflict term matrix; The vector encoding module is used to perform semantic vector encoding on all medical terms to obtain the normalized sparse semantic vectors corresponding to each term. The cross-validation module is used to perform four-way cross-validation calculations on candidate confusion pairs based on semantic vectors to obtain the cross-validation results. The risk stratification module is used to perform risk stratification based on cross-validation results and clinical modification information to obtain the risk stratification results. The report generation module is used to generate a quality review report based on the fusion of the conflict pair list and risk classification results.

[0015] The present invention also proposes an electronic device comprising a memory and a processor coupled to each other, wherein the memory stores program instructions and the processor executes the program instructions to implement the above-mentioned intelligent review method for standardized quality of medical terminology.

[0016] The present invention also proposes a computer-readable storage medium storing program instructions that can be executed by a processor, the program instructions being used to implement the above-mentioned intelligent review method for standardized quality of medical terminology.

[0017] Compared with existing technologies, the present invention provides a method, device, electronic device and storage medium for intelligent quality review of medical terminology standardization. By automatically acquiring multi-source standardized data, constructing a conflict term matrix, performing standardized preprocessing, generating a sorted list of conflict pairs, performing semantic vector encoding, implementing four-way cross-validation calculation, completing risk classification and generating a quality review report, it achieves efficient and accurate quality review of medical terminology standardization, and has the advantages of improving the efficiency and accuracy of medical terminology standardization quality review.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0020] Figure 1 This is a schematic diagram of a medical terminology standardization quality intelligent review method provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of a medical terminology standardization quality intelligent review device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of this invention. The components of this invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0023] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] like Figure 1 As shown, this embodiment provides a method for intelligent review of the standardization quality of medical terminology. First, multi-source standardized data is acquired. This data can be collected from different information systems, databases, or through manual annotation. For example, standardized tables provided by different hospitals can be manually merged to form a preliminary dataset containing the original terms, their corresponding standardized results, and source tags.

[0025] Optionally, a conflict term matrix can be constructed based on the multi-source standardized data. After acquiring the multi-source standardized data, a preliminary scan can be performed to identify the original terms with different standardized results. For example, by traversing all standardized records, each original term and its corresponding standardized results can be recorded in a simple list. If an original term is found to correspond to two or more different standardized results, it is marked as a conflict term, and all its standardized results are recorded together. This process can form a preliminary set of conflict terms, where each conflict term is accompanied by all its different standardized results.

[0026] Subsequently, the original terms in the conflict terminology matrix undergo normalization preprocessing to extract standardized terminology cores and associated clinical modification information. Basic text cleaning operations can be performed on the original terms in the conflict terminology matrix. For example, common special characters, numbers, or general stop words are removed. This processing aims to simplify terminology expression, making it easier for subsequent analysis. Through this processing, a relatively clean terminology form can be obtained as the terminology core. For some terms containing descriptive words, such as "left-sided pneumonia," "left-sided" can be manually or through simple rules identified as clinical modification information, with "pneumonia" serving as the terminology core.

[0027] Optionally, a sorted list of conflict pairs is generated based on the conflict term matrix. After the conflict term matrix is ​​constructed, for each original term with multiple standardized results, all its different standardized results can be combined pairwise to generate a series of conflict pairs. For example, if the original term A has standardized results B and C, then conflict pair (B, C) is generated. These conflict pairs can be simply listed to form a preliminary list of conflict pairs. The sorting of this list can be based on the frequency of occurrence of the conflict pairs; for example, conflict pairs that occur more frequently are listed first.

[0028] Simultaneously, semantic vector encoding is performed on all medical terms to obtain normalized sparse semantic vectors for each term. To capture the semantic information of medical terms, a basic vectorization method can be employed. For example, a simple bag-of-words model vector can be created for each term, where each dimension of the vector corresponds to a word in the vocabulary and records the frequency of that word in the term. Subsequently, these bag-of-words vectors undergo simple normalization, such as dividing each element of the vector by its maximum value, to obtain a preliminary normalized vector. This vector can be considered a sparse semantic vector because the vocabulary is typically large, while a single term contains only a limited number of words.

[0029] Optionally, four-way cross-validation is performed on the candidate confusion pairs based on the semantic vector to obtain the cross-validation results. After obtaining the semantic vectors of the terms, potential confusing term pairs can be identified. For example, term pairs with close semantic vector distances can be selected as candidate confusion pairs. For each candidate confusion pair, such as the original term A and its standardized result A', and the original term B and its standardized result B', four basic similarities can be calculated: the similarity between A and A', the similarity between A and B', the similarity between B and B', and the similarity between B and A'. These similarities can be obtained by calculating the cosine similarity or Euclidean distance between the corresponding semantic vectors. These calculation results constitute the preliminary cross-validation results.

[0030] Subsequently, risk grading is performed based on the cross-validation results and the extracted clinical modification information to obtain the risk grading results. Combining the cross-validation results and the previously extracted clinical modification information, a risk assessment can be performed on potential standardization issues. For example, if cross-validation shows that two terms have low semantic similarity and their clinical modification information differs significantly (e.g., one term describes "Type I" and the other describes "Type II"), they can be marked as high-risk. Conversely, if the semantic similarity is high and the clinical modification information is consistent or the differences do not affect the core semantics, the risk is low. This assessment can be based on pre-set simple rules, such as setting similarity thresholds and modification information matching rules to classify issues into high, medium, and low risk levels.

[0031] Finally, a quality review report is generated by merging the conflict pair list with the risk grading results. The previously generated conflict pair list and risk grading results are integrated to produce a quality review report. This report can simply list all identified conflict and confusion pairs, along with their corresponding risk levels. For example, the report could contain a table where each row represents an issue, listing the original term, its conflict or confusion standardization result, and the risk level of the issue. The report aims to provide an overview so that human reviewers can quickly understand the problems existing in the standardization process.

[0032] In some possible embodiments, a specific method for constructing a conflict term matrix based on multi-source standardized data is proposed, including the following steps: First, batch parsing of the multi-source standardized data is performed to extract the original terms, standardized results, and source file identifiers from each standardized record, constructing an initial dataset. This step aims to uniformly extract key information from heterogeneous multi-source standardized data. Multi-source standardized data may originate from different databases, text files, or API interfaces, and its format may include CSV, JSON, XML, or custom text formats. Batch parsing can employ predefined parsing rules, regular expression matching, or machine learning-based entity extraction techniques to identify and extract the original terms, their corresponding standardized results, and unique identifiers of the data sources. The extracted information is then organized into a structured initial dataset, such as stored in a table, where each row represents a standardized record, containing fields such as the original terms, standardized results, and source file identifiers, laying the data foundation for subsequent processing.

[0033] Secondly, the standardized results undergo one-to-many mapping standardization, splitting the composite standardized results containing delimiters into independent standardized terms. After deduplication and lexicographical sorting, these terms are recombined into a standardized string. This process aims to eliminate inconsistencies in the representation of standardized results, ensuring that identical standardized results from different sources can be accurately identified. For example, an original term might be standardized as "hypertension / primary hypertension" or "primary hypertension; hypertension". This step identifies delimiters such as " / " or ";", splitting the composite result into two independent terms: "hypertension" and "primary hypertension". Subsequently, deduplication is performed on these independent terms to avoid duplicate counting. Finally, the deduplicated independent standardized terms are sorted lexicographically and recombined into a unique standardized string (e.g., "hypertension; primary hypertension"), ensuring a consistent standardized representation even if the order or delimiters of the original standardized results differ, as long as they contain the same set of independent terms.

[0034] Next, the standardized data is aggregated using the original terms as keys, and the number of unique standardized results, the entropy value of the standardized result distribution, the master standardized result, and the list of source documents are calculated for each original term. The core of this step lies in the quantitative analysis of the standardization status of each original term. The system traverses the initial dataset, grouping data by original terms as unique identifiers. For each group of original terms, the number of types of its corresponding standardized strings (i.e., the results after one-to-many mapping standardization) is counted; this is the number of unique standardized results. Simultaneously, the entropy value of the standardized result distribution is calculated. This entropy value reflects the degree of uncertainty or diversity in the standardized results of the original term; a higher entropy value indicates a greater degree of conflict. Furthermore, the most frequently occurring standardized result is identified as the master standardized result, and all source document identifiers involving that original term are compiled to form a list of source documents, providing a basis for subsequent conflict tracing.

[0035] Finally, records with more than one unique standardized result are selected to construct a conflict term matrix. After completing the above aggregation and quantitative analysis, the system will filter according to preset conditions. Specifically, only when the number of unique standardized results corresponding to a certain original term is greater than one is the original term considered to have a standardization conflict. These conflicting original terms and their corresponding quantitative analysis results (including the number of unique standardized results, the entropy value of the standardized result distribution, the main standardized result, and the list of source documents, etc.) will be extracted and organized into a structured conflict term matrix. This matrix clearly shows which original terms have multiple standardized results and the quantitative characteristics of these conflicts, providing clear review targets for subsequent normalization preprocessing and quality review.

[0036] Optionally, after constructing a conflict term matrix based on multi-source standardized data, the original terms may contain questionable information, redundant characters, or non-core descriptions. These factors can interfere with subsequent semantic analysis and conflict pair generation. Without effective processing, this may lead to inaccurate identification of core terms, affecting the accuracy of standardized quality review. Therefore, a mechanism is needed to refine these original terms to extract their core semantics and separate useful clinical modification information.

[0037] To address the aforementioned issues, this embodiment performs normalization preprocessing on the original terms in the conflict terminology matrix to extract standardized terminology cores and associated clinical modification information. Specifically, this preprocessing includes the following steps: First, the original terms are truncated for interrogative information. Medical terminology often contains parentheses, question marks, or other separators representing interrogative or supplementary information, which is not necessarily the core semantics of the term. Therefore, this step identifies the position of the first interrogative marker in the term and truncates it accordingly, retaining only the core descriptive part before the interrogative marker. For example, the term "hypertension (is it primary?)" will, after interrogative information truncation, become simply "hypertension," thus focusing on the core concept of the term.

[0038] Secondly, recursive noise reduction is performed on the truncated terms. This step aims to further purify the terms and remove non-semantic interference elements. This includes removing any numerical prefixes (e.g., "1." in "1. pneumonia"), punctuation marks (e.g., "." in "pneumonia."), and other non-semantic suffixes (e.g., "-diagnosis" in "pneumonia-diagnosis"). Simultaneously, clinical modifiers conforming to preset medical dimension patterns are separated and extracted during this process. These modifiers (such as "acute," "chronic," "left," and "right"), while not forming the core of the terminology, carry important clinical contextual information, such as the course, location, or severity of the disease. Through preset medical dimension knowledge or pattern matching, these modifiers are identified and stored as independent related information for subsequent refined analysis.

[0039] Finally, to ensure the robustness of the preprocessing, if the terminology core obtained after the above truncation and noise reduction is empty, the system will revert to the original complete terminology. This revert mechanism avoids the complete loss of terminology information due to overprocessing, ensuring that the original data is preserved even under extreme or anomalous terminology formats, thereby obtaining standardized terminology cores and associated clinical modification information.

[0040] In some possible embodiments, a sorted list of conflict pairs is proposed based on a conflict term matrix. Specifically, this includes: for each original term with multiple standardized results, combining all its different standardized results in pairs to generate conflict pairs; calculating a priority score for each conflict pair, the calculation parameters of which include the entropy value of the original term, the frequency of occurrence of the original term, the frequency difference of the standardized results of the conflict pair, and the distribution of the source documents of the conflict results; and sorting all conflict pairs in descending order according to the priority score to obtain a sorted list of conflict pairs.

[0041] Specifically, for each original term identified in the conflict term matrix that has multiple standardized results, the system combines all its different standardized results in pairs to generate specific conflict pairs. For example, if an original term has standardized results A, B, and C, then conflict pairs (A,B), (A,C), and (B,C) will be generated. This pairwise combination method refines the multiple conflicts under an original term into a series of smallest conflict units that can be independently analyzed and reviewed, making subsequent review work more focused and specific.

[0042] To quantify the importance of each conflict pair, this embodiment calculates a priority score for each pair. This score comprehensively considers multiple dimensions to ensure a thorough assessment of the conflict. The entropy value of the original term reflects the degree of disorder or uncertainty in the distribution of the standardized results of the original term. A higher entropy value indicates a more dispersed standardized result for the original term, with no clearly dominant standardized result. This usually means that the standardization problem of the term is more complex and widespread, and therefore its corresponding conflict pair should have a higher priority. The frequency of occurrence of the original term represents the total number of times the original term appears in the multi-source standardized data. The higher the frequency of occurrence of the original term, the wider the scope of its standardized conflict impact and the greater its impact on the overall data quality; therefore, its corresponding conflict pair should have a higher priority. The frequency difference of the standardized results of the conflict pair measures the difference in frequency of occurrence of the two standardized results under the original term within the conflict pair. For example, for the conflict pair (A, B), the absolute difference between the frequency of occurrence of standardized result A and the frequency of occurrence of standardized result B is calculated. Smaller frequency differences may indicate that both standardized results have some support, making the conflict more "stubborn" or difficult to resolve, thus potentially requiring higher priority for manual review. The distribution of source documents for conflicting results statistically indicates how many different source documents or systems each standardized result in the conflict pair originates from. If the two standardized results of a conflict pair come from multiple different source documents, it suggests that the conflict is a widespread, cross-system, and cross-departmental issue with a wider impact and potentially greater difficulty in resolution, thus warranting higher priority. These parameters can be weighted and combined using preset weights or machine learning models to generate the final priority score.

[0043] After calculating the priority scores for all conflict pairs, the system sorts them in descending order based on these scores. This means that the conflict pairs with the highest priority scores will be placed at the top of the list, allowing reviewers to focus on and address the most important and urgent quality issues first. The final result is a sorted list of conflict pairs, providing clear and efficient guidance for subsequent quality reviews.

[0044] In some possible embodiments, a method is proposed to perform semantic vector encoding on all medical terms to obtain normalized sparse semantic vectors corresponding to each term. This process includes: Initialize a zero vector of a specified dimension and perform validity checks on the input terms; The input terms are subjected to multi-level feature hashing and weighting. Character-level features, N-gram-level features, and prefix-suffix structure-level features are hashed and mapped and their corresponding weights are accumulated. The weighted vector is subjected to Top-K sparse compression, retaining only the K dimensions with the largest weights and setting the remaining dimensions to zero. The sparse compressed vector is normalized using the L2 norm to obtain a normalized sparse semantic vector.

[0045] Specifically, when performing semantic vector encoding on all medical terms, it is first necessary to initialize a zero vector of a specified dimension and then validate the input terms. The purpose of initializing the zero vector is to provide each medical term with a fixed-length, initially zero-space representation. Its dimension can be flexibly set according to the size of the medical terminology corpus, the semantic complexity, and available computing resources. Validation aims to ensure that the input medical terms conform to preset data formats and content specifications. For example, it checks whether the term is an empty string, contains illegal special characters, or exceeds a predetermined length limit, thereby preventing invalid or abnormal data from interfering with the subsequent semantic encoding process and ensuring the accuracy and stability of the encoding.

[0046] Subsequently, the input terms, after validity verification, undergo multi-level feature hashing and weighting. This step is the core of semantic vector encoding, aiming to comprehensively capture the semantic information of medical terms from different granularities. Specifically, it includes: Hash mapping is applied to character-level features and corresponding weights are accumulated: Character-level features refer to individual characters or Chinese characters that constitute a term. By hash mapping these basic constituent elements, the most fundamental morphological information of the term can be captured, and even spelling errors or subtle variations can be partially identified.

[0047] Hash mapping and weighting of N-gram features: N-gram features refer to N consecutive characters or segments of words in a term. For example, for the term "myocardial infarction," its 2-gram features might include "myocardium," "myocardial infarction," and "infarction." This feature can capture local sequence patterns and common phrase structures within a term, which is of great significance for identifying word roots and components of medical terminology.

[0048] Hash mapping and weighting of prefix and suffix structural features: Medical terms often contain prefixes (such as "super" and "sub") and suffixes (such as "-inflammation" and "-symptom") with specific semantics. By identifying and hashing these prefix and suffix structures, we can effectively capture the derivation relationships of terms and specific medical dimension information.

[0049] During the hash mapping process, corresponding weights are accumulated. These weights can be determined based on the frequency of features in the corpus, TF-IDF values, or through expert knowledge, machine learning model training, etc., to enhance the representational power of key features in the semantic vector, making important semantic information occupy a more prominent position in the vector.

[0050] Building upon this, Top-K sparse compression is performed on the weighted vector. This step aims to optimize the storage and computational efficiency of the vector while preserving the most important semantic information. Specifically, only the K dimensions with the largest weights in the weighted vector are retained, while the remaining dimensions are set to zero. The selection of the value of K is a trade-off, typically requiring empirical setting or experimental optimization based on the actual application scenario, data sparsity, and tolerance for information loss. In this way, the dimensionality of the vector can be significantly reduced, decreasing storage space and subsequent computational complexity, while ensuring that the most representative semantic features are preserved.

[0051] Finally, the sparsely compressed vectors are normalized using the L2 norm to obtain normalized sparse semantic vectors. The purpose of L2 norm normalization is to unify the length of all semantic vectors to 1. This means that in subsequent vector similarity calculations (such as cosine similarity), the similarity result will depend only on the direction of the vectors, and will no longer be affected by the vector length (i.e., the sum of feature weights). This normalization process improves the accuracy and comparability of similarity calculations between semantic vectors of different terms, making semantic similarity measurement more fair and effective.

[0052] In some possible embodiments, it is proposed that semantic vector encoding can also be implemented using any of the following: a pre-trained medical domain word vector model, TF-IDF combined with N-gram features, or a character-level neural network encoder.

[0053] Specifically, when using a pre-trained medical domain word vector model, this model is obtained through large-scale training on massive medical text corpora (e.g., medical journals, clinical reports, electronic medical records, etc.). Each medical term is mapped to a fixed-dimensional dense vector, which captures the semantic similarity, association, and contextual information between terms. For example, classic word vector models such as Word2Vec, GloVe, and FastText can be used, or pre-trained language models based on the Transformer architecture (such as BERT, BioBERT, ClinicalBERT, etc.) can be trained. When encoding a medical term, the corresponding vector representation of the term can be directly queried from the pre-trained model; for out-of-vocabulary words, they can be combined using sub-word or character-level models, or processed through transfer learning. This method can leverage rich medical domain knowledge to provide high-quality semantic representations.

[0054] When using TF-IDF combined with N-gram features, this method first decomposes medical terms into N-gram sequences, where N-grams can be character-level (e.g., 2-gram, 3-gram) or word-level. Then, for each N-gram, its term frequency (TF) in the current term and its inverse document frequency (IDF) across the entire medical terminology corpus are calculated. The TF-IDF value reflects the importance of the N-gram in the term and its discriminative power within the corpus. Finally, each medical term is represented as a high-dimensional vector, where each dimension corresponds to a unique N-gram feature, the value of which is the TF-IDF score of that N-gram. This method effectively captures the local structure and keyword information of terms, demonstrating good performance in identifying medical terms with specific patterns.

[0055] When using a character-level neural network encoder, this encoder treats medical terms as a sequence of characters and learns the term's representation directly from the character level using a neural network model (e.g., a convolutional neural network (CNN) or a recurrent neural network (RNN), such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU)). CNNs can capture local patterns (such as prefixes, suffixes, and roots) in the character sequence using convolutional kernels of different sizes; RNNs can handle the sequential information of the character sequence, capturing the dependencies between characters. By encoding the character sequence, a fixed-dimensional vector is ultimately generated to represent the medical term. This method is robust to spelling errors, abbreviations, and out-of-vocabulary words because it does not rely on a predefined vocabulary but learns semantics from combinations of characters.

[0056] By introducing pre-trained medical term vector models, TF-IDF combined with N-gram features, or character-level neural network encoders as alternatives or supplementary schemes for semantic vector encoding, this embodiment significantly enhances the flexibility and accuracy of semantic representation of medical terms. The pre-trained model utilizes prior knowledge from massive amounts of medical text, enabling it to capture deep semantic relationships and improve the understanding of complex medical concepts. TF-IDF combined with N-gram features excels at identifying key information and local structures within terms, exhibiting higher discriminative power for specific keywords and phrases. The character-level neural network encoder demonstrates stronger robustness to spelling variations and out-of-vocabulary words, effectively addressing the dynamic changes in medical terminology. These diverse encoding strategies allow the system to select the most appropriate encoding method based on different medical terminology characteristics and application scenarios, thereby improving the accuracy and reliability of subsequent conflict detection and risk grading, and ensuring the comprehensiveness and efficiency of medical terminology standardization quality review.

[0057] In some possible embodiments, by acquiring multi-source standardized data and constructing a conflict term matrix, original terms with multiple standardized results can be identified. These original terms are then further normalized and preprocessed to extract core terminology and clinical modification information, generating a sorted list of conflict pairs. Furthermore, by semantically vectorizing all medical terms, normalized sparse semantic vectors corresponding to each term can be obtained. However, while the above methods can identify conflict terms with multiple standardized results and encode them semantically, effectively identifying and quantifying the potential confusion relationships between different original terms and between them and their respective standardized results remains a challenge. Simply comparing the semantic vectors of individual terms is insufficient to fully reveal this complex confusion risk; a more refined verification mechanism is needed.

[0058] Optionally, a four-way cross-validation calculation based on semantic vectors is proposed to obtain the cross-validation results. Specifically, the original terms are first screened based on character-level similarity, and original term pairs with character similarity exceeding a preset threshold are listed as candidate confusion pairs. This initial screening step aims to quickly identify original term pairs that have a certain degree of similarity in surface form, and these term pairs are more likely to have semantic confusion risks. Character-level similarity can be calculated using various algorithms. For example, edit distance (such as Levenshtein distance) can be used to measure the minimum number of operations required to convert between two strings, or Jaccard similarity can be used to evaluate the degree of overlap of the character N-gram sets of two terms. By setting a preset threshold, such as 0.7 or 0.8, only original term pairs with character similarity exceeding this threshold are selected as candidate confusion pairs for subsequent in-depth analysis, thereby effectively reducing the amount of computation and focusing on high-risk areas.

[0059] Subsequently, for each selected candidate confusion pair, a four-way cross-similarity is calculated based on the corresponding sparse semantic vector. The sparse semantic vector, obtained in previous steps through multi-level feature hashing and weighting, effectively captures the deep semantic information of medical terms. Four-way cross-similarity is a multi-dimensional semantic relationship evaluation method, comprising four key similarity values: a first similarity, measuring the semantic consistency between the original term A and its standardized result A'; a second similarity, measuring the semantic similarity between the original term A and the standardized result B' of the original term B, to assess whether term A has a more suitable standardized result B'; a third similarity, measuring the semantic consistency between the original term B and its standardized result B'; and a fourth similarity, measuring the semantic similarity between the original term B and the standardized result A' of the original term A, to assess whether the original term B has a more suitable standardized result A'. These similarity values ​​are typically obtained by calculating the cosine similarity of the corresponding semantic vectors, revealing the semantic associations and potential confusion between the original term and its standardized result from different perspectives.

[0060] Based on this, the fusion similarity of candidate confusing pairs is calculated, which is a weighted fusion result of character similarity and semantic similarity. This fusion mechanism aims to comprehensively utilize the surface formal features and deep semantic features of terms to provide a more comprehensive and robust basis for confusion judgment. For example, the fusion similarity can be expressed as a linear weighted sum of the semantic information embodied in character similarity and four-way cross-similarity (e.g., through some aggregation or weighting of the four-way cross-similarity), where the weight coefficients can be optimized and adjusted according to the actual application scenario or through machine learning methods. In this way, the bias that may be caused by a single similarity measure can be avoided, thereby more accurately identifying potential confusing term pairs.

[0061] In some possible embodiments, when performing risk stratification based on cross-validation results and clinical modification information, a pre-defined knowledge base of legal differences in medical terminology is proposed. The knowledge base contains clinically recognized legal variation dimensions and their value sets. For the original terms and standardized results in the candidate confusion pairs, the corresponding dimension words are extracted, dimension consistency is checked, and dimension consistency, dimension conflict, or no dimension information is marked.

[0062] Specifically, the system invokes a pre-built knowledge base of legally valid medical terminology variations. This knowledge base contains clinically recognized legal variation dimensions and their value sets. It refers to the system accessing a pre-built database or knowledge graph specifically designed to store rules and patterns of legal variations in medical terminology. The purpose of this knowledge base is to identify which variations in medical terminology are clinically acceptable and which may indicate genuine standardization errors. The construction of this knowledge base can rely on the experience of medical experts, through in-depth analysis and refinement of a large amount of medical literature, clinical guidelines, existing terminology standards (such as SNOMEDCT, ICD, etc.), and historical standardization data. The knowledge base contains a series of key "dimensions," such as "anatomical location," "disease stage," "treatment method," "diagnostic method," and "pathological type," and each dimension is associated with a "value set." For example, the "anatomical location" dimension may include specific values ​​such as "left," "right," and "bilateral." This knowledge base is typically stored in the form of structured data, such as relational databases, graph databases, or XML / JSON files, and provides corresponding interfaces for the system to query and call, ensuring the accuracy, completeness, and maintainability of its content.

[0063] Based on this, for the original terms and standardized results in the candidate confusion pairs, the corresponding dimensional vocabulary is extracted, and dimensional consistency is checked. Dimensional consistency, conflict, or lack of dimensional information are marked. The aim is to determine whether the differences between the original terms and their standardized results belong to the legitimate variations defined in the knowledge base by comparing the clinical dimensional information contained in them. First, for each candidate confusion pair, the system uses Natural Language Processing (NLP) techniques, such as Named Entity Recognition (NER), keyword matching, rule matching, or deep learning-based text understanding models, to identify and extract words or phrases related to the dimensions defined in the aforementioned "Medical Terminology Legitimate Difference Dimension Knowledge Base" from the original terms and their corresponding standardized result text. For example, for the term "left lung adenocarcinoma stage III," the system might identify "left lung" as an anatomical location dimensional vocabulary, "adenocarcinoma" as a pathological type dimensional vocabulary, and "stage III" as a disease stage dimensional vocabulary. After extraction, the system compares the dimensional vocabulary extracted from the original terms and standardized results respectively. If both terms contain information in a certain dimension, and this information is considered equivalent or legitimate variation in the knowledge base (e.g., "left side" and "left" are considered equivalent in the "anatomical location" dimension), then it is marked as "dimensional consistent." If both terms contain information in a certain dimension, but this information is considered contradictory or incompatible in the knowledge base (e.g., the original term is "stage I lung cancer," and the standardized result is "stage II lung cancer"), then it is marked as "dimensional conflict." If no information is extracted from a term in a particular dimension, or if only one of the two terms contains information in a certain dimension, then it is marked as "no dimensional information." This verification process can employ rule-based matching algorithms or combine machine learning models for more complex semantic comparisons to ensure the accuracy of the verification.

[0064] In some possible embodiments, a risk score is calculated based on the four-way cross-similarity and dimensionality consistency verification results, and the risk score satisfies: (, dimensional consistency flag), where The first similarity score, For the second similarity, The third similarity The fourth similarity is used; based on the risk score range, candidate confusion pairs are marked as high risk, medium risk, or low risk.

[0065] Specifically, calculating a risk score is a key indicator for quantifying the degree of potential error or inconsistency in assessing candidate confusion. This risk score comprehensively considers semantic similarity and clinical consistency, providing a numerical basis for subsequent risk stratification. The similarity between the original term A and its standardized result A' reflects the semantic fit between the term and its standard form; Represents the similarity between the standardized result B' of the original term A and the original term B, to evaluate whether there is a more appropriate standardized result B' for term A; Represents the similarity between the original term B and its own standardized result B', similar to SimAA, reflecting the semantic fit between term B and its standard form; Represents the similarity between the original term B and the standardized result A' of the original term A, to evaluate whether there is a more appropriate standardized result A' for the original term B. The dimension consistency flag comes from the result of the above dimension consistency check, which can be a boolean value (e.g., 1 for consistent, 0 for inconsistent or conflicting), or an enumeration value (e.g., consistent, conflicting, no information), or even a quantified dimension difference score. The function f can be in various forms. For example, a linear weighted summation model can be adopted, i.e., ` =w1 +w2 +w3 +w4 +w5 Dimension consistency flag`, where w1 to w5 are preset weights, and these weights can be optimized according to expert experience, historical data analysis or through machine learning methods to reflect the contribution degree of each parameter to the final risk. In addition, the function f can also be a non - linear decision tree, neural network or rule - based expert system, which comprehensively considers various input parameters through complex logical judgments.

[0066] On this basis, according to the interval of the risk score, the candidate confusion pairs are marked as high - risk, medium - risk or low - risk. This is a process of converting the quantified risk score into an understandable and operable risk level, which is convenient for reviewers to quickly identify and handle problems of different severity levels. This process is usually achieved by setting predefined risk score thresholds. For example, two thresholds T1 and T2 can be set such that when ` >=T1`, the candidate confusion pair is marked as "high - risk"; when `T2 <= <T1`, it is marked as "medium - risk"; when ` <T2`, it is marked as "low - risk". These thresholds can be trained and optimized according to historical review data, domain expert experience or through statistical analysis and machine learning models to ensure that the risk grading can accurately reflect the actual severity of quality problems and review priorities.

[0067] In some possible embodiments, a method for generating a quality review report by fusion is proposed, specifically including: deduplicating the intra-term conflict list and the inter-term confusion list respectively; merging the conflict list records using the original terms as keys; merging the confusion list records using the normalized original term pairs as keys; sorting the issue records based on preset priority rules, including intra-term conflicts taking precedence over inter-term confusion, larger issue clusters having higher priority, and issues containing multiple types of evidence taking precedence over issues containing a single type of evidence; and generating a structured quality review report based on the aggregated issue records, the report including issue cluster identifiers, core issue descriptions, issue details, and terminology statistics.

[0068] Specifically, the deduplication process for the intra-term conflict list and the inter-term confusion list aims to eliminate duplicate or overlapping records that may have arisen during the initial analysis. The intra-term conflict list records cases where a single original term has multiple standardized results. This embodiment uses the original term as a unique identifier to aggregate all conflict records related to that original term, ensuring that the conflict information for each original term is presented only once, but includes all conflict details. For example, if the original term "hypertension" conflicts with both the standardized result "Hypertension" and the standardized result "HTN," these conflict information will be merged into a single record with "hypertension" as the key. The inter-term confusion list records potential confusion relationships between different original terms. This embodiment uses the normalized pairs of original terms as the key for merging. This means that if original term A and original term B are identified as confused, and original term B and original term A are also identified as confused, or if the same normalized pair of terms is identified as confused through different paths, these records will be merged into one to avoid repeatedly reporting the same confusion issue. The normalized original term pair usually refers to the pair composed of term cores after preprocessing (such as removing modifiers, standardizing capitalization, etc.), which ensures that term pairs with different expressions but the same semantics are regarded as the same object of confusion.

[0069] Building upon this foundation, to further optimize review efficiency, this embodiment sorts the aggregated issue records based on preset priority rules. These priority rules are set based on experience and importance in the review of medical terminology standardization. First, intra-terminal conflicts are given higher priority than inter-terminal confusions. This is because the existence of multiple standardized results for an original term usually indicates fundamental ambiguity in its own definition or standardization process; resolving such issues can often eliminate more potential confusions at the source. Second, the larger the issue cluster, the higher its priority. The size of an issue cluster can be measured in various ways. For example, the more conflicting standardized results an original term involves, or the higher the frequency of a confusion pair appearing in different data sources, the wider the scope and the stronger the prevalence of the issue, and therefore, it should be prioritized. Furthermore, issues containing multiple types of evidence take precedence over issues with a single type of evidence. For example, an issue supported simultaneously by semantic similarity analysis, inconsistencies in clinical modification information, and high-risk grading results is more likely to be considered true and serious, and should be reviewed more frequently than issues supported by only a single type of evidence. These types of evidence can include semantic vector encoding results, cross-validation results, clinical modification information, etc.

[0070] Finally, this embodiment generates a structured quality review report based on the sorted aggregated issue records. This report is not a simple list, but rather presented in a clear and easy-to-understand structure, designed to provide reviewers with comprehensive decision support information. The report includes issue cluster identifiers, assigning a unique identifier to each independent issue cluster for easy tracking and management. The core issue description summarizes the essence of the issue in concise language, such as "The original term 'X' has ambiguous standardization" or "The original terms 'A' and 'B' are easily confused." The issue details provide all specific data and analysis results supporting the core issue, including all conflicting standardization results, detailed information on confused terms, relevant similarity scores, risk grading results, identification of relevant source documents, and consistency verification results for clinical modification information. Furthermore, the report includes terminology statistics, such as the frequency of occurrence of the original terms, the distribution of different standardization results, and the co-occurrence frequency of confused pairs, providing reviewers with quantitative background information to help them assess the severity and scope of the issue.

[0071] like Figure 2 As shown, this embodiment provides an intelligent review device for the standardization quality of medical terminology. First, the data acquisition module acquires multi-source standardized data. This data can be collected from different information systems, databases, or through manual annotation. For example, the data acquisition module can manually merge standardized tables provided by different hospitals to form a preliminary dataset containing the original terms, their corresponding standardized results, and source tags.

[0072] Optionally, the conflict construction module constructs a conflict term matrix based on the multi-source standardized data. After acquiring the multi-source standardized data, the conflict construction module can perform a preliminary scan of the data to identify the original terms with different standardized results. For example, by traversing all standardized records, each original term and its corresponding standardized results can be recorded in a simple list. If an original term is found to correspond to two or more different standardized results, it is marked as a conflict term, and all its standardized results are recorded together. This process can form a preliminary set of conflict terms, where each conflict term is accompanied by all its different standardized results.

[0073] Subsequently, the preprocessing module performs normalization preprocessing on the original terms in the conflict terminology matrix, extracting standardized terminology cores and associated clinical modification information. For the original terms in the conflict terminology matrix, the preprocessing module can perform basic text cleaning operations. For example, it removes common special characters, numbers, or general stop words from the terms. This processing aims to simplify terminology expression, making it easier for subsequent analysis. Through this processing, a relatively clean terminology form can be obtained as the terminology core. For some terms containing descriptive words, such as "left-sided pneumonia," the preprocessing module can manually or through simple rules identify "left-sided" as clinical modification information, and "pneumonia" as the terminology core.

[0074] Based on this, the conflict sorting module generates a sorted list of conflict pairs using the conflict term matrix. After the conflict term matrix is ​​constructed, for each original term with multiple standardized results, the conflict sorting module can combine all its different standardized results pairwise to generate a series of conflict pairs. For example, if the original term A has standardized results B and C, then the conflict pair (B, C) is generated. These conflict pairs can be simply listed to form a preliminary list of conflict pairs. The sorting of this list can be based on the frequency of occurrence of the conflict pairs; for example, conflict pairs that occur more frequently are listed first.

[0075] Simultaneously, the vector encoding module performs semantic vector encoding on all medical terms, obtaining normalized sparse semantic vectors for each term. To capture the semantic information of medical terms, the vector encoding module can employ a basic vectorization method. For example, a simple bag-of-words model vector can be created for each term, where each dimension of the vector corresponds to a word in the vocabulary, recording the frequency of that word in the term. Subsequently, these bag-of-words vectors undergo simple normalization, such as dividing each element of the vector by its maximum value, to obtain a preliminary normalized vector. This vector can be considered a sparse semantic vector because the vocabulary is typically large, while a single term contains only a limited number of words.

[0076] Optionally, the cross-validation module performs four-way cross-validation on the candidate confusion pairs based on the semantic vector to obtain the cross-validation results. After obtaining the semantic vectors of the terms, the cross-validation module can identify potential confusion term pairs. For example, term pairs with close semantic vector distances can be selected as candidate confusion pairs. For each candidate confusion pair, such as the original term A and its standardized result A', and the original term B and its standardized result B', the cross-validation module can calculate four basic similarities: the similarity between A and A', the similarity between A and B', the similarity between B and B', and the similarity between B and A'. These similarities can be obtained by calculating the cosine similarity or Euclidean distance between the corresponding semantic vectors. These calculation results constitute the preliminary cross-validation results.

[0077] Subsequently, the risk grading module performs risk grading based on the cross-validation results and the clinical modification information, obtaining the risk grading result. Combining the cross-validation results and the previously extracted clinical modification information, the risk grading module can assess the risk of potential standardization issues. For example, if cross-validation shows that two terms have low semantic similarity and their clinical modification information differs significantly (e.g., one term describes "Type I" and the other "Type II"), they can be marked as high-risk. Conversely, if the semantic similarity is high and the clinical modification information is consistent or the differences do not affect the core semantics, the risk is low. This assessment can be based on preset simple rules, such as setting similarity thresholds and modification information matching rules to classify issues into high, medium, and low risk levels.

[0078] Finally, the report generation module merges the conflict pair list with the risk grading results to generate a quality review report. The previously generated conflict pair list and risk grading results are integrated to produce a quality review report. This report can simply list all identified conflict and confusion pairs, along with their corresponding risk levels. For example, the report could contain a table where each row represents an issue, listing the original term, its conflict or confusion standardization result, and the risk level of the issue. The report aims to provide an overview so that human reviewers can quickly understand the problems existing in the standardization process.

[0079] This embodiment, through the collaborative work of a data acquisition module, a conflict resolution module, and a preprocessing module, can systematically identify original terms with multiple standardization results. Deep semantic understanding of medical terms is achieved through a vector encoding module and a cross-validation module. Optionally, a risk grading module combines clinical modification information to perform risk grading, effectively distinguishing between legitimate differences and genuine errors, and discovering hidden mapping errors. Therefore, this embodiment can significantly improve the efficiency and accuracy of quality inspection of standardized medical terminology results, fundamentally ensuring the quality of standardized terminology resources and solving the problems of high manual costs, inconsistent standards, and insufficient semantic understanding in traditional methods.

[0080] This embodiment also discloses an electronic device, including a memory and a processor coupled to each other. The memory stores program instructions, and the processor executes the program instructions to implement the above-mentioned intelligent review method for standardized quality of medical terminology.

[0081] The innovation of this embodiment lies in the fact that by performing an intelligent review method through the electronic device, multi-source standardized data conflict analysis and deep semantic vector encoding are combined with a four-way cross-validation method, thereby effectively distinguishing between clinically legitimate differences and real standardized errors and discovering hidden mapping errors, thus improving the efficiency and accuracy of quality inspection.

[0082] Specifically, during operation, the electronic device first acquires multi-source standardized data, which includes standardized records from different sources. Each standardized record contains the original term, the standardized result, and the source identifier. Based on this multi-source standardized data, a conflict term matrix is ​​constructed. This matrix contains the original terms with multi-standardized results and their corresponding quantitative analysis results. The original terms in the conflict term matrix undergo normalization preprocessing to extract standardized terminology cores and associated clinical modification information. A sorted list of conflict pairs is generated based on the conflict term matrix. Semantic vector encoding is performed on all medical terms to obtain normalized sparse semantic vectors for each term. Four-way cross-validation is performed on candidate confusion pairs based on these semantic vectors to obtain cross-validation results. Risk grading is performed based on the cross-validation results and the clinical modification information to obtain risk grading results. Finally, a quality review report is generated by fusing the conflict pair list and the risk grading results.

[0083] To address the technical problem of lacking effective automated solutions for quality control of medical terminology standardization results, this embodiment provides a computer-readable storage medium. This computer-readable storage medium stores program instructions executable by a processor, which are used to implement the aforementioned intelligent quality review method for medical terminology standardization. The core innovation of this embodiment lies in combining the construction of a conflict terminology matrix from multi-source standardized data with four-way cross-validation of normalized sparse semantic vectors using specific logic, and introducing a risk grading mechanism for clinical modification information. This effectively distinguishes between legitimate clinical differences and genuine standardization errors, achieving the effect of automatically discovering hidden mapping errors and ensuring the quality of standardized terminology resources.

[0084] Specifically, when the processor executes the program instructions, it first acquires multi-source standardized data, which includes standardized records from different sources. Each standardized record contains the original term, the standardized result, and the source identifier. Then, based on this multi-source standardized data, a conflict term matrix is ​​constructed. This matrix contains the original terms with multi-standardized results and their corresponding quantitative analysis results. Subsequently, the original terms undergo normalization preprocessing to extract standardized terminology cores and associated clinical modification information. A sorted list of conflict pairs is generated based on the conflict term matrix. Medical terms are semantically vectorized to obtain normalized sparse semantic vectors for each term. Four-way cross-validation is performed on candidate confusion pairs based on these semantic vectors to obtain cross-validation results. The cross-validation results are combined with the clinical modification information to perform risk grading, resulting in a risk grading result. Finally, a quality review report is generated by fusing the conflict pair list and the risk grading result.

[0085] The embodiments and examples presented herein are provided to best illustrate embodiments of the invention and its particular applications, thereby enabling those skilled in the art to practice and use the invention. However, those skilled in the art will understand that the above description and examples are provided merely for ease of illustration and example. The descriptions presented are not intended to cover all aspects of the invention or to limit the invention to the precise forms disclosed.

[0086] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent quality review of standardized medical terminology, characterized in that, include: Acquire multi-source standardized data, which includes standardized records from different sources, and each standardized record includes the original terminology, the standardized result, and the source identifier; A conflict term matrix is ​​constructed based on the multi-source standardized data. The conflict term matrix includes the original terms with multi-standardization results and their corresponding quantitative analysis results. The original terms in the conflict terminology matrix are preprocessed for standardization to extract the standardized terminology core and associated clinical modification information. A sorted list of conflict pairs is generated based on the conflict term matrix; Semantic vector encoding is performed on all medical terms to obtain normalized sparse semantic vectors corresponding to each term. Based on the semantic vector, four-way cross-validation is performed on the candidate confusion pairs to obtain the cross-validation results; Risk stratification is performed based on the cross-validation results and the clinical modification information to obtain the risk stratification results; A quality review report is generated by merging the conflict pair list with the risk classification results.

2. The method according to claim 1, characterized in that, The construction of the conflict term matrix based on the multi-source standardized data includes: Batch parsing of the multi-source standardized data is performed to extract the original terms, standardized results and source file identifiers from each standardized record, and an initial dataset is constructed. The normalization results are subjected to one-to-many mapping normalization processing, which splits the composite normalization results containing delimiters into independent normalization terms, and after deduplication and lexicographical sorting, they are recombined into normalized strings; Aggregate the standardized data using the original terms as keys, and calculate the number of unique standardized results, the distribution entropy of the standardized results, the main standardized results, and the list of source documents for each original term; Records with a unique number of standardized results greater than 1 are selected to construct a conflict term matrix.

3. The method according to claim 2, characterized in that, The normalization preprocessing of the original terms in the conflict term matrix includes: The original terms are truncated by the position of the first question mark in the term, retaining the core description part before the question mark; Recursive noise removal is performed on the truncated terms, removing the numerical codes, punctuation marks, and non-semantic suffixes at the beginning and end of the terms, while separating and extracting clinical modifiers that conform to the preset medical dimension pattern. If the terminology core is empty after preprocessing, then revert to the original complete terminology to obtain the standardized terminology core and associated clinical modification information.

4. The method according to claim 3, characterized in that, The process of generating a sorted list of conflict pairs based on the conflict term matrix includes: For each original term with multiple standardized results, all its different standardized results are combined pairwise to generate conflict pairs; For each conflict pair, a priority score is calculated, and the calculation parameters of the priority score include the entropy value of the original term, the frequency of occurrence of the original term, the frequency difference of the conflict pair standardization results, and the distribution of the source documents of the conflict results. All conflict pairs are sorted in descending order based on the priority score to obtain a sorted list of conflict pairs.

5. The method according to claim 4, characterized in that, Semantic vector encoding is performed on all medical terms to obtain normalized sparse semantic vectors for each term, including: Initialize a zero vector of a specified dimension and perform validity checks on the input terms; The input terms are subjected to multi-level feature hashing and weighting. Character-level features, N-gram-level features, and prefix-suffix structure-level features are hashed and mapped and their corresponding weights are accumulated. The weighted vector is subjected to Top-K sparse compression, retaining only the K dimensions with the largest weights and setting the remaining dimensions to zero. The sparse compressed vector is normalized using the L2 norm to obtain a normalized sparse semantic vector.

6. The method according to claim 1, characterized in that, The step of performing four-way cross-validation on candidate confusion pairs based on the semantic vector includes: The original terms are initially screened based on character-level similarity, and the original term pairs with character similarity exceeding a preset threshold are listed as candidate confusion pairs. For each candidate confusion pair, a four-way cross-similarity is calculated based on the corresponding sparse semantic vector. The four-way cross-similarity includes the first similarity between the original term A and its own standardized result A', the second similarity between the original term A and the standardized result B' of the original term B, the third similarity between the original term B and its own standardized result B', and the fourth similarity between the original term B and the standardized result A' of the original term A. Calculate the fusion similarity of the candidate confusion pairs, where the fusion similarity is a weighted fusion result of character similarity and semantic similarity; The risk stratification based on the cross-validation results and the clinical modification information includes: The system invokes a pre-defined knowledge base of legally recognized medical terminology variation dimensions, which includes clinically recognized legal variation dimensions and their value sets. For the original terms and standardized results in the candidate confusion pairs, extract the corresponding dimensional vocabulary, perform dimensional consistency verification, and mark the dimensional consistent, dimensional conflicting, or dimensional information missing. Based on the four-way cross-similarity and dimensionality consistency verification results, a risk score is calculated, and the risk score satisfies: (, dimensional consistency flag), where The first similarity score, For the second similarity, The third similarity The fourth similarity; Based on the range of the risk score, the candidate confusion pairs are marked as high risk, medium risk, or low risk.

7. The method according to claim 6, characterized in that, The process of generating a quality review report by fusing the conflict pair list with the risk classification results includes: Deduplication is performed on the intra-term conflict list and the inter-term confusion list respectively. The conflict list records are merged using the original terms as the key, and the confusion list records are merged using the normalized original term pairs as the key. The problem records are sorted based on preset priority rules, which include intra-term conflicts taking precedence over inter-term confusions, larger problem clusters having higher priority, and problems containing multiple types of evidence taking precedence over problems containing a single type of evidence. A structured quality review report is generated based on the aggregated issue records. The report includes issue cluster identifiers, core issue descriptions, issue details, and terminology statistics.

8. A medical terminology standardization quality intelligent review device, characterized in that, include: The data acquisition module is used to acquire multi-source standardized data, which includes standardized records from different sources. Each standardized record includes the original terminology, the standardized result, and the source identifier. The conflict construction module is used to construct a conflict term matrix based on the multi-source standardized data. The conflict term matrix contains the original terms with multi-standardization results and their corresponding quantitative analysis results. The preprocessing module is used to perform normalization preprocessing on the original terms in the conflict term matrix, and extract the standardized term core and associated clinical modification information. The conflict sorting module is used to generate a sorted list of conflict pairs based on the conflict term matrix. The vector encoding module is used to perform semantic vector encoding on all medical terms to obtain the normalized sparse semantic vectors corresponding to each term. The cross-validation module is used to perform four-way cross-validation calculation on the candidate confusion pairs based on the semantic vector to obtain the cross-validation result. The risk grading module is used to perform risk grading based on the cross-validation results and the clinical modification information to obtain the risk grading results. The report generation module is used to generate a quality review report by fusing the conflict pair list with the risk classification results.

9. An electronic device, characterized in that, The method includes a memory and a processor coupled to each other, wherein the memory stores program instructions and the processor executes the program instructions to implement the medical terminology standardization quality intelligent review method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The system stores program instructions that can be executed by a processor, the program instructions being used to implement the medical terminology standardization quality intelligent review method according to any one of claims 1 to 7.

Images