Intelligent retrieval method and device fusing multi-source medical knowledge and storage medium

Abstract

The invention discloses an intelligent retrieval method and device fusing multi-source medical knowledge and a storage medium, and relates to the technical field of intelligent retrieval. The method comprises the following steps: constructing a multi-query variant set; obtaining a multi-source candidate set based on the multi-query variant set; carrying out fusion screening and high-level evidence retention on the multi-source candidate set to obtain an effective candidate set, generating a retrieval answer and carrying out source labeling; the retrieval coverage range is widened through query entity normalization and multi-query variant construction, the high-level evidence priority and semantic matching accuracy are enhanced by means of parallel vector retrieval, metadata filtering and fusion division screening, non-redundant effective candidates are generated through duplicate removal processing, and the semantic matching accuracy is improved. And the retrieval reliability is guaranteed through a retrieval abnormity monitoring and retry mechanism, so that the problems that the query coverage is not comprehensive, evidence screening lacks evidence-based priority consideration, candidate results have redundant conflicts and retrieval abnormity cannot be effectively handled in the existing medical intelligent retrieval are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent retrieval technology, specifically to an intelligent retrieval method, device, and storage medium that integrates multi-source medical knowledge. Background Technology

[0002] While existing intelligent medical retrieval serves as a crucial support for clinical evidence-based decision-making, its lack of deep adaptation to clinical needs in actual medical information retrieval reveals the following shortcomings: First, existing retrieval systems mostly match directly based on the original query text, resulting in a single semantic dimension of the query and making it prone to missing core relevant evidence due to differences in terminology, leading to incomplete query coverage. Second, existing evidence screening largely relies on a single similarity score, causing high-level, high-value evidence to not be prioritized, resulting in a lack of evidence-based priority considerations in evidence screening. Third, existing retrieval systems lack targeted deduplication and conflict resolution mechanisms, resulting in a large amount of redundant information and contradictory evidence in the candidate results, affecting the usability of the retrieval. Finally, existing retrieval systems do not perform real-time monitoring and anomaly assessment of core processes, leading to retrieval service interruptions or excessively low efficiency, resulting in a lack of effective handling of retrieval anomalies.

[0003] Therefore, there is an urgent need for an intelligent retrieval method that integrates multi-source medical knowledge, which can achieve comprehensive semantic coverage of queries, strengthen the priority of evidence-based evidence, eliminate redundant and conflicting content, and ensure retrieval stability, in order to solve the above-mentioned technical bottlenecks and meet the needs of precise and efficient clinical evidence-based decision-making. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent retrieval method, device, and storage medium that integrates multi-source medical knowledge, solving the problems of incomplete query coverage, lack of evidence-based priority consideration in evidence screening, redundancy and conflict in candidate results, and lack of effective handling of retrieval anomalies in existing medical intelligent retrieval systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent retrieval method integrating multi-source medical knowledge, comprising the following steps: S1. Identify and normalize the query entities of disease, drug and examination categories in the original query to form a standardized query text. Based on the standardized query text, construct effective query variants and combine them with the standardized query text to form a set of multiple query variants.

[0006] S2. For the clinical guideline set and the case data set, parallel vector retrieval is performed based on the multi-query variant set, and filtering and composite relevance sorting are performed to obtain a multi-source candidate set.

[0007] S3. The multi-source candidate set is fused, screened, and high-level evidence is retained to obtain an effective candidate set.

[0008] S4. Generate search answers for the valid candidate set according to the preset template and add source annotations.

[0009] S5. Evaluate the time-consuming data generated in S2-S3 and determine the anomaly index. If it is greater than the set threshold, the search is retried. If the anomaly index is not greater than the set threshold after a single retry, the search answer is obtained in sequence. If the anomaly index is still greater than the set threshold after multiple retryes, S3 is disabled, and the search answer is generated based on the multi-source candidate set and the preset template.

[0010] Intelligent retrieval devices that integrate multi-source medical knowledge include: The query variant acquisition module is used to identify query entities of disease, drug and examination categories in the original query and normalize them to form standardized query text. Based on the standardized query text, effective query variants are constructed and combined with the standardized query text to form a set of multiple query variants.

[0011] The candidate answer output module is used to perform parallel vector retrieval based on multiple query variant sets for the clinical guideline set and the case data set, and to perform filtering and relevance ranking to obtain a multi-source candidate set.

[0012] The effective answer generation module is used to perform fusion score calculation based on evidence level and source type weights on the multi-source candidate set, prioritize the retention of high-level evidence, and deduplicate and conflict adjudicate to obtain the effective candidate set.

[0013] The search results output module is used to generate search answers from the valid candidate set according to a preset template and to annotate the sources.

[0014] The retrieval quality monitoring module is used to evaluate the time-consuming data generated in the candidate answer output module and the valid answer generation module, determine the anomaly index, and if it is greater than a set threshold, the retrieval is retried. If the anomaly index is not greater than the set threshold after a single retry, the retrieval answers are obtained in sequence. If the anomaly index is still greater than the set threshold after multiple retries, the valid answer generation module is disabled, and the retrieval answers are generated based on the multi-source candidate set and the preset template.

[0015] The present invention has the following beneficial effects: This invention broadens the retrieval coverage by query entity normalization and constructing multiple query variants. It enhances the priority of high-level evidence and the accuracy of semantic matching by using parallel vector retrieval, metadata filtering, and fusion-based quantitative screening. After deduplication, it generates effective candidates without redundancy. Then, it ensures the reliability of retrieval through retrieval anomaly monitoring and retry mechanisms. This invention solves the problems of incomplete query coverage, lack of evidence-based priority consideration in evidence screening, redundancy and conflict in candidate results, and lack of effective handling of retrieval anomalies in existing medical intelligent retrieval.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0017] Figure 1 This is a flowchart of the intelligent retrieval method that integrates multi-source medical knowledge according to the present invention.

[0018] Figure 2 This is a module connection diagram of the intelligent retrieval device that integrates multi-source medical knowledge according to the present invention. Detailed Implementation

[0019] Please see Figure 1 The present invention provides a technical solution: an intelligent retrieval method integrating multi-source medical knowledge, comprising the following steps: S1. Identify and normalize the query entities of disease, drug and examination categories in the original query to form a standardized query text. Based on the standardized query text, construct effective query variants and combine them with the standardized query text to form a set of multiple query variants.

[0020] The specific process is as follows: A fine-tuned BiLSTM-CRF biomedical entity recognition model is used to extract query entities related to diseases, drugs, and examinations from raw clinical data. Based on a publicly available terminology dictionary, a bidirectional mapping algorithm is employed to normalize the query entities, that is, to uniformly map synonyms, abbreviations, or colloquialisms to standardized terms, resulting in a normalized entity set. For example, T2DM is mapped to type 2 diabetes. Furthermore, meaningless interjections and redundant modifiers are removed from the normalized entity set, and standardized query text can be generated through sentence reconstruction.

[0021] The query entities related to diseases, drugs, and examinations in the standardized query text are replaced with preset equivalent expressions to generate multiple sets of query texts after entity replacement, and the scenarios are limited based on the clinical scenarios associated with the query entities.

[0022] It should be noted that equivalent expression replacement can be performed through a pre-defined expression library. The content of the expression library comes from the thesaurus of the Unified Medical Language System (UMLS), the set of commonly used terms in clinical guidelines, etc., and includes equivalent expressions for various entities in the query entity. For example, the equivalent expression for the disease entity type 2 diabetes is non-insulin-dependent diabetes. In addition, clinical scenarios include the acute or chronic attributes of disease entities, the medication scenarios of drug entities, and the diagnostic scenarios of examination entities. The attributes are derived from the ICD-11 classification standard and the National Medical Products Administration's drug catalog.

[0023] The standardized query text, after being defined by the scenario, is transformed into an equivalent sentence structure for clinical retrieval. At the same time, the limiting words are adjusted or supplemented to generate candidate query variants.

[0024] In this embodiment, three types of clinical search equivalent sentence templates can be preset, including: ① Keyword combination type, which combines the core entity with the search intent, such as emergency treatment plan for acute myocardial infarction; ② Question-oriented type, which is guided by interrogative words, such as what are the emergency treatment plans for acute myocardial infarction; ③ Specific demand type, which supplements limiting conditions, such as emergency thrombolytic therapy plan for adult acute myocardial infarction.

[0025] Substitute each standardized query text into one of the three sentence templates and perform the following operations: ① Sentence structure transformation, such as transforming declarative sentences into keyword combinations; ② Supplementation of reasonable limiting words, adding core limiting words such as diagnostic criteria, treatment timing, and medication dosage based on the search intent, without adding irrelevant limiting words; ③ Deletion of duplicate or meaningless limiting words, such as retaining only one if both "emergency" and "urgent" appear simultaneously.

[0026] A fine-tuned BiLSTM-CRF biomedical entity recognition model is used to transform the standardized query text and each text in the candidate query variant set into a vector representation. Semantic similarity calculation is used to calculate the semantic similarity between each candidate query variant and the standardized query text. Candidate query variants with semantic similarity greater than the semantic similarity threshold are selected as valid query variants.

[0027] It should be noted that the fine-tuning process of the BiLSTM-CRF biomedical entity recognition model is as follows: a UMLS-annotated medical corpus is used as the training set, the learning rate is set to 1e-5, the number of iterations is 50 rounds, the batch size is 32, and the fine-tuning target is an entity recognition F1 score ≥ 0.92, where the F1 score is the harmonic mean of precision and recall.

[0028] The semantic similarity calculation can be performed using the cosine similarity algorithm, with the semantic similarity threshold set between 0.7 and 0.85, preferably 0.8.

[0029] Equivalent expression replacement based on UMLS thesaurus and clinical guideline terminology sets unifies the description of medical entities, avoiding the omission of relevant evidence due to terminology differences; limiting the query to the clinical scenarios associated with the query entities makes the query more relevant to actual application scenarios; three types of clinical search equivalent sentence transformation adapt to the search habits of clinicians, and the addition of core limiting words improves the targeting of the search; semantic similarity screening ensures that effective variants do not deviate from the core semantics of the standardized query, ultimately broadening the semantic coverage dimension of the search and improving the recall and precision of subsequent searches.

[0030] S2. For the clinical guideline set and the case data set, parallel vector retrieval is performed based on the multi-query variant set, and filtering and composite relevance sorting are performed to obtain a multi-source candidate set.

[0031] It should be noted that the specific process is as follows: Each query variant in the set of multiple query variants is transformed into a query vector with the same dimension as the document vector. A unified query vector is generated by weighted fusion of all query vectors. The weight of the standardized query text is C1, and the weight of each effective query variant is (1-C1) / k, where k is the number of effective query variants.

[0032] The unified query vector is compared with the cosine similarity of all document vectors in the clinical guidelines set and the case data set, and the similarity score is output.

[0033] For the documents corresponding to the document vectors of the clinical guideline set and the case data set, they are sorted in descending order of similarity score, and the top N documents of each set are selected as initial candidate results. In this invention, the value of N ranges from 10 to 50, preferably 30, to ensure that the initial candidate results cover potentially relevant documents.

[0034] The initial candidate results are filtered field by field based on the metadata filtering fields to obtain the filtered candidate set. The metadata filtering fields include document type, evidence level, publication time interval, disease entity, and chapter path prefix.

[0035] It should be noted that the screening process is as follows: Candidate guidelines must be clinical guidelines, and candidate cases must be inpatient or outpatient cases, excluding other irrelevant document types; the evidence level of candidate documents must be A, B, or C, excluding documents below C level to ensure the quality of the evidence base; the default range is five years prior to the current search time up to the current year, which can be customized for clinical scenarios, and the document publication year must be within this closed range; the disease entities contained in the document must completely cover the disease entities in the normalized core entity set, with no missing core diseases and no additional irrelevant core diseases; the document chapter path prefix must belong to a preset set that matches the query topic, such as treatment recommendations, drug treatment, etc. when the query topic is treatment.

[0036] If the initial candidate results are empty after metadata filtering, the following search suggestion will be returned: No matching results under the current search conditions. It is recommended to adjust one of the following conditions: ① Expand the publication time range; ② Lower the evidence level requirement (such as allowing Class C evidence).

[0037] For each candidate result in the filtered candidate set, calculate the chapter consistency score and the source authority score, and then perform a weighted sum to obtain the composite relevance score.

[0038] The chapter consistency score is the ratio of the number of chapter path prefixes in the candidate results to the total number of preset prefixes corresponding to the query topic. The source authority score is used to quantify the document source and the level of evidence. The base score for clinical guidelines is 0.8, and the base score for case data sources is 0.5. The weighting coefficient for the level of evidence is: 0.2 for level A, 0.1 for level B, and 0.0 for level C. For example, the source authority score for a level A guideline is 0.8 + 0.2 = 1.0.

[0039] Based on the composite correlation, the candidate results are sorted in descending order, and the top K candidate results are selected to form a multi-source candidate set.

[0040] For example, the value of K is in the range of 15-30, preferably 20, that is, 20 candidate structures are retained, and a multi-source candidate set including 40 candidate results is obtained by combining them.

[0041] The weighted fusion of multiple query variants into a unified query vector integrates the core semantics of all variants, ensuring the comprehensiveness of the retrieval benchmark. Parallel vector retrieval processes two types of data sources simultaneously, significantly improving retrieval efficiency. Metadata multi-field (document type, evidence level, publication time, etc.) is filtered field by field, accurately eliminating irrelevant and low-quality documents and ensuring the basic quality of candidate results. The composite relevance score combines chapter consistency score and source authority score to quantify the correlation strength and evidence reliability between candidates and queries. The top K results are sorted by score, providing a high-quality and highly relevant candidate pool for subsequent effective candidate screening.

[0042] S3. The multi-source candidate set is fused, screened, and high-level evidence is retained to obtain an effective candidate set.

[0043] Furthermore, the specific process for obtaining the effective candidate set is as follows: Obtain the preset base weights for evidence level and source type, determine the weight adjustment coefficient based on the multi-query variant set, and multiply the base weights for evidence level and source type by the weight adjustment coefficient to obtain the revised evidence level weight. and source type adjustment weight .

[0044] Perform deep medical semantic matching on a set of multiple query variants and a set of multiple source candidates to generate a semantic matching score for each candidate text. .

[0045] Combining composite correlation analysis Calculate the fusion score for each candidate text The calculation formula is: .

[0046] in, For composite term weighting factors, For semantic item weighting factors, the value of a ranges from 0.6 to 0.8, and the value of b ranges from 0.2 to 0.4, and a+b=1, where a=0.6 and b=0.4 in the emergency scenario.

[0047] For guideline candidates in the multi-source candidate set whose evidence level is not less than a set level, evidence-based relevance verification is performed based on the multi-query variant set to obtain the matching confidence level, and guideline candidates whose matching confidence level is not less than the set confidence level threshold are selected as priority candidates.

[0048] It should be noted that the matching confidence score refers to the percentage of intersection between the core entity set in the multi-query variant set and the core entity set of the guide candidates. The confidence score threshold is set to 0.75-0.85, with 0.8 being preferred.

[0049] Identify clinical scenarios with multiple query variant sets and generate dynamic fusion difference thresholds.

[0050] The process of generating the dynamic fusion difference threshold is as follows: Iterate through all text in the multi-query variant set, and match it against a pre-defined clinical scenario keyword library (containing 10+ core keywords such as emergency, first aid, emergency treatment, routine, long-term management, and follow-up) to extract all matched scenario keywords, forming a scenario keyword set K_scene (an empty set if no match is found). Extract the core disease entity D (which is from the same source as the S1 entity recognition) from the multi-query variant set, query the urgency attribute in the ICD-11 disease classification standard, and label it as acute severe illness or non-acute disease (if no core disease entity is found, it defaults to non-acute disease).

[0051] If K_scene contains any of the keywords "emergency", "first aid", or "urgent treatment", or if Tag_emer is "acute critical illness", it is determined to be an emergency scene; if K_scene contains any of the keywords "routine", "long-term management", or "follow-up", or if Tag_emer is "non-acute disease" and K_scene is an empty set, it is determined to be a routine scene; if a priority candidate satisfies that the difference between its fusion score and the fusion scores of all other priority candidates is not less than the dynamic fusion score difference threshold, then the priority candidate is retained.

[0052] The dynamic fusion difference threshold is determined based on the set threshold mapping rules.

[0053] For example, if the scene is an emergency scenario, Δ=0.03 is generated (lowering the threshold for prioritizing high-level evidence and quickly focusing on key evidence); if the scene is a routine scenario, Δ=0.05 is generated (raising the priority judgment standard and ensuring the accuracy of evidence).

[0054] If the number of preferred candidates to be retained is greater than M, then the top M candidates are retained in descending order of fusion score.

[0055] If the number of preferred candidates retained is less than M, then the non-preferred candidates are merged with the preferred candidates, and the top M are retained in descending order of fusion score to obtain the preliminary candidate set.

[0056] It should be noted that the value of M ranges from 3 to 10, with 5 being the preferred value, to meet the need for rapid acquisition of key evidence in clinical diagnosis and treatment.

[0057] The initial candidate set is deduplicated to obtain the effective candidate set.

[0058] Weight correction based on multi-query variant sets makes the weights of evidence level and source type more aligned with retrieval needs. The fusion score calculation of comprehensive weight, composite relevance score and semantic matching score improves the accuracy of screening. Prioritizing the retention of high-level evidence aligns with the core principles of evidence-based medicine, ensuring that authoritative evidence takes precedence. The dynamic fusion score difference threshold generated by clinical scenario identification adapts to the retrieval needs of different scenarios such as emergency and routine. Deduplication removes redundant candidates, and the final effective candidate set is authoritative, accurate and non-redundant, fully meeting the actual needs of clinical evidence-based decision-making.

[0059] Furthermore, the process of determining the weight adjustment coefficient based on the set of multiple query variants is as follows: Iterate through the standardized query text and all valid variations in the multi-query variant set, extract keywords that express the search purpose, count the frequency of keyword occurrences, and determine the topic tags based on the purpose corresponding to the most frequent keyword.

[0060] Based on the BiLSTM-CRF biomedical entity recognition model, disease entities, drug entities, and examination entities are extracted from a multi-query variant set. The urgency of the disease is then determined based on these entities. Disease incidence rate and drug management category value .

[0061] It should be noted that the urgency of the disease Disease incidence rate and drug management category value The acquisition process is as follows: Query the ICD-11 classification code corresponding to the disease entity. If the code belongs to the acute disease category, then... If the value is 1, then it belongs to the chronic disease category. If the value is 0, then there is no disease entity. It is 0.

[0062] According to publicly available disease incidence data from the World Health Organization (WHO), a disease is considered common if its incidence rate is ≥1‰. The incidence rate is 1; if the incidence rate is <1‰, it is considered a rare disease. If the value is 0, then there is no disease entity. =0; If a drug is listed in the drug catalog published by the National Medical Products Administration, and the drug is classified as a specially controlled drug (narcotic drugs, psychotropic drugs, etc.), then... If the value is 1, then it belongs to ordinary medicines. If the value is 0, then there is no drug entity. It is 0.

[0063] Based on the given topic tags, determine the corresponding association strength. and the priority quantification value corresponding to the topic tags. Priority quantization value These are pre-set, exemplary, treatment topics in the database. For version 1.0, medication theme 0.667, diagnostic topic It is 0.0.

[0064] Furthermore, the strength of the association The calculation process is as follows: Using publicly available entries from the UpToDate Evidence-Based Medicine Database and the NCCN Clinical Practice Guidelines Database as statistical benchmarks, a topic-entity association statistical database was constructed: For the treatment topic, count the total number of treatment recommendation items Nt for disease entity D and drug entity, and the total number of all treatment drug recommendation items N_b corresponding to disease entity D; For the diagnostic topic, count the total number of diagnostic recommendation items N_d for disease entities and examination entities, and the total number of all diagnostic and examination recommendation items N_bd corresponding to disease entity D; For the medication topic, count the total number of indication matching entries between the drug entity and the disease entity, Nm, and the total number of all indication entries corresponding to the drug entity, N_b. The formula for calculating the correlation strength is as follows: If the topic tag is a treatment topic: R=N / N_b (N_b≠0; R=0 if N_b=0 or there is no M); If the topic tag is a diagnostic topic: R=N_d / N_bd (N_bd≠0; R=0 if N_bd=0 or there is no E); If the topic tag is medication-related: R = N / N_b (N_b ≠ 0; R = 0 if N_b = 0 or there is no D / M). The calculation formula for the scenario quantification index is as follows: .

[0065] in, and All are weighting factors.

[0066] Retrieve valid search data stored in the database, and quantify and calibrate each valid search data value according to the corresponding scenario. Correlation strength calibration value Manual calibration correction factor and adaptation accuracy .

[0067] It should be noted that the corresponding scenario quantization calibration value Correlation strength calibration value The clinical search cases were derived from manual annotation. Each set of cases was jointly annotated by 3 senior clinicians and 2 medical informatics experts, ensuring a high accuracy rate. This is the consistency coefficient of the calibration results.

[0068] The similarity of each valid search result is calculated using the Euclidean distance formula for the scene quantification index, association strength, corresponding scene quantification calibration value, and association strength calibration value. .

[0069] Valid search data with similarity values ​​greater than a set similarity threshold and an absolute difference between the association strength and the calibrated association strength value not greater than a set absolute threshold are selected. An initial correction coefficient is then calculated based on the selected valid search data. The calculation formula is: .

[0070] It should be noted that the initial correction coefficient is designed to be a weighted average of historical valid search data, resulting in a correction coefficient that fits the current search scenario. If the number of valid search results is 0, the initial correction coefficient... Use the preset default correction factor directly, such as 1.0.

[0071] For the initial correction coefficient By imposing range constraints on the values, the weight correction coefficients are obtained.

[0072] For example, if the initial correction coefficient If the value is within the set range, the initial correction coefficient will be output directly. This is the weighting correction coefficient; if the initial correction coefficient is... If the value is less than the lower limit of the range, then modify it to the lower limit of the range; if the initial correction coefficient is... If the value is greater than the upper limit of the range, then it will be modified to the upper limit of the range.

[0073] By statistically analyzing keyword frequency to determine topic tags, extracting entity attributes such as disease urgency and incidence, and combining association strength to construct a scenario-based quantitative index, the scenario characteristics of search needs are comprehensively depicted. Weighted calculations based on similarity from historical search data ensure that correction coefficients are supported by objective data, avoiding subjective assumptions. Value range constraints ensure the rationality of the coefficients, and the final weighted correction coefficients accurately adapt to the current search topic, entity attributes, and scenario characteristics, improving the targeting and accuracy of subsequent fusion score calculations and making the screening results more aligned with actual needs.

[0074] Furthermore, the process of performing deep medical semantic matching on the multi-query variant set and the multi-source candidate set to generate a semantic matching score for each candidate text is as follows: Obtain the unified query vector of the multi-query variant set and the candidate text vector set of the multi-source candidate set.

[0075] Calculate the cosine similarity value, entity matching degree, and topic similarity between the unified query vector and each candidate text vector in the candidate text vector set.

[0076] It should be noted that the entity matching degree is calculated as follows: the disease entity set, drug entity set, and examination entity set corresponding to the unified query vector are obtained respectively, and the candidate disease entity set, candidate drug entity set, and candidate examination entity set of the candidate text vector are calculated based on the set formula.

[0077] Taking the disease entity set corresponding to the unified query vector and the candidate disease entity set of the candidate text vector as examples, the disease entity matching rate is calculated using the formula: Disease entity matching rate = |Disease entity set ∩ Candidate disease entity set| / |Disease entity set|. The drug entity matching rate and examination entity matching rate are calculated using the same formula. The entity matching degree is obtained by weighted summation of the disease entity matching rate, drug entity matching rate and examination entity matching rate.

[0078] Topic similarity can be calculated using the Jaccard similarity algorithm by comparing the keyword ratios between the unified query vector and the candidate text vector.

[0079] The semantic fit is obtained by weighted summation of cosine similarity, entity matching degree and topic similarity, and the semantic fit is normalized to obtain semantic matching score.

[0080] A unified query vector and candidate text vector ensure consistency in the matching baseline; cosine similarity measures the overall semantic fit, entity matching focuses on the overlapping adaptation of core medical entities, and topic similarity characterizes the fit between the search intent and the candidate topic. These three dimensions comprehensively cover the key dimensions of semantic matching; weighted summation followed by normalization ensures that the semantic matching score is within a unified range, accurately and comprehensively reflecting the fine-grained semantic fit between the set of multiple query variants and the candidate text. This provides reliable semantic dimension support for the calculation of the fusion score and further improves the accuracy of effective candidate selection.

[0081] The process of deduplicating the initial candidate set to obtain the effective candidate set is as follows: A two-dimensional hash signature algorithm is used to generate a hash signature for each candidate text in the initial candidate set, and Hamming distance is used to calculate the similarity of the signature hash values ​​of any two candidate texts.

[0082] If the similarity of the signature hash values ​​is greater than the set hash similarity threshold, it is judged as a suspected duplicate candidate.

[0083] It should be noted that the two-dimensional hash signature algorithm uses the SimHash algorithm to convert candidate text into a 64-bit hash signature. The Hamming distance is calculated by XOR operation and counting the number of 1s. The hash similarity is calculated as (64 - Hamming distance) / 64. The hash similarity threshold is set to 0.85-0.95, with 0.9 being preferred.

[0084] For suspected duplicate candidates, semantic similarity is calculated based on the text vectors of the two candidates. If the semantic similarity is greater than the set semantic similarity threshold, the candidate is determined to be a duplicate candidate.

[0085] For two candidate texts that are duplicate candidates, retain the one with the higher evidence level; if the evidence levels are the same, retain the one with the higher fusion score; if the fusion scores are also the same, retain the one with the shorter chapter path length, thus obtaining a valid candidate set.

[0086] A dual-dimensional hash signature, combining text content and chapter path, quickly generates a unique identifier. Hamming distance calculation efficiently filters out suspected duplicate candidates, improving deduplication efficiency. Semantic similarity verification based on text vectors accurately determines whether a candidate is a true duplicate, avoiding misjudgments. The retention rule of evidence level > fusion score > chapter path length ensures that the retained candidates are high-quality content with higher authority, stronger relevance, and more focused themes from the duplicate set. Ultimately, redundancy is effectively eliminated, making the effective candidate set more refined and of higher quality, and preventing duplicate information from interfering with clinical decision-making.

[0087] S4. Generate search answers for the valid candidate set according to the preset template and add source annotations.

[0088] Specifically, the system generates structured answers containing key points, suggestions, and precautions for the effective candidate set according to three preset templates: diagnosis, treatment, and medication. The system also extracts the source information of the candidate results in the effective candidate set through a preset source information extraction component and marks them in the format of [document name (including publication year) | chapter path (accurate to sub-chapter) | evidence level]. Each structured answer can be marked with a maximum of 5 sources.

[0089] S5. Evaluate the time-consuming data generated in S2-S3 and determine the anomaly index. If it is greater than the set threshold, the search is retried. If the anomaly index is not greater than the set threshold after a single retry, the search answer is obtained in sequence. If the anomaly index is still greater than the set threshold after multiple retryes, S3 is disabled, and the search answer is generated based on the multi-source candidate set and the preset template.

[0090] The process of determining the abnormal index is as follows: The conversion time for obtaining a set of multiple query variants into a unified query vector, the sorting time for sorting based on composite relevance scores, and the filtering time for fusion-based filtering of multi-source candidate sets.

[0091] The anomaly index is obtained by averaging the conversion time, sorting time, and filtering time.

[0092] It should be noted that the number of retries is set to 3. After the first retrieval is abnormal, it will retry once. If the abnormality index still exceeds the standard, it will retry again. If the standard is still not met after a total of 2 retries, S3 will be disabled.

[0093] By collecting time-consuming data from three key stages—converting multiple query variants into a unified vector, ranking by composite relevance, and filtering by fusion—the efficiency of the core retrieval process is comprehensively reflected. The anomaly index obtained through mean processing can objectively quantify the degree of retrieval anomalies. The retry mechanism based on the index can repair temporary anomalies, and redundant steps are disabled after multiple unsuccessful retries to ensure uninterrupted retrieval service. Ultimately, this ensures the stability and reliability of the intelligent retrieval method, avoids retrieval failures or low efficiency due to process anomalies, and meets the needs of clinical scenarios for continuous retrieval services.

[0094] Intelligent retrieval devices that integrate multi-source medical knowledge, such as Figure 2 As shown, it includes: The query variant acquisition module is used to identify query entities of disease, drug and examination categories in the original query and normalize them to form standardized query text. Based on the standardized query text, effective query variants are constructed and combined with the standardized query text to form a set of multiple query variants.

[0095] The candidate answer output module is used to perform parallel vector retrieval based on multiple query variant sets for the clinical guideline set and the case data set, and to perform filtering and relevance ranking to obtain a multi-source candidate set.

[0096] The effective answer generation module is used to perform fusion score calculation based on evidence level and source type weights on the multi-source candidate set, prioritize the retention of high-level evidence, and deduplicate and conflict adjudicate to obtain the effective candidate set.

[0097] The search results output module is used to generate search answers from the valid candidate set according to a preset template and to annotate the sources.

[0098] The retrieval quality monitoring module is used to evaluate the time-consuming data generated in the candidate answer output module and the valid answer generation module, determine the anomaly index, and if it is greater than a set threshold, the retrieval is retried. If the anomaly index is not greater than the set threshold after a single retry, the retrieval answers are obtained in sequence. If the anomaly index is still greater than the set threshold after multiple retries, the valid answer generation module is disabled, and the retrieval answers are generated based on the multi-source candidate set and the preset template.

[0099] A computer-readable storage medium for storing a program that, when executed by a processor, implements the intelligent retrieval method for integrating multi-source medical knowledge as described above.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0105] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent retrieval method integrating multi-source medical knowledge, characterized in that, Includes the following steps: S1. Identify and normalize the query entities of disease, drug and examination categories in the original query to form a standardized query text. Based on the standardized query text, construct effective query variants and combine them with the standardized query text to form a set of multiple query variants. S2. For the clinical guideline set and the case data set, parallel vector retrieval is performed based on the multi-query variant set, and filtering and composite relevance sorting are performed to obtain a multi-source candidate set; S3. The multi-source candidate set is fused, screened, and high-level evidence is retained to obtain an effective candidate set; S4. Generate search answers for the valid candidate set according to the preset template and add source annotations; S5. Evaluate the time-consuming data generated in S2-S3 and determine the anomaly index. If it is greater than the set threshold, the search is retried. If the anomaly index is not greater than the set threshold after a single retry, the search answer is obtained in sequence. If the anomaly index is still greater than the set threshold after multiple retryes, S3 is disabled, and the search answer is generated based on the multi-source candidate set and the preset template.

2. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 1, characterized in that, The process of constructing effective query variants based on standardized query text is as follows: Replace disease, drug, and examination-related query entities in the standardized query text with preset equivalent expressions, and limit the scenario based on the clinical scenario associated with the query entity; The standardized query text, after being defined by the scenario, is transformed into an equivalent sentence structure in clinical retrieval. At the same time, the limiting words are adjusted or supplemented to generate candidate query variants. Semantic similarity calculation is used to calculate the semantic similarity between each candidate query variant and the standardized query text, and candidate query variants with semantic similarity greater than the semantic similarity threshold are selected as valid query variants.

3. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 1, characterized in that, The process of obtaining the multi-source candidate set is as follows: Each query variant in the set of multiple query variants is transformed into a query vector with the same dimension as the document vector, and a unified query vector is generated by weighted fusion of all query vectors; The cosine similarity of the unified query vector with all document vectors in the clinical guidelines set and the case data set is calculated, and the similarity score is output. For the documents corresponding to the document vectors of the clinical guideline set and the case data set, sort them in descending order of similarity score, and extract the top N documents as the initial candidate results. The initial candidate results are filtered field by field based on the metadata filtering fields to obtain the filtered candidate set. The metadata filtering fields include document type, level of evidence, publication time interval, disease entity and chapter path prefix. For each candidate result in the filtered candidate set, calculate the chapter consistency score and the source authority score, and then perform a weighted sum to obtain the composite relevance score; Based on the composite correlation, the candidate results are sorted in descending order, and the top K candidate results are selected to form a multi-source candidate set.

4. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 1, characterized in that, The process of fusing and screening the multi-source candidate set and preserving high-level evidence to obtain the effective candidate set is as follows: Obtain the preset base weights for evidence level and source type, determine the weight adjustment coefficient based on the multi-query variant set, and multiply the base weights for evidence level and source type by the weight adjustment coefficient to obtain the adjusted evidence level weight. and source type adjustment weight ; Deep medical semantic matching is performed on the multi-query variant set and the multi-source candidate set to generate a semantic matching score for each candidate text. ; Combining composite correlation analysis Calculate the fusion score for each candidate text The calculation formula is: ； in, For composite term weighting factors, For semantic item weighting factors; For guideline candidates in the multi-source candidate set whose evidence level is not less than a set level, evidence-based relevance verification is performed based on the multi-query variant set to obtain the matching confidence level, and guideline candidates whose matching confidence level is not less than a set confidence threshold are selected as priority candidates. Identify clinical scenarios with multiple query variant sets and generate dynamic fusion difference thresholds; If a priority candidate satisfies that the difference between its fusion score and the fusion scores of all other priority candidates is not less than the dynamic fusion score difference threshold, then the priority candidate is retained. If the number of preferred candidates to be retained is greater than M, then the top M candidates are retained in descending order of fusion score. If the number of preferred candidates retained is less than M, then the non-preferred candidates are merged with the preferred candidates, and the top M are retained in descending order of fusion score to obtain the preliminary candidate set. The initial candidate set is deduplicated to obtain the effective candidate set.

5. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 4, characterized in that, The process of determining the weight correction coefficient based on the multi-query variant set is as follows: Iterate through the standardized query text and all valid variations in the multi-query variant set, extract keywords that express the search purpose, count the frequency of keyword occurrences, and determine the topic tags based on the purpose corresponding to the most frequent keyword; Based on a biomedical entity recognition model, disease entities, drug entities, and examination entities are extracted from a multi-query variant set. The urgency of the disease is then determined based on these entities. Disease incidence rate and drug management category value ; Based on the given topic tags, determine the corresponding association strength. and the priority quantification value corresponding to the topic tags. ; The calculation formula for the scenario quantification index is as follows: ； in, and All are weighting factors; Retrieve valid search data stored in the database, and quantify and calibrate each valid search data value according to the corresponding scenario. Correlation strength calibration value Manual calibration correction factor and adaptation accuracy ; The similarity of each valid search result is calculated using the Euclidean distance formula for the scene quantification index, association strength, corresponding scene quantification calibration value, and association strength calibration value. ; Valid search data with similarity values ​​greater than a set similarity threshold and an absolute difference between the association strength and the calibrated association strength value not greater than a set absolute threshold are selected. An initial correction coefficient is then calculated based on the selected valid search data. The calculation formula is: ； For the initial correction coefficient By imposing range constraints on the values, the weight correction coefficients are obtained.

6. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 4, characterized in that, The process of performing deep medical semantic matching on the multi-query variant set and the multi-source candidate set to generate a semantic matching score for each candidate text is as follows: Obtain the unified query vector of the multi-query variant set and the candidate text vector set of the multi-source candidate set; Calculate the cosine similarity value, entity matching degree, and topic similarity between the unified query vector and each candidate text vector in the candidate text vector set; The semantic fit is obtained by weighted summation of cosine similarity, entity matching degree and topic similarity, and the semantic fit is normalized to obtain semantic matching score.

7. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 4, characterized in that, The process of deduplicating the initial candidate set to obtain the effective candidate set is as follows: A two-dimensional hash signature algorithm is used to generate a hash signature for each candidate text in the initial candidate set, and Hamming distance is used to calculate the similarity of the signature hash values ​​of any two candidate texts. If the similarity of the signature hash values ​​is greater than the set hash similarity threshold, it is judged as a suspected duplicate candidate; For suspected duplicate candidates, semantic similarity is calculated based on the text vectors of the two candidates. If the semantic similarity is greater than the set semantic similarity threshold, the candidate is determined to be a duplicate candidate. For two candidate texts that are duplicate candidates, retain the one with the higher evidence level; if the evidence levels are the same, retain the one with the higher fusion score; if the fusion scores are also the same, retain the one with the shorter chapter path length, thus obtaining a valid candidate set.

8. The intelligent retrieval method for integrating multi-source medical knowledge according to claim 1, characterized in that, The process of evaluating the time-consuming data generated in S2-S3 and determining the anomaly index is as follows: The conversion time for obtaining a set of multiple query variants into a unified query vector, the sorting time for sorting based on composite relevance scores, and the filtering time for filtering multi-source candidate sets through fusion and scoring. The anomaly index is obtained by averaging the conversion time, sorting time, and filtering time.

9. An intelligent retrieval device integrating multi-source medical knowledge, used in the intelligent retrieval method integrating multi-source medical knowledge as described in any one of claims 1-8, characterized in that, include: The query variant acquisition module is used to identify query entities of disease, drug and examination categories in the original query and normalize them to form standardized query text. Based on the standardized query text, effective query variants are constructed and combined with the standardized query text to form a set of multiple query variants. The candidate answer output module is used to perform parallel vector retrieval based on multiple query variant sets for the clinical guideline set and the case data set, and to perform filtering and relevance ranking to obtain a multi-source candidate set; The effective answer generation module is used to perform fusion score calculation based on evidence level and source type weights on the multi-source candidate set, prioritize the retention of high-level evidence, and deduplicate and conflict adjudicate to obtain the effective candidate set; The search results output module is used to generate search answers from the valid candidate set according to a preset template and to annotate the sources. The retrieval quality monitoring module is used to evaluate the time-consuming data generated in the candidate answer output module and the valid answer generation module, determine the anomaly index, and if it is greater than a set threshold, the retrieval is retried. If the anomaly index is not greater than the set threshold after a single retry, the retrieval answers are obtained in sequence. If the anomaly index is still greater than the set threshold after multiple retries, the valid answer generation module is disabled, and the retrieval answers are generated based on the multi-source candidate set and the preset template.

10. A computer-readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the intelligent retrieval method for integrating multi-source medical knowledge as described in any one of claims 1 to 8.

Images