Evidence-based medicine literature retrieval methods, devices, electronic devices, and media based on hybrid semantic understanding

By employing a hybrid semantic understanding approach, the PICOS structured retrieval system is constructed using large language models and vector models. This solves the problems of automatic parsing accuracy and structured constraints in evidence-based medicine literature retrieval in existing technologies, enabling automated retrieval from natural language queries to evidence level constraints, thereby improving retrieval accuracy and recall.

CN123115387APending Publication Date: 2026-11-17INST OF MEDICAL INFORMATION CHINESE ACAD OF MEDICAL SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611362029.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing evidence-based medicine literature retrieval systems suffer from problems such as low accuracy of automatic natural language parsing, non-standard semantic expansion of medical terminology, lack of PICOS structured constraints in hybrid retrieval, high cost of manual screening of evidence levels, and difficulty in balancing retrieval accuracy and recall.

Method used

A hybrid semantic understanding-based approach is adopted, which extracts PICOS quintuples through a large language model, combines offline-built semantic extended index and vector model hybrid index to perform hybrid retrieval, and configures a three-level fallback mechanism. Finally, it sorts and filters based on evidence level labels to achieve end-to-end automated retrieval.

Benefits of technology

When users input natural language queries, the system automatically completes query understanding and structured conversion, ensuring that the search results meet user needs in the five dimensions of PICOS, improving search accuracy and recall, and reducing manual screening costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115387A_ABST
    Figure CN123115387A_ABST
Patent Text Reader

Abstract

This application discloses a method, apparatus, electronic device, and medium for evidence-based medicine literature retrieval based on hybrid semantic understanding. The method and apparatus are applied to electronic devices. Specifically, the method involves first classifying and labeling medical literature using multiple tags, generating evidence level tags based on an evidence level pyramid mapping; constructing a semantically extended index oriented towards PICOS based on the existing CMeSH thesaurus; generating dense and sparse vectors of the literature using a two-stage fine-tuned vector model to construct a hybrid index; extracting the PICOS quintuple of the user query using a large language model, and performing BM25 retrieval, dense vector retrieval, and sparse vector retrieval on each element after CMeSH expansion; fusing these three approaches and implementing PICOS forced intersection with a three-level fallback mechanism; and finally sorting and filtering the search results according to the evidence level tags to obtain the final search results. Through the synergy of the above operations, the system can automatically complete query understanding and structured conversion to obtain the final search results when the user only inputs a natural language query, thus achieving end-to-end automated evidence-based medicine literature retrieval from natural language to evidence level constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of medical information retrieval technology, and more specifically, to a method, device, electronic device, and medium for retrieving evidence-based medical literature based on hybrid semantic understanding. Background Technology

[0002] Evidence-based medicine emphasizes that clinical decisions must be based on the best available scientific evidence. In evidence-based medicine practice, literature is graded according to research type and quality, with systematic reviews, meta-analyses, and randomized controlled trials at the top of the evidence pyramid, while case reports and expert opinions are at lower levels. However, with millions of new papers added to mainstream medical literature databases each year, clinicians face severe information overload when searching.

[0003] To address this challenge, various retrieval aids have been developed. Keyword matching based on Boolean logic is the most widely used method, exemplified by systems like PubMed's Boolean query function. However, this method requires users to manually construct complex search queries, demanding a high level of expertise and retrieval experience. Some systems have begun to introduce the PICOS structured query aid tool, helping users to standardize their clinical questions according to five dimensions: population, intervention, control, outcome, and study design. Furthermore, a few studies have attempted to use machine learning to classify the level of evidence in literature. Some commercial systems and open-source frameworks employ hybrid retrieval models combining traditional retrieval with vector retrieval, such as the BM25 plus SBERT architecture.

[0004] While the aforementioned technologies have improved literature retrieval efficiency to some extent, several key shortcomings remain. First, evidence level screening heavily relies on manual intervention, requiring users to manually sift through massive amounts of literature, resulting in low efficiency and a high risk of omissions. Second, a vocabulary gap exists between Chinese and English searches, leading to insufficient utilization of existing thesauruses. Third, the conversion from natural language to PICOS is immature, either requiring manual input or only supporting template filling, failing to automatically and accurately complete the conversion. Fourth, hybrid searches lack the structured constraints of PICOS, making it impossible to guarantee that results simultaneously meet user needs across all five dimensions.

[0005] In summary, existing evidence-based medicine literature retrieval suffers from multiple defects, including low accuracy of automatic PICOS parsing of natural language, non-standard semantic expansion of medical terminology, lack of PICOS structured constraints in mixed retrieval, reliance on manual screening for evidence levels, and an inability to balance retrieval accuracy and recall. It also fails to achieve end-to-end retrieval from query understanding, structure conversion, semantic expansion, and evidence quality constraints when users input natural language queries. Summary of the Invention

[0006] In view of this, this application provides a method, device, electronic device and medium for evidence-based medicine literature retrieval based on hybrid semantic understanding, aiming to solve the technical problems of existing technologies that are difficult to accurately interpret natural language clinical questions, lack standardized medical semantic expansion, lack of structured constraints in hybrid retrieval, high cost of manual screening of evidence levels, and difficulty in balancing retrieval accuracy and literature recall, so as to realize end-to-end fully automated evidence-based medicine literature retrieval with evidence level constraints.

[0007] To achieve the above objectives, the following solution is proposed:

[0008] A method for retrieving evidence-based medicine literature based on hybrid semantic understanding, applied to electronic devices, includes the following steps:

[0009] Receive natural language queries input by users and extract PICOS quintuples through a large language model;

[0010] Based on the offline-built semantic expansion index, the semantics of each element in the PICOS quintuple are expanded to generate an expanded set of query elements.

[0011] Based on the offline constructed vector model hybrid index, a hybrid retrieval is performed on each element in the query element set to obtain the candidate document set corresponding to each element;

[0012] PICOS forced intersection is performed on the candidate document set corresponding to each element. If the number of intersection results is less than the preset threshold, a three-level backoff mechanism is executed to obtain the final candidate document set.

[0013] Based on the evidence level labels generated offline, the final candidate document set is sorted and filtered to obtain the search results.

[0014] Optionally, receiving a natural language query input by the user and extracting PICOS quintuples using a large language model includes the following steps:

[0015] Using an example-guided prompting approach, the large language model is required to output PICOS quintuples in strict JSON format, and it is explicitly required that if a certain element is not explicitly mentioned in the user input, the field should output an empty string, and speculation or supplementation is prohibited.

[0016] Optionally, the step of receiving a natural language query input by the user and extracting PICOS quintuples through a large language model further includes the following steps:

[0017] Each element in the PICOS quintuple output by the large language model is semantically retrieved and verified in the CMeSH thesaurus. If an element fails to reach a preset semantic similarity threshold with any topic word in CMeSH, it is determined that there is a risk of hallucination, and the extraction step is repeated.

[0018] If the preset semantic similarity threshold is still not reached after re-extraction, the system automatically marks the element as pending confirmation.

[0019] The PICOS quintuple is displayed in an editable form on the search interface, allowing users to correct elements in the pending confirmation state and any issues they discover.

[0020] Optionally, the hybrid index based on the offline constructed vector model performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, including the following steps:

[0021] For the type of each element in the currently edited PICOS quintuple, the user-input text is encoded into a query vector using the BGE-M3 basic model;

[0022] Retrieve multiple candidate terms with the highest cosine similarity to the query vector from the CMeSH vector index;

[0023] For each candidate term, a weighted similarity is calculated based on the weight of its type with the current element;

[0024] Sort the terms by weighted similarity from high to low, return a list of sorted terms for the user to select, and add the selected terms as extended synonyms of the element to obtain the query element set.

[0025] Optionally, the hybrid index based on the offline constructed vector model, which performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, further includes the following steps:

[0026] By utilizing the Chinese-English glossary of CMeSH, when the user enters Chinese text, the corresponding English CMeSH terms are also displayed in the recommendation list;

[0027] For CMeSH terms with multiple candidate meanings, disambiguation is performed using other elements already entered by the user.

[0028] Optionally, the hybrid index based on the offline constructed vector model performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, including the following steps:

[0029] For each non-empty element in the candidate document set, sparse search and dense vector search are performed respectively, and the top-ranked documents are selected from each to obtain the sparse search candidate set and the dense search candidate set.

[0030] The multiple documents are ranked according to the weighted sum of the lexical retrieval score and the semantic retrieval score as the fusion score;

[0031] The top few documents with the highest fusion scores under each element are selected as the candidate document set corresponding to each element.

[0032] Optionally, the hybrid index based on the offline constructed vector model, which performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, further includes the following steps:

[0033] Parameter search is performed based on a small-scale labeled query set, and the default values ​​of the fusion weights of lexical retrieval and semantic retrieval are obtained by optimizing the retrieval quality indicators.

[0034] Based on the semantic clarity and terminology of the user query, the fusion weights are dynamically adjusted within a preset range: when the query is semantically ambiguous or is a general description, the weight of lexical retrieval is reduced to emphasize semantic retrieval; when the query contains rare professional terms, the weight of lexical retrieval is increased to emphasize lexical retrieval.

[0035] Optionally, the three-level rollback mechanism includes a first-level rollback, a second-level rollback, and a third-level rollback, wherein:

[0036] The first-level rollback is as follows: if the number of intersection results is less than a preset value, then remove the research design element and calculate the intersection of the remaining four candidate sets of elements;

[0037] The second-level rollback is as follows: if the number of intersection results after the first-level rollback is still less than the preset value and the control element exists, then the control element is further removed, and the intersection of the three candidate sets of elements, population, intervention, and outcome, is calculated.

[0038] The three-level fallback is as follows: if the number of intersection results after the two-level fallback is still less than the preset value, then the forced intersection is abandoned and a weighted union is used instead. The union of all candidate sets of elements is taken, and the score of each document is the weighted sum of the fusion scores of each element. The documents with the highest ranking are returned, along with a relaxed matching prompt.

[0039] Optionally, the final candidate document set is sorted and filtered based on the evidence level labels generated offline to obtain search results, including:

[0040] Sort by evidence level in ascending order, and within the same level, sort by PICOS fusion score in descending order;

[0041] If the research design type specified by the user is not empty, the ranking weight will be dynamically adjusted according to the research design type specified by the user: the research design type specified by the user will be promoted to the highest temporary level, and other types in the original evidence level sequence will be moved to the next level in the original order.

[0042] The system receives filtering instructions input by the user, performs real-time filtering on the current list based on evidence level labels, and obtains the search results.

[0043] Optionally, it also includes an offline preprocessing operation, which includes the following steps:

[0044] Each document in the document database is annotated at multiple levels to generate document type labels and corresponding evidence level labels.

[0045] The semantically extended index is constructed based on CMeSH;

[0046] The vector model hybrid index is constructed based on the fine-tuned BGE-M3 base model, which supports both dense vector retrieval and sparse vector retrieval.

[0047] Optionally, the step of performing multi-level annotation on each document in the document database to generate document type labels and corresponding evidence level labels includes the following steps:

[0048] We use PubMedBERT, which is pre-trained in the medical field, as the base model. We add two fully connected network layers to the output of BERT. The number of nodes in the output layer is the total number of predefined document types. We use the Sigmoid activation function for multi-label classification.

[0049] We use weighted binary cross-entropy as the loss function. For each document type, we calculate the weight based on the frequency of that type in the training set and truncate the weight within a preset range.

[0050] Constructing difficult-to-response examples for training: For each document with a type label, randomly select one document with a different type label from the same CMeSH main subject terms as a difficult-to-response example.

[0051] Optionally, the step of performing multi-level annotation on each document in the document database to generate document type labels and corresponding evidence level labels further includes the following steps:

[0052] The model prediction results and the hard rule results are combined by voting. The hard rules include: if the title of the document contains preset keywords, the corresponding document type label is forcibly added; if the journal to which the document belongs is a preset journal, the corresponding document type label is forcibly added.

[0053] If the rule label conflicts with the model label, the rule label shall prevail.

[0054] For documents whose document type labeling module outputs nothing or whose confidence levels for all types are below the preset threshold, a default evidence level is automatically assigned.

[0055] Optionally, constructing the semantically extended index based on CMeSH includes:

[0056] Read the Chinese and English entries and inter-word relationships officially released by CMeSH, and concatenate the Chinese name, English name and common aliases of each entry. Then, input the BGE-M3 basic model that has been fine-tuned by medical corpus to generate dense vectors.

[0057] The dense vectors of all terms are stored in the Faiss index to obtain the semantic extended index.

[0058] Optionally, the construction of the vector model hybrid index based on the fine-tuned BGE-M3 base model includes the following steps:

[0059] The BGE-M3 basic model undergoes two-stage domain fine-tuning. The first stage involves pre-training using a general medical corpus for masked language modeling and contrastive learning. The second stage uses constructed positive and negative pairs for retrieval task fine-tuning. Positive pairs include titles and abstracts of the same document, abstract pairs with citation relationships, and different reports from the same clinical trial. Negative pairs include random irrelevant document pairs, irrelevant document pairs from the same issue of the same journal, document pairs sharing multiple CMeSH keywords but with significantly different levels of evidence, and document pairs with high lexical retrieval scores but are actually irrelevant. The loss function employs a combination of contrastive loss and ranking loss.

[0060] A literature retrieval device for evidence-based medicine based on hybrid semantic understanding, applied to electronic devices, the retrieval device comprising:

[0061] The element extraction module is configured to receive natural language queries input by the user and extract PICOS quintuples through a large language model;

[0062] The semantic extension module is configured to perform semantic extension on each element in the PICOS quintuple based on offline construction, and generate an extended set of query elements.

[0063] The hybrid retrieval module is configured to perform a hybrid index based on an offline-built vector model, and performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element.

[0064] The intersection processing module is configured to perform PICOS forced intersection on the candidate document set corresponding to each element. If the number of intersection results is less than the preset threshold, a three-level backoff mechanism is executed to obtain the final candidate document set.

[0065] The sorting and filtering module is configured to sort and filter the final candidate document set based on offline generated evidence level labels to obtain search results.

[0066] Optionally, it also includes an offline preprocessing module, which is configured to perform the following steps:

[0067] Each document in the document database is annotated at multiple levels to generate document type labels and corresponding evidence level labels.

[0068] The semantically extended index is constructed based on CMeSH;

[0069] The vector model hybrid index is constructed based on the fine-tuned BGE-M3 base model, which supports both dense vector retrieval and sparse vector retrieval.

[0070] An electronic device includes at least one processor and a memory connected to the processor, wherein:

[0071] The memory is used to store computer programs or instructions;

[0072] The processor is used to execute the computer program or instructions to enable the electronic device to implement the evidence-based medicine literature retrieval method as described above.

[0073] A computer-readable storage medium is applied to an electronic device, the storage medium carrying one or more computer programs that can be executed by the electronic device to enable the electronic device to perform the evidence-based medicine literature retrieval method as described above.

[0074] As can be seen from the above technical solution, this application discloses a method, device, electronic device, and storage medium for evidence-based medicine literature retrieval based on hybrid semantic understanding. The method and device are applied to electronic devices, specifically involving: first, multi-label classification and annotation of medical literature; generating evidence level labels based on the evidence level pyramid mapping; constructing a semantically extended index for PICOS based on the existing CMeSH thesaurus; generating dense and sparse vectors of the literature using a vector model fine-tuned in two stages to construct a hybrid index; extracting the PICOS quintuple of the user query using a large language model, and performing BM25 retrieval, dense vector retrieval, and sparse vector retrieval on each element after CMeSH expansion; implementing PICOS forced intersection and configuring a three-level fallback mechanism after fusion of the three approaches; and sorting and filtering the retrieval results according to the evidence level labels to obtain the retrieval results. Through the synergy of the above operations, the system can automatically complete query understanding and structured conversion to obtain the final retrieval results when the user only inputs a natural language query, thus achieving end-to-end automated evidence-based medicine literature retrieval from natural language queries to evidence level constraints. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a flowchart illustrating an evidence-based medicine literature retrieval method according to an embodiment of this application;

[0077] Figure 2 This is a flowchart illustrating an offline preprocessing operation according to an embodiment of this application;

[0078] Figure 3 This is a block diagram of an evidence-based medicine literature retrieval device according to an embodiment of this application;

[0079] Figure 4 This is a block diagram of another evidence-based medicine literature retrieval device according to an embodiment of this application;

[0080] Figure 5 This is a block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0081] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0082] This application applies to the field of medical information retrieval based on natural language processing, enabling users to achieve end-to-end retrieval processes including query understanding, structured conversion, semantic extension, and evidence quality constraints from the input of natural language queries. The specific technical solution is described below.

[0083] Figure 1 This is a flowchart illustrating an evidence-based medicine literature retrieval method according to an embodiment of this application.

[0084] like Figure 1 As shown, the evidence-based medicine literature retrieval method based on hybrid semantic understanding provided in this embodiment is applied to an electronic device, which can be understood as a computer, server, or cloud platform with data computing and information processing capabilities. It is used to retrieve literature based on pre-trained document type labels and corresponding evidence level labels, semantic expansion indexes, and a hybrid index of vector models. The retrieval method specifically includes the following steps:

[0085] S1. Receive natural language queries and extract PICOS quintuples using a large language model.

[0086] The system receives natural language query requests and corresponding query content from users, and extracts the PICOS quintuple through a large language model. The PICOS quintuple includes five elements: population (P), intervention (I), control (C), outcome (O), and study design (S).

[0087] PICOS is a standardized tool for constructing clinical questions in evidence-based medicine. Here, P (Population) refers to the study population; I (Intervention) refers to the intervention taken in the study population; C (Comparison) refers to the control group or comparison measure; O (Outcome) refers to the outcome measure; and S (Study Design) refers to the type of study design (such as randomized controlled trials, systematic reviews, etc.). Using these five dimensions, ambiguous clinical questions can be clearly expressed using standardized methods, forming the basis for evidence-based medicine searches and systematic reviews.

[0088] Specifically, the system uses a few-shot prompting approach to invoke large language models such as GLM-4, DeepSeek-V3, and GPT-4. The prompts explicitly require the model to output strictly JSON-formatted PICOS quintuples, and explicitly require that if a certain element is not explicitly mentioned in the user input, that field should output an empty string, prohibiting speculation or supplementation. Simultaneously, the system sets the model inference temperature parameter to 0.1 to limit output randomness and reduce the probability of hallucinations.

[0089] For example, when a user inputs "Cardiovascular outcome risk of patients with type 2 diabetes using dapagliflozin: Meta-analysis of a randomized controlled trial", the model outputs a 5-tuple in JSON format: P = "Patients with type 2 diabetes", I = "dapagliflozin", C = "Placebo or standard treatment", O = "Cardiovascular outcome risk", S = "Randomized controlled trial, Meta-analysis".

[0090] To further eliminate potential illusions arising from large language models, this application performs semantic retrieval verification on each element of the PICOS quintuple output by the model in the CMeSH thesaurus. If an element fails to reach a preset semantic similarity threshold (≥0.70) with any CMeSH thesaurus, it is considered to pose a risk of illusion, and the extraction step is repeated. If the preset semantic similarity threshold is still not reached after re-extraction, the system automatically marks the element as pending confirmation. The PICOS quintuple is displayed in an editable format on the search interface, allowing users to correct elements in the pending confirmation state and any issues they discover. Only quintuples confirmed by the user are allowed to proceed to the next search stage.

[0091] S2. Based on the offline-built semantic expansion index, semantic expansion processing is performed on each element in the PICOS quintuple to generate a set of query elements.

[0092] Specifically, for the currently edited PICOS element type T, the system encodes the user-input text q into a query vector v_q using the BGE-M3 basic model. It then retrieves the top K terms with the highest cosine similarity to v_q from the CMeSH vector index (Faiss), e.g., K=15, resulting in a candidate list L={(c_i, sim_i)}. For each candidate term c_i, the system queries the weight w_T(c_i) of that term on the current element type T from the offline stored weight vector, calculating the weighted similarity sim_weighted(q, c_i, T)=sim_i×(1+λ×w_T(c_i)), where λ is the weighting strength coefficient, with an empirical value of 0.3. The system sorts the terms by weighted similarity from high to low and displays the top 10 terms as checkboxes to the user. The terms selected by the user are added to the query as expanded synonyms for that element, resulting in an expanded set of query elements.

[0093] Meanwhile, leveraging CMeSH's Chinese-English comparison function, when the user inputs Chinese text, the corresponding English CMeSH words are simultaneously displayed in the recommendation list, supporting cross-language lookup. For CMeSH entries with multiple candidate meanings, such as "cold" which can correspond to "common cold" or "low temperature," disambiguation is performed using other PICOS elements already entered by the user. Specifically, for each candidate meaning of the word to be disambiguated, the semantic consistency score between that meaning and the set of PICOS elements already determined by the user is calculated, and the meaning with the highest score is selected as the final mapping result.

[0094] S3. Based on the offline constructed vector model hybrid index, perform hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element.

[0095] Specifically, for each non-empty element e (an element with at least one value among P, I, C, O, and S) in the query element set, the system performs two types of retrieval: First, it performs a BM25 sparse search on the title and abstract fields stored in Elasticsearch, retrieving the top 200 documents to obtain a sparse search candidate set; second, it encodes the extended text of element e into a query vector using a fine-tuned BGE-M3 base model, and retrieves the 200 most similar documents in the Faiss dense vector index to obtain a dense search candidate set. The system sorts the documents according to the fusion score Score_e(d) = α × BM25_score(e, d) + (1-α) × cosine_sim(v_q_e, v_d_dense), where α is the fusion weight of lexical retrieval and semantic retrieval. The system selects the top 150 documents with the highest fusion scores for each element as the candidate document set corresponding to that element.

[0096] The method for determining and adjusting the fusion weight α is as follows: The system performs a grid search based on a small-scale labeled query set (500 clinical questions), and optimizes α to 0.3 as the system default value using NDCG@10 as the indicator. Simultaneously, the system dynamically adjusts α within the range of 0.2 to 0.5 based on the semantic clarity (word vector dispersion) and terminology specialization (proportion of rare words) of the user query: when the query is semantically ambiguous or a general description, α is reduced to emphasize semantic retrieval; when the query contains rare specialized terms, α is increased to emphasize lexical retrieval.

[0097] S4. Perform PICOS forced intersection on the candidate document set corresponding to each element. If the number of intersection results is less than the preset threshold, execute the three-level rollback mechanism.

[0098] A forced intersection (PICOS) is performed on the candidate literature sets corresponding to each element. This involves taking the intersection of the candidate literature sets for the five elements: population (P), intervention (I), control (C), outcome (O), and study design (S) (operating only on elements with actual user values): R_picos = S_P ∩ S_I ∩ S_C ∩ S_O ∩ S_S. This forced intersection strategy ensures that the final literature meets the user's needs simultaneously across all five dimensions.

[0099] If the number of intersection results is less than the preset value (e.g., 5 articles), the system will execute a three-level rollback mechanism in sequence: first-level rollback, second-level rollback, and third-level rollback.

[0100] The first-level rollback is as follows: Remove the research design element S, calculate R'=∩_{e∈{P, I, C, O}}S_e, and record the removal prompt at the same time.

[0101] The second-level rollback is as follows: if the number of results after the first-level rollback is still less than the preset value and the control element C exists, then the control element C is further removed, and R''=S_P∩S_I∩S_O is calculated.

[0102] The third-level fallback is as follows: if the number of results after the second-level fallback is still less than the preset value, then the forced intersection is abandoned and the weighted union is used instead, that is, the union of all candidate sets of elements is taken, the score of each document is Σ_{e∈active}score_e(d), and the top-ranked documents (such as the top 20 documents) are returned, along with a relaxed matching prompt.

[0103] The system attempts the above steps sequentially, and stops backtracking once a sufficient number of results are obtained, thus obtaining the final candidate document set.

[0104] S5. Based on the evidence level labels generated offline, the final candidate document set is sorted and filtered to obtain the search results.

[0105] Specifically, results are sorted in ascending order of evidence level (Level 1 is the highest, decreasing with each subsequent level). Within the same level, results are sorted in descending order of PICOS fusion score. If the user-specified research design type is not empty, the system dynamically adjusts the sorting weights based on the user-specified research design type: the user-specified research design type is promoted to the temporary highest level, such as RCT being promoted to Level 1. Other types in the original evidence level sequence are shifted to the next level in their original order, with a minimum level of 6. This adjustment only applies to the current search result ranking and does not change the original evidence level label of the literature in the database.

[0106] The system receives filtering instructions from the user, such as "show only level 1 evidence" or "hide level 3 and below evidence", filters the current list in real time according to the evidence level labels, and finally outputs the search results.

[0107] The present invention will be further illustrated by a complete retrieval example below.

[0108] Scenario setting: Clinicians need to search for evidence from randomized controlled trials on the effect of empagliflozin on all-cause mortality in patients with heart failure.

[0109] Step 1: The user enters a natural language query on the search interface: "The effect of empagliflozin on all-cause mortality in patients with heart failure requires a randomized controlled trial."

[0110] Step 2: The system invokes a large language model to extract the PICOS quintuple using example-guided prompts. The output is: P = "heart failure patients", I = "empagliflozin", C = "the user did not explicitly mention a control, so the model outputs an empty string as instructed", O = "all-cause mortality", and S = "randomized controlled trial". The model output is then semantically validated against the CMeSH lexicon, and all elements meet the similarity threshold. The system displays the quintuple on the search interface. The user confirms that element C is empty (no need to add it) and proceeds to the next step.

[0111] Step 3: The user clicks the expand button next to the P element, causing the system to perform a search based on the CMeSH semantic expansion index, recommending terms such as Heart Failure, Chronic Heart Failure, HFrEF3, etc. The user selects Heart Failure and HFrEF. The O element is also expanded accordingly. Finally, the expanded set of query elements is generated.

[0112] Step 4: Perform a hybrid search on the four non-empty elements P, I, O, and S. That is, perform BM25 sparse search and BGE-M3 dense vector search on each element separately, and select the top 150 documents from each after merging and sorting as the candidate set. Since element C is empty, the system does not perform a search on element C.

[0113] Step 5: Perform PICOS forced intersection on the candidate sets of the four elements: R=S_P∩S_I∩S_O∩S_S, resulting in 38 documents.

[0114] Step 6: Sort the 38 articles according to the level of evidence label. Because the user specified S="Randomized Controlled Trial", RCT articles are temporarily upgraded to Level 1 evidence, while systematic reviews are downgraded to Level 2. The sorted results list is output, with the first item being the EMPEROR-Reduced trial published in the *New England Journal of Medicine* in 2025, tagged as RCT, and displayed as Level 1 (RCT, boosted). Users can select to display only Level 1 evidence on the results screen; the system will filter in real time, showing only articles with Level 1 evidence.

[0115] In addition, the base model for document type labeling can be other pre-trained biomedical language models such as BioLinkBERT; the vector index can be other approximate nearest neighbor retrieval libraries such as HNSW; the large language model can be any model that supports Few-shot Prompting, such as the GLM series, GPT series, Claude series, etc.; the preset thresholds (intersection number less than 5 documents, semantic similarity threshold 0.70, confidence threshold 0.3) are all exemplary values ​​and can be flexibly adjusted according to the actual application scenario.

[0116] As can be seen from the above technical solution, this embodiment provides a method for evidence-based medicine literature retrieval based on hybrid semantic understanding. This method is applied to electronic devices. Specifically, it involves first classifying and labeling medical literature using multiple tags, generating evidence level tags based on an evidence level pyramid mapping; constructing a semantically extended index for PICOS based on the existing CMeSH thesaurus; generating dense and sparse vectors of the literature using a vector model fine-tuned in two stages to construct a hybrid index; extracting the PICOS quintuple of the user query using a large language model, and performing BM25 retrieval, dense vector retrieval, and sparse vector retrieval on each element after CMeSH expansion; fusing these three approaches and implementing PICOS forced intersection with a three-level fallback mechanism; and finally sorting and filtering the search results according to the evidence level tags to obtain the search results. Through the synergy of the above operations, the system can automatically complete query understanding and structured conversion to obtain the final search results when the user only inputs a natural language query, thus achieving end-to-end automated evidence-based medicine literature retrieval from natural language queries to evidence level constraints.

[0117] In one specific embodiment of this application, the application further includes an offline preprocessing operation. This operation forms the basis of the retrieval method of this invention, and the following offline preprocessing operation is performed before executing the online retrieval step. The specific process is as follows: Figure 2 As shown.

[0118] S101. Perform multi-level annotation on each document in the document database to generate document type labels and corresponding evidence level labels.

[0119] Specifically, the system uses microsoft / BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext as the base model. This model is a BERT model pre-trained on the PubMed corpus (containing approximately 30 million biomedical literature abstracts). BERT is a pre-trained language model based on the Transformer architecture. Its core mechanisms include masked language modeling and next-sentence prediction, enabling it to examine the left and right contexts of words during training, thereby generating vector representations containing rich semantic information. PubMedBERT, based on the BERT architecture, is specifically pre-trained on biomedical literature, thus exhibiting excellent representation capabilities for biomedical texts. The system adds two fully connected layers to the BERT output, with a structure of 768 dimensions → 256 dimensions → L dimensions, where L is the total number of predefined document types, such as 15 categories, including Systematic Review, Meta-Analysis, Randomized Controlled Trial, Controlled Clinical Trial, Case Reports, etc. The output layer uses the Sigmoid activation function for multi-label classification, meaning a document can belong to multiple types simultaneously.

[0120] This application employs weighted binary cross-entropy as the loss function. For each document type c, the frequency f_c of that type in the training set is statistically analyzed, and the weight w_c = median(f) / f_c is calculated. w_c is then truncated within the range [0.5, 5.0] to address the problem of uneven distribution of document types. During training, the system constructs difficult-to-fail examples: for each document with type label T, a document with a type label other than T is randomly selected from the same CMeSH main topic terms as a difficult-to-fail example, forcing the model to learn the fine-grained ability to distinguish different research types within the same clinical topic.

[0121] After the model predictions are completed, the prediction results are merged with the hard rule results. Hard rules include: if the document title contains preset keywords (e.g., systematic review and meta-analysis), the corresponding document type tag is forcibly added; if the journal to which the document belongs is a preset journal (e.g., Cochrane Database Syst Rev), the SystematicReview tag is forcibly added. The final tag is the union of the model tag and the rule tag. If the rule tag conflicts with the model tag, the rule tag prevails.

[0122] The system pre-establishes a mapping table between document types and levels of evidence: Level 1 (highest) corresponds to Systematic Review, Meta-Analysis, Guideline, etc.; Level 2 corresponds to Randomized-Controlled Trial; Level 3 corresponds to Controlled Clinical Trial, Clinical Study, etc.; Level 4 corresponds to Case Reports, Observational Study, etc.; Level 5 corresponds to Comment, Editorial, Letter, etc.; and Level 6 corresponds to Veterinary related studies. For documents where the annotation module output is empty or the confidence level for all types is below the preset threshold (e.g., 0.3), the system automatically assigns a default level of evidence 6 and records it in the audit log.

[0123] S102. Construct a semantic extended index based on CMeSH.

[0124] The system constructs a semantically extended index based on the Chinese Medical Subject Headings (CMeSH) published by the Institute of Medical Information, Chinese Academy of Medical Sciences. CMeSH is a bilingual (Chinese-English) subject heading list, based on the standard Chinese translation of the US National Library of Medicine's MeSH (Medical Subject Headings), with localized extensions to subject terms and headings to meet the needs of applications in my country's medical and health fields. CMeSH integrates all content and functions of the medical subject heading annotation table, tree structure table, and wheel-order table, including hierarchical relationships (tree structure), synonym relationships, and correlation relationships between subject terms.

[0125] Specifically, the system reads the Chinese and English entries and inter-word relationships published by CMeSH, concatenates the Chinese name, English name and common aliases of each entry, inputs them into the BGE-M3 basic model that has been fine-tuned by medical corpus to generate dense vectors, and stores all entry vectors in the Faiss index for semantic nearest neighbor retrieval.

[0126] Faiss (Facebook AI Similarity Search) is a large-scale vector similarity search library developed by Facebook AI Research. Its core idea is to perform efficient similarity searches through vector quantization and inverted indexing. The system uses the IVF4096_PQ index structure—IVF (Inverted File) is a cluster-based index structure that divides the vector space into multiple cluster centers (4096 in this example). During the search, it first finds the cluster centers closest to the query vector, and then precisely searches in the corresponding inverted list, thus significantly narrowing the search range. PQ (Product Quantization) divides the high-dimensional vector into multiple sub-vectors and quantizes and encodes them separately, compressing vector storage space and reducing computational complexity while ensuring retrieval accuracy.

[0127] Furthermore, using an annotated literature database (at least 100,000 articles, each annotated with PICOS elements and the CMeSH indexing terms used), the frequency f_t(c) of each CMeSH term c appearing in the literature as an indexing term for each PICOS element (P, I, C, O, S) is counted. For each element type t, a comprehensive weight w_t(c) = [f_t(c) / (Σ_{t'}f_{t'}(c)+ε)]×log(N / f_total(c)) is calculated, which incorporates global importance, where ε is the smoothing term, N is the total number of articles in the database, and f_total(c) is the total frequency of term c among all PICOS elements, resulting in a weight vector W_c = (w_P, w_I, w_C, w_O, w_S). This weight vector is stored in Redis, supports millisecond-level queries, and is incrementally updated based on user interaction feedback.

[0128] S103. Construct a hybrid index based on the fine-tuned BGE-M3 base model.

[0129] Two-stage domain fine-tuning was performed on the BGE-M3 basic model. The BGE-M3 basic model is a general embedding model released by the Beijing Academy of Artificial Intelligence. It supports three functions: dense vector retrieval, sparse vector retrieval, and multi-vector retrieval. It supports more than 100 languages ​​and can handle inputs with a maximum of 8192 tokens.

[0130] Embedding vectors are numerical representations that map text to a high-dimensional vector space. Dense vectors have low dimensionality (e.g., 1024 dimensions), with all dimensions being non-zero. They are generated by models such as BERT and are commonly used for semantic similarity calculation. Traditional sparse vectors (such as word frequency vectors generated by the BM25 algorithm) are high-dimensional with most dimensions having zero values, making them suitable for keyword matching. The BGE-M3 base model generates "learned" sparse vectors, calculating importance weights for each word by applying the ReLU activation function to a linear layer, thus incorporating contextual semantics into the sparse representation and combining the advantages of accurate matching and semantic understanding. BM25 (OkapiBM25) is a ranking function based on the bag-of-words model, which improves upon the TF-IDF method by introducing a word frequency saturation mechanism and document length normalization, and is widely used in the field of information retrieval.

[0131] Phase 1: Using PubMed Central open-source paper abstracts (approximately 3 million), we continued pre-training using masked language modeling and contrastive learning (positive example: the title and abstract of the same article) to adapt the model to the style of medical texts.

[0132] Phase Two: Constructing positive and negative example pairs for fine-tuning the retrieval task. Positive example pairs include: titles and abstracts of the same document, abstract pairs with citation relationships, and different reports from the same clinical trial. Negative example pairs include: random, unrelated document pairs, two unrelated documents from the same issue of the same journal (semantically similar but unrelated), document pairs sharing at least two CMeSH keywords but with evidence levels differing by more than two levels (the aim is to widen the evidence level distance in the embedding space), and samples with high BM25 scores but actually irrelevant. The loss function uses a combination of InfoNCE contrastive loss (temperature τ=0.07) and MarginRankingLoss (margin=0.2). Training hyperparameters: batch size 256 (using gradient accumulation), optimizer AdamW, learning rate 2e-5, linear warmup 10%, 3 training epochs.

[0133] After fine-tuning, the title, abstract, and document type tag list of each document are concatenated and input into the fine-tuned BGE-M3 base model, outputting both dense and sparse vectors. The mechanism for generating sparse vectors in BGE-M3 is as follows: a linear transformation and ReLU activation function are applied to the contextualized embedding of each token, and the importance weight w_t=ReLU(W_lex^T·H[i]) of each word is calculated, ultimately obtaining a sparse mapping from word to weight. Dense vectors are stored in the Faiss index, sparse vectors are stored in the inverted index, and the document's metadata (title, author, evidence level, document type list, DOI) is stored in Elasticsearch. These three are linked through the document ID, thus constructing a hybrid index that simultaneously supports dense and sparse vector retrieval.

[0134] The above operations enable offline preprocessing, providing the necessary tools for online retrieval.

[0135] Although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous.

[0136] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0137] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including but not limited to object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer.

[0138] Figure 3 This is a block diagram of an evidence-based medicine literature retrieval device according to an embodiment of this application.

[0139] like Figure 3As shown, the evidence-based medicine literature retrieval device based on hybrid semantic understanding provided in this embodiment is applied to an electronic device, which can be understood as a computer, server, or cloud platform with data computing and information processing capabilities. It is used to retrieve literature based on pre-trained document type labels and corresponding evidence level labels, semantic expansion indexes, and vector model hybrid indexes. Specifically, the retrieval device includes an element extraction module 10, a semantic expansion module 20, a hybrid retrieval module 30, an intersection processing module 40, and a sorting and filtering module 50.

[0140] The element extraction module is used to receive natural language queries and extract PICOS quintuples through a large language model.

[0141] The system receives natural language query requests and corresponding query content from users, and extracts the PICOS quintuple through a large language model. The PICOS quintuple includes five elements: population (P), intervention (I), control (C), outcome (O), and study design (S).

[0142] The semantic expansion module is used to perform semantic expansion processing on each element in the PICOS 5-tuple based on the offline-built semantic expansion index, generating a set of query elements. The specific process has been described in detail above and will not be repeated here.

[0143] The hybrid retrieval module is used for hybrid indexing based on an offline-built vector model. It performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element. The specific process has been described in detail above and will not be repeated here.

[0144] The intersection processing module is used to perform PICOS forced intersection on the candidate document sets corresponding to each element. If the number of intersection results is less than the preset threshold, a three-level rollback mechanism is executed.

[0145] A forced intersection (PICOS) is performed on the candidate literature sets corresponding to each element. This involves taking the intersection of the candidate literature sets for the five elements: population (P), intervention (I), control (C), outcome (O), and study design (S) (operating only on elements with actual user values): R_picos = S_P ∩ S_I ∩ S_C ∩ S_O ∩ S_S. This forced intersection strategy ensures that the final literature meets the user's needs across all five dimensions. The specific process has been detailed above and will not be repeated here.

[0146] The sorting and filtering module is used to sort and filter the final candidate document set based on the evidence level labels generated offline, thereby obtaining the search results. The specific process has been described in detail above and will not be repeated here.

[0147] As can be seen from the above technical solution, this embodiment provides an evidence-based medicine literature retrieval device based on hybrid semantic understanding. This device is applied to electronic devices. Specifically, it first performs multi-label classification and annotation on medical literature, generating evidence level labels based on the evidence level pyramid mapping; it then constructs a semantically extended index for PICOS based on the existing CMeSH thesaurus; it uses a vector model fine-tuned in two stages to generate dense and sparse vectors for the literature, constructing a hybrid index; it extracts the PICOS quintuples of the user query using a large language model, expands them with CMeSH, and performs BM25 retrieval, dense vector retrieval, and sparse vector retrieval on each element respectively; after fusion of these three approaches, it implements PICOS forced intersection and configures a three-level fallback mechanism; finally, it sorts and filters the retrieval results according to the evidence level labels to obtain the retrieval results. Through the synergy of the above operations, it can automatically complete query understanding and structured conversion to finally obtain the retrieval results when the user only inputs a natural language query, thus achieving end-to-end automated evidence-based medicine literature retrieval from natural language queries to evidence level constraints.

[0148] In one specific embodiment of this application, an offline preprocessing module 60 is also included, such as... Figure 4 As shown. This module serves as the foundation of the retrieval method of this invention, performing the following offline preprocessing operations before executing the online retrieval step. The specific pre-training process has been described in detail above and will not be repeated here.

[0149] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The name of a unit does not necessarily limit the unit itself; for example, the first acquisition unit can also be described as "a unit that acquires at least two Internet Protocol addresses".

[0150] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0151] Figure 5 This is a block diagram of an electronic device according to an embodiment of this application.

[0152] The following is for reference. Figure 5This document illustrates a structural diagram suitable for implementing the electronic device in the embodiments of this disclosure. The terminal device in the embodiments of this disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. This electronic device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this disclosure.

[0153] The electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from an input device 506 into a random access memory (RAM) 503. The RAM also stores various programs and data required for the operation of the electronic device. The processing unit, ROM, and RAM are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0154] Typically, the following devices can be connected to the I / O interface: input devices including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 508 including, for example, magnetic tapes, hard disks, etc.; and communication devices 509. Communication device 509 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although electronic devices with various devices are shown in the figures, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0155] This application also provides a computer-readable medium embodiment.

[0156] The aforementioned computer-readable medium is applied to an electronic device and carries one or more computer programs. When these programs are executed by the electronic device, the device performs multi-label classification and annotation of medical literature, generates evidence level labels based on the evidence level pyramid mapping, constructs a semantic extension index for PICOS based on the existing CMeSH thesaurus, generates dense and sparse vectors for the literature using a two-stage fine-tuned vector model, and constructs a hybrid index. A large language model is used to extract the PICOS quintuples from the user query, and after CMeSH expansion, BM25 retrieval, dense vector retrieval, and sparse vector retrieval are performed on each element respectively. After the three-way fusion, PICOS forced intersection is implemented and a three-level fallback mechanism is configured. The search results are sorted and filtered according to the evidence level labels to obtain the search results. Through the synergy of the above operations, the system can automatically complete query understanding and structured transformation to obtain the final search results when the user only inputs a natural language query, thus achieving end-to-end automated evidence-based medicine literature retrieval from natural language queries to evidence level constraints.

[0157] It should be noted that the computer-readable medium described above in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0158] In this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0159] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0160] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0161] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0162] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for retrieving evidence-based medicine literature based on hybrid semantic understanding, applied to electronic devices, characterized in that: The retrieval method includes the following steps: Receive natural language queries input by users and extract PICOS quintuples through a large language model; Based on the offline-built semantic expansion index, the semantics of each element in the PICOS quintuple are expanded to generate an expanded set of query elements. Based on the offline constructed vector model hybrid index, a hybrid retrieval is performed on each element in the query element set to obtain the candidate document set corresponding to each element; PICOS forced intersection is performed on the candidate document set corresponding to each element. If the number of intersection results is less than the preset threshold, a three-level backoff mechanism is executed to obtain the final candidate document set. Based on the evidence level labels generated offline, the final candidate document set is sorted and filtered to obtain the search results.

2. The evidence-based medicine literature retrieval method as described in claim 1, characterized in that, The process of receiving natural language queries input by the user and extracting PICOS quintuples using a large language model includes the following steps: Using an example-guided prompting approach, the large language model is required to output PICOS quintuples in strict JSON format, and it is explicitly required that if a certain element is not explicitly mentioned in the user input, the field should output an empty string, and speculation or supplementation is prohibited.

3. The evidence-based medicine literature retrieval method as described in claim 2, characterized in that, The process of receiving a natural language query input from a user and extracting PICOS quintuples using a large language model also includes the following steps: Each element in the PICOS quintuple output by the large language model is semantically retrieved and verified in the CMeSH thesaurus. If an element fails to reach a preset semantic similarity threshold with any topic word in CMeSH, it is determined that there is a risk of hallucination, and the extraction step is repeated. If the preset semantic similarity threshold is still not reached after re-extraction, the system automatically marks the element as pending confirmation. The PICOS quintuple is displayed in an editable form on the search interface, allowing users to correct elements in the pending confirmation state and any issues they discover.

4. The evidence-based medicine literature retrieval method as described in claim 1, characterized in that, The hybrid index based on the offline-built vector model performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, including the following steps: For the type of each element in the currently edited PICOS quintuple, the user-input text is encoded into a query vector using the BGE-M3 basic model; Retrieve multiple candidate terms with the highest cosine similarity to the query vector from the CMeSH vector index; For each candidate term, a weighted similarity is calculated based on the weight of its type with the current element; Sort the terms by weighted similarity from high to low, return a list of sorted terms for the user to select, and add the selected terms as extended synonyms of the element to obtain the query element set.

5. The evidence-based medicine literature retrieval method as described in claim 4, characterized in that, The hybrid index based on the offline-built vector model performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, and also includes the following steps: By utilizing the Chinese-English glossary of CMeSH, when the user enters Chinese text, the corresponding English CMeSH terms are also displayed in the recommendation list; For CMeSH terms with multiple candidate meanings, disambiguation is performed using other elements already entered by the user.

6. The evidence-based medicine literature retrieval method as described in claim 1, characterized in that, The hybrid index based on the offline-built vector model performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, including the following steps: For each non-empty element in the candidate document set, sparse search and dense vector search are performed respectively, and the top-ranked documents are selected from each to obtain the sparse search candidate set and the dense search candidate set. The multiple documents are ranked according to the weighted sum of the lexical retrieval score and the semantic retrieval score as the fusion score; The top few documents with the highest fusion scores under each element are selected as the candidate document set corresponding to each element.

7. The evidence-based medicine literature retrieval method as described in claim 6, characterized in that, The hybrid index based on the offline-built vector model performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element, and also includes the following steps: Parameter search is performed based on a small-scale labeled query set, and the default values ​​of the fusion weights of lexical retrieval and semantic retrieval are obtained by optimizing the retrieval quality indicators. Based on the semantic clarity and terminology of the user query, the fusion weights are dynamically adjusted within a preset range: when the query is semantically ambiguous or is a general description, the weight of lexical retrieval is reduced to emphasize semantic retrieval; when the query contains rare professional terms, the weight of lexical retrieval is increased to emphasize lexical retrieval.

8. The evidence-based medicine literature retrieval method as described in claim 1, characterized in that, The three-level rollback mechanism includes a first-level rollback, a second-level rollback, and a third-level rollback, wherein: The first-level rollback is as follows: if the number of intersection results is less than a preset value, then remove the research design element and calculate the intersection of the remaining four candidate sets of elements; The second-level rollback is as follows: if the number of intersection results after the first-level rollback is still less than the preset value and the control element exists, then the control element is further removed, and the intersection of the three candidate sets of elements, population, intervention, and outcome, is calculated. The three-level fallback is as follows: if the number of intersection results after the two-level fallback is still less than the preset value, then the forced intersection is abandoned and a weighted union is used instead. The union of all candidate sets of elements is taken, and the score of each document is the weighted sum of the fusion scores of each element. The documents with the highest ranking are returned, along with a relaxed matching prompt.

9. The evidence-based medicine literature retrieval method as described in claim 1, characterized in that, The evidence level labels generated offline are used to sort and filter the final candidate document set to obtain search results, including: Sort by evidence level in ascending order, and within the same level, sort by PICOS fusion score in descending order; If the research design type specified by the user is not empty, the ranking weight will be dynamically adjusted according to the research design type specified by the user: the research design type specified by the user will be promoted to the highest temporary level, and other types in the original evidence level sequence will be moved to the next level in the original order. The system receives filtering instructions input by the user, performs real-time filtering on the current list based on evidence level labels, and obtains the search results.

10. The evidence-based medicine literature retrieval method according to any one of claims 1 to 9, characterized in that, It also includes offline preprocessing operations, which include the following steps: Each document in the document database is annotated at multiple levels to generate document type labels and corresponding evidence level labels. The semantically extended index is constructed based on CMeSH; The vector model hybrid index is constructed based on the fine-tuned BGE-M3 base model, which supports both dense vector retrieval and sparse vector retrieval.

11. The evidence-based medicine literature retrieval method as described in claim 10, characterized in that, The step of performing multi-level annotation on each document in the literature database to generate document type labels and corresponding evidence level labels includes the following steps: We use PubMedBERT, which is pre-trained in the medical field, as the base model. We add two fully connected network layers to the output of BERT. The number of nodes in the output layer is the total number of predefined document types. We use the Sigmoid activation function for multi-label classification. We use weighted binary cross-entropy as the loss function. For each document type, we calculate the weight based on the frequency of that type in the training set and truncate the weight within a preset range. Constructing difficult-to-response examples for training: For each document with a type label, randomly select one document with a different type label from the same CMeSH main subject terms as a difficult-to-response example.

12. The evidence-based medicine literature retrieval method as described in claim 11, characterized in that, The step of performing multi-level annotation on each document in the document database to generate document type labels and corresponding evidence level labels also includes the following steps: The model prediction results are merged with the hard rule results. The hard rules include: if the title of a document contains preset keywords, the corresponding document type label is forcibly added; if the journal to which the document belongs is a preset journal, the corresponding document type label is forcibly added. If the rule label conflicts with the model label, the rule label shall prevail. For documents whose document type labeling module outputs nothing or whose confidence levels for all types are below the preset threshold, a default evidence level is automatically assigned.

13. The evidence-based medicine literature retrieval method as described in claim 10, characterized in that, The construction of the semantically extended index based on CMeSH includes: Read the CMeSH Chinese and English entries and the relationships between them. For each entry, concatenate its Chinese name, English name and common aliases, and input them into the BGE-M3 basic model that has been fine-tuned by the medical corpus to generate a dense vector. The dense vectors of all terms are stored in the Faiss index to obtain the semantic extended index.

14. The evidence-based medicine literature retrieval method as described in claim 10, characterized in that, The construction of the vector model hybrid index based on the fine-tuned BGE-M3 base model includes the following steps: The BGE-M3 basic model undergoes two-stage domain fine-tuning. The first stage involves pre-training using a general medical corpus for masked language modeling and contrastive learning. The second stage uses constructed positive and negative pairs for retrieval task fine-tuning. Positive pairs include titles and abstracts of the same document, abstract pairs with citation relationships, and different reports from the same clinical trial. Negative pairs include random irrelevant document pairs, irrelevant document pairs from the same issue of the same journal, document pairs sharing multiple CMeSH keywords but with significantly different levels of evidence, and document pairs with high lexical retrieval scores but are actually irrelevant. The loss function employs a combination of contrastive loss and ranking loss.

15. A device for retrieving evidence-based medicine literature based on hybrid semantic understanding, applied to electronic devices, characterized in that, The retrieval device includes: The element extraction module is configured to receive natural language queries input by the user and extract PICOS quintuples through a large language model; The semantic extension module is configured to perform semantic extension on each element in the PICOS quintuple based on offline construction, and generate an extended set of query elements. The hybrid retrieval module is configured to perform a hybrid index based on an offline-built vector model, and performs a hybrid retrieval on each element in the query element set to obtain the candidate document set corresponding to each element. The intersection processing module is configured to perform PICOS forced intersection on the candidate document set corresponding to each element. If the number of intersection results is less than the preset threshold, a three-level backoff mechanism is executed to obtain the final candidate document set. The sorting and filtering module is configured to sort and filter the final candidate document set based on offline generated evidence level labels to obtain search results.

16. The evidence-based medicine literature retrieval device as described in claim 15, characterized in that, It also includes an offline preprocessing module, which is configured to perform the following steps: Each document in the document database is annotated at multiple levels to generate document type labels and corresponding evidence level labels. The semantically extended index is constructed based on CMeSH; The vector model hybrid index is constructed based on the fine-tuned BGE-M3 base model, which supports both dense vector retrieval and sparse vector retrieval.

17. An electronic device, characterized in that, The electronic device includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs or instructions; The processor is used to execute the computer program or instructions to enable the electronic device to implement the evidence-based medicine literature retrieval method as described in any one of claims 1 to 14.

18. A computer-readable medium for use in an electronic device, characterized in that, The storage medium carries one or more computer programs that can be executed by the electronic device, thereby enabling the electronic device to perform the evidence-based medicine literature retrieval method as described in any one of claims 1 to 14.