Methods, computing devices, computer program products, and computer-readable storage media for constructing a knowledge graph of medical literature

Abstract

This disclosure provides a method, computing device, computer program product, and computer-readable storage medium for constructing a knowledge graph of medical literature. The method includes: extracting multiple granularity knowledge objects from the medical literature using a medical domain-specific cue word template-driven large language model, wherein the multiple granularity knowledge objects include entities, relationships between entities, and declarations of each entity and / or relationship; determining key excerpts from the medical literature for each knowledge object using a progressively narrowing matching algorithm based on preceding and following anchor points; and classifying the knowledge objects according to the key excerpts of each knowledge object and the context of the corresponding medical literature using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature.

Description

Technical Field

[0001] This disclosure generally relates to the field of machine learning, and more specifically, to a method, computing device, computer program product, and computer-readable storage medium for constructing knowledge graphs of medical literature. Background Technology

[0002] The construction of medical knowledge bases heavily relies on the structured extraction and evidence-based grading of heterogeneous literature such as clinical guidelines, drug instructions, and gene testing reports. Traditional knowledge base construction techniques primarily employ a Retrieval-augmented Generation (RAG) architecture. This architecture segments documents into fixed-length text blocks, establishes vector indexes, retrieves relevant text blocks through semantic similarity retrieval, and then generates answers using a large language model. Some systems have introduced knowledge graph technology, utilizing large language models to extract entities and relationships from text, forming a graph-structured knowledge representation. The recently emerging GraphRAG method builds upon this foundation by adding community detection and community report generation, producing multi-level summaries through graph structure clustering, thus enhancing the ability to answer global questions.

[0003] In the field of evidence-based medicine, the Oxford Centre for Evidence-Based Medicine has established an evidence grading system from Level 1a (systematic reviews of multiple randomized controlled trials) to Level 5 (expert opinions), and a recommendation strength grading from Level A (strong recommendation) to Level D (weak recommendation). This system is the core standard for assessing the credibility of evidence in clinical decision-making.

[0004] However, the aforementioned background technology has the following structural defects:

[0005] First, the knowledge granularity of traditional RAG systems stops at the text block level. Specifically, text blocks are mechanically divided into fixed number of characters. A text block may contain multiple entities, multiple relationships, and multiple statements. However, when a user asks a question, they can only locate the text block as a whole and cannot distinguish which entity attribute, which relationship between entities, or which clinical statement a specific statement in the answer originates from. This coarse-grained mapping leads to a lack of knowledge traceability.

[0006] Secondly, even with the introduction of knowledge graph extraction, existing methods only preserve the semantic descriptions of graph nodes and edges, failing to establish a precise mapping from knowledge objects to specific sentences in the original documents. Because large language models often suffer from rewriting, truncation, or encoding deviations when generating key original text extracts, the extracted text is not entirely consistent with the original text, making it impossible to accurately locate and highlight it within the original text block. This deviation breaks the "answer to original text" tracing chain.

[0007] Third, traditional systems lack structured assessments of evidence levels after generating knowledge graphs, failing to distinguish between high-level evidence from systematic reviews and low-level evidence from expert opinions. Clinical users, when viewing relationships or claims within the knowledge graph, are unable to determine the strength of evidence, severely limiting the clinical usability of medical knowledge bases.

[0008] Therefore, there is a need to construct a medical knowledge graph that combines high traceability and evidence-based grading capabilities in an automated and fine-grained manner. Summary of the Invention

[0009] To address at least one of the aforementioned problems, this disclosure provides a method for constructing knowledge graphs of medical literature, achieving evidence-based knowledge graph construction through a three-stage pipeline. This method decomposes medical literature into three fine-grained knowledge objects: entities, relations, and claims. It ensures accurate matching between key excerpts and the original text, and automatically assigns structured evidence classifications to relations and claims according to the Oxford Grading System, thereby overcoming the triple shortcomings of traditional techniques: coarse knowledge granularity, inaccurate original text matching, and lack of evidence classification.

[0010] According to one aspect of this disclosure, a method for constructing a knowledge graph of medical literature is provided. The method includes: extracting multiple granular knowledge objects from the medical literature using a medical domain-specific cue word template-driven large language model, wherein the multiple granular knowledge objects include entities, relationships between entities, and declarations of each entity and / or relationship; determining key excerpts from the medical literature for each knowledge object using a progressive reduction matching algorithm based on preceding and following anchor points; and classifying the knowledge objects according to the key excerpts of each knowledge object and the context of the corresponding medical literature using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature.

[0011] In some embodiments, extracting knowledge objects of multiple granularities from the medical literature includes: dividing the medical literature into multiple text blocks, wherein adjacent text blocks in the multiple text blocks have overlapping areas; and using the prompt word template to drive a large language model to extract the multiple granularities of knowledge objects from the multiple text blocks.

[0012] In some embodiments, extracting knowledge objects at multiple granularities from the medical literature further includes: for the same entity in the multiple text blocks, generating a comprehensive summary of the entity based on multiple descriptions of the entity in different text blocks; performing entity clustering for all entities in the multiple granularities of knowledge objects to generate at least one community; and generating a community report for each community, wherein each community report includes a community name, an executive summary, an impact score, and multiple key insights, wherein each key insight is accompanied by a data source.

[0013] In some embodiments, dividing the medical literature into multiple text blocks includes: extracting text from the medical literature and preserving paragraph structure and chart position information; using a visual language model to convert images in the medical literature into structured text descriptions and inserting them into the positions of the images to obtain a text file corresponding to the medical literature; dividing the text file into multiple text blocks of a fixed length, wherein adjacent text blocks in the multiple text blocks have overlapping areas; and assigning a document identifier of the medical literature to each text block.

[0014] In some embodiments, using the prompt word template to drive a large language model to extract knowledge objects of multiple granularities from the multiple text blocks includes: based on a pre-determined list of entity types, using the prompt word template to drive a large language model to extract knowledge objects of multiple granularities from the multiple text blocks, wherein for entity type knowledge objects, at least one of entity name, entity type, and entity description is extracted; for relation type knowledge objects, relation description and relation strength score are extracted; and for declaration type knowledge objects, at least one of subject, object, declaration type, declaration state, declaration description, and original text citation is extracted.

[0015] In some embodiments, determining key excerpts from the medical literature for each knowledge object using a matching algorithm that progressively reduces anchor points includes: for each knowledge object, selecting several candidate text blocks from all text blocks based on the keyword hit rate and text block length of the knowledge object and each text block in all text blocks of the medical literature; and using a large language model to select the candidate text block with the highest relevance to the knowledge object from the several candidate text blocks as the key evidence text block, and determining the key excerpts associated with the key evidence text block.

[0016] In some embodiments, determining key excerpts for each knowledge object from the medical literature using a matching algorithm that progressively reduces anchor points further includes: determining whether the key excerpt is a continuous substring of the key evidence text block; and if it is determined that the key excerpt is not a continuous substring of the key evidence text block, repairing the key excerpt using the matching algorithm that progressively reduces anchor points to obtain repaired key excerpts.

[0017] In some embodiments, the key text excerpt is repaired using a matching algorithm that progressively reduces the number of preceding and following anchor points to obtain the repaired key text excerpt. This includes: taking the first n characters of the key text excerpt (E) as the preceding anchor point P, and taking the last n characters of the key text excerpt (E) as the following anchor point S; determining whether the starting position i of the preceding anchor point P can be found in the key evidence text block (T); if the starting position i of the preceding anchor point P can be found in the key evidence text block (T), then searching for the last occurrence of the following anchor point S starting from position i in the key evidence text block (T), and determining whether its ending position j can be found; if the ending position j of the following anchor point S can be found, then taking the continuous substring from the starting position i to the ending position j in the key evidence text block (T) as the repaired key text excerpt; if the starting position i or the ending position j of the preceding anchor point P cannot be found in the key evidence text block (T), then decrementing the value of n and repeating the matching.

[0018] In some embodiments, the key text excerpt is repaired using a matching algorithm that gradually reduces the anchor points before and after, in order to obtain the repaired key text excerpt, the following steps are also taken: if the starting position i or ending position j of the previous anchor point P cannot be found in the key evidence text block (T) when n drops to a preset threshold, then a key phrase is extracted from the key text excerpt (E), and the shortest continuous interval containing the key phrase is located in the key evidence text block (T) as the repaired key text excerpt.

[0019] In some embodiments, the knowledge objects are classified using a predefined evidence-based medicine grading system to generate a knowledge graph of the medical literature based on key original text extracts and the context of the corresponding medical literature for each knowledge object. This includes inputting the context of the medical literature corresponding to the knowledge object, key evidence text blocks, key original text extracts, and the definition of the predefined evidence-based medicine grading system into a large language model to output a structured grading result.

[0020] In some embodiments, the method of classifying the knowledge objects based on key excerpts of each knowledge object and the context of the corresponding medical literature using a predefined evidence-based medicine grading system to generate a knowledge graph of the medical literature further includes: if the output structured grading results contain a thinking process, then the thinking content is extracted from the grading results, and only the structured grading conclusions are retained.

[0021] In some embodiments, the predefined evidence-based medicine grading system is a two-track system, including: a first track, which is the Oxford evidence grading system, outputting structured results containing evidence level, major level of evidence, and recommendation level; and a second track, which is the evidence quality and recommendation strength system, outputting structured results containing evidence quality and recommendation strength.

[0022] In some embodiments, a matching algorithm with progressively reduced anchor points is used to determine key original text extracts for each knowledge object from the medical literature. This is performed in batches using a first concurrency configuration, and intermediate results are periodically persisted during the processing. Based on the key original text extracts of each knowledge object and the context of the corresponding medical literature, the knowledge objects are classified using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature. This is performed in batches using a second concurrency configuration lower than the first concurrency configuration, and checkpoints are periodically saved.

[0023] According to another aspect of this disclosure, a computing device is provided. The computing device includes: at least one processor; and at least one memory coupled to the at least one processor and storing instructions for execution by the at least one processor, the instructions, when executed by the at least one processor, causing the computing device to perform the steps according to the method described above.

[0024] According to another aspect of this disclosure, a computer-readable storage medium is provided having computer program code stored thereon, which, when run, performs the methods described above.

[0025] According to another aspect of this disclosure, a computer program product is provided having computer program code stored thereon, which, when run, performs the method as described above. Attached Figure Description

[0026] This disclosure will be better understood by referring to the following description of specific embodiments given in the accompanying drawings, and other objects, details, features, and advantages of this disclosure will become more apparent.

[0027] Figure 1 A schematic diagram is shown for implementing a computational system for constructing a knowledge graph of medical literature according to embodiments of the present disclosure.

[0028] Figure 2 A flowchart of a method for constructing a knowledge graph of medical literature according to some embodiments of the present disclosure is shown.

[0029] Figure 3 A detailed flowchart of a method for multi-granularity knowledge object extraction according to some embodiments of the present disclosure is shown.

[0030] Figure 4 A detailed flowchart of a method for document preprocessing and text block generation according to some embodiments of the present disclosure is shown.

[0031] Figure 5 A schematic diagram of a community report according to some embodiments of the present invention is shown.

[0032] Figure 6 A detailed flowchart of a method for determining key text extracts according to some embodiments of the present disclosure is shown.

[0033] Figure 7 A detailed flowchart of a method for repairing critical excerpts according to some embodiments of the present disclosure is shown.

[0034] Figure 8 A structural block diagram of a computing device suitable for implementing embodiments of the present disclosure is shown. Detailed Implementation

[0035] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0036] In the following description, certain specific details are set forth for the purpose of illustrating embodiments of the various inventions to provide a thorough understanding of these embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known apparatuses, structures, and techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0037] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0038] Throughout this specification, references to "one embodiment" or "some embodiments" indicate that a particular feature, structure, or characteristic described in connection with an embodiment is included in at least one embodiment. Therefore, the appearance of "in one embodiment" or "in some embodiments" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0039] Furthermore, the terms first, second, third, fourth, etc., used in the specification and claims are used only for the purpose of clarity of description to distinguish between different objects, and do not limit the size or other order of the objects they describe.

[0040] Figure 1 A schematic diagram is shown for implementing a computational system 1 for constructing a knowledge graph of medical literature according to embodiments of the present disclosure. Figure 1 As shown, system 1 includes one or more data sources 10, computing devices 20, and a network 30. Data sources 10 and computing devices 20 can interact via the network 30. Here, data source 10 may be, for example, a device or repository for providing various medical documents (such as clinical guidelines, drug instructions, research papers, etc., in PDF, Word, or image formats). Computing device 20 can acquire these medical documents and process them to construct a knowledge graph. Computing device 20 may include at least one processor 22 and at least one memory 24 coupled to the at least one processor 22, the memory 24 storing instructions 26 executable by the at least one processor 22, which, when executed by the at least one processor 22, perform at least a portion of the methods described below. A specific structure of computing device 20 may be combined, for example, as follows. Figure 8 As stated above.

[0041] According to the scheme disclosed herein, medical literature is first preprocessed and segmented, then a large language model is used to extract knowledge objects (entities, relations, claims) of multiple granularities from the text blocks, then the key original text excerpts for each knowledge object are precisely located, and finally the knowledge objects are automatically classified to generate a knowledge graph with precise original text citations and structured evidence levels.

[0042] Figure 2 A flowchart of a method 200 for constructing a knowledge graph of medical literature according to some embodiments of the present disclosure is shown. Method 200 may, for example, be derived from... Figure 1 The system 1 shown in the diagram is executed by computing device 20. The following example demonstrates execution within computing device 20. Figures 1 to 8 Method 200 is described.

[0043] like Figure 2As shown, method 200 includes step 210, wherein computing device 20 uses medical domain-specific cue word templates to drive a large language model to extract knowledge objects of multiple granularities from medical literature. The multiple granularities of knowledge objects include entities, relationships between entities, and claims for each entity and / or relationship. Here, the large language model can be any currently known or future-developed large language model, such as, but not limited to, DeepSeek V4, Qwen3, Gemini 3 Pro, Doubao 1.5, and their future further development versions.

[0044] Figure 3 A detailed flowchart of a method (step 210) for multi-granularity knowledge object extraction according to some embodiments of the present disclosure is shown. Method 210 may further include sub-steps 212 and 214.

[0045] like Figure 3 As shown, in sub-step 212, computing device 20 divides medical literature (e.g., from data source 10) into multiple text blocks, wherein adjacent text blocks in the multiple text blocks have overlapping areas.

[0046] Figure 4 A detailed flowchart of a method for segmenting text blocks (sub-step 212) according to some embodiments of the present disclosure is shown. Sub-step 212 may further include sub-steps 412, 414, 416, and 418.

[0047] In sub-step 412, computing device 20 can extract text from the input medical document (such as files in PDF, Word, image, etc. formats). For example, a layout-aware OCR (Optical Character Recognition) engine can be used for text extraction. During text extraction, the paragraph structure and figure / table position information of the original document are preserved.

[0048] In sub-step 414, for images (such as charts, flowcharts, medical images, etc.) in the input medical literature, computing device 20 can use a Visual-Language Model (VLM) to convert the image into a structured text description. For example, a bar chart showing the distribution of left ventricular ejection fraction in heart failure patients can be transcribed into the structured text description "This bar chart shows the distribution of left ventricular ejection fraction in heart failure patients, where HFrEF accounts for 42% and HFpEF accounts for 38%." Then, computing device 20 can insert these structured text descriptions into the location of the image in the medical literature.

[0049] Through sub-steps 412 and 414, computing device 20 can obtain the text file corresponding to the medical document. Note that the execution of sub-steps 412 and 414 is not limited to the order shown in the figure and described above, but can be executed in other orders, such as in parallel, or before sub-step 414.

[0050] In sub-step 416, the computing device 20 divides the text file obtained after the above processing into multiple text blocks of a fixed length. For example, the blocks can be divided into blocks of 1200 characters each. Adjacent text blocks have an overlap area, for example, 150 characters. In this way, semantic coherence across text blocks can be ensured, and critical information can be avoided from being cut off.

[0051] In sub-step 418, computing device 20 also assigns a document identifier to each text block, such as a document title, a Digital Object Identifier (DOI), or an internal ID, for subsequent document tracing and source management.

[0052] return Figure 3 In sub-step 214, computing device 20 can use the medical-specific prompt word template to drive a large language model to extract knowledge objects of multiple granularities from multiple text blocks divided in sub-step 212.

[0053] In some embodiments, the cue word template can be constructed based on a pre-defined list of medical entity types. This list may, for example, cover more than eighty medical classifications such as site of occurrence, clinical course, severity, detection technology, active ingredient, and route of administration (e.g., based on the proactive health terminology system published by the Open Medical and Healthcare Alliance (OMAHA)). For entity-type knowledge objects, the cue word template can guide the large language model to extract at least one of the entity name, entity type, and entity description. For relation-type knowledge objects, where two entities in the linked text have a clear association, the cue word template can guide the large language model to extract the relation description and selectively extract the relation strength score (e.g., an integer from 0 to 10). For claim-type knowledge objects, the cue word template can guide the large language model to extract at least one of the following: subject, object, claim type (e.g., "efficacy claim," "safety claim"), claim status (e.g., "proven," "disproven," "pending verification"), claim description, and original text citation. For example, using the OMAHA entity type column as described above, a total of 9,650 entities, 18,026 relations, and 5,111 claims were extracted from the heart failure knowledge base, and a total of 3,681 entities, 6,563 relations, and 2,447 claims were extracted from the IGA nephropathy knowledge base.

[0054] In some embodiments, step 210 may further include sub-steps 216, 218, and 219 (e.g., ...). Figure 3 The dashed box indicates that it is an optional or further sub-step.

[0055] In sub-step 216, for the same entity in multiple text blocks, computing device 20 can generate a comprehensive summary of the entity based on multiple descriptions of the entity in different text blocks. Specifically, the list of descriptions of the entity appearing in all text blocks can be input into a large language model, which outputs a comprehensive summary that integrates all information and eliminates contradictions and redundancy.

[0056] In sub-step 218, for all entities in the knowledge objects at multiple granularities, computing device 20 can perform entity clustering to generate at least one community. For example, computing device 20 can utilize a graph community detection algorithm (e.g., the Leiden algorithm) to cluster these entities according to the principle of maximizing modularity to form one or more communities. Each community represents a set of semantically closely related entities. For example, for a heart failure database, the top-level community could include communities such as "heart failure classification and ejection fraction," "natriuretic peptides and cardiac function assessment," and "SGLT2 inhibitors and heart failure prognosis."

[0057] In step 219, computing device 20 can generate a community report for each community. For example, computing device 20 can use a large language model to generate a report for each community, which may include the community name, executive summary, impact score, and several (e.g., 5-10) key insights, each of which is accompanied by its data source (i.e., the original text block or knowledge object supporting the insight).

[0058] Figure 5 A schematic diagram of a community report according to some embodiments of the present invention is shown. Figure 5 The diagram illustrates four community reports, 510, 520, 530, and 540. Each community report includes multiple semantically related entities.

[0059] like Figure 5 As shown, community report 510 may include entities 511, 512, 513, 514, and 515, where entity 511 is, for example, the entity "heart failure classification," entity 512 is, for example, the entity "ejection fraction," entity 513 is, for example, the entity "HFrEF," and entity 514 is, for example, the entity "HFpEF," etc. Therefore, community report 510 can also be called a community report on "heart failure classification and ejection fraction." Similarly, community report 520 may include entities 521, 522, 523, and 524, community report 530 may include entities 531, 532, 533, and 534, and community report 540 may include entities 541, 542, and 543. Furthermore, community reports may be related; for example, entities in different community reports may have "relationships." Figure 5 As shown, there is a relationship between entity 514 of community report 510 and entity 541 of community report 540, and so on.

[0060] return Figure 2 In step 220, computing device 20 can use a matching algorithm that progressively reduces anchor points to determine key excerpts from the medical literature for each knowledge object.

[0061] Figure 6 A detailed flowchart of a method (step 220) for determining key text extracts according to some embodiments of the present disclosure is shown. Step 220 may further include at least sub-steps 222 and 224.

[0062] In sub-step 222, for each knowledge object (entity, relation, or statement), computing device 20 can select several candidate text blocks from all text blocks based on the keyword hit rate and text block length of each text block in all text blocks of the medical literature. For example, the frequency of keywords in the knowledge object appearing in all text blocks (keyword hit rate) can be calculated, and combined with the length of the text blocks for scoring and sorting, selecting the top K text blocks with the highest scores (e.g., K=5) as candidate text blocks.

[0063] In sub-step 224, computing device 20 can use a large language model to select the candidate text block with the highest relevance to the knowledge object from the several candidate text blocks selected in sub-step 222 as the key evidence text block, and determine the key excerpt associated with the key evidence text block. Based on its understanding of the content of the candidate text blocks, the large language model can determine which text block best supports or explains the knowledge object, and extract the most relevant original text as the key excerpt.

[0064] In some embodiments, step 220 may further include sub-steps 226 and 228 (e.g. Figure 6 (As shown in the dashed box).

[0065] In sub-step 226, computing device 20 can determine whether the key original text extract output by the large language model in sub-step 224 is a continuous substring of the key evidence text block. Because the large language model may rewrite, summarize, or truncate the original text, the output extract may not be completely consistent with the original text.

[0066] In sub-step 228, if it is determined in sub-step 226 that the key original text excerpt is not a continuous substring of the key evidence text block, then the computing device 20 can use a matching algorithm that gradually reduces the anchor points to repair the key original text excerpt in order to obtain the repaired key original text excerpt.

[0067] Figure 7 A detailed flowchart of a method 700 (i.e., sub-step 228) for repairing critical original text extracts according to some embodiments of this disclosure is shown. Figure 7 In the illustrated embodiment, the method for key text extraction and repair is described using a matching algorithm that gradually reduces the number of anchor points. However, those skilled in the art will understand that this disclosure is not limited to this, but other similar matching algorithms can be used, such as matching based on minimizing edit distance (Levenshtein distance) or matching based on the longest common substring of a sliding window.

[0068] Here, we assume that the key original text extract output by the large language model in sub-step 224 is string E, and the selected key evidence text block is string T.

[0069] like Figure 7 As shown, in step 710, the first n characters of the key text extract E are taken as the front anchor point P, i.e., P=E[0:n], and the last n characters of the key text extract E are taken as the back anchor point S, i.e., S=E[-n:]. Initially, n can be set to a large value, for example, n=10.

[0070] In step 720, determine whether the starting position i of the previous anchor point P can be found in the key evidence text block T, i.e., i = T.find(P).

[0071] If the starting position i of the previous anchor point P is found in step 720 (i.e., i>= 0), then in step 730, the last occurrence of the next anchor point S is searched in the key evidence text block T starting from position i (i.e., the substring T[i:] in T starting from position i), and it is determined whether its ending position j can be found. Here, i and j are greater than 0 and less than the length of the key evidence text block T, and j>i.

[0072] If the end position j of the back anchor point S is found in step 730, then in step 740, the continuous substring from the start position i to the end position j in the key evidence text block T (i.e., T[i: i+j]) is used as the repaired key original text extract.

[0073] On the other hand, if the starting position i (i.e., the judgment in step 720 is "no") or the ending position j (i.e., the judgment in step 730 is "no") of the previous anchor point P cannot be found in the key evidence text block T, then in step 750, the value of n is decremented (for example, the value of n is decremented by 1, i.e., n = n - 1), and steps 720 to 740 of the above matching process are repeated.

[0074] Furthermore, Figure 7 The process may further include step 760 after step 750, in which it may be determined whether n is greater than or equal to a preset threshold (e.g., 5). If it is determined in step 760 that n is greater than or equal to the threshold, then steps 720 to 740 of the matching process described above are repeated.

[0075] On the other hand, if it is determined in step 760 that n is less than the threshold, then step 770 can be proceeded to, where method 700 can be terminated or additional easing remedies can be performed.

[0076] Such lenient remedial measures may include, for example, extracting several key phrases (core words retained after word segmentation and removal of stop words) from the key original text extract E, and then locating the shortest continuous interval in the key evidence text block T that can contain these key phrases, and using the shortest continuous interval as the repaired key original text extract.

[0077] In this way, it can be ensured to the greatest extent that the final key excerpt is a continuous string that actually exists in the original text block, thus supporting accurate sentence-level highlighting on the front end.

[0078] return Figure 2 In step 230, the computing device 20 can classify each knowledge object based on key original text extracts and the context of the corresponding medical literature using a predefined evidence-based medicine hierarchy to generate a knowledge graph of medical literature. Typically, entities themselves are not classified; classification primarily targets relations and declarations.

[0079] In some embodiments, the predefined evidence-based medicine grading system can be a dual-track system, and the grading results using the dual-track system can be output in parallel.

[0080] Here, the first track is the Oxford gradation of evidence system, which outputs structured results including the level of evidence (e.g., 1a, 1b, 1c, 2a, 2b, 2c, 3a, 3b, 4, 5), the major level of evidence (1-5, where 1 is the highest), and the recommendation level (A / B / C / D, where A represents a level 1 study with good consistency, and D represents a level 5 study or a study with inconsistent results).

[0081] The second track is the evidence quality and recommendation strength system, which outputs a structured result that includes evidence quality (A / B / C / D, corresponding to high / medium / low / very low quality respectively) and recommendation strength (1=strong recommendation, 2=weak recommendation).

[0082] Specifically, computing device 20 inputs the context of the medical literature corresponding to the knowledge object (usually centered on the key original text, taking several thousand characters before and after to provide sufficient background), key evidence text block T, key original text extract E, and the definition of the aforementioned dual-track hierarchical system into the large language model. Based on this information, the large language model outputs a structured hierarchical result, including specific levels, strengths, and explanations of the hierarchical basis. If the output generated by the reasoning-based large language model contains a chain-of-thought process, the thought content can be further extracted from the hierarchical result, retaining only the structured hierarchical conclusion. The hierarchical result is written into the evidence index of the knowledge graph and stored in association with the key original text extract to form the knowledge graph.

[0083] In some embodiments, the anchoring process of key text extraction in step 220 can be performed in batches using a first concurrency configuration to improve efficiency, such as processing 100 knowledge objects per batch using 100 concurrent threads. Meanwhile, to prevent data loss due to unexpected interruptions in long processes, intermediate results are periodically (e.g., every 50 objects processed) persisted to disk during processing.

[0084] In some embodiments, the hierarchical process in step 230 is performed in batches using a second concurrency configuration lower than the first concurrency configuration, for example, processing 20 objects per batch using 20 concurrent threads. This is because hierarchical processes typically involve more complex inference, and excessive concurrency can lead to unstable model calls or timeouts. Similarly, the process periodically saves checkpoints to prevent the loss of all data due to abnormal model calls.

[0085] Ultimately, the knowledge graph constructed using the above method 200 has fine-grained traceability (can be located to specific sentences in the original text) and a clear level of evidence for each knowledge object (entity, relation, statement), which greatly improves the reliability, credibility and clinical usability of the medical knowledge base.

[0086] For example, in building a heart failure knowledge base, inputting 10 clinical guidelines and research articles can yield tens of thousands of entities, relationships, and statements after processing. For the relationship "SGLT2 inhibitors reduce the risk of heart failure hospitalization," the system will locate the supporting sentences (e.g., "The EMPEROR-Reduced trial showed that empagliflozin significantly reduced the risk of heart failure hospitalization by 25%) and automatically assign it an Oxford evidence level of "1b" (high-quality randomized controlled trial) and a recommendation level of "A." When using the knowledge base, users can click on the relationship to view the highlighted supporting evidence and its level information.

[0087] Figure 8 A structural block diagram of a computing device 800 suitable for implementing embodiments of the present disclosure is shown. The computing device 800 may be, for example, the computing device 20 as described above.

[0088] like Figure 8 As shown, computing device 800 may include one or more central processing units (CPUs) 810 (only one is schematically shown in the figure), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 820 or loaded from storage unit 880 into random access memory (RAM) 830. Various programs and data required for the operation of computing device 800 may also be stored in RAM 830. CPU 810, ROM 820, and RAM 830 are interconnected via bus 840. Input / output (I / O) interface 850 is also connected to bus 840.

[0089] Multiple components in the computing device 800 are connected to the I / O interface 850, including: an input unit 860, such as a keyboard and mouse; an output unit 870, such as various types of displays and speakers; a storage unit 880, such as a hard disk and optical disk; and a communication unit 890, such as a network interface card (NIC), a modem, or a wireless transceiver. The communication unit 890 allows the computing device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0090] The method 200 described above may be executed, for example, by the CPU 810 of the computing device 800. For example, in some embodiments, the method 200 may be implemented as a computer software program tangibly included in a machine-readable medium, such as storage unit 880. In some embodiments, part or all of the computer program may be loaded and / or installed on the computing device 800 via ROM 820 and / or communication unit 890. When the computer program is loaded into RAM 830 and executed by the CPU 810, one or more operations of the method 200 described above may be performed. Furthermore, the communication unit 890 may support wired or wireless communication functions.

[0091] Those skilled in the art will understand that Figure 8 The computing device 800 shown is merely illustrative. In some embodiments, the computing device 20 may contain more or fewer components than the computing device 800.

[0092] The method 200 for constructing a knowledge graph of medical literature according to the present disclosure and the computing device 800 that can be used as a computing device 20 have been described above with reference to the accompanying drawings. However, those skilled in the art will understand that the execution of the steps and sub-steps of method 200 is not limited to the order shown in the figures and described above, but can be performed in any other reasonable order. Furthermore, the computing device 800 does not necessarily include… Figure 8 All components shown may include only some of the components necessary to perform the functions described in this disclosure, and the connection of these components is not limited to the form shown in the figures.

[0093] This disclosure can be a method, apparatus, system, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of this disclosure.

[0094] In one or more exemplary designs, the functions described herein may be implemented using hardware, software, firmware, or any combination thereof. For example, if implemented in software, the functions may be stored as one or more instructions or code on a computer-readable medium, or transmitted as one or more instructions or code on a computer-readable medium.

[0095] The various units of the apparatus disclosed herein can be implemented using discrete hardware components or integrated on a single hardware component, such as a processor. For example, they can be implemented or perform the various exemplary logic blocks, modules, and circuits described herein using a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof for performing the functions described herein.

[0096] Those skilled in the art should also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in connection with embodiments of this disclosure can be implemented as electronic hardware, computer software, or a combination of both.

[0097] The foregoing description of this disclosure is intended to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the spirit and scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the broadest scope of the principles and novel features disclosed herein.

Claims

1. A method for constructing a knowledge graph of medical literature, comprising: A large language model is driven by prompt word templates tailored to the medical field to extract knowledge objects of multiple granularities from the medical literature. These multiple granularities of knowledge objects include entities, relationships between entities, and declarations of each entity and / or relationship. Using a matching algorithm that progressively reduces anchor points, key excerpts are extracted from each knowledge object in the medical literature. as well as Based on key excerpts of the original text for each knowledge object and the context of the corresponding medical literature, the knowledge objects are classified using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature.

2. The method of claim 1, wherein extracting knowledge objects of multiple granularities from the medical literature includes: The medical document is divided into multiple text blocks, wherein adjacent text blocks have overlapping areas. as well as The prompt word template is used to drive a large language model to extract knowledge objects of multiple granularities from the multiple text blocks.

3. The method of claim 2, wherein extracting knowledge objects of multiple granularities from the medical literature further includes: For the same entity in the multiple text blocks, a comprehensive summary of the entity is generated based on multiple descriptions of the entity in different text blocks; For all entities in the knowledge objects at the multiple granularities, perform entity clustering to generate at least one community; as well as A community report is generated for each community, which includes the community name, executive summary, impact score, and multiple key insights, each with its own data source.

4. The method of claim 2, wherein dividing the medical literature into multiple text blocks comprises: The medical literature was subjected to text extraction, and paragraph structure and figure / table location information were preserved. The images in the medical literature are converted into structured text descriptions using a visual language model and then inserted into the positions of the images to obtain a text file corresponding to the medical literature. The text file is divided into multiple text blocks of fixed length, wherein adjacent text blocks have overlapping areas. as well as Assign a document identifier to each text block of the medical document.

5. The method of claim 2, wherein using the prompt word template to drive a large language model to extract knowledge objects of multiple granularities from the multiple text blocks includes: Based on a pre-determined list of entity types, the prompt word template drives a large language model to extract knowledge objects of multiple granularities from the multiple text blocks. For knowledge objects of entity type, at least one of entity name, entity type, and entity description is extracted. For knowledge objects of relation type, relation description and relation strength score are extracted. For knowledge objects of declaration type, at least one of subject, object, declaration type, declaration state, declaration description, and original text citation is extracted.

6. The method of claim 1, wherein determining key excerpts from the medical literature for each knowledge object using a matching algorithm that progressively reduces anchor points includes: For each knowledge object, several candidate text blocks are selected from all text blocks based on the keyword hit rate and text block length of the knowledge object and each text block in all text blocks of the medical literature; as well as Using a large language model, the candidate text block with the highest relevance to the knowledge object is selected from the several candidate text blocks as the key evidence text block, and the key original text excerpt associated with the key evidence text block is determined.

7. The method of claim 6, wherein determining key excerpts from the medical literature for each knowledge object using a matching algorithm that progressively reduces anchor points further comprises: Determine whether the key original text excerpt is a continuous substring of the key evidence text block; as well as If it is determined that the key original text excerpt is not a continuous substring of the key evidence text block, the key original text excerpt is repaired using a matching algorithm that gradually reduces the anchor points before and after, so as to obtain the repaired key original text excerpt.

8. The method of claim 7, wherein the key text extract is repaired using a matching algorithm that progressively reduces anchor points to obtain the repaired key text extract, comprising: Take the first n characters of the key original text excerpt (E) as the first anchor point P, and take the last n characters of the key original text excerpt (E) as the last anchor point S; Determine whether the starting position i of the previous anchor point P can be found in the key evidence text block (T); If the starting position i of the previous anchor point P can be found in the key evidence text block (T), then search for the last occurrence of the subsequent anchor point S starting from position i in the key evidence text block (T) and determine whether its ending position j can be found. If it is determined that the end position j of the back anchor point S can be found, then the continuous substring from the start position i to the end position j in the key evidence text block (T) is taken as the repaired key original text extract. If the start position i or end position j of the previous anchor point P cannot be found in the key evidence text block (T), then the value of n is decremented and the matching is repeated.

9. The method of claim 8, wherein repairing the key text extract using a matching algorithm that progressively reduces anchor points to obtain the repaired key text extract further includes: If the starting position i or ending position j of the previous anchor point P cannot be found in the key evidence text block (T) when n drops to a preset threshold, then the key phrase is extracted from the key original text extract (E), and the shortest continuous interval containing the key phrase is located in the key evidence text block (T) as the key original text extract for repair.

10. The method of claim 1, wherein classifying the knowledge objects using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature based on key excerpts of each knowledge object and the context of the corresponding medical literature comprises: The context of the medical literature corresponding to the knowledge object, key evidence text blocks, key original text excerpts, and the definition of the predefined evidence-based medicine hierarchy are all input into the large language model to output a structured hierarchical result.

11. The method of claim 10, wherein, based on key original text extracts of each knowledge object and the context of the corresponding medical literature, classifying the knowledge objects using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature further includes: If the output structured classification results contain the thinking process, then the thinking content is extracted from the classification results, and only the structured classification conclusions are retained.

12. The method of claim 10, wherein the predefined evidence-based medicine hierarchy is a dual-track system, comprising: The first track uses the Oxford grading system for evidence, which outputs structured results including the level of evidence, the major level of evidence, and the recommendation level. The second track is the evidence quality and recommendation strength system, which outputs structured results that include evidence quality and recommendation strength.

13. The method of claim 1, wherein Using a matching algorithm that gradually reduces the anchor points, key original text extracts are determined for each knowledge object from the medical literature. The first concurrency configuration is used for batch parallel processing, and intermediate results are periodically persisted during the processing. Based on the key original text extracts of each knowledge object and the context of the corresponding medical literature, the knowledge objects are classified using a predefined evidence-based medicine hierarchy to generate a knowledge graph of the medical literature. The knowledge objects are processed in batches using a second concurrency configuration lower than the first concurrency configuration, and checkpoints are saved periodically.

14. A computing device, comprising: At least one processor; as well as At least one memory coupled to the at least one processor and storing instructions for execution by the at least one processor, the instructions, when executed by the at least one processor, causing the computing device to perform the steps of the method according to any one of claims 1 to 13.

15. A computer program product having stored computer program code thereon, which, when run, performs the method as described in any one of claims 1 to 13.

16. A computer-readable storage medium having stored thereon computer program code, which, when run, performs the method as described in any one of claims 1 to 13.

Images