Medical literature reference link tracking method and system

Abstract

The invention specifically discloses a medical literature citation link tracking method and system, relates to medical health information processing, and solves the problems of large citation network scale, data isomerism, incomplete metadata analysis and low tracking efficiency in clinical evidence-based decision and drug safety monitoring. The method comprises the steps that heterogeneous medical literatures are collected and normalized, and evidence-based medical evidence levels are marked; analyzing the reference literature list, and complementing the DOI / PMID metadata; constructing a reference topology network of the graph database, clustering nodes by using a MeSH tree hierarchical structure, and extracting PICO elements; citation intensity is calculated by integrating citation frequency, position weight, context motivation, evidence level and PICO similarity, and key paths are screened; traceability and diffusion tracking are supported, draft withdrawal events and adverse drug reaction signals are dynamically monitored, and real-time early warning is updated through streaming increment. According to the method, visual tracking and risk assessment of the reference relationship of the medical literature are realized, and efficient and accurate evidence tracing support is provided for clinical evidence-based decision and drug safety assessment.

Description

Technical Field

[0001] This invention belongs to the field of medical and health information processing technology, specifically relating to a method and system for tracing medical literature citation links. Background Technology

[0002] In today's medical research field, with the deepening of academic research and the rapid development of medical technology, the output of medical literature has experienced explosive growth, forming a complex global citation network. Citation relationships in medical literature not only record the evolution of scientific discoveries but also serve as crucial evidence for evaluating research value, tracing the source of knowledge, and ensuring academic integrity. By linking and mining massive amounts of literature data, it is possible to effectively reveal the dynamics of the discipline's forefront and assist in clinical evidence-based decision-making, particularly in scenarios such as drug efficacy evaluation, updating treatment guidelines, and detecting adverse reaction signals.

[0003] Among them, the medical literature citation link tracing method and system, as a core technical means to support academic evaluation and knowledge traceability, is based on the principle of constructing a link model that reflects the logical inheritance relationship between documents by accurately identifying and associating metadata of literature citations. This technology aims to extend from single document nodes to multi-dimensional citation networks, thereby providing researchers with a clear, intuitive knowledge dissemination path map with a time dimension.

[0004] Traditional medical literature tracking technologies have revealed several significant shortcomings in practical applications: First, existing solutions often focus on specific negative scenarios such as retractions or miscitations, relying excessively on explicit retraction timelines, resulting in a narrow scope of application and difficulty in systematically covering normal and widespread medical citation behavior. Second, some technologies emphasize content analysis based on semantic similarity, focusing primarily on the diffusion of influence at the content level rather than explicit citation relationships in a strict sense. Due to a lack of deep, structured parsing capabilities for key metadata such as reference lists, digital object identifiers, and medical literature index numbers, the tracking results often contain a large number of unrelated non-citations, severely impacting the accuracy and professionalism of the citation chain. Third, when processing massive amounts of dynamically updated literature data, existing systems lack efficient link reconstruction mechanisms and multi-level visualization capabilities, making it difficult to reflect the full evolution of medical knowledge in real time and accurately. Furthermore, existing technologies generally lack the ability to perceive the levels of evidence in evidence-based medicine, failing to distinguish the differentiated value of systematic reviews and case reports in knowledge dissemination, and failing to introduce a PICO element matching mechanism to assess the clinical relevance of citation relationships, making it difficult for tracking results to directly serve clinical evidence-based decision-making. The existence of these problems means that existing literature tracking methods cannot meet the urgent needs of medical research for high-quality, high-precision, and full-chain knowledge tracing. There is an urgent need to develop an intelligent tracking innovation solution that can integrate evidence-based medicine methodology and deep structured processing. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a method and system for medical literature citation link tracing to solve problems such as large scale of medical citation networks, heterogeneous data sources, incomplete parsing of citation metadata, and insufficient link tracing efficiency.

[0006] To address the technical challenges arising from the rapid growth in the volume of medical literature output, such as complex citation network structures, reliance of traditional tracking technologies on specific negative scenarios, insufficient deep structured parsing capabilities of metadata, and inefficient link reconstruction mechanisms, this invention constructs a medical literature knowledge graph containing multi-dimensional metadata. By combining a high-precision structured parsing engine and a citation strength assessment model, it achieves accurate identification, deep correlation, and multi-level visual tracking of medical literature citation links. Furthermore, through evidence-based medicine evidence level stratification and PICO element matching, it provides systematic evidence chain support for clinical evidence-based decision-making, drug safety monitoring, and traceability of treatment guidelines.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In one aspect, a medical literature citation link tracing system, comprising the following components: The clinical literature collection and evidence level labeling module is used to obtain heterogeneous literature data sources from biomedical literature index databases, clinical trial registration databases and drug regulatory literature databases, perform data cleaning, deduplication and format standardization processing, and label the evidence-based medicine evidence level according to the research design type of the literature report to generate a unified literature stream to be parsed. The PICO element extraction and citation metadata parsing module is used to extract key metadata from the document stream to be parsed using a pre-defined regular expression library and natural language processing model. The key metadata includes Digital Object Identifier (DOI), Medical Literature Index Number (PMID), Clinical Trial Registration Number, Document Title, Author Information, Publication Date, Abstract, and Reference List. The reference list is then split item by item to form independent citation entities. Simultaneously, PICO elements are extracted from the document abstract based on the medical structured abstract format. The PICO elements include patient population characteristics, intervention measures, control protocols, and clinical outcome indicators. The disease-drug dimension medical evidence network construction module is used to establish document node entities and their directional citation relationships based on the extracted key metadata, and construct a medical literature knowledge graph. Each document node is associated with a time attribute label, a subject classification code mapped based on the MeSH tree hierarchical structure, an evidence-based medicine evidence level, and a PICO element vector. The module uses the hierarchical relationships in the MeSH tree hierarchical structure to perform hierarchical clustering of document nodes, so that document nodes under the same disease lineage or the same pharmacology category form a hierarchical subgraph. The evidence-based medicine evidence weight assessment module is used to perform multi-dimensional weight calculation on the directional citation relationship edge. It comprehensively considers the frequency of the citation in the text, the weight of the chapter position of the citation, the position polarity of the citation context, the academic influence index of the cited literature, the evidence-based medicine evidence level weight factor, and the PICO element similarity between the citing and cited literature as a clinical relevance adjustment coefficient to generate a comprehensive citation intensity score. The clinical evidence chain tracing and visualization module is used to perform depth-first traversal or breadth-first traversal algorithms in the medical literature knowledge graph based on the seed literature or keywords specified by the user, and extract and generate a multi-level citation chain tracing topology graph with time dimension by combining the comprehensive citation strength score. The retraction impact assessment and drug safety signal early warning module is used to monitor the update status of nodes in the medical literature knowledge graph in real time. When a retraction, correction, or abnormal fluctuation in citation frequency is detected in a core node in the link, the link reconstruction mechanism is triggered and an early warning is sent to the user. When a retraction event is detected, the potential impact of the retracted literature on downstream clinical research conclusions and related treatment guidelines is assessed based on the MeSH tree hierarchy and PICO element matching relationship, and a risk assessment report is generated that includes the affected disease type, the name of the drug involved, and the associated treatment guideline number. When a new adverse drug reaction report is detected in the citation link, a drug safety signal early warning is triggered.

[0008] On the other hand, the present invention also provides a method for tracing medical literature citation links, the method comprising the following steps: Step S110: Access multiple medical literature databases through a distributed crawler architecture or application programming interface to obtain a heterogeneous medical literature dataset containing full-text information or abstract information, and perform deduplication and normalization processing on the heterogeneous medical literature dataset to eliminate data format differences. Step S120: The structured parsing engine is invoked to parse the normalized literature data. The pattern matching algorithm is used to locate the reference area and extract fields such as DOI, PMID, clinical trial registration number, and publication year for each reference to construct a structured citation metadata object. When the DOI or PMID is missing, it is completed through an external authoritative index library interface and a similarity verification mechanism. At the same time, PICO elements are extracted from the literature abstract based on the medical structured abstract format. Step S130: Map the structured citation metadata object to a graph database. Using individual documents as nodes and citation relationships as directed edges, establish a medical literature citation relationship topology network covering the entire sample. Assign a globally unique identifier to each node. Node attributes include subject classification codes based on the MeSH tree hierarchy, evidence-based medicine evidence levels, and PICO element vectors. Use the hierarchical relationships in the MeSH tree hierarchy to perform hierarchical clustering of document nodes. Step S140: Extract the distribution characteristics of citations in the main text of the document, assign differentiated position weights according to the specific chapter in which the citation appears, calculate the basic citation score by combining the frequency of citation mentions, citation context motivation, academic influence indicators and time decay factor, and superimpose the evidence-based medicine evidence level weight factor and PICO element similarity as clinical relevance adjustment coefficient on the basic citation score to calculate the comprehensive citation intensity score of each directed edge. Step S150: Receive a tracing request instruction, take the target document node as the starting point, and perform recursive tracing or downstream tracing along the directed edge. Based on the comprehensive citation strength score and time window parameters, filter key paths and construct a multi-dimensional citation link tracing model. Step S160: Use a streaming computing framework to process new literature data and status change events in real time, dynamically maintain the medical literature citation relationship topology network, and perform risk assessment and status labeling on abnormal nodes in the link according to preset medical citation logic rules.

[0009] In an optional implementation, the heterogeneous medical literature dataset acquisition process in step S110 includes: establishing an adapter set for medical databases from different sources, wherein the adapter set supports parsing Extensible Markup Language, Hypertext Markup Language, and Portable Document Format; during the data acquisition process, using a hash algorithm to extract fingerprints from the literature titles and author lists, and eliminating duplicate literature records by comparing fingerprint information to ensure the uniqueness and accuracy of the dataset.

[0010] To further improve the structured parsing effect of the reference list, the extraction process of the reference list in step S120 involves a sequence labeling algorithm based on deep learning. By training a preset named entity recognition model, the algorithm automatically identifies fields such as author, title, journal name, volume number, issue number, and page number in the citation string. When the DOI or PMID cannot be obtained directly, the system calls the search interface of an external authoritative index library and uses the key fields of the citation to perform fuzzy matching and verification, and reversely completes the missing identifiers.

[0011] Meanwhile, the medical literature citation relationship topology network established in step S130 can be stored using an attribute graph model. Each literature node contains at least the following attribute fields: literature type, subject classification code based on the MeSH tree-like hierarchical structure of the medical subject thesaurus, evidence-based medicine evidence level, PICO element vector, medical core keyword vector, and citation frequency distribution. In addition to recording citation relationships, the directed edges also store citation hierarchy depth information to support subsequent fast link retrieval.

[0012] In another embodiment, the method for calculating the comprehensive citation intensity score in step S140 includes: counting the total number of times the target citation is mentioned in the text as a frequency parameter; determining the position weight based on the chapter in which the citation is located; performing semantic and position analysis on the citation context to determine the citation motivation and assigning a context coefficient; assigning an evidence level weight factor based on the evidence-based medical evidence level of the cited literature, and calculating the PICO element similarity between the citing and cited literature as a clinical relevance moderating coefficient; and combining the academic influence index and time decay factor of the cited literature to calculate the comprehensive citation intensity score.

[0013] Furthermore, the citation link tracing model in step S150 supports a bidirectional tracing mode, namely, a source tracing mode that traces the knowledge source upwards and a diffusion mode that traces the research evolution downwards. When constructing the tracing model, the system adopts a heuristic search algorithm, which prioritizes expanding branches with higher citation strength scores. At the same time, a time window limitation parameter is introduced to only display the link evolution within the user-specified time period, so as to reduce the complexity and redundancy of the topology.

[0014] Meanwhile, the citation link tracing topology graph generated in step S150 has multi-level interactive features, allowing users to view detailed metadata, abstracts, and contribution indicators of the document in the entire link by clicking on nodes; the system uses force-directed layout algorithm or hierarchical layout algorithm to dynamically render the topology graph, reflecting the influence of the document through the size of the nodes and reflecting the strength of the citation through the thickness of the lines.

[0015] Furthermore, in step S160, the dynamic maintenance mechanism can adopt an incremental update strategy: whenever a newly published medical article is detected entering the system, only the direct citation relationships involved in the article are calculated and the graph database is partially updated, without rescanning the entire dataset; the system has a built-in status monitor that regularly synchronizes the latest status of the document nodes already in the database with the external database. If an article is found to be marked as retracted, its position in all citation links is automatically identified, and the affected downstream articles are marked with risk; at the same time, based on the MeSH code and PICO elements of the retracted articles, the potential impact on downstream clinical research conclusions and related treatment guidelines is assessed, and a risk assessment report containing the affected disease type, the name of the drug involved, and the associated treatment guideline number is generated.

[0016] Furthermore, the retraction impact assessment and drug safety signal early warning module may also include an anomaly identification algorithm based on link topology features. By analyzing whether there are abnormal patterns such as circular citations, excessive self-citations, or abnormal growth in citations within a specific time period in the citation link, it can identify potential medical academic misconduct or clinical data fabrication risks and generate a detailed anomaly diagnosis report. When a new adverse drug reaction report is detected in the citation link, a drug safety signal early warning is triggered.

[0017] In addition, the medical literature citation link tracking system may also include a clinical evidence-based decision support interface. This interface allows clinicians or researchers to input specific disease names, drug names, or clinical problem descriptions. The system automatically maps the input content to the corresponding MeSH subject terms and PICO element framework, locates the core literature links in the field, extracts sample size, study design type, efficacy indicators, adverse reaction rate, and evidence-based medicine evidence level from the literature, and outputs evidence summaries arranged from high to low evidence level, providing hierarchical data support for the formulation of clinical pathways or the demonstration of research topics.

[0018] When processing full-text PDF documents, if the document can be directly extracted and searched, text parsing is preferred for extracting the text content. If the document is a scanned image or the text cannot be directly extracted, optical character recognition technology is used to extract the text content, and a rule-based paragraph segmentation algorithm is used to identify the logical structure of the document, ensuring that the citation extraction process is not affected by the document's formatting. For non-English medical documents, the system integrates a machine translation engine to uniformly translate key fields into the target language before metadata extraction, so as to achieve cross-language citation link association.

[0019] To improve the consistency of node entities, the PICO element extraction and citation metadata parsing module can also have an author disambiguation function. By establishing an author feature vector space, it comprehensively considers the consistency of the author's institution, collaborator network, research field, and publication history to solve the node confusion problem caused by different authors with the same name or one person having multiple authors, and ensures the uniqueness of node entities in the citation chain.

[0020] To reflect the timeliness of medical knowledge, the evidence-based medicine weight assessment module can also introduce a citation half-life parameter. Taking into account the speed of updating medical knowledge, literature that was published a long time ago and whose citation intensity is gradually decreasing will be weakened in the link display, thereby highlighting current research hotspots and cutting-edge trends.

[0021] For interdisciplinary link analysis, the clinical evidence chain tracking and visualization module supports automatic labeling of interdisciplinary intersections based on the hierarchical distance of the nearest common ancestor node in the MeSH tree hierarchy when a citation link is detected to cross different medical disciplines. The module labels the disease type, drug category, or treatment technology involved in the intersection and analyzes the role of the intersection in promoting interdisciplinary integration and innovation. When there are literature nodes in the tracking path that report opposite outcomes for the same clinical problem, the literature nodes are labeled as clinical evidence dispute nodes, and their respective sample sizes, study design types, and outcome indicators are extracted for users to compare.

[0022] As an optional setting, the retraction impact assessment and drug safety signal early warning module supports user-defined early warning rules. Users can set specific monitoring thresholds according to their own research needs. For example, when the citation intensity of a core reference decreases by a preset percentage threshold within a preset time period, the system will push an analysis report to the user. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical architecture of a medical literature citation link tracing method proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-dimensional citation strength assessment model in this invention; Figure 3 This is a statistical analysis chart of the citation metadata structured parsing results in this invention, used to display the accuracy, recall and F1 score of DOI / PMID identification under different data sources; Figure 4 This is a data analysis chart showing the distribution and comparison of citation intensity scores in this invention. It is used to demonstrate the influence of different chapter position weights and citation motivation adjustment coefficients on the distribution of citation intensity scores, as well as their correlation with influence indicators. Figure 5 This is a performance evaluation data analysis graph of link tracing and topology generation used in this invention to show the link retrieval time, throughput and incremental update latency under different graph sizes and concurrent request conditions; Figure 6 This is a data analysis chart evaluating the effectiveness of the dynamic monitoring and early warning mechanism in this invention, used to demonstrate the accuracy, recall, and early warning response latency of the anomaly detection algorithm for circular references, excessive self-citations, and retraction events. Detailed Implementation

[0024] This embodiment uses a large-scale comprehensive medical academic resource platform as an application scenario. It implements a method and system for tracing medical literature citation links to address the complex citation relationships generated during the circulation of massive amounts of medical literature. In this scenario, the system needs to process real-time updated data from major global biomedical literature index databases, clinical trial registration databases, and drug regulatory literature databases. It also provides academic integrity monitoring, knowledge evolution analysis, and treatment guideline tracing services to research management departments, clinical evidence-based decision-making bodies, and drug safety monitoring departments.

[0025] The medical literature citation chain tracking system consists of a clinical literature collection and evidence level labeling module, a PICO element extraction and citation metadata parsing module, a disease-drug dimension medical evidence network construction module, an evidence-based medicine evidence weight assessment module, a clinical evidence chain tracking and visualization module, and a retraction impact assessment and drug safety signal early warning module.

[0026] The clinical literature collection and evidence level labeling module is deployed on a server cluster. It can employ multiple collection nodes for parallel data acquisition or access multiple medical literature databases via a standard application programming interface (API), and data acquisition can be performed within the scope of data source access policies and authorization. In one embodiment, the request interval and single-batch request size can be configured according to the external interface rate limiting policy to ensure the stability of the collection process. During the collection phase, hash fingerprints are generated for metadata such as literature titles and author lists, and duplicate entries are removed. For example, the SHA-256 algorithm can be used to generate a 256-bit hash fingerprint from the normalized title and author strings, and duplicate entries are eliminated by comparing the fingerprint database. Normalization processing includes unified encoding, field mapping, and time field standardization to ensure consistency between subsequent parsing and database entry. During the normalization process, the system labels each article with an evidence-based medicine level based on the research design type of the literature report (such as systematic reviews and meta-analyses, randomized controlled trials, cohort studies, case-control studies, case reports, etc.). For example, systematic reviews and meta-analyses correspond to evidence level I, randomized controlled trials correspond to evidence level II, cohort studies correspond to evidence level III, case-control studies correspond to evidence level IV, and case reports correspond to evidence level V.

[0027] The PICO element extraction and citation metadata parsing module receives the normalized document stream to be parsed. This module preferably uses a regular expression library to locate DOIs and PMIDs. DOIs can be matched using a string pattern starting with "10.", containing a 4-9 digit registration number separated by " / ", while PMIDs can be matched using a 7-8 digit rule (optionally with a "PMID:" prefix). Clinical trial registration numbers can be matched using a pattern of "NCT" followed by 8 digits. Simultaneously, the system extracts PICO elements from the document abstracts based on the medical structured abstract format. This is achieved by recognizing field tags such as "Participants," "Intervention," "Comparison," and "Outcome," or by utilizing a pre-trained medical natural language processing model to extract patient group characteristics, intervention measures, control protocols, and clinical outcome indicators, respectively. Furthermore, the system integrates a medical terminology normalization module based on the Unified Medical Language System (UMLS), mapping disease names, drug names, and surgical names appearing in the documents to Standard Concept Unique Identifiers (CUIs) to eliminate entity ambiguity caused by synonyms and abbreviations. For reference lists with complex formatting, the module can further integrate a named entity recognition model to identify fields such as author, title, journal name, volume, issue, and page. When processing full-text PDF documents, if the document can be directly extracted and searchable, text parsing is preferred to extract the text content. If it is a scanned image or the text cannot be directly extracted, optical character recognition technology is used to obtain the text content. For citation records for which DOI or PMID cannot be directly extracted, the system calls the search interface of an external authoritative index, uses fuzzy matching based on key citation fields, and performs similarity verification between the matching results and the original citation string. Preferably, the similarity threshold can be set to 0.95, and missing identifiers are added when the threshold is reached.

[0028] The disease-drug dimension medical evidence network construction module establishes document node entities in a graph database based on extracted key metadata. Each document node includes document type, subject classification code mapped from the MeSH tree hierarchy, evidence-based medicine evidence level, PICO element vector, medical core keyword vector, and citation frequency distribution. The system utilizes the hierarchical relationships in the MeSH tree hierarchy to perform hierarchical clustering of document nodes, forming hierarchical subgraphs of document nodes under the same disease spectrum (e.g., C04.588 specific tumor type under the C04 tumor category) or the same pharmacology category (e.g., specific drug subclasses under the D27 chemical behavior and pharmacology category), facilitating subsequent focused tracking by disease or drug dimension. The system establishes a medical document citation relationship topology network covering the entire sample, using citation relationships as directed edges, and assigns a globally unique identifier to each node. The directed edges not only record the citation pointer but also store the hierarchical depth information of the citation and the specific coordinates of the citation appearing in the text.

[0029] The evidence-based medicine weighting module performs multi-dimensional combined calculations on the edges of directional citation relationships. The system counts the total number of times a citation appears in the text as a frequency parameter; it analyzes the chapter position of the citation and assigns position weights according to a preset position weight table, such as 0.10 for Introduction, 0.20 for Methods, 0.30 for Results, and 0.40 for Discussion. The system identifies citation motivation through medical semantics and stance analysis, generating contextual coefficients, such as 1.2 for supportive citations, 1.0 for neutral citations, and 0.8 for skeptical citations. The system also assigns evidence level weight factors based on the evidence-based medicine evidence level of the cited literature; for example, Level I evidence (systematic reviews and meta-analyses) corresponds to a weight factor of 1.5, Level II evidence (randomized controlled trials) corresponds to a weight factor of 1.3, Level III evidence (cohort studies) corresponds to a weight factor of 1.1, Level IV evidence (case-control studies) corresponds to a weight factor of 0.9, and Level V evidence (case reports) corresponds to a weight factor of 0.7. Simultaneously, the PICO similarity between citing and cited documents is calculated as a clinical relevance moderating coefficient. This similarity is obtained by calculating the cosine similarity of the text vectors for four dimensions: patient population, intervention, control protocol, and clinical outcome, and then taking a weighted average. Furthermore, an academic influence index and a time decay factor are introduced to reflect academic impact and timeliness (the time decay rate can be set to 0.05 per year). The citation intensity score is calculated as follows:

[0030] Where S is the overall citation intensity score; f is the total number of times the citation appears in the text; w is the position weight coefficient corresponding to the chapter where the citation is located; α is the context coefficient corresponding to the citation motivation; β is the academic influence index of the cited literature; and γ is the time decay factor.

[0031] The Clinical Evidence Chain Tracing and Visualization module receives tracing request instructions and, starting from the target document node, recursively traces the source or downstream along directed edges. The system employs a heuristic search algorithm, prioritizing the expansion of critical paths with citation strength scores exceeding a preset threshold. This module supports bidirectional tracing modes: an upward tracing mode to the knowledge source and a downward diffusion mode to trace the evolution of research. The system introduces a time window constraint parameter, displaying only the chain evolution within a user-specified time period. During path traversal, the system utilizes the hierarchical relationships in the MeSH tree structure. When it detects that the first letter of the MeSH code of adjacent nodes in a citation path differs (e.g., from a Class C disease to a Class D medicinal chemistry), it automatically labels the path as an interdisciplinary citation path and quantifies the degree of interdisciplinarity based on the hierarchical distance of the nearest common ancestor node. When the tracing results include cited literature on the same PICO clinical question (same patient population and intervention) but reporting opposite clinical outcomes, the system labels the path as a conflicting clinical evidence path and extracts the sample size, study design type, and outcome indicators of each literature for user comparison. The generated citation link tracing topology is dynamically rendered using a force-directed layout algorithm. The size of the nodes reflects the influence of the literature, and the thickness of the lines reflects the strength of the citations.

[0032] The retraction impact assessment and drug safety signal early warning module utilizes a streaming computing framework to process newly added literature data in real time. The system employs an incremental update strategy, only partially updating the direct citation relationships involved in newly published literature. A built-in status monitor periodically synchronizes the retraction or correction status of literature in external databases. When a retraction is detected at a core node in the citation chain, the system automatically identifies its position in all citation chains and marks the affected downstream literature as risk-prone. Simultaneously, based on the MeSH code and PICO elements of the retracted literature, it assesses the potential impact of the retracted literature on downstream clinical research conclusions and related treatment guidelines, generating a risk assessment report including the affected disease type, involved drug name, and associated treatment guideline number, and sending an early warning to the user. Furthermore, when the system detects a new adverse drug reaction report in the citation chain, it automatically triggers a drug safety signal early warning, notifying relevant clinical and drug regulatory users.

[0033] This embodiment provides a method for tracing medical literature citation links, the execution steps of which are described as follows: In step S110, the system accesses the medical literature database through a distributed crawler architecture or a standard application programming interface to obtain heterogeneous medical literature datasets containing full-text information or abstract information. During the data access phase, the system performs deduplication and normalization processing to eliminate data format differences between different platforms, and standardizes metadata such as titles, authors, and publishing institutions to ensure that time fields are uniformly converted to standard timestamps.

[0034] In step S120, the system calls the structured parsing engine to parse the normalized literature data. The parsing engine first uses a pattern matching algorithm to locate the reference area, then splits the reference strings one by one, extracting fields such as DOI, PMID, and publication year. It can further extract information such as author, title, journal name, volume, issue, and page number by combining it with a named entity recognition model. For citations missing DOI or PMID, the system performs search completion and similarity verification through an external index database interface. Optionally, data collection and parsing statistics examples are shown in Table 1.

[0035] Table 1. Statistical Examples of Data Collection and Metadata Parsing

[0036] In step S130, the system maps structured citation metadata objects to a graph database, establishing a medical literature citation relationship topology network with individual documents as nodes and citation relationships as directed edges. Each node is assigned a globally unique identifier; in one embodiment, a 128-bit identifier can be used to reduce the probability of conflicts. Node attributes include subject classification codes based on the MeSH tree-like hierarchical structure, evidence-based medicine evidence levels, PICO element vectors, and medical core keyword vectors, used to support hierarchical clustering based on disease lineages or drug categories, cross-disciplinary link analysis, and semantic retrieval.

[0037] In step S140, the system extracts the distribution characteristics of citations within the main text of the document. The system determines the specific chapter in which the citation appears by parsing the coordinate information of the PDF or the paragraph tags of the XML and assigns differentiated positional weights. Simultaneously, it identifies the citation motivation through semantic analysis and calculates the comprehensive citation strength score for each directed edge by combining citation frequency, academic influence, and time decay factors. Example calculation results of the comprehensive citation strength are shown in Table 2.

[0038] Table 2. Calculation results of example of comprehensive citation strength

[0039] In step S150, the system receives the tracking command input by the user. Starting from the target document node, the system performs a depth-first or breadth-first traversal algorithm along the directed edges, and filters key paths based on the comprehensive citation strength score; preferably, the strength threshold can be set to 0.005 for key path filtering. The system further combines time window parameters to generate a citation link tracing model and displays a topology map with time axis attributes in the visualization interface, allowing users to click on nodes to view document details and contribution indicators.

[0040] In step S160, the system uses a streaming computing framework to monitor new documents and status change events. When a new document is added to the database, the system triggers an incremental update mechanism; when a node in the citation chain is detected to have changed its status to retraction or correction, the system uses anomaly identification rules to assess its impact on the citation chain and marks the affected paths with risks and provides early warnings. Examples of experimental metrics for link tracing and dynamic monitoring are shown in Table 3.

[0041] Table 3 Examples of performance metrics for link tracing and dynamic monitoring (values ​​vary with data size and deployment environment)

[0042] This invention is technically adaptable to various healthcare information application scenarios. Whether for clinical evidence-based decision-making, drug safety monitoring, traceability of treatment guidelines, or supervision of medical academic integrity, this invention can provide high-precision, full-chain medical literature citation tracking services through evidence-based medicine evidence level stratification, PICO element matching, MeSH level-driven graph clustering, deep structured analysis, multi-dimensional strength assessment, and dynamic topology tracing. The system's distributed architecture and streaming computing mechanism ensure processing efficiency and response speed in massive data environments.

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any brief modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for tracing citation links in medical literature, characterized in that, Includes the following steps: Step 1: Collect heterogeneous medical literature data and normalize it, then label the evidence-based medicine evidence level; Step 2: Parse the reference list using a structured parsing engine to extract citation metadata and complete missing identifiers, while also extracting PICO elements; Step 3: Construct a medical literature citation relationship topology network. The node attributes include subject classification codes, PICO element vectors, disease codes, and drug codes mapped based on the MeSH tree hierarchy structure. Cluster the nodes using the MeSH tree hierarchy structure. Step 4: Combine citation frequency, position weight, citation motivation, evidence level weight factor and PICO element similarity as clinical relevance moderating coefficients to calculate the citation intensity score; Step 5: Based on the citation strength score, perform link tracing, construct a clinical evidence chain tracing model, and visualize it; Step six: Dynamically monitor and issue warnings for abnormalities in the pipeline, and assess the impact of retractions and drug safety risks.

2. The medical literature citation link tracing method according to claim 1, characterized in that, The process of performing in-depth analysis on the normalized literature data in step two includes: It integrates a deep learning-based sequence labeling algorithm and uses a bidirectional long short-term memory network combined with a conditional random field model to automatically identify the author, title, journal name, volume number, issue number, and page number fields in medical citation strings; For records where DOI or PMID cannot be directly extracted, the search interface of an external authoritative index is called to perform fuzzy matching using the identified citation key fields. The matching results are then compared with the original citation string for similarity verification. If the similarity exceeds a preset threshold, reverse completion of the missing DOI or PMID is performed. When the structured parsing engine performs the parsing task in step two, the processing method for the full-text document is as follows: text extraction is used first to obtain the text content; when the full-text document is a scanned image or the text cannot be directly extracted, optical character recognition technology is used to extract the text content, and a rule-based paragraph segmentation algorithm is used to identify the logical structure of the document in order to complete the reference list location and PICO element extraction. By using the medical terminology normalization module based on the Unified Medical Language System (UMLS), disease names, drug names, and surgical names appearing in the literature are mapped to standard concept unique identifiers (CUI) to eliminate entity ambiguity caused by synonyms and abbreviations. For non-English medical literature, an integrated machine translation engine is invoked before metadata extraction to uniformly translate key fields into the target language, thereby enabling cross-language medical citation link association and ensuring the integrity of the evidence chain between clinical research literature in different languages.

3. The medical literature citation link tracing method according to claim 1, characterized in that, The method for calculating the comprehensive citation intensity score in step four includes: The total frequency of the target citation mentioned in the text is counted, and a frequency parameter is generated. Based on the general logical structure of medical papers, the main body is divided into an introduction, methods, results, discussion, and conclusion, and each part is assigned a corresponding preset position weight coefficient. Using a pre-defined medical citation stance dictionary, a stance polarity analysis is performed on the sentence containing the citation to determine whether the citation is a supportive, neutral, or questioning citation, and a corresponding motivation adjustment coefficient is assigned. Evidence level weighting factors are assigned based on the evidence-based medicine evidence level indicated by the cited literature, and the PICO element similarity between the citing and cited literature is calculated as a clinical relevance moderating coefficient. The citation half-life parameter is introduced to take into account the timeliness of medical knowledge, and the link weakening process is performed on literature that was published a long time ago and whose citation intensity is gradually decreasing. The frequency parameter, position weight coefficient, motivation adjustment coefficient, evidence level weight factor, and clinical relevance adjustment coefficient are combined and calculated, and combined with the academic influence index of the cited literature and the time decay factor, to generate a quantitative comprehensive citation intensity score.

4. The medical literature citation link tracing method according to claim 1, characterized in that, The process of constructing a multi-dimensional clinical evidence chain tracing model in step five includes: It provides a two-way tracking mode, including an upstream tracking mode that traces back to the source of knowledge and a downstream diffusion mode that traces back to the evolution of research and clinical translation; When building the tracking model, a heuristic search algorithm is used to prioritize expanding branch paths with reference strength scores higher than a preset score threshold. Introduce a time window limitation parameter to limit the display of the evolution of clinical evidence within a specific time period based on user instructions; The clinical evidence chain tracing topology has multi-level interactive features, supporting the viewing of literature's metadata, abstract, evidence-based medicine evidence level, PICO elements, and clinical evidence contribution indicators of the literature in the chain by clicking on nodes. When a reference link is detected to cross different medical disciplines, the cross-disciplinary intersection is automatically labeled based on the hierarchical distance of the nearest common ancestor node in the MeSH tree hierarchy, and the role of the intersection in promoting the integration of disciplines is analyzed. Key experimental conclusions and clinical evidence are extracted, and the disease type, drug category or treatment technology involved in the intersection is labeled. When there are literature nodes in the tracking path that report opposite outcomes for the same clinical problem, the literature nodes are marked as clinical evidence dispute nodes, and their respective sample sizes, study design types, and outcome indicators are extracted for clinical decision-makers to compare.

5. A medical literature citation tracing system, serving clinical evidence-based decision support and drug safety signal monitoring, characterized in that, include: The clinical literature collection and evidence level labeling module is used to access multiple medical literature databases through a distributed crawler architecture or application programming interface. The medical literature databases include at least one of biomedical literature index databases, clinical trial registration databases, or drug regulatory literature databases. It acquires heterogeneous medical literature datasets containing at least one of clinical research reports, drug trial records, disease diagnosis and treatment literature, or adverse drug reaction reports. It performs normalization processing on the heterogeneous medical literature datasets, including deduplication, field standardization, and format conversion, to eliminate differences in literature data formats between different platforms and ensure that the time field is uniformly set to a standard timestamp. During the normalization process, each article is labeled with its level of evidence-based medicine according to the research design type reported in the literature, which is used for subsequent stratified evaluation of clinical evidence. The PICO element extraction and citation metadata parsing module is used to extract PICO elements from the abstract of each document based on the medical structured abstract format. The PICO elements include patient population characteristics, intervention measures, control protocols, and clinical outcome indicators. Simultaneously, it calls a structured parsing engine to parse the normalized document data, uses a pattern matching algorithm to locate the reference area, and extracts the DOI, PMID, clinical trial registration number, publication year, author, title, journal name, volume number, issue number, and page number information from each reference to construct a structured citation metadata object. When the DOI or PMID is missing, it calls an external authoritative index library interface to perform fuzzy matching and verification using key citation fields to complete the missing identifier. The disease-drug dimension medical evidence network construction module is used to map the structured citation metadata objects to a graph database. With individual documents as nodes and citation relationships as directed edges, a medical literature citation relationship topology network covering the entire sample is established, and a globally unique identifier is assigned to each node. The node attributes include subject classification codes based on the MeSH tree-like hierarchical structure mapping, evidence-based medicine evidence levels, PICO element vectors, and disease and drug codes involved in the literature. The hierarchical relationship in the MeSH tree structure is used to perform hierarchical clustering of document nodes according to disease lineage and pharmacological category to support focused retrieval of clinical evidence by disease or drug dimension; The Evidence-Based Medicine Evidence Weight Assessment Module is used to extract the distribution characteristics of citations in the main text of a document, analyze the paragraph tags or coordinate information of the document to determine the specific chapter in which the citation appears, assign differentiated position weight coefficients to citations in different chapters according to a preset position weight table, identify the citation motivation by combining the medical semantic features of the sentence in which the citation is located, and superimpose the evidence-based medicine evidence level weight factor and the similarity of PICO elements as a clinical relevance adjustment coefficient on the basic citation score to calculate the comprehensive citation strength score for each directed edge to quantify the degree of clinical evidence support. The clinical evidence chain tracing and visualization module is used to receive tracing request instructions from clinical evidence-based queries or drug safety tracking. Starting from the target document node, it recursively traces the source or downstream diffusion along the directed edges. During the path traversal, it uses the hierarchical relationship of the MeSH tree structure to identify evidence propagation paths across disease spectrums or pharmacological categories, filters key evidence paths with citation intensity scores higher than a preset threshold, constructs a multi-dimensional clinical evidence chain tracing model, and uses a layout algorithm to dynamically render the topology graph. The node size reflects the influence of the document, and the line thickness reflects the citation intensity. The visualization rendering results are used to provide evidence propagation paths for clinicians, drug regulators, or treatment guideline developers. The retraction impact assessment and drug safety signal early warning module is used to process newly added clinical research literature, adverse drug reaction reports or updated treatment guidelines in real time using a streaming computing framework, execute incremental update strategies to dynamically maintain the medical literature citation relationship topology network, and monitor the node status in the link in real time according to preset medical citation logic rules. When a change in the status of a core node is detected, a link reconstruction mechanism is triggered and an early warning prompt is sent to the user terminal. When a retraction event is detected in cited literature, the potential impact of the retracted literature on downstream clinical research conclusions and related treatment guidelines is assessed based on the matching relationship between the MeSH tree hierarchy and PICO elements, and a clinical risk assessment report is generated that includes the affected disease type, the name of the drug involved, and the associated treatment guideline number; when a new adverse drug reaction report is detected in the citation chain, a drug safety signal warning is triggered and clinical and drug regulatory users are notified.

6. A medical literature citation link tracing system according to claim 5, characterized in that, During the data acquisition process, the clinical literature collection and evidence level labeling module uses a hash algorithm to extract fingerprints from the literature titles and author lists, and eliminates duplicate literature records by comparing the fingerprint information. The PICO element extraction and citation metadata parsing module has an author disambiguation function. By establishing an author feature vector space, it comprehensively analyzes the consistency of the author's institution, collaborator network, research field, and publication history to determine the consistency of the author entity and solve the node confusion caused by the same name of different people or one person having multiple names. The PICO element extraction and citation metadata parsing module also integrates a medical terminology normalization module based on the Unified Medical Language System (UMLS) to map the disease names, drug names, and surgical names appearing in the literature to the Standard Concept Unique Identifier (CUI). The disease-drug dimension medical evidence network construction module uses an attribute graph model to store medical literature citation relationships. In addition to recording the citation pointer, the directed edges also store the citation level depth information and the physical coordinate information of the citation in the text. The graph database employs sharding technology. Based on the top-level subject classification of the MeSH tree hierarchy, it distributes literature nodes of different medical disciplines across different storage nodes and simultaneously calculates and stores the topological distance between nodes when establishing directed edges.

7. A medical literature citation link tracing system according to claim 5, characterized in that, The retraction impact assessment and drug safety signal early warning module integrates an anomaly identification algorithm based on link topology features. By analyzing whether there are abnormal patterns such as circular citations, excessive self-citations, or abnormal growth in citations within a specific time period in the citation link, it identifies medical academic misconduct or the risk of clinical data fabrication and generates an abnormal diagnosis report. The retraction impact assessment and drug safety signal early warning module supports user-defined early warning rules. Users can set specific monitoring thresholds according to their needs. When the citation intensity of core references decreases by a preset percentage within a preset time period, an analysis report is pushed to the user. The retraction impact assessment and drug safety signal early warning module has a built-in status monitor. The status monitor periodically synchronizes the retraction or correction status of literature in external databases. When a retraction is detected at a core node in the citation chain, it automatically identifies its position in all citation chains and marks the affected downstream literature for risk. At the same time, based on the PICO elements and MeSH codes of the retracted literature, it assesses the potential impact of the clinical conclusions involved on current treatment guidelines or drug use protocols and generates a guideline-level risk transmission report.

8. A medical literature citation link tracing system according to claim 5, characterized in that, It also includes a clinical evidence-based decision support interface, which is used to receive disease names, drug names or clinical problem descriptions input by users, automatically map the input content to the corresponding MeSH subject terms and PICO element framework, locate the core literature links in the related fields, extract the sample size, study design type, efficacy indicators, adverse reaction rate and evidence-based medicine evidence level from the literature, and output evidence summaries in descending order of evidence level, providing hierarchical data support for clinical pathway development or research topic argumentation; When calculating the motivation adjustment coefficient, the evidence-based medicine evidence weight assessment module uses an attention mechanism to capture the core verbs and adjectives in the quoted sentences and identifies words with medical academic orientation to adjust the weights. The clinical evidence chain tracing and visualization module adopts a progressive loading strategy when rendering large-scale networks, pre-displaying the core links and dynamically loading secondary branches as the user interacts, ensuring smooth interaction in a large data environment. When processing massive amounts of heterogeneous data, the clinical literature collection and evidence level labeling module uses a distributed message queue for task scheduling, pushing the data acquired by each collection node to a unified cleaning center for normalization mapping.

Images