A biomedical literature close reading content generation system and generation method

Abstract

The application discloses a biomedical literature intensive reading content generation system and method, and belongs to the field of biomedical literature processing. The system solves the problems of inaccurate type identification, incomplete long text analysis, non-standard terminology, untrustworthy data and poor transmission adaptability in the prior art. The system comprises preprocessing, type discrimination routing, modular extraction, data verification and synthesis output modules in sequence. The method comprises five steps of preprocessing, type discrimination, modular extraction, data verification and synthesis output. Through long text segmentation integration, literature type self-adaptive analysis, professional term standardization and key data verification and correction, the structured intensive reading tweet is generated. The application improves the integrity, accuracy and professionalism of the intensive reading content, shortens the generation cycle to the minute level, and the output content can be directly used for medical public number release, meeting the needs of scientific research assistance and popular science dissemination.

Description

Technical Field

[0001] This invention belongs to the field of biomedical literature data processing and content generation, specifically relating to a biomedical literature in-depth reading content generation system and generation method. Background Technology

[0002] With the rapid development of biomedical research, a massive amount of literature is emerging, posing a challenge to researchers, medical practitioners, and science communicators to efficiently obtain the core information from these documents. Literature parsing and content generation technologies based on large language models are increasingly being applied in academic support and science communication. Existing tools generate literature summaries or interpretations through semantic understanding, improving reading efficiency to some extent.

[0003] However, existing technologies have significant shortcomings in generating content for in-depth reading of biomedical literature: The accuracy of document type identification is insufficient. A uniform processing strategy or simple keyword judgment is often used, which makes it difficult to distinguish between different paradigms of documents such as basic research and clinical research, resulting in a mismatch between the generated content and the characteristics of document type. The lack of type-driven differentiated processing logic means that different types of literature have significant differences in research objectives, evidence levels, and structural characteristics. However, existing technologies have not achieved dynamic adaptation between the analysis process and the generation logic, which can easily lead to deviations in focus. The use of medical terminology and drug names is not standardized, lacks professional reference database constraints and verification, and has problems such as translation errors and inconsistent expressions, which affect the professionalism. Limited long text processing capabilities, constrained by the model's context window, make it difficult to fully cover long documents, and easily lead to omissions of key research background, experimental methods, or core results; There is a risk of fabricated data; the model can easily infer experimental data or conclusions that do not exist in the original text, and there is a lack of effective verification mechanisms, making it difficult to check and correct. The generated content has poor adaptability, is not designed for dissemination scenarios such as medical public accounts, and is difficult to balance professionalism and readability, thus limiting its practical application value.

[0004] The core reasons for the above-mentioned defects are: existing technologies use a general large language model and a unified processing flow, but lack refined modeling of biomedical literature types; the identification of literature types and subsequent processing flows are not effectively linked; professional terms lack standardized constraints; long texts lack a systematic segmentation, parsing and integration mechanism; and the content generation process lacks key data verification and validation.

[0005] In summary, existing technologies are insufficient to meet the requirements of completeness, accuracy, professionalism, and dissemination adaptability for in-depth reading of biomedical literature. There is an urgent need for a technical solution that is adaptive to document type and has the ability to fully parse long texts and verify data. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as inaccurate document type identification, incomplete parsing of long texts, non-standard terminology usage, low data credibility, and poor dissemination adaptability, this application designs a biomedical literature in-depth reading content generation system and method. Through technical means such as adaptive document type parsing, long text segmentation and integration, standardization of professional terminology, and key data verification, it generates complete, accurate, professional, credible, and dissemination-adaptable in-depth reading content, meeting the needs of scientific research assistance and popular science dissemination.

[0007] To achieve the above-mentioned objectives, this invention provides a biomedical literature in-depth reading content generation system, comprising a literature input and preprocessing module, a literature type identification and routing module, a modular content extraction module, a data verification and quality control module, and an in-depth reading content synthesis and output module that work in sequence and in collaboration. The document input and preprocessing module is used to receive biomedical literature data in PDF or Word format and user-specified term variables (standard English drug names and corresponding standard Chinese translations, such as Rituximab), perform text cleaning, segmentation, term standardization, structure recognition and metadata extraction, and output structured text based on Markdown hierarchical specifications and a set of standard terms stored in the form of a JSON dictionary. The document type identification and routing module is used to analyze structured text features and generate document type identification results. The structured text features include at least the research object and research design elements, methodological keywords, article genre identifiers, and length features. According to the preset "document type - processing branch" mapping rules, the workflow is routed to the corresponding processing branch. Basic research, clinical research, case reports, reviews, and communications are matched with the basic research processing flow, clinical research processing flow, case report processing flow, review processing flow, and communications processing flow, respectively. Those that do not match are routed to other processing branches. Specifically, the analysis of the structured text features includes extracting and comparing the following core text feature indicators. 1) Review-type characteristic indicators: include identifying keywords such as meta-analysis and literature review, and have exclusive content characteristics of comprehensively analyzing existing research progress and not reporting new original experimental or clinical data; 2) Basic research characteristic indicators: These include research objectives that focus on molecular or cellular mechanisms, as well as methodological descriptions involving in vitro experiments, animal models, or molecular detection. If the text contains both laboratory experimental and patient data characteristics, this indicator will be used to prioritize the determination of basic research. 3) Clinical research characteristic indicators: including the characteristics of research subjects involving human intervention or observation, as well as the keyword characteristics of research design terms such as randomized controlled trials, cohort studies, and cross-sectional analyses; 4) Case report feature indicators: include the number of detailed clinical descriptions for patients within a preset threshold (e.g., ≤5 patients), including keyword features such as case report and rare disease, and possessing the exclusive feature of not including systematic analysis or statistical comparison; 5) Feature indicators for communication: including keyword features of articles such as letters and briefings, as well as structural features of articles with a length below the preset page threshold and focusing on preliminary results reporting; 6) Catch-up Judgment Criteria: If the extracted structured text features do not meet the criteria in items 1) to 5) above after traversal and comparison, the catch-up logic is triggered, and the document type is determined to be "other". The analysis of the structured text features specifically includes extracting and comparing the following core text feature indicators. The modular content extraction module contains parallel independent processing flows adapted to different document types. The independent processing flows include at least the basic research processing flow, clinical research processing flow, case report processing flow, review processing flow, and communication processing flow. The trigger condition for calling the independent processing flow is the document type determination result output by the document type discrimination and routing module. Based on the determination result, the corresponding parsing strategy and content extraction logic are called to generate type-adapted modular core content. The modular core content consists of multiple content units. The content unit includes at least one of the following: document title, abstract, research background, research methods / experimental design, main results, discussion / conclusion, and references. A corresponding subset is selected according to the document type. The data verification and quality control module is used to compare the key data of the generated content with the original document and perform quality judgment. The key data refers to the numerical data appearing in the original document. The accuracy requirement for the comparison is that the numerical level is consistent. During the comparison, boundary constraints are used for matching percentages, intervals, and decimals, and equivalent representations are supported. The judgment rule for verification is as follows: extract key data and the sentences they contain from the generated content and compare them with the corresponding content in the original document. If any key data cannot be matched in the original document, the corresponding sentence is judged as an error and a set of error sentences is output. Otherwise, the data is judged to be correct. The correction priority of the error data is to make accurate corrections with the original document as the highest priority. If the corresponding data description does not exist in the original document, the relevant data description in the generated content is deleted while maintaining the logical coherence of the rest of the content. At the same time, terminology consistency verification is performed and user-specified terminology variables are used as the highest priority for mandatory uniformity. The in-depth reading content synthesis and output module is used to integrate modular core content, generate structured, directly disseminated in-depth reading content of the document, and output it. The output content is structured text organized according to a preset Markdown hierarchical template, including a recommended title, introduction, background, method / treatment process, main results, discussion / conclusion, and references. The output content is stored as a string and integrated by a variable aggregator as a unified text result for subsequent display and publication. The export method is to output in text format, which is suitable for medical public account publishing scenarios and supports direct copying and use.

[0008] Preferably, the processing flow of the document input and preprocessing module includes: locating the starting position of the references based on text structure features, and truncating and removing all text from the starting position to the end of the document as irrelevant text to obtain clean text for large language model processing; dividing the long text into multiple processing segments; performing segmentation on the long text after removing references, based on a preset character count threshold range of 4500-5000 characters and using a period followed by a space "." as sentence boundary anchors for dynamic addressing; if a sentence boundary is not found within the threshold range, it is allowed to extend backward to no more than 5500 characters to obtain a complete sentence boundary; outputting text segments in array form; extracting drug names and medical terms, and then... The professional reference library performs standardized validation to generate a set of standard terms; it identifies document subheadings and converts them into Markdown format; it extracts metadata such as document titles, abstracts, and publishing journals. The document metadata includes at least the corresponding author and their affiliation, publication date, full journal name, journal abbreviation, DOI number, volume number (issue number), all authors, and page range. The user-specified term variable is the generic drug product name, which serves as the highest priority term constraint. Its specific execution rule is as follows: when the drug product description corresponding to the user-specified term variable has a synonym or translation difference from the drug name extracted from the document, the standard product description of the user-specified term variable replaces the other drug descriptions and synonyms, and the entire text is unified.

[0009] Preferably, the document type identification and routing module classifies documents into six categories: basic research, clinical research, case reports, reviews, communications, and others. Reviews are systematic or descriptive summaries of existing research progress without generating new experimental / clinical data; basic research includes in vitro experiments involving cell, animal, or molecular mechanisms; clinical research involves human subjects and has research designs such as randomized controlled trials, cohort studies, cross-sectional studies, or diagnostic accuracy studies, and reports results; case reports are narratives of the diagnosis and treatment process of a single or small number of cases; communications are articles containing genre identifiers such as Correspondence, Communication, Letter, or Brief report and have a concise and refined structure; documents not belonging to the above categories are classified as "other." The modular content extraction module's independent processing flows include basic research processing flows, clinical research processing flows, case report processing flows, review processing flows, and communications processing flows, each adapted to the structural characteristics and information extraction focus of the corresponding document type.

[0010] Preferably, the clinical research processing flow focuses on extracting clinical trial design, inclusion / exclusion criteria, statistical methods, and core results. The clinical trial design types include at least randomized controlled trials, cohort studies, case-control studies, prospective / retrospective studies, and single-arm / double-blind / placebo-controlled and multicenter studies, with the design type identified based on the literature methodology section. The core results extraction scope includes at least the results data corresponding to the primary and secondary endpoints. The case report processing flow extracts case descriptions, treatment processes, and clinical outcomes, reconstructs the diagnostic and treatment logic, and organizes the treatment process in chronological order based on the start and end times of each stage and corresponding treatment details to reconstruct a timeline. The data verification and quality control module includes a key data verification unit and a terminology calibration unit. The key data verification unit compares experimental data, statistical indicators, and other numerical information, while the terminology calibration unit verifies the consistency of Chinese and English terminology and corrects non-standard expressions.

[0011] Preferably, the intensive reading content synthesis output module integrates content according to a preset Markdown template, and the main text is organized using a structure such as "#first-level heading, ##second-level heading"; the preface is output with ">" in the citation format, and paragraphs are separated by two line breaks; a structured tweet containing recommended titles, preface, main text, and references is generated, and integrated into a unified text result by a variable aggregator, which can be directly used in dissemination scenarios such as medical public accounts.

[0012] For generating structured tweets, the system / module constrains each component based on a preset word count threshold and content template. Specific generation rules include: 1) Recommended title generation rules: Only one title can be generated. The content must extract core points from the introduction and literature metadata, and dynamically match a pre-defined title template based on the number of corresponding authors and their affiliated institutions. Specific templates include: a single-institution corresponding author template (format: "[Journal Abbreviation] Corresponding Author Institution and Name: Main Findings"), a multi-institutional corresponding author template (format: "[Journal Abbreviation] Names of Each Corresponding Author: Main Findings"), and a mechanism-based research degradation template (if the complexity of the main finding mechanism exceeds a preset threshold, "Main Findings" will be replaced with "Research Theme"). Title generation must forcibly retain the original author names and provide a Chinese translation of the institution names.

[0013] 2) Preface generation rules: The overall word count is limited to the preset threshold range (around 400 words). The content is forced to adopt a two-part structure: the first part is a background description extracted from the disease, phenotypic features and literature content; the second part is a combination of the corresponding author's original text, the Chinese translation institution, the original journal article and the original English title of the literature in a fixed word order (with an automatic addition of a literature citation mark such as "[1]" at the end), and at the end of the paragraph, the main findings and innovations of the study are summarized in 1 to 2 sentences.

[0014] During the translation and terminology processing of the preface, the system strictly calls the preset professional terminology dictionary and follows specific industry standards (such as disease names following the ICD-10 standard and drug names following the CADN standard); at the same time, the output text is formatted using preset markup languages ​​(such as adding specific quotation marks ">" before paragraphs and bolding the affiliation, author and main findings).

[0015] 3) Text generation rules: The overall word count is constrained within a preset threshold range (approximately 2000 characters). The main text content does not use a fixed template but is controlled by the document type determination result output by the upstream module, performing adaptive assembly of the structure. If it is determined to be basic research, the main text structure is assembled as follows: background, results, and discussion; If it is determined to be a clinical study, the main text structure should be: background, methods, results, and discussion. If it is determined to be a case report, the main text structure is assembled as follows: background, treatment process, and discussion.

[0016] A method for generating in-depth reading content for biomedical literature, applied to the aforementioned biomedical literature in-depth reading content generation system, includes the following steps: Step S1, Literature Input and Global Preprocessing: Receive biomedical literature and user-specified terminology variables, perform text cleaning, segmentation, terminology standardization, structure recognition and metadata extraction, and output structured text and a set of standard terms. Step S2, Document Type Identification and Routing: Analyze the features of structured text, and first determine whether it is a communication document according to the discrimination rules. If it is a communication document, the communication type is directly output; otherwise, it is determined in the categories of review, basic research, clinical research, and case report. Unmatched documents are marked as "other" and routed to the corresponding processing branch. Step S3, Modular Content Extraction: Based on the document type, the appropriate processing flow is called to extract the modular core content. Step S4, Multiple Data Validation and Quality Control: Extract key values ​​and their corresponding sentences from the generated content and compare them with the original literature; if erroneous sentences are found during the comparison, correction is triggered: make precise corrections based on the data in the original text, delete the data description if the corresponding data is not found in the original text and maintain sentence coherence; at the same time, perform terminology consistency checks and unify them to standard terminology. Step S5, Synthesis and Output of In-depth Reading Content: Integrate modular core content, generate structured in-depth reading tweets, and output them.

[0017] Preferably, in step S1, terminology standardization specifically involves: matching the extracted drug names and medical terms with a pre-set professional reference library to generate standard translations; if there is no corresponding term in the professional reference library, the original translation is used (the large model will generate corresponding translations simultaneously during extraction; if there is no corresponding drug name or medical term in the standard library, this translation is used); the structure recognition specifically involves: extracting the structural and semantic features of text lines that end without punctuation and are general phrases, and performing contextual verification in conjunction with the pre-acquired document abstract to dynamically identify document subheadings; further, the identified subheadings are divided into different levels of Markdown structured text, wherein subheadings containing preset macro-skeleton keywords such as "Introduction", "Background", "Methods", "Results", and "Discussions" are mapped to second-level headings, and micro-specific subheadings are mapped to third-level headings; The process of segmenting the long text into four processing segments specifically includes: first, locating and removing the text at the end of references that contain the identifiers "References" or "REFERENCES"; then, performing sliding segmentation based on a preset threshold range of 4500-5000 characters, dynamically addressing and truncating the text when it reaches this threshold range using the combination of English periods and spaces (".") as the natural sentence boundary anchor point, and allowing the threshold to be dynamically extended to a preset maximum character limit (such as 5500 characters) to obtain complete sentence boundaries, thereby segmenting the long text into several semantically complete text blocks and constructing them into an array; finally, performing sequential equal division calculation based on the number of array elements, dividing the text block array into four processing segments, and merging the text blocks with remainders that cannot be divided evenly into the fourth segment, in order to adapt to the downstream model context window constraints and achieve parallel load balancing.

[0018] Preferably, in step S2, the document type determination is achieved by analyzing the research purpose, experimental design, and result presentation characteristics of the document. The research purpose includes at least whether it is a systematic or descriptive summary of existing research without generating new experimental / clinical data; the experimental design includes at least whether it involves human subjects and research designs such as randomized controlled trials, cohort studies, case-control studies, cross-sectional studies, or diagnostic accuracy studies, or in vitro experimental methodological elements such as cell / animal / molecular mechanisms; and the result presentation characteristics include at least whether it reports original experimental / clinical data, whether it is a narrative of the diagnosis and treatment process for a single case or a small number of cases, and whether it includes correspondence, communication, letter, or brief. The document type is identified by genre identifiers such as report, and the document type determination results are output in the order of judging communication type first and then other categories; in step S3, the basic research processing flow extracts the research background, experimental results, discussion and conclusions, and retains experimental data and statistical indicators. The experimental data and statistical indicators are retained in accordance with the original text and include integer, decimal, percentage and interval forms; the review processing flow performs full text structure analysis and generates recommended titles, introductions, core viewpoints and references in parallel, wherein the parallel generation is realized by the parallel execution mechanism of the workflow engine.

[0019] Preferably, in step S4, the key data verification includes extracting core information such as experimental values, statistical results, and treatment plans from the generated content and their corresponding statements, comparing them with the corresponding content in the original literature, and outputting the verification result; when the verification result is not "data is correct", the incorrect statements are precisely corrected by referring to the correct data content in the original literature fragment; if the corresponding data description cannot be found in the original literature, the relevant data description in the generated content is deleted and the statement is kept coherent; the terminology consistency verification includes the correspondence verification between English abbreviations and Chinese full names, and is unified according to the rule of outputting "Chinese full name (standard abbreviation)" when it first appears, and uniformly outputting the standard abbreviation when it appears later, while user-specified terminology variables are used as the highest priority terminology constraints to enforce uniform use.

[0020] Preferably, in step S4, the key data verification includes core information such as experimental values, statistical results, and treatment plans, as well as the statements they contain, which are compared with the corresponding content in the original literature, and the verification results are output. When the verification result is not "data is correct", the incorrect statements are precisely corrected by referring to the correct data content in the original literature fragment. If the corresponding data description cannot be found in the original literature, the relevant data description in the generated content is deleted and the sentence is kept coherent. The terminology consistency verification includes the correspondence verification between English abbreviations and Chinese full names, and is unified according to the rule of outputting "Chinese full name (standard abbreviation)" when it first appears and uniformly outputting the standard abbreviation when it appears thereafter. At the same time, the user-specified terminology variables are forced to be used uniformly in the title and body text.

[0021] Preferably, the document data format in step S1 includes PDF and Word, and the input and parsing adaptation specifically includes: Format recognition and routing: Read the metadata and file header features of the input document, identify the document format, and route it to the corresponding parsing adapter; Scanned version and standard PDF adaptation: If it is recognized as a PDF format, it is checked whether it contains a text character layer; if it does, the underlying character stream is directly extracted and paragraph coordinates are mapped; if it is determined to be a scanned PDF that does not contain a text character layer, the optical character recognition (OCR) component is triggered to perform layout analysis and extract structured text from image features; Word format structured adaptation: If it is recognized as a Word format, its underlying Extensible Markup Language (XML) nodes or Document Object Model (DOM) are parsed to extract paragraph and heading text that maintain the hierarchical characteristics of the original text; Encryption status interception mechanism: Document permission status is verified before executing the PDF or Word parsing; if the document is identified as being in an encrypted or Digital Rights Management (DRM) protected state, the parsing process is intercepted, and an abnormal interruption signal containing an encryption identifier is output to trigger a user interaction request; the output content of step S5 contains a clear information hierarchy and standardized reference format, which can be directly used for publication on medical public accounts; the reference format is APA format, which includes at least the author's last name and first letter, year, article title, journal name, volume number (issue number), page number, and DOI link, and the journal name and volume number are in italics, while the issue number is not in italics within parentheses; the DOI is provided in the form of https: / / doi.org / xxx when it exists; when the number of authors is 1-20, all authors are listed and connected with the last author using '&'; when the number of authors is greater than 20, the first 19 authors are listed followed by an ellipsis... and the last author is listed; the user-specified terminology variable is the common English and Chinese name of the drug, and the standard expression is enforced uniformly in the generated content.

[0022] The advantages and effects of this application are as follows: 1. Improve the accuracy of document type matching and the rationality of content structure: The technical features are accurate document type identification and type-driven differentiated processing flow. Through the refined modeling and independent analysis logic of six types of documents, the structure and focus of the generated content are highly matched with the document type, which solves the problems of unreasonable structure and bias of focus in the existing technology and meets the needs of in-depth reading of different types of documents.

[0023] 2. Achieve complete parsing of long texts and reduce the risk of missing key information: The technical features include long text segmentation, structured recognition and information integration mechanism. Long documents are split into segments that are adapted to the model for processing. Combined with the structured subheadings and cross-segment information fusion, the limitations of the context window are broken, and the complete coverage of the whole text is achieved, which significantly reduces the rate of missing key information.

[0024] 3. Standardize the use of professional terminology and enhance the professionalism of the content: The technical features include a built-in professional comparison library of medical terms and drug names and a priority constraint mechanism. User-designated terms are given the highest priority to ensure accurate translation and consistent expression of professional terms, eliminate mistranslation and non-standard use, and achieve professional publishing standards.

[0025] 4. Reduce the risk of data fabrication and enhance the credibility of content: The technical feature is a key data closed-loop verification and correction mechanism. By comparing the generated content with the core values ​​of the original documents, it automatically corrects erroneous data, effectively curbs fabricated information caused by model illusion, improves the reliability of the carefully read content, and avoids academic misleading.

[0026] 5. Significantly improves the efficiency of generating in-depth reading content: The technology features fully automated processing, requiring no manual intervention from document input to post output, reducing the traditional hours of manual reading and writing to minutes, and significantly reducing the cost of scientific research and popular science work.

[0027] 6. Adapt to different communication scenarios and enhance practical application value: The technical feature is a structured tweet generation template. The generated content has a clear information hierarchy, standardized format and good readability. It can be directly used for new media dissemination such as medical public accounts, which solves the problem of limited application scenarios of existing technologies.

[0028] 7. High flexibility and scalability: The technical features include a modular architecture and parameterized control, which supports rapid adaptation to different application scenarios. It can cover more document types and professional fields by expanding the processing flow and reference library, and has strong reusability and scalability.

[0029] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0030] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0032] Figure 1 Overview of the technical roadmap for generating biomedical literature reading content designed for this application; Figure 2 A detailed flowchart of the biomedical literature reading content generation system designed for this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0034] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0035] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0036] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0037] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0038] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0039] Example 1: This invention constructs a closed-loop workflow of input preprocessing - type discrimination - modular extraction - data verification - synthetic output. Through the collaborative work of six major functional modules, combined with the adaptive processing logic of document type, it realizes the automated and high-quality generation of in-depth reading content of biomedical literature. The specific implementation of the technical solution is described in detail below with reference to the accompanying drawings.

[0040] Specific implementation of system modules like Figure 2 The diagram shown is a detailed flowchart of the biomedical literature in-depth reading content generation system of this invention. The system adopts a "layered processing + type adaptation" architecture, and each module works collaboratively according to the data flow sequence. The specific implementation principle and process are as follows: The document input and preprocessing module, as the core of the system's input and preprocessing layers, receives user-uploaded PDF / Word format biomedical documents and user-specified generic drug product names (global terminology variables) (maximum text length limited to 100 characters). These user-specified terminology variables are located in the system's input layer (Start node). The system provides a custom text input box, allowing users to manually type in "English (Chinese)" key-value pairs. First, irrelevant text is removed, locating the "References" or "REFERENCES" field in the original text and removing subsequent reference paragraphs. The long text after removing references undergoes a two-stage segmentation and splicing process: The first stage uses a sliding scan within a preset character threshold range of 4500-5000 characters, dynamically addressing sentence boundaries using a period followed by a space ".". If no sentence boundary is found within the threshold range, the process extends to a maximum of 5500 characters to obtain a complete sentence boundary, outputting text segments in array format. The second stage divides the text segments into four processing pieces in their original order, and further processes the text within each processing piece. The process sequentially performs subheading recognition and Markdown heading hierarchy processing, then concatenates the subheadings into text segments. These four subheadings are then sequentially concatenated into a full-text summary. Drug names and medical terms are extracted from the literature, and standardized validation is performed using drug name lookup libraries and medical terminology lookup libraries respectively, generating a standard terminology set (user-specified terms have the highest priority and are enforced for uniform use). Subheadings (such as "Introduction" and "Results") are identified through semantic understanding and converted into Markdown hierarchical format to form structured text. Metadata such as the literature title, abstract, journal, and DOI are extracted in parallel to provide foundational information for subsequent content generation.

[0041] Document type identification and routing module: Based on preprocessed structured text and metadata, it calls a large language model to analyze document characteristics (research purpose, experimental design, result presentation method, etc.) and accurately classifies the document into one of the six categories: basic research, clinical research, case report, review, communication, and others. Through the "problem classifier", the workflow is routed to the corresponding processing branch in the modular content extraction module to achieve accurate matching between type and processing logic.

[0042] Modular content extraction module: As the core processing layer of the system, it contains five independent parallel processing streams, each adapted to different document types. This tool reads only one document at a time and determines the document type through condition nodes, then enters the unique branch of the corresponding type for processing. Therefore, there is no parallel processing among the five independent processing streams.

[0043] Basic research processing flow: Traverse structured text, identify and extract research background, experimental results, discussion, and conclusion modules, and focus on retaining core information such as experimental data and statistical significance indicators, and aggregate them into structured content blocks; Clinical research processing flow: The method module has been added for independent extraction, focusing on identifying clinical trial design types (RCT / cohort study), inclusion criteria, exclusion criteria, and statistical methods. The background, methods, results, and discussion modules have been integrated to adapt to the CONSORT standard. Case report processing workflow: Extract case description, treatment process, and clinical outcome modules through a dedicated Prompt template, reconstruct the diagnosis and treatment logic according to the timeline, and simultaneously extract rare disease or special treatment analysis from the discussion section; Furthermore, the core design points (technical features) are reflected in the following four aspects: 1. Structured Instruction Framework: The Prompt template uses a highly structured format (containing modules for "Task," "Steps," "Restrictions," "Requirements," and "Output Example"). It mandates that LLM map and deconstruct unstructured case text into three independent data slots: case description (e.g., patient baseline, present medical history), treatment process (interventions, dosage, duration), and clinical outcome (efficacy, adverse reactions).

[0044] 2. Time-Sequence Reconstruction Instructions: Within the

Steps

Requirements

[0045] 3. Specific Focusing and Noise Reduction Rules: For the "Discussion" section specific to case reports, the Prompt template incorporates specific focusing instructions. This forces LLM to filter out general disease background information, extracting only analysis content strongly correlated with "rare disease characteristics" or "mechanisms of action / prognosis of specific therapies," thereby improving the signal-to-noise ratio of core information.

[0046] Review processing flow: Perform full-text macro-structure analysis, identify hierarchical outlines, and generate recommended titles, introductions, core viewpoints of the main text, and references in parallel to synthesize review-type in-depth reading content; The rules for generating recommended titles are based on a set of conditional triggering rules. The basis and logic for title generation mainly include the following dimensions: 1. Data source basis: The input content for title generation is strictly limited to the <Introduction> (used to extract the main findings or research topics) and <Document metadata> (used to extract journal abbreviations, corresponding author names and affiliations) extracted by the system from the preceding nodes.

[0047] 2. Dynamic Conditional Routing Basis: The system dynamically matches and applies different construction principles based on the number of corresponding authors and the complexity of the research mechanism. Principle 1 (Single Corresponding Affiliation): Match the format "[Journal Abbreviation] Corresponding Author's Affiliation, Name, etc.: Main Findings".

[0048] Principle 2 (Multiple Corresponding Authorities): Matching format "[Journal Abbreviation] Correspondence A / Correspondence B / Correspondence C: Main Findings".

[0049] Principle 3 (Catch-all for complex mechanism research): When the "key findings" cannot be concisely expressed in a single sentence, it is automatically downgraded to a refined research topic, matching the format "[Journal Abbreviation] Corresponding party et al.: Revealing [research topic]".

[0050] Communication processing flow: Extract basic information and core academic viewpoints from literature, perform clinical significance assessment, analyze the potential impact of viewpoints on clinical practice, and generate commentary content focusing on academic debate.

[0051] Data verification and quality control module: As the core of the content optimization layer, it performs dual verification and correction: The key data verification unit extracts core information such as experimental values, statistical results, and treatment plans from the generated content and compares them one by one with the original literature fragments. When a deviation is found, it automatically calls the original literature to make corrections; The terminology calibration unit verifies the consistency of terminology used throughout the text, ensures that English abbreviations correspond to Chinese full names in a standardized manner, forces the use of user-specified standard terminology, and corrects non-standard expressions.

[0052] The in-depth reading content synthesis and output module receives verified and corrected modular core content, integrates it according to a preset Markdown template, and automatically generates a unique and highly condensed recommended title based on the literature metadata and introduction content through a condition-triggered rule engine. It then assembles the preface, structured body text, and standardized references, ultimately generating a complete in-depth reading article. The output format is adapted to the dissemination scenarios of medical public accounts, possessing clear information hierarchy and good readability. Users can use it directly or make minor adjustments before publishing.

[0053] Specific implementation of the generation method like Figure 1 The diagram shown is an overview of the technical route for generating biomedical literature reading content according to the present invention. The method steps and system modules correspond one-to-one, forming a standardized processing pipeline. The specific implementation details are as follows: Step S1, Literature Input and Global Preprocessing: The user uploads biomedical literature (PDF / Word) and inputs standard drug terminology; the system removes irrelevant content such as references, and segments the long text into four parts; extracts drugs / terminology and standardizes them through a professional reference library; identifies subheadings and converts them into Markdown format; extracts metadata and outputs structured text and a set of standard terminology; if the literature format is invalid, processing is terminated and feedback is sent to the user. If the content is invalid or abnormal, the system operation log will synchronously record the node position, timestamp, and exception stack information of the task termination, realizing full-chain traceability of error handling.

[0054] Step S2, Document Type Identification and Routing: The system analyzes the features of structured text to determine the document type (one of six categories); based on the type, the workflow is routed to the corresponding processing branch to ensure that the subsequent extraction logic is compatible with the document type.

[0055] Step S3, Modular Content Extraction: Call the corresponding processing flow according to the document type: Basic research extracts background, results, discussion, and conclusions; Clinical research adds method module extraction; Case reports reconstruct the diagnosis and treatment process according to timeline; Reviews generate core viewpoints and outlines; Communications assess clinical significance and extract academic viewpoints, outputting modular core content.

[0056] Step S4, Multiple Data Validation and Quality Control: The system compares the key data of the generated content with the original document and automatically corrects erroneous values; it verifies the consistency of terminology, standardizes the use of English abbreviations and full Chinese names, and enforces the uniformity of user-specified terminology; it generates a validation report and marks ambiguous information that requires manual review (if any).

[0057] Step S5, Content Synthesis and Output: The system integrates modular content according to the dissemination adaptation template, generating a structured post containing a recommended title, introduction, main text, and references. Users can directly copy or export this content for publication in scenarios such as medical public accounts, completing the generation of detailed reading content. Regarding export format, the system defaults to supporting multi-level headings and formatted Markdown (MD) format, while also supporting structured JSON data output for easy reuse in different scenarios. In terms of export implementation, a multi-dimensional delivery system is provided: users can either copy or download the generated structured content as a standard file with one click through the web front-end interface, or automatically push it to external third-party business platforms via a standard interface, achieving seamless end-to-end integration and efficient content distribution.

[0058] The specific implementation of this invention, through modular architecture and type adaptive logic, realizes the automated and high-quality generation of in-depth reading content for biomedical literature, specifically addresses the core pain points of existing technologies, provides an efficient tool for scientific research assistance and medical popularization, significantly reduces labor costs, and improves content reliability and dissemination adaptability.

[0059] Clinical research literature application cases To further illustrate the practical application effect of the system of the present invention in the context of clinical research literature, the following uses an English clinical research literature as an example to give the entire generation process: (1) Input and variable setting: The user uploads the clinical research literature file to be analyzed at the workflow input terminal and enters the generic product name of the drug as a global term variable; (2) Title and abstract extraction: The system identifies and extracts the main title and full text of the abstract from the input documents, and uses them as title and abstract metadata for subsequent content generation; (3) Structured and metadata extraction: The system identifies subheadings in the segmented text and converts them into Markdown heading level format, and then splices the four segments in order into a full-text summary text; in parallel, it extracts metadata such as document title, abstract, corresponding author, institution, journal information, DOI number, volume, issue and page number; (4) Extraction of key content from documents: The system extracts key content from the full text summary and outputs a structured outline and key points to assist in subsequent chapter generation and subheading recognition. (5) Extraction and verification of medical terms and drug names: The system extracts important medical terms and drug names from the literature and translates them into Chinese to form a set of standard terms and a set of drug names; the set of drug names is further checked and replaced with the built-in drug comparison table to unify the standard translations; the English abbreviations and full Chinese names in the terms are checked and replaced for consistency, and the rule of "outputting the full Chinese name (standard abbreviation) for the first appearance and uniformly outputting the standard abbreviation for subsequent appearances" is implemented. (6) Segmentation and splicing summary: The system divides the text segment array into four segments in the original order to form four processing segments (the remainder is assigned to the last segment); the subheadings in each segment are identified and converted into Markdown heading levels and then spliced ​​into segment text in order. Then, the four segment texts are spliced ​​into the full text summary text in order. (7) Type identification and routing: The system classifies literature based on the characteristics of literature research design and result presentation, outputs the literature type identification results and routes them to the clinical research processing flow; (8) Module division and classification: The system is based on the Markdown heading module with "#" at the beginning of each line in the full-text summary text. The modules are divided one by one. After obtaining the segmented text units, the system combines the features extracted from the pre-acquired global content of the document, performs contextual semantic intent analysis, and accurately maps and classifies each text unit into the six preset standard functional modules. The specific classification judgment rules include: Background module determination: Identify whether the text unit is located in the introduction area at the beginning of the document, and extract whether it contains semantic features that characterize the current state of the disease / clinical situation, unmet needs, or research motivation (such as keywords such as "no consensus" or "urgent need"). Method module determination: Identify whether the text unit has a specific structured subheading (such as study design, subjects, statistical analysis, etc.), or extract whether it contains entity content such as experimental methods, sample grouping, intervention treatment and endpoint indicators; The results module determines whether the extracted text contains core original data, semantic references to chart numbers (as shown in Figure X), and quantitative expressions (such as data features like effectiveness and confidence intervals). Discussion module judgment: Extract whether the text contains self-evaluative words that explain the results, speculate on the mechanism, and discuss the limitations or innovations of this study, as well as semantic expressions that are comparative with previous studies; Conclusion module judgment: Detect whether the text is an independent summary paragraph, or located at the end of the discussion module and has the characteristics of summarizing the final research conclusion; Last-line classification: If no feature match is found after the above rules are followed in sequence, the last-line logic is triggered, and the input text unit is uniformly classified as "other" module (such as author statement, data availability and other auxiliary text).

[0060] After the above classification, the system outputs a structured text unit array with clear functional module labels to enable accurate extraction and information fusion of structured content across segments.

[0061] (9) Modular content extraction: Based on the descriptions in the methods and results sections of the clinical research processing flow, extract the study design type, study subjects and inclusion and exclusion criteria, treatment regimens and administration methods, statistical methods, primary and secondary endpoints and their corresponding results data to form modular core content; (10) Data verification and terminology unification: The system extracts key values ​​and their corresponding sentences from the generated content, compares them with the original document fragments, and outputs the verification results; if the verification is not "data is correct", the erroneous sentences are corrected according to the original data, and the data description is deleted and the sentence is kept coherent when there is no corresponding data in the original text; at the same time, the terminology is verified for consistency, and the "full Chinese name (standard abbreviation)" is output for the first appearance, and the standard abbreviation is output uniformly for subsequent appearances, and the user-specified terminology variable is used as the highest priority to force the use of unified terms; (11) Synthetic output: The system integrates modular content according to the preset Markdown hierarchical template. The preface is output in the ">" citation format and paragraphs are separated by two newline characters. References are generated in APA format and attached to the end of the article. The output is a uniform string result, which users can directly copy for publication on medical public accounts or import into the front-end page for rendering and layout according to Markdown titles and citation symbols.

[0062] Processing time: The tool took 300 seconds. Accuracy: The translation mapping accuracy of medical entities, core drug names and abbreviations reached 100% (with no errors in proper nouns).

[0063] The above description is merely a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter alterations to these embodiments within the spirit and principles of the present invention, achieved through conventional substitutions or by achieving the same function without departing from the principles and spirit of the present invention, fall within the scope of protection of the present invention.

Claims

1. A biomedical literature in-depth reading content generation system, characterized in that, It includes a document input and preprocessing module, a document type identification and routing module, a modular content extraction module, a data verification and quality control module, and a close reading content synthesis and output module that work in sequence. The document input and preprocessing module is used to receive biomedical literature data and user-specified terminology variables, perform text cleaning, segmentation, terminology standardization, structure recognition and metadata extraction, and output structured text and a set of standard terms. The document type identification and routing module is used to analyze the features of structured text, determine the document type, and route it to the corresponding processing branch; The modular content extraction module includes independent processing streams adapted to different document types. Based on the document type, it calls the corresponding parsing strategy and content extraction logic to generate modular core content that is adapted to the document type. The data verification and quality control module is used to compare the key data of the generated content with the original document, and to perform terminology consistency verification and error data correction. The in-depth reading content synthesis and output module is used to integrate modular core content, generate structured, directly disseminated in-depth reading content of documents, and output it.

2. The biomedical literature in-depth reading content generation system according to claim 1, characterized in that, The processing flow of the document input and preprocessing module includes: removing irrelevant text such as references, dividing long texts into multiple processing segments; extracting drug names and medical terms, performing standardized verification through a professional reference library, and generating a standard terminology set; identifying document subheadings and converting them into Markdown format; extracting metadata such as document titles, abstracts, and publishing journals; and using the user-specified terminology variable, which is the generic product name of the drug, as the highest priority term constraint.

3. The biomedical literature in-depth reading content generation system according to claim 1, characterized in that, The document type identification and routing module divides documents into six categories: basic research, clinical research, case reports, reviews, communications, and others. The modular content extraction module has independent processing flows including basic research processing flow, clinical research processing flow, case report processing flow, review processing flow, and communications processing flow, which are adapted to the structural features and information extraction focus of the corresponding document types.

4. The biomedical literature in-depth reading content generation system according to claim 3, characterized in that, The clinical research processing flow focuses on extracting clinical trial design, inclusion / exclusion criteria, statistical methods, and core results; the case report processing flow extracts case descriptions, treatment processes, and clinical outcomes, and reconstructs the diagnosis and treatment logic; the data verification and quality control module includes a key data verification unit and a terminology calibration unit. The key data verification unit compares experimental data, statistical indicators, and other numerical information, while the terminology calibration unit verifies the consistency of Chinese and English terminology and corrects non-standard expressions.

5. The biomedical literature in-depth reading content generation system according to claim 1, characterized in that, The in-depth reading content synthesis and output module integrates content according to a preset Markdown template to generate a structured tweet containing a recommended title, preface, main text, and references, which can be directly used in dissemination scenarios such as medical public accounts.

6. A method for generating in-depth reading content of biomedical literature, applied to the biomedical literature in-depth reading content generation system according to any one of claims 1-5, characterized in that, Includes the following steps: Step S1, Literature Input and Global Preprocessing: Receive biomedical literature and user-specified terminology variables, perform text cleaning, segmentation, terminology standardization, structure recognition and metadata extraction, and output structured text and a set of standard terms. Step S2, Document Type Identification and Routing: Analyze the features of structured text, determine the document type, and route it to the corresponding processing branch; Step S3, Modular Content Extraction: Based on the document type, call the appropriate processing flow to extract the core modular content; Step S4, Multiple Data Validation and Quality Control: Compare the key data of the generated content with the original document, correct errors, and verify the consistency of terminology usage; Step S5, Synthesis and Output of In-depth Reading Content: Integrate modular core content, generate structured in-depth reading tweets, and output them.

7. The method for generating biomedical literature in-depth reading content according to claim 6, characterized in that, In step S1, terminology standardization specifically involves matching the extracted drug names and medical terms with a pre-set professional reference library to generate standard translations; structure recognition specifically involves identifying keywords such as "Introduction" and "Methods" to convert the text into Markdown hierarchical structured text; and long texts are segmented into four processing pieces to adapt to the model's context window limitations.

8. The method for generating in-depth biomedical literature content according to claim 6, characterized in that, In step S2, the document type determination is achieved by analyzing the research purpose, experimental design, and result presentation characteristics of the document; in step S3, the basic research processing stream extracts the research background, experimental results, discussion and conclusions, and retains experimental data and statistical indicators; the review processing stream performs full-text structure analysis and generates recommended titles, introductions, core viewpoints and references in parallel.

9. The method for generating in-depth biomedical literature content according to claim 6, characterized in that, In step S4, key data verification includes comparing core information such as experimental values, statistical results, and treatment plans, and automatically calling original literature excerpts to correct data deviations when they are found; terminology consistency verification includes the correspondence between English abbreviations and full Chinese names, and the mandatory uniform use of user-specified terminology variables.

10. The method for generating biomedical literature in-depth reading content according to claim 6, characterized in that, The document data formats in step S1 include PDF and Word; the output content in step S5 contains a clear information hierarchy and a standardized reference format, which can be directly used for publication on medical public accounts; the user-specified terminology variable is the common English and Chinese names of drugs, and the standard expression is forcibly used in the generated content.

Images