A hospital scientific research content management method and system based on generative AI

Abstract

The application relates to the technical field of artificial intelligence, and discloses a hospital scientific research content management method and system based on generative AI, aiming to solve the technical problems of difficult identification of AI-generated content in scientific research content, difficult verification of factual errors, and untraceable process. The method comprises the following steps: capturing user editing behavior and content increment data in real time, fusing behavior and language features through a double-flow spatiotemporal attention network to identify the source of the content; using a medical knowledge graph to perform consistency verification and literature tracing on the extracted medical entities and factual assertions; and generating tamper-proof full-process traceability logs based on a hash chain. The system comprises six modules: data capture, feature extraction, source identification, fact verification, risk disposal, and log generation. The application realizes real-time, multi-dimensional, and traceable intelligent supervision of scientific research content, effectively improving the authenticity and credibility of hospital scientific research achievements.

Description

Technical Field

[0001] This application relates to the field of AI technology, and in particular to a method and system for managing hospital research content based on generative AI. Background Technology

[0002] With the continuous evolution of artificial intelligence (AI) technology, generative AI is developing rapidly and profoundly changing the scientific research ecosystem. Generative AI, represented by GPT, BERT, PaLM, and GLM, has demonstrated outstanding capabilities in understanding and generating natural language, and is widely used in research processes such as biomedical analysis, academic writing, literature reviews, data analysis, academic translation, and code generation. Researchers can use these tools to quickly complete tasks such as writing initial drafts of papers, constructing research backgrounds, and expanding experimental ideas, thereby significantly improving research efficiency and technological innovation capabilities. However, behind these technological benefits lie serious ethical challenges in scientific research.

[0003] The construction of a scientific research integrity system is a crucial guarantee for promoting the healthy development of science. Especially in the medical and health field, scientific research integrity is not only related to people's lives and health, but also key to enhancing independent scientific and technological innovation capabilities. Because generative AI's generation mechanism is based on probabilistic language prediction, its output may contain erroneous information, false citations, or spliced ​​text. If researchers use model-generated content without verification, it is highly likely to mislead research directions or lead to academic misconduct. More concerningly, some researchers fail to acknowledge AI involvement in their academic achievements, submit AI-generated paragraphs as original work, or even construct false data and experimental results through models, thereby challenging fundamental ethical requirements such as the authenticity and verifiability of scientific research.

[0004] Currently, generative AI inherently possesses "black box" characteristics, making it difficult to trace its origins and explain. This is particularly prominent in medical research, which demands a high degree of interpretability, further exacerbating the difficulties of academic oversight and accountability. Most platforms have not yet established identification standards and usage rules for AI-generated content, resulting in blurred ethical boundaries and a negative ecosystem. If the misuse of generative AI spreads, it will not only undermine the foundation of trust within the academic community but also affect the long-term quality of scientific and technological innovation. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems that traditional single text detection methods are easily circumvented in professional fields, and to enhance the transparency and accountability of the research process.

[0006] To achieve the above objectives, this application provides a hospital research content management method and system based on generative AI, which adopts the following technical solution:

[0007] Firstly, this application provides a hospital research content management method based on generative AI, including:

[0008] In a scientific research text editing environment, the system captures the sequence of user behavior events and incremental content data in real time during the text editing process. The sequence of behavior events includes keyboard input events, mouse operation events, clipboard operation events, and window focus switching events. The incremental content data includes the character content involved in text insertion, deletion, and modification operations.

[0009] The behavioral event sequence is subjected to feature engineering to extract and form a user behavior temporal feature vector; linguistic features are extracted from the incremental content data to form a content incremental temporal feature vector.

[0010] The user behavior time-series feature vector and the content increment time-series feature vector are input into a pre-trained human-computer collaborative recognition model to generate a content source probability distribution for the current text editing activity. The content source probability distribution indicates the probability that the current content is created by humans, generated by artificial intelligence, or imported from external sources.

[0011] Real-time grammatical parsing is performed on incremental content data to identify and extract medical entities, numerical data, and factual assertions. Medical entities include disease names, drug names, gene targets, surgical procedures, and diagnostic indicators. Factual assertions are statements of logical relationships connecting two or more medical entities.

[0012] The extracted medical entities and factual assertions are queried in a pre-built medical knowledge graph database to perform factual consistency checks and citation source verification. The factual consistency check determines whether the factual assertions conflict with existing knowledge in the knowledge graph. The citation source verification determines whether the reference information declared in the content is true and matches the content.

[0013] Based on the probability distribution of content sources, the results of factual consistency verification, and the results of source tracing verification of cited literature, a research content risk assessment report is generated, and preset regulatory intervention actions are triggered.

[0014] The sequence of behavioral events, incremental content data, probability distribution of content sources, results of factual consistency verification, results of citation source verification, and records of regulatory intervention actions are encrypted and appended using a chain hash algorithm to form a research process traceability log that is temporally correlated and tamper-proof.

[0015] Preferably, real-time capture of user behavior event sequences and incremental content data during text editing specifically includes:

[0016] By hooking the text editor's kernel event listener, and according to the set sampling frequency, the system captures the timestamp of each key press, the timestamp of each key release, and the key code information; it also captures the timestamp of each clipboard paste operation, the data hash value of the pasted content, and the character length of the pasted content; and it captures the coordinates and timestamps of each mouse click, drag, and scroll wheel operation.

[0017] Establish a fixed time window, aggregate the behavioral events captured within the fixed time window, and calculate the behavioral statistical features. The behavioral statistical features include the average typing speed per unit time, the entropy value of the keystroke interval time distribution, the frequency of use of the delete key and backspace key, the frequency of paste operation and the average paste length, and the pause and burst ratio in the text input process.

[0018] Preferably, extracting linguistic features from incremental content data to form a temporal feature vector of incremental content specifically includes:

[0019] The incremental content data is segmented into words, and each word is converted into a high-dimensional real number vector using a word vector model pre-trained on a massive medical literature corpus.

[0020] Calculate the perplexity score of incremental content data within its context;

[0021] The statistical incremental data includes the N-gram frequency distribution, syntactic complexity index, and density of medical terminology; the syntactic complexity index includes average sentence length, subordinate clause depth, and average branching factor of the syntax tree.

[0022] The high-dimensional real vector, perplexity score, N-gram frequency distribution, syntactic complexity index, and density of medical terminology are concatenated into an incremental temporal feature vector.

[0023] Preferably, the human-machine collaborative recognition model is a two-stream spatiotemporal attention network;

[0024] The dual-stream spatiotemporal attention network includes a behavioral feature encoding branch and a content feature encoding branch. The behavioral feature encoding branch uses a long short-term memory network to encode the temporal feature vector of user behavior in order to capture the long-term dependence of user operation habits. The content feature encoding branch uses a one-dimensional convolutional neural network to encode the temporal feature vector of incremental content in order to extract local language patterns of text content.

[0025] The outputs of the behavioral feature encoding branch and the content feature encoding branch are fused through a cross-modal attention mechanism to calculate the mutual weights between behavioral features and content features, and generate a fused spatiotemporal feature representation.

[0026] The fused spatiotemporal feature representation is passed through a fully connected layer and a flexible maximum classifier to output the probability distribution of content sources.

[0027] Preferably, identifying and extracting medical entities, numerical data, and factual assertions specifically includes:

[0028] A named entity recognizer based on a conditional random field model is adopted. The named entity recognizer uses medical dictionary and context features to identify and classify medical entities from text.

[0029] Design rule templates based on regular expressions and dependency parsing to extract quantitative descriptions containing numerical values ​​and units, as well as statements of key relationships between entities, from text, thereby forming structured factual assertion triples.

[0030] Preferably, the medical knowledge graph database is constructed by integrating the following multiple data sources:

[0031] Publicly available biomedical literature databases;

[0032] An authoritative clinical trial registry and clinical practice guideline database;

[0033] The hospital's internal electronic medical record data, clinical test database, and previous research project database were used, and all sensitive data was anonymized and desensitized.

[0034] By performing entity linking, relation extraction, and knowledge fusion on the above data sources, a weighted directed graph containing nodes and edges is constructed, where nodes represent medical entities, edges represent relationships between entities, and the weight of an edge represents the credibility of that relationship.

[0035] Preferably, the factual consistency verification and citation source tracing verification specifically include:

[0036] The factual assertion triples to be verified are converted into graph query statements. The database of medical knowledge graphs is searched to see if there are identical or semantically equivalent paths. If they exist, the evidence sources and credibility scores that support the assertion in the knowledge graph are returned. If they do not exist or there are conflicting paths, they are marked as factually questionable.

[0037] The document analyzes the factual assertion triplet citations to extract key information such as author, year, journal, and unique identifier of the digital object; it then uses the extracted key information to query the document database through the application programming interface to verify its authenticity.

[0038] Obtain the abstract or full text of the verified literature, and calculate its matching degree with relevant statements in the content to be verified using a text similarity algorithm to confirm the accuracy of the citation.

[0039] Preferred regulatory intervention actions include:

[0040] Real-time highlighting; based on the probability distribution of content sources, text fragments that are determined to be generated by artificial intelligence or imported from external sources are marked with different transparency or background color in the text editor.

[0041] Risk warning pop-up: When the fact consistency verification result is that the fact is questionable, or the source verification of cited literature fails, a non-blocking warning window will pop up in the scientific text editing environment, detailing the detected problems and potential risks.

[0042] Content locking and review: For paragraphs that repeatedly trigger high-risk warnings or are determined by the system to contain academic misconduct, they are automatically set to read-only status, and a review request is generated and pushed to the review queue of the hospital's scientific research management department.

[0043] Mandatory AI interaction recording: The system has a built-in proxy module for interacting with external generative artificial intelligence models. All query requests initiated by users through this proxy module, the complete responses of the models, and the user's adoption behavior are forcibly recorded in the traceability log.

[0044] Secondly, this application provides a hospital research content management system based on generative AI, applied to a hospital research content management method based on generative AI as described in the first aspect, comprising:

[0045] The data capture module is used to capture the sequence of user behavior events and incremental content data in real time during the text editing process in a scientific research text editing environment.

[0046] The feature extraction module is used to perform feature engineering on the behavioral event sequence to extract and form a user behavior temporal feature vector; and to perform linguistic feature extraction on the incremental content data to form a content incremental temporal feature vector.

[0047] The source identification module is equipped with a pre-trained human-machine collaborative identification model, which is used to receive user behavior time-series feature vectors and content increment time-series feature vectors, and generate a content source probability distribution;

[0048] The fact verification module is used to perform real-time grammatical parsing on incremental content data, identify and extract medical entities, numerical data and factual assertions, and compare them with a pre-built medical knowledge graph database to perform fact consistency verification and citation source verification.

[0049] The risk management module is used to generate a research content risk assessment report based on the probability distribution of content sources, the results of factual consistency verification, and the results of citation literature source tracing verification, and to trigger preset regulatory intervention actions.

[0050] The log generation module is used to encrypt and append all the output data from the data capture module, source identification module, fact verification module, and risk management module using a chain hash algorithm to form a time-series-related and tamper-proof scientific research process traceability log.

[0051] Preferably, the data capture module is used for:

[0052] By hooking the text editor's kernel event listener, the system captures the timestamps of each keystroke, the timestamps of each keystroke, and the key code information at a sampling frequency of no less than 100Hz; it also captures the timestamps of each clipboard paste operation, the data hash value of the pasted content, and the character length of the pasted content; and it captures the coordinates and timestamps of each mouse click, drag, and scroll wheel operation.

[0053] Establish a fixed time window, aggregate the original behavioral events captured within the time window, and calculate a set of behavioral statistical features. These features include the average typing speed per unit time, the entropy value of the keystroke interval time distribution, the frequency of use of the delete and backspace keys, the frequency and average length of the paste operation, and the pause and burst ratio during text input.

[0054] Compared with existing technologies, this invention provides a hospital research content management method and system based on generative AI, which has the following beneficial effects:

[0055] 1. This invention proposes a multimodal fusion real-time monitoring mechanism. By combining user keyboard input dynamics, editing behavior patterns and other behavioral characteristics with the linguistic features of text content, a dual-stream spatiotemporal neural network is constructed to identify the source of content. This significantly improves the recognition accuracy of generative AI content that has been polished or is highly similar to human writing style, and solves the technical problem that traditional single text detection methods are easily circumvented in professional fields.

[0056] 2. An immutable research process traceability log based on a chain hash algorithm was designed, which solidifies every user operation, every AI interaction, and every system detection result in the form of encrypted timestamps. This provides an objective, complete, and unalterable chain of evidence for research integrity review, contribution definition, and traceability of academic achievements, thereby enhancing the transparency and accountability of the research process.

[0057] 3. The regulatory function is seamlessly embedded into the daily writing environment of researchers. Through non-intrusive intervention methods such as highlighting and risk pop-ups, real-time feedback and guidance are provided without interrupting the research process. This achieves an effective balance between research efficiency and academic norms, transforming the traditional post-review model into a new regulatory paradigm of in-process prevention and real-time assistance. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the execution steps of a hospital research content management method based on generative AI, according to an embodiment of this application.

[0059] Figure 2 This is a flowchart illustrating the steps of extracting linguistic features from incremental content data to form a temporal feature vector of incremental content in a hospital research content management method based on generative AI, as described in an embodiment of this application.

[0060] Figure 3 This is a flowchart illustrating the steps of a regulatory intervention action in a hospital research content management method based on generative AI, according to an embodiment of this application. Detailed Implementation

[0061] The following is in conjunction with the appendix Figure 1-3 This application will be described in further detail.

[0062] This application discloses a method and system for managing hospital research content based on generative AI.

[0063] Firstly, referring to Figure 1 This embodiment discloses a hospital research content management method based on generative AI, including the following steps:

[0064] S1. In a scientific research text editing environment, capture in real time the sequence of user behavior events and incremental content data during the text editing process;

[0065] The behavioral event sequence includes keyboard input events, mouse operation events, clipboard operation events, and window focus switching events; the incremental content data includes the character content involved in text insertion, deletion, and modification operations.

[0066] S2. Perform feature engineering on the behavioral event sequence to extract and form a user behavior temporal feature vector; perform linguistic feature extraction on the incremental content data to form a content incremental temporal feature vector;

[0067] S3. Input the user behavior time-series feature vector and the content increment time-series feature vector into a pre-trained human-computer collaborative recognition model to generate a content source probability distribution for the current text editing activity;

[0068] Among them, the content source probability distribution indicates the probability that the current content is created by humans, generated by artificial intelligence, or imported from external sources;

[0069] S4. Perform real-time grammatical parsing on incremental content data to identify and extract medical entities, numerical data, and factual assertions.

[0070] Among them, medical entities include disease names, drug names, gene targets, surgical procedures, and diagnostic indicators; factual assertions are statements of logical relationships connecting two or more medical entities;

[0071] S5. Based on the extracted medical entities and factual assertions, query a pre-built medical knowledge graph database to perform factual consistency verification and citation source verification;

[0072] Among them, the factual consistency verification judges whether the factual assertions conflict with the existing knowledge in the knowledge graph; the citation source verification judges whether the reference information declared in the content is real and the content matches.

[0073] S6. Based on the probability distribution of content sources, the results of factual consistency verification, and the results of source tracing verification of cited literature, generate a research content risk assessment report and trigger preset regulatory intervention actions;

[0074] S7. The sequence of behavioral events, incremental content data, probability distribution of content sources, results of factual consistency verification, results of citation source verification, and records of regulatory intervention actions are encrypted and appended using a chain hash algorithm to form a research process traceability log that is temporally related and tamper-proof.

[0075] Furthermore, the real-time capture of user behavior event sequences and incremental content data during the text editing process specifically includes:

[0076] By hooking the text editor's kernel event listener, and according to the set sampling frequency, the system captures the timestamp of each key press, the timestamp of each key release, and the key code information; it also captures the timestamp of each clipboard paste operation, the data hash value of the pasted content, and the character length of the pasted content; and it captures the coordinates and timestamps of each mouse click, drag, and scroll wheel operation.

[0077] Establish a fixed time window, aggregate the behavioral events captured within the fixed time window, and calculate the behavioral statistical features;

[0078] Behavioral statistical characteristics include average typing speed per unit time, entropy value of keystroke interval distribution, frequency of use of delete and backspace keys, frequency of paste operation and average paste length, and pause and burst ratio during text input.

[0079] The kernel event listener continuously monitors all user interactions at a sampling frequency of no less than 100Hz.

[0080] For keyboard input events, the system records the press timestamp, release timestamp, and key code information for each key press, thereby accurately reconstructing the user's keystroke dynamics.

[0081] For mouse operation events, the system captures the screen coordinates, operation type, and timestamp of each click, drag, and scroll wheel operation;

[0082] For clipboard operation events, the system extracts the original string of the pasted content each time a paste occurs, calculates its data hash value, and records the timestamp and character length of the paste operation.

[0083] For window focus switching events, the system monitors the switching behavior between the text editor window and other application windows, and records the time points when focus enters and leaves;

[0084] In this process, all behavioral event sequence data are written to a circular buffer in real time and aggregated according to a fixed time window; the length of the fixed time window is set to 1s to 5s, and the sliding step of the fixed time window is 0.5s.

[0085] Within each fixed time window, the system calculates a set of behavioral statistical features, including average typing speed per unit time, entropy of the keystroke interval distribution, frequency of use of the delete and backspace keys, frequency and average paste length, and the ratio of pauses to bursts during text input. A pause is defined as a time interval of more than 0.8 seconds between two consecutive input events, while a burst refers to intensive input behavior of typing more than ten characters consecutively within 0.5 seconds. These statistical features constitute a quantitative representation of the behavioral event sequence, providing basic data for subsequent feature engineering.

[0086] In step S2 above, the engineering processing of the behavioral event sequence specifically includes normalizing the above-mentioned behavioral statistical features, smoothing with a sliding window, and performing difference operations to eliminate individual differences and highlight dynamic trends.

[0087] The normalization method uses the Z-score method to calculate the mean and standard deviation based on the behavioral data of the same user over the past thirty days.

[0088] The sliding window smoothing employs a five-point weighted average filter to suppress high-frequency noise. Differential operations are used to extract the rate of change of behavioral features, such as the acceleration of typing speed or abrupt changes in pasting frequency.

[0089] Through the above process, the final user behavior time-series feature vector has 32 dimensions, with each dimension corresponding to a processed behavioral statistical indicator.

[0090] Simultaneously, linguistic features are extracted from the incremental content data. Incremental content data refers to text segments added, deleted, or modified during the text editing process.

[0091] The system first performs word segmentation on the incremental text, using a bidirectional maximum matching algorithm based on a medical dictionary to ensure accurate segmentation of technical terms.

[0092] Among them, reference Figure 2 The process of extracting linguistic features from incremental content data to form a temporal feature vector of incremental content data specifically includes:

[0093] S21. Perform word segmentation on the incremental content data, and use a word vector model pre-trained on medical literature corpus to convert each word into a high-dimensional real number vector;

[0094] S22. Calculate the perplexity score of incremental content data in context;

[0095] S23. Statistical analysis of N-gram frequency distribution, syntactic complexity index, and density of medical terminology in incremental data; syntactic complexity index includes average sentence length, subordinate clause depth, and average branching factor of syntax tree.

[0096] S24. Concatenate the high-dimensional real number vector, perplexity score, N-gram frequency distribution, syntactic structure complexity index, and medical terminology density into a content incremental temporal feature vector.

[0097] Specifically, using a word vector model pre-trained on a massive medical literature corpus, each word is converted into a 512-dimensional high-dimensional real-number vector. The system further calculates the perplexity score of the incremental text within its context. This score is derived using a medical domain language model based on a transformer architecture and is used to measure the text's linguistic fluency and semantic rationality. In addition, the system statistically analyzes the frequency distribution of N-grams in the incremental text, with N ranging from two to four, covering phrase-level language patterns; calculates syntactic complexity indicators, including average sentence length, subordinate clause depth, and the average branching factor of the syntax tree; and calculates the density of medical terminology, i.e., the ratio of the number of medical terms to the total number of words. All these features are concatenated into a comprehensive content-incremental temporal feature vector with 1024 dimensions, constituting the content-incremental temporal feature vector.

[0098] Furthermore, the human-machine collaborative recognition model is a two-stream spatiotemporal attention network;

[0099] The dual-stream spatiotemporal attention network includes a behavioral feature encoding branch and a content feature encoding branch. The behavioral feature encoding branch uses a long short-term memory network to encode the temporal feature vector of user behavior in order to capture the long-term dependence of user operation habits. The content feature encoding branch uses a one-dimensional convolutional neural network to encode the temporal feature vector of incremental content in order to extract local language patterns of text content.

[0100] The outputs of the behavioral feature encoding branch and the content feature encoding branch are fused through a cross-modal attention mechanism to calculate the mutual weights between behavioral features and content features, and generate a fused spatiotemporal feature representation.

[0101] The fused spatiotemporal feature representation is passed through a fully connected layer and a flexible maximum classifier to output the probability distribution of content sources;

[0102] The dual-stream spatiotemporal attention network includes a behavioral feature encoding branch and a content feature encoding branch.

[0103] Among them, the behavioral feature encoding branch adopts a two-layer stacked long short-term memory network with 256 hidden units in each layer. It is used to encode the temporal feature vector of user behavior to capture the long-term dependence of user operation habits, such as the typing rhythm pattern or editing preferences of a specific user.

[0104] The content feature encoding branch uses a one-dimensional convolutional neural network with three convolutional layers, kernel sizes of 3, 5, and 7, and channel numbers of 128, 256, and 512, respectively. It is used to encode the incremental temporal feature vector of the content to extract local language patterns of the text content, such as repetitive structures or overly fluent sentences commonly found in AI-generated text.

[0105] Therefore, the output feature sequences of the behavior feature encoding branch and the content feature encoding branch are fused through a cross-modal attention mechanism. The cross-modal attention mechanism first calculates the dot product similarity matrix between the behavior features and the content features, then generates attention weights through a softmax function, and finally performs a weighted summation of the content features to obtain the behavior-guided content representation.

[0106] The fused spatiotemporal feature representation has a dimension of 1024 and is input into a three-layer fully connected network with 512, 256, and 256 neurons in each layer. The final layer connects to a flexible maximum value classifier, outputting a content source probability distribution. This distribution contains three probability values, corresponding to three source categories: original content, AI-generated content, and externally copied / imported content. The model training employs a cross-entropy loss function and introduces class balancing weights to address uneven data distribution. The training data comes from real-world records of hospital researchers writing using text from different sources in a controlled environment, encompassing over 100,000 hours of editing activity and corresponding content tags.

[0107] In step S4, identifying and extracting the medical entities, numerical data, and factual assertions includes:

[0108] S41. A named entity recognizer based on a conditional random field model is adopted. The named entity recognizer uses medical dictionary and context features to identify and classify medical entities from text.

[0109] S42. Using a pre-defined rule template based on regular expressions and dependency parsing, quantitative descriptions containing numerical values ​​and units, as well as statements of key relationships between entities, are extracted from the text and structured factual assertion triples are formed.

[0110] Among them, the medical entity recognition uses a named entity recognizer based on a conditional random field model. The feature functions of the named entity recognizer include word form features, part-of-speech tagging, prefixes and suffixes, medical dictionary matching results, and the average word vector within the context window.

[0111] The medical dictionary is compiled from the International Classification of Diseases, 10th Edition, the Generic Names Database of Drugs, the Human Genome Nomenclature Committee's Standard List, and the National Clinical Practice Terminology Collection.

[0112] The named entity recognizer outputs each recognized entity and its category label, including disease name, drug name, gene target, surgical procedure and diagnostic indicators. Numerical data extraction is achieved through regular expression-based rule templates, which cover common medical numerical expressions, such as "blood pressure 120 mmHg", "tumor diameter 3.5 cm", "hemoglobin concentration 120 g / L", etc. The system extracts the numerical values, units and associated medical entities to form structured records.

[0113] In this process, factual assertion extraction employs a method that combines dependency parsing with rule templates.

[0114] First, the system constructs a dependency syntax tree for the incremental text. Then, it traverses the predicate nodes in the tree to identify the subject and object they govern. If both the subject and object are identified medical entities, and the predicate belongs to a predefined set of relational verbs (including "treatment," "inhibition," "induction," "diagnosis," "association," "cause," "relief," etc.), then the triple (head entity, relation, tail entity) is output as a factual assertion. For example, the assertion (aspirin, inhibit, platelet aggregation) is extracted from the sentence "Aspirin can inhibit platelet aggregation."

[0115] Furthermore, the medical knowledge graph database is constructed by integrating the following multiple data sources, including:

[0116] Publicly available biomedical literature databases, such as the US National Library of Medicine Biomedical Literature Retrieval System Abstract Database, the European Molecular Biology Laboratory European Bioinformatics Institute Database, the World Health Organization International Clinical Trial Registry Platform, and the Chinese Medical Association Clinical Practice Guidelines Database, etc.

[0117] An authoritative clinical trial registry and clinical practice guideline database;

[0118] The hospital's internal electronic medical record data, clinical test database, and past research project database are all anonymized and desensitized. All sensitive data undergoes strict anonymization and desensitization before being entered into the database, removing privacy fields such as patient identity information and doctor's employee number.

[0119] By performing entity linking, relation extraction, and knowledge fusion on the above data sources, a weighted directed graph containing nodes and edges is constructed, where nodes represent medical entities, edges represent relationships between entities, and the weight of an edge represents the credibility of that relationship.

[0120] The knowledge graph construction employs entity linking, relation extraction, and knowledge fusion techniques, ultimately forming a weighted directed graph containing hundreds of millions of nodes and edges. Nodes represent medical entities, edges represent relationships between entities, and the weight of an edge represents the credibility of that relationship, ranging from zero to one, determined by the quantity, quality, and consistency of the evidence sources. Fact consistency verification transforms the factual assertion triples to be verified into a graph query statement, searching the knowledge graph for identical or semantically equivalent paths.

[0121] Semantic equivalence judgment is based on the synonym mapping and relational reasoning rules of the medical ontology. If a supporting path exists, a list of evidence sources and a comprehensive credibility score are returned; if a conflicting path exists (e.g., there is a high credibility relation in the knowledge graph that "aspirin induces gastric bleeding", while the assertion to be verified is "aspirin treats gastric bleeding"), it is marked as factually questionable, and the conflict strength is calculated.

[0122] The citation source verification analyzes the reference annotations of factual assertion triples to extract key information such as author, year, journal, volume and issue, page number, and unique identifier of the numeric object.

[0123] The system uses the application programming interface (API) of the literature database to perform precise searches using the extracted information. If the literature exists, its abstract or full text is obtained, and a BERT-based semantic similarity model is used to calculate the match between the abstract / text and relevant statements in the content to be verified. If the match is lower than a preset threshold (e.g., 0.7), the citation is deemed inaccurate.

[0124] Specifically, the verification of factual consistency and the source tracing of referenced documents include:

[0125] S101. Convert the factual assertion triples to be verified into graph query statements, and search the medical knowledge graph database for the existence of identical or semantically equivalent paths; if they exist, return the evidence sources and credibility scores that support the assertion in the knowledge graph; if they do not exist or conflicting paths exist, mark them as factually questionable.

[0126] S102. Parse the factual assertion triplet citations and extract key information such as author, year, journal, and unique identifier of the digital object; use the extracted key information to query and verify the authenticity of the document by calling the application programming interface of the document database;

[0127] S103. Obtain the abstract or full text of the verified document, and calculate its matching degree with the relevant statements in the content to be verified through a text similarity algorithm to confirm the accuracy of the citation.

[0128] The research content risk assessment report adopts a three-level classification system: low risk, medium risk, and high risk. Low risk refers to content that is original work by humans and has passed fact-checking; medium risk refers to content that is generated by artificial intelligence but has passed fact-checking, or original work by humans but contains minor inaccuracies in citation; high risk refers to content that is generated by artificial intelligence and whose facts are questionable, or contains fictitious citations, or is judged to be externally copied and pasted for three or more consecutive paragraphs.

[0129] Regulatory intervention actions are dynamically triggered based on risk levels. For low-risk content, the system only records the information in the background and does not intervene through the interface.

[0130] For medium-risk content, the system adjusts the transparency of relevant text fragments or highlights them with different background colors in the text editor. For example, it marks AI-generated content as light blue and semi-transparent, and marks inaccurately cited content as light yellow.

[0131] For high-risk content, the system will display a non-blocking warning window, detailing the detected problem type, specific location, potential risks, and suggested modifications. If the cumulative number of high-risk content in the same document exceeds five, or if the intensity of a single factual conflict exceeds the threshold of 0.9, the system will automatically set the relevant paragraphs to read-only status and generate a review request, which will be pushed to the review queue of the hospital's research management department.

[0132] Among them, reference Figure 3 Regulatory intervention actions include:

[0133] S61. Real-time highlighting: Based on the probability distribution of content sources, text fragments that are determined to be generated by artificial intelligence or imported from external sources are marked with different transparency or background color in the text editor.

[0134] S62. Risk warning pop-up: When the fact consistency verification result is that the fact is questionable, or the source tracing verification of cited literature fails, a non-blocking warning window will pop up in the scientific text editing environment, detailing the detected problems and potential risks.

[0135] S63. Content locking and review: For paragraphs that continuously trigger high-risk warnings or are judged by the system to have academic misconduct, they are automatically set to read-only status and a review work order is generated and pushed to the review queue of the hospital's scientific research management department.

[0136] S64. Mandatory AI Interaction Recording: The system has a built-in agent module for interacting with external generative artificial intelligence models. All query requests initiated by users through this agent module, the complete responses of the models, and the user's adoption behavior are forcibly recorded in the traceability log to ensure that all AI interaction behaviors are traceable.

[0137] In step S7, the sequence of behavioral events, incremental content data, probability distribution of content sources, results of factual consistency verification, results of citation source verification, and records of regulatory intervention actions are encrypted and appended using a chain-like hash algorithm to form a time-series-related and tamper-proof research process traceability log. This research process traceability log adopts a blockchain-like structure, with each record containing the hash value of the previous record, the data content of the current record, a timestamp, user identification, and digital signature. The hash algorithm uses the national cryptographic standard SM3 to ensure data integrity. The log is stored in a secure storage server within the hospital using an append-only method, with strictly limited access permissions. Any attempt to tamper with historical records will cause all subsequent hash chains to break, thus being immediately detected by the system. This research process traceability log provides an objective, complete, and unalterable chain of evidence for research integrity review, contribution determination, and the tracing of academic achievements.

[0138] Secondly, this embodiment discloses a hospital research content management system based on generative AI, applied to the hospital research content management method based on generative AI as described in the first aspect, including:

[0139] The data capture module is used to capture the sequence of user behavior events and incremental content data in real time during the text editing process in a scientific research text editing environment.

[0140] The feature extraction module is used to perform feature engineering on the behavioral event sequence to extract and form a user behavior temporal feature vector; and to perform linguistic feature extraction on the incremental content data to form a content incremental temporal feature vector.

[0141] The source identification module is equipped with a pre-trained human-machine collaborative identification model, which is used to receive user behavior time-series feature vectors and content increment time-series feature vectors, and generate a content source probability distribution;

[0142] The fact verification module is used to perform real-time grammatical parsing on incremental content data, identify and extract medical entities, numerical data and factual assertions, and compare them with a pre-built medical knowledge graph database to perform fact consistency verification and citation source verification.

[0143] The risk management module is used to generate a research content risk assessment report based on the probability distribution of content sources, the results of factual consistency verification, and the results of citation literature source tracing verification, and to trigger preset regulatory intervention actions.

[0144] The log generation module is used to encrypt and append all the output data from the data capture module, source identification module, fact verification module, and risk management module using a chain hash algorithm to form a time-series-related and tamper-proof scientific research process traceability log.

[0145] The data capture module serves the following purpose:

[0146] By hooking the text editor's kernel event listener, the system captures the timestamps of each keystroke, the timestamps of each keystroke, and the key code information at a sampling frequency of no less than 100 Hz; it also captures the timestamps of each clipboard paste operation, the data hash value of the pasted content, and the character length of the pasted content; and it captures the coordinates and timestamps of each mouse click, drag, and scroll wheel operation.

[0147] Establish a fixed time window, aggregate the original behavioral events captured within the time window, and calculate a set of behavioral statistical features. These features include the average typing speed per unit time, the entropy value of the keystroke interval time distribution, the frequency of use of the delete and backspace keys, the frequency and average length of the paste operation, and the pause and burst ratio during text input.

[0148] The data capture module is deployed on the client side of the scientific text editing environment and is responsible for capturing user behavior event sequences and incremental content data in real time.

[0149] The feature extraction module runs on the backend server, receiving raw data and performing feature engineering and linguistic feature extraction. The source identification module loads a pre-trained human-machine collaborative recognition model, infers from the feature vectors, and outputs a probability distribution of the content source.

[0150] The fact verification module integrates a medical named entity recognizer, a dependency parser, and a knowledge graph query engine to perform entity extraction and fact verification.

[0151] The risk management module generates a risk assessment report based on the multi-source detection results and triggers corresponding regulatory intervention actions through the client interface.

[0152] The log generation module is responsible for solidifying all intermediate results and final decision records into an immutable traceability log using a chain hash algorithm.

[0153] The modules communicate asynchronously via encrypted message queues, ensuring system stability and responsiveness under high concurrency scenarios. The overall system architecture supports horizontal scaling, allowing for the addition of computing nodes to meet the concurrent demands of a large number of research users within the hospital.

[0154] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A hospital research content management method based on generative AI, characterized in that, The method includes: In a scientific research text editing environment, capture in real time the sequence of user behavior events and incremental content data during the text editing process; Feature engineering is performed on the behavioral event sequence to extract and form a user behavior temporal feature vector; linguistic features are extracted from the incremental content data to form a content incremental temporal feature vector. Extracting linguistic features from incremental content data to form incremental content temporal feature vectors specifically includes segmenting the incremental content data into words and using a word vector model pre-trained on medical literature corpus to convert each word into a high-dimensional real number vector. Calculate the perplexity score of incremental content data within its context; The statistical incremental data includes the N-gram frequency distribution, syntactic complexity index, and density of medical terminology; the syntactic complexity index includes average sentence length, subordinate clause depth, and average branching factor of the syntax tree. The high-dimensional real number vector, perplexity score, N-gram frequency distribution, syntactic structure complexity index, and medical terminology density are concatenated into an incremental temporal feature vector. The user behavior time-series feature vector and the content increment time-series feature vector are input into a pre-trained human-computer collaborative recognition model to generate a content source probability distribution for the current text editing activity. Real-time grammatical parsing of incremental content data identifies and extracts medical entities, numerical data, and factual assertions. Based on the extracted medical entities and factual assertions, a pre-built medical knowledge graph database is queried to perform factual consistency verification and citation source verification. Based on the probability distribution of content sources, the results of factual consistency verification, and the results of source tracing verification of cited literature, a research content risk assessment report is generated, and preset regulatory intervention actions are triggered. The sequence of behavioral events, incremental content data, probability distribution of content sources, results of factual consistency verification, results of citation source verification, and records of regulatory intervention actions are encrypted and appended using a chain hash algorithm to form a research process traceability log that is temporally correlated and tamper-proof.

2. The hospital research content management method based on generative AI according to claim 1, characterized in that, The real-time capture of user behavior event sequences and incremental content data during text editing includes: By hooking the text editor's kernel event listener, and according to the set sampling frequency, the system captures the timestamp of each key press, the timestamp of each key release, and the key code information; it also captures the timestamp of each clipboard paste operation, the data hash value of the pasted content, and the character length of the pasted content; and it captures the coordinates and timestamps of each mouse click, drag, and scroll wheel operation. Establish a fixed time window, aggregate the behavioral events captured within the fixed time window, and calculate the behavioral statistical features. The behavioral statistical features include the average typing speed per unit time, the entropy value of the keystroke interval time distribution, the frequency of use of the delete key and backspace key, the frequency of paste operation and the average paste length, and the pause and burst ratio in the text input process.

3. The hospital research content management method based on generative AI according to claim 2, characterized in that, The human-machine collaborative recognition model is a two-stream spatiotemporal attention network; The dual-stream spatiotemporal attention network includes behavioral feature encoding branches and content feature encoding branches; The behavioral feature encoding branch uses a long short-term memory network to encode the temporal feature vectors of user behavior in order to capture the long-term dependence of user operating habits; The content feature encoding branch uses a one-dimensional convolutional neural network to encode the incremental temporal feature vectors of the content in order to extract local language patterns of the text content; The outputs of the behavioral feature encoding branch and the content feature encoding branch are fused through a cross-modal attention mechanism to calculate the mutual weights between behavioral features and content features, and generate a fused spatiotemporal feature representation. The fused spatiotemporal feature representation is passed through a fully connected layer and a flexible maximum classifier to output the probability distribution of content sources.

4. The hospital research content management method based on generative AI according to claim 3, characterized in that, Identifying and extracting medical entities, numerical data, and factual assertions specifically includes: A named entity recognizer based on a conditional random field model is used to identify and classify medical entities from text by utilizing a medical dictionary and contextual features. Using pre-defined rule templates based on regular expressions and dependency parsing, quantitative descriptions containing numerical values ​​and units, as well as statements of key relationships between entities, are extracted from the text and structured factual assertion triples are formed.

5. The hospital research content management method based on generative AI according to claim 4, characterized in that, The medical knowledge graph database is built by integrating the following multiple data sources: Publicly available biomedical literature databases; An authoritative clinical trial registry and clinical practice guideline database; In addition to the hospital's internal electronic medical record data, clinical laboratory database, and database of past research projects; By performing entity linking, relation extraction, and knowledge fusion on the data source, a weighted directed graph containing nodes and edges is constructed. Nodes represent medical entities, edges represent relationships between entities, and the weight of an edge represents the credibility of that relationship.

6. The hospital research content management method based on generative AI according to claim 5, characterized in that, Fact consistency verification and citation source verification include: The factual assertion triples to be verified are converted into graph query statements. The database of medical knowledge graphs is searched to see if there are identical or semantically equivalent paths. If they exist, the evidence sources and credibility scores that support the assertion in the knowledge graph are returned. If they do not exist or there are conflicting paths, they are marked as factually questionable. The document analyzes the factual assertion triplet citations to extract key information, including author, year, journal, and unique identifier of the digital object. By calling the application programming interface of the document database, the extracted key information is used to query and verify the authenticity of the document. Obtain the abstract or full text of the verified literature, and calculate its matching degree with the statements in the content to be verified using a text similarity algorithm to confirm the accuracy of the citation.

7. A hospital research content management method based on generative AI according to claim 6, characterized in that, Regulatory intervention actions include: Real-time highlighting: Based on the probability distribution of content sources, text fragments that are determined to be generated by artificial intelligence or imported from external sources are marked with different transparency or background color in the text editor; Risk warning pop-up: When the fact consistency verification result is that the fact is questionable, or the source verification of cited literature fails, a non-blocking warning window will pop up in the scientific text editing environment, and the detected problems and potential risks will be displayed in the window; Content locking and review: For paragraphs that repeatedly trigger high-risk warnings or are determined by the system to contain academic misconduct, they are automatically set to read-only status, and a review request is generated and pushed to the review queue of the hospital's scientific research management department. Mandatory AI Interaction Recording: Through a built-in proxy module that interacts with external generative artificial intelligence models, all user query requests initiated through this proxy module, the model's complete response, and the user's adoption behavior are forcibly recorded in the traceability log.

8. A hospital research content management system based on generative AI, characterized in that, The hospital research content management method based on generative AI as described in any one of claims 1-7 includes: The data capture module is used to capture the sequence of user behavior events and incremental content data in real time during the text editing process in a scientific research text editing environment. The feature extraction module performs feature engineering on the behavioral event sequence to extract user behavior temporal feature vectors; it also performs linguistic feature extraction on incremental content data to form incremental content temporal feature vectors; it segments the incremental content data and uses a word vector model pre-trained on medical literature corpora to convert each word into a high-dimensional real vector; it calculates the perplexity score of the incremental content data in context; it statistically analyzes the N-gram frequency distribution, syntactic complexity index, and medical terminology density in the incremental content data; the syntactic complexity index includes average sentence length, subordinate clause depth, and average branching factor of the syntax tree; and it concatenates the high-dimensional real vectors, perplexity scores, N-gram frequency distribution, syntactic complexity index, and medical terminology density into an incremental content temporal feature vector. The source identification module is equipped with a pre-trained human-machine collaborative identification model, which is used to receive user behavior time-series feature vectors and content increment time-series feature vectors, and generate a content source probability distribution; The fact verification module is used to perform real-time grammatical parsing on incremental content data, identify and extract medical entities, numerical data and factual assertions, and compare them with a pre-built medical knowledge graph database to perform fact consistency verification and citation source verification. The risk management module is used to generate a research content risk assessment report based on the probability distribution of content sources, the results of factual consistency verification, and the results of citation literature source tracing verification, and to trigger preset regulatory intervention actions. The log generation module is used to encrypt and append all the output data from the data capture module, source identification module, fact verification module, and risk management module using a chain hash algorithm to form a time-series-related and tamper-proof scientific research process traceability log.

9. The hospital research content management system based on generative AI according to claim 8, characterized in that, The data capture module is used for: By hooking the text editor's kernel event listener, the system captures the timestamp of each key press, the timestamp of each key release, and the key code information; it also captures the timestamp of each clipboard paste operation, the data hash value of the pasted content, and the character length of the pasted content; and it captures the coordinates and timestamps of each mouse click, drag, and scroll wheel operation. Establish a fixed time window, aggregate the original behavioral events captured within the time window, and calculate a set of behavioral statistical features. These features include the average typing speed per unit time, the entropy value of the keystroke interval time distribution, the frequency of use of the delete and backspace keys, the frequency and average length of the paste operation, and the pause and burst ratio during text input.

Images