Medical patent evaluation method and system based on multi-modal large model and big data

Abstract

The invention discloses a medical patent evaluation method and system based on a multi-modal large model and big data, belongs to the technical field of artificial intelligence, and solves the problems of insufficient multi-source data fusion, lack of dynamic value evaluation and low evaluation efficiency of an existing evaluation mode. Extracting technical elements and legal elements in the text, the image and the structured data, associating multi-dimensional data by utilizing a dynamic knowledge graph technology and based on a multi-modal analysis result of a patent document, performing multi-dimensional parallel evaluation on the patent multi-source data, and generating a dynamic weight comprehensive score and a risk report based on an evaluation result; according to the invention, by integrating the multi-modal medical data and the big data analysis technology, the medical patent is intelligently evaluated from multiple dimensions by means of the multi-modal big model, the patent evaluation is more comprehensive and accurate, and the patent can be comprehensively analyzed from multiple dimensions, so that the efficiency and objectivity of the medical patent evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence, and particularly relates to a medical patent evaluation method and system based on a multi-modal large model and big data. BACKGROUND

[0002] In recent years, with the continuous development of medical science, especially the gradual popularization of the application of precision medicine, genomics, artificial intelligence and big data technology, a large number of innovative patent applications have emerged in the field of medical technology. Medical patents not only involve technological innovation, but also involve huge market potential, implementation feasibility and social benefits and many other aspects. Therefore, how to comprehensively, systematically and accurately evaluate these patents has become a problem to be solved.

[0003] At present, the traditional medical patent evaluation method mainly relies on artificial experience and statistical models to evaluate the innovation, practicality and market potential of patents, and the evaluation efficiency is low, which is difficult to meet the rapid growth of medical patent evaluation demand. At present, although there are some patent evaluation methods involving machine learning, expert scoring and other methods, these methods often cannot consider multiple dimensions of analysis, and cannot give accurate conclusions for some complex medical technology evaluation. Therefore, how to use emerging technical means (such as artificial intelligence (multi-modal large model), big data analysis and deep learning) to improve the automation and accuracy of patent evaluation has become an important research direction in the field of current medical patent evaluation.

[0004] The existing medical patent evaluation methods can be generally divided into the following categories: expert scoring-based evaluation methods, database search-based evaluation methods and machine learning-based evaluation methods.

[0005] Among them, the expert scoring method is a traditional evaluation method. This method relies on experts in the field of patents to evaluate patents, including technical innovation, market prospects, and patent implementation difficulty. Experts score based on various technical indicators of patents to obtain a comprehensive evaluation result. The advantage of this method is that it has strong subjectivity and experience, and is suitable for evaluating simple patents, but it has the disadvantages of large subjective bias, low efficiency, and poor scalability. The evaluation method based on database retrieval mainly relies on information in the patent database for evaluation. By searching relevant patents, academic literature, technical reports, and other materials, the technical innovation and market demand of the patent are evaluated. Common databases include WIPO, USPTO, CNIPA, etc. This method can preliminarily evaluate the novelty of the technology and market competition by comparing the technical solutions of existing patents and market demand. However, the database retrieval method has the problem of limited information sources, and cannot comprehensively consider market, scientific research, social, and other multi-dimensional factors, and cannot deeply mine complex technical innovations. In recent years, with the development of machine learning technology, automatic patent evaluation methods based on machine learning have been widely applied. This method usually uses natural language processing (NLP) technology to extract technical features of patents based on patent text data, and uses classification algorithms, clustering algorithms, etc. to classify and evaluate patents. For example, using deep learning models such as convolutional neural networks (CNN) or long short-term memory networks (LSTM) to analyze and evaluate the technical innovation of patents.

[0006] Although the existing machine learning methods can improve the evaluation efficiency, most of them only rely on patent text data, ignoring market demand, implementation feasibility, and multi-modal data sources, and lack dynamic value evaluation, which cannot real-time correlate the latest medical research results, market data, and regulatory policy changes, resulting in certain limitations of existing methods.

[0007] In summary, current medical patent evaluation mainly relies on a combination of manual evaluation and statistical models, and the evaluation method has the following limitations:

[0008] 1) Existing evaluation methods are highly subjective: Traditional medical patent evaluation relies on expert experience, and there are problems of non-uniform evaluation standards and large subjective bias.

[0009] 2) Insufficient multi-source data fusion: Medical patents often involve a large amount of multi-modal information such as text, images related to clinical and experimental data, and traditional methods are difficult to effectively fuse and process, resulting in biased evaluation results.

[0010] 3) Lack of dynamic value evaluation: unable to real-time correlate the latest medical research results, market data, and regulatory policy changes.

[0011] 4) Evaluation efficiency is low: manual evaluation takes a long time and it is difficult to meet the rapidly growing demand for medical patent evaluation.

[0012] 5) Lack of intelligent evaluation system: current known tools do not fully use artificial intelligence technology to conduct multi-dimensional intelligent evaluation of the innovation, coverage, applicability and future value of patents.

[0013] In view of the above problems, we propose a medical patent evaluation method and system based on multi-modal large model and big data. SUMMARY

[0014] The present application aims to overcome the shortcomings of the prior art and provide a medical patent evaluation method and system based on multi-modal large model and big data, which solves the problems of insufficient multi-source data fusion, lack of dynamic value evaluation and low evaluation efficiency of the existing evaluation methods.

[0015] The present application is implemented as follows: a medical patent evaluation method based on multi-modal large model and big data, the method comprising:

[0016] S10, receiving a patent document uploaded by a user and target market information, wherein the patent document includes a claim, a specification, and an attached drawing;

[0017] S20, multi-modal analysis of the patent document based on a multi-modal large model, extracting technical elements and legal elements in the text, image, and structured data, and aligning the extracted patent multi-source data;

[0018] S30, pre-constructing a dynamic knowledge graph, correlating patent-literature-clinical data nodes, using dynamic knowledge graph technology and correlating multi-dimensional data based on the multi-modal analysis results of the patent document;

[0019] S40, obtaining multi-dimensional data and patent multi-source data, performing multi-dimensional parallel evaluation of the patent multi-source data by the multi-modal large model, and generating a dynamic weight comprehensive score and a risk report based on the evaluation results;

[0020] S50, updating the multi-modal large model weight and knowledge graph nodes in real time through a dynamic optimization mechanism according to user feedback.

[0021] Preferably, when extracting technical elements and legal elements in the text, image, and structured data, the technical elements include indications, technical problems / solutions, examples / experimental data, and the legal elements include claim scope, priority information, and review history.

[0022] Preferably, the multi-modal large model training method comprises:

[0023] The multi-modal large model is trained by using multi-source heterogeneous data, wherein the multi-source heterogeneous data includes patent data, non-patent literature, market data, technical trends, medical image and chart data sets, a legal regulation library, and a medical device registration database.

[0024] The data types of the multi-source heterogeneous data are identified, and the multi-source heterogeneous data is fused based on the data types, wherein a BioBERT model is used to extract technical feature vectors for text modal data, a VisionTransformer model is used to identify medical image structure features for image modal data, and a CNN combined with a BiLSTM model is used to analyze gene sequence biological features for sequence modal data.

[0025] The multi-source heterogeneous data after fusion processing is loaded, a cross-modal alignment network is constructed, text, chemical structural formula, and biological sequence are mapped to a unified vector space, and feature alignment is realized.

[0026] The data vector after feature alignment is obtained, and a dynamic knowledge graph is constructed based on the data vector after feature alignment, the dynamic knowledge graph takes medical technology entities as nodes, and is associated with technical attributes, business attributes, and clinical attributes, and learns the relevance of newly published literature / patents in real time based on a graph neural network (GNN) model.

[0027] The base model of the multi-modal large model is selected, wherein ChatGLM3 combined with LLaMA3 is used as the base model of the multi-modal large model, the multi-modal large model is trained and fine-tuned, a loss function of the multi-modal large model is defined, and a converged multi-modal large model is output.

[0028] Preferably, when the multi-modal large model is trained and fine-tuned, the multi-modal large model is fine-tuned based on LoRA / QLoRA fine-tuning technology, and the loss function of the multi-modal large model is represented as:

[0029] L=a1·CrossEntropy+a2·(1 / ClaimCoverage)

[0030] wherein a1 and a2 represent two weight coefficients for balancing the importance of two loss terms CrossEntropy and ClaimCoverage, respectively.

[0031] Preferably, when the multi-modal large model performs multi-dimensional parallel evaluation on patent multi-source data, it evaluates based on technical dimensions, clinical values, registration dimensions, legal dimensions, and market dimensions, respectively, and when a dynamic weight comprehensive score and a risk report are generated based on the evaluation results, the dynamic weight comprehensive score formula is represented as:

[0032] Score=α·Tech+β·Clinical+γ·Regulatory+δ·Legal+ε·Market

[0033] wherein Score represents a dynamic weight comprehensive score, and a, b, g, d, and e are weights of the technical dimension, the clinical value, the registration dimension, the legal dimension, and the market dimension, respectively.

[0034] In another aspect, the present application also provides a medical patent evaluation system based on a multi-modal large model and big data, comprising:

[0035] a user interaction layer for providing a Web / mini-program / APP multi-terminal portal, receiving a patent document and target market information uploaded by a user based on the Web / mini-program / APP multi-terminal portal;

[0036] an application layer deploying a medical patent AI evaluation engine including five evaluation dimensions of technology, clinic, registration, market, and law, and the medical patent AI evaluation engine performing multi-dimensional parallel evaluation on patent multi-source data based on a multi-modal large model, and generating a dynamic weight comprehensive score and a risk report based on the evaluation results;

[0037] a large model layer for real-time updating of multi-modal large model weights and knowledge graph nodes through a dynamic optimization mechanism;

[0038] a data processing layer for obtaining cross-modal data, fusing and processing the cross-modal data, and generating a dynamic knowledge graph;

[0039] a big data layer for integrating global patent databases, medical literature databases, medical image data sets, medical device registration databases, market analysis reports, and legal regulation databases.

[0040] Preferably, the large model layer comprises:

[0041] a multi-model collaborative engine for scheduling a multi-modal large model dedicated to the medical field;

[0042] a medical large model fine-tuning module for fine-tuning training based on the multi-modal large model using fine-tuning technology and medical field data samples;

[0043] a dynamic optimization mechanism module for real-time updating of multi-modal large model weights and knowledge graph nodes according to evaluation feedback.

[0044] Preferably, the data processing layer comprises:

[0045] a multi-modal feature fusion engine for extracting technical elements and legal elements from text, images, and structured data, and aligning the extracted patent multi-source data;

[0046] A knowledge graph construction module is configured to associate patent-literature-clinical data nodes and construct a dynamic knowledge graph.

[0047] Preferably, the multi-model collaborative engine comprises:

[0048] An innovation evaluation layer is configured to generate a technical scheme comparison report.

[0049] A value prediction layer is configured to predict patent value based on an XGBoost integrated model.

[0050] A infringement risk detection layer is configured to identify potential infringement patents based on subgraph matching of the dynamic knowledge graph.

[0051] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0052] In the embodiments of the present application, multi-modal medical data and big data analysis technology are integrated, and a multi-modal large model is used to intelligently evaluate medical patents from multiple dimensions, making the patent evaluation more comprehensive and accurate, and enabling comprehensive analysis of patents from multiple dimensions, thereby avoiding the problems of single data and one-sided evaluation in the prior art.

[0053] In the embodiments of the present application, multi-modal large model multi-dimensional parallel evaluation of patent multi-source data can effectively eliminate subjective factors in the artificial review process, ensuring the fairness and consistency of the evaluation results, and can be dynamically updated according to the latest patent and market information, so that the evaluation method always remains cutting-edge and timely.

[0054] In the embodiments of the present application, a multi-modal large model is provided, which can efficiently fuse information from different data sources (such as patent literature, medical images, clinical trial data, etc.) based on multi-modal data fusion technology and analyze in multiple dimensions to improve the comprehensiveness and accuracy of the evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A flowchart of a multi-modal analysis process of the patent document based on a multi-modal large model in Embodiment 1 of the present application is shown.

[0056] Figure 2 A dynamic updating mechanism diagram of the dynamic knowledge graph in Embodiment 1 of the present application is shown.

[0057] Figure 3 A structural schematic diagram of a medical patent evaluation system based on a multi-modal large model and big data provided by the present application is shown. DETAILED DESCRIPTION

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application; the description and the drawings are to be regarded as illustrative in nature and are not intended to limit the application. The terms "comprising," "having," "including," and "containing" are to be construed as open-ended terms (i.e., meaning "including, but not limited to,") unless otherwise noted. The terms "first," "second," and the like, as used herein do not have any specific meaning unless otherwise noted.

[0059] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily all directed to the same embodiment, or to a single alternative embodiment. It is expressly understood that any of the embodiments described herein can be combined with any of the other embodiments.

[0060] The existing evaluation methods have problems of insufficient fusion of multi-source data, lack of dynamic value evaluation and low evaluation efficiency. In view of the above problems, the present application provides a medical patent evaluation method and system based on multi-modal large model and big data. In brief, the method comprises the following steps: receiving the patent document and target market information uploaded by the user, extracting the technical elements and legal elements in the text, image and structured data, correlating the patent-literature-clinical data nodes, using the dynamic knowledge graph technology and based on the multi-modal analysis results of the patent document, correlating multi-dimensional data, using the multi-modal large model to perform multi-dimensional parallel evaluation on the multi-source data of the patent, and generating a dynamic weight comprehensive score and a risk report based on the evaluation results. In the embodiment of the present application, by integrating multi-modal medical data and big data analysis technology, the multi-modal large model is used to intelligently evaluate the medical patent from multiple dimensions, and the patent evaluation is more comprehensive and accurate. The comprehensive analysis of the patent from multiple dimensions can avoid the problems of single data and one-sided evaluation in the prior art. At the same time, the automatic evaluation process can greatly improve the evaluation efficiency, and is suitable for rapid evaluation of large-scale patents. Especially in the background of rapid development of medical technology, the application can keep up with the rapid speed of patent application, thereby improving the efficiency, accuracy, real-time performance and objectivity of medical patent evaluation.

[0061] Embodiment 1

[0062] The embodiment of the present application provides a medical patent evaluation method based on multi-modal large model and big data. The medical patent evaluation method based on multi-modal large model and big data comprises the following steps:

[0063] S10, receiving the patent document and target market information uploaded by the user, wherein the patent document comprises the claims, the specification and the drawings;

[0064] S20, performing multi-modal analysis on the patent document based on the multi-modal large model, extracting the technical elements and legal elements in the text, image and structured data, and aligning the extracted patent multi-source data;

[0065] In the embodiment of the present application, Figure 1 The embodiment 1 of the present application shows a multi-modal analysis process of the patent document based on the multi-modal large model. When extracting the technical elements and legal elements in the text, image and structured data, the technical elements comprise indications, technical problems / solutions, embodiments / experimental data, and the legal elements comprise claim scope, priority information and review history. When aligning the extracted patent multi-source data, the alignment means comprise but are not limited to medical image recognition instrument structure / biomarker, clinical experiment matching, academic paper verification and market data correlation.

[0066] S30, pre-constructing a dynamic knowledge graph, correlating patent-document-clinical data nodes, using dynamic knowledge graph technology and based on the multi-modal analysis results of the patent documents to correlate multi-dimensional data;

[0067] S40, obtaining multi-dimensional data and patent multi-source data, the multi-modal large model performing multi-dimensional parallel evaluation on the patent multi-source data, and generating a dynamic weight comprehensive score and a risk report based on the evaluation results;

[0068] It should be noted that when the multi-modal large model performs multi-dimensional parallel evaluation on the patent multi-source data, it is based on the evaluation of the technical dimension, clinical value, registration dimension, legal dimension and market dimension, and when the dynamic weight comprehensive score and risk report are generated based on the evaluation results, the weight of each dimension can be adaptively adjusted, for example: a high-risk market gives a higher weight to the registration dimension, and the dynamic weight comprehensive score formula is represented as:

[0069] Score = alpha * Tech + beta * Clinical + gamma * Regulatory + delta * Legal + epsilon * Market

[0070] Wherein, Score represents the dynamic weight comprehensive score, alpha, beta, gamma, delta and epsilon are the weights of the technical dimension, clinical value, registration dimension, legal dimension and market dimension, and the coefficients alpha-epsilon are dynamically optimized by the large model layer, wherein Tech, Clinical, Regulatory, Legal and Market represent the scores of the technical dimension, clinical value, registration dimension, legal dimension and market dimension.

[0071] Wherein, the technical dimension: comparing existing patent technology features, quantifying innovation levels (such as improvement type / disruptive type), clinical value: searching for associated clinical trial results, calculating efficacy evidence strength (such as RCT number, sample size), registration dimension: matching target market device classification rules, predicting registration cycle / success rate, legal dimension: semantic analysis of claim scope, evaluating infringement risk probability; market dimension: analyzing market share of competitors, predicting market penetration rate.

[0072] In the embodiment of the application, the multi-modal large model can effectively eliminate subjective factors in the artificial evaluation process, ensure the fairness and consistency of the evaluation results, and dynamically update according to the latest patent and market information, so that the evaluation method always remains cutting-edge and timely. The application is not only suitable for patent evaluation in the medical field, but also can be popularized to other technical fields, and has a wide market application prospect.

[0073] S50, according to user feedback, the multi-modal large model weight and knowledge graph node are updated in real time through a dynamic optimization mechanism.

[0074] In the embodiments of the present application, by integrating multi-modal medical data and big data analysis technology, the multi-modal large model is used to intelligently evaluate medical patents from multiple dimensions, and the patent evaluation is more comprehensive and accurate. The comprehensive analysis of the patent from multiple dimensions can avoid the problem of single data and one-sided evaluation in the prior art. At the same time, the automated evaluation process can greatly improve the evaluation efficiency, and is suitable for rapid evaluation of large-scale patents. Especially in the background of rapid development of medical technology, it can keep up with the rapid speed of patent application, thereby improving the efficiency, accuracy, real-time and objectivity of medical patent evaluation.

[0075] It should be noted that the core of the multi-modal large model is "cross-modal learning". Traditional models usually focus on a single data type (such as pure text or pure image), while multi-modal large models map different modal data to a unified semantic space through joint training, achieving information complementation. For example, the model can combine objects in pictures and textual descriptions to infer more accurate semantic information. In the embodiments of the present application, the multi-modal large model has a multi-modal fusion strategy, which is represented as: the model uses attention mechanisms, cross-modal encoders and other technologies to align and fuse features from different modalities. For example, the CLIP model realizes the matching of text and image through contrastive learning, and the GPT-4V (Vision) supports joint generation of images and text. When training the large model, a large amount of multi-modal data (such as pictures with text descriptions, video subtitles, etc.) is used for pre-training, so that the model learns the general representation ability across modalities, reducing the data requirements of downstream tasks.

[0076] In the embodiments of the present application, a multi-modal large model is provided, which can efficiently fuse information from different data sources (such as patent literature, medical images, clinical trial data, etc.) based on multi-modal data fusion technology, and analyze in multiple dimensions to improve the comprehensiveness and accuracy of the evaluation. Based on the feature-aligned data vector, a dynamic knowledge graph can be constructed. By combining multi-modal data fusion, big data analysis and automation technology, the multi-modal large model is used to intelligently evaluate medical patents from multiple dimensions (technology dimension, clinical dimension, registration dimension, market dimension and legal dimension, etc.), thereby constructing an intelligent evaluation system with the ability of real-time updating and optimization.

[0077] Embodiment 2

[0078] The embodiments of the present application provide a medical patent evaluation method based on a multi-modal large model and big data. The present embodiment takes cardiovascular interventional instrument patent evaluation as an example.

[0079] The embodiment of the application provides a multi-modal large model training method, which is used for evaluating a cardiovascular intervention instrument patent.

[0080] S101, early data preparation and large model fine-tuning training: multi-modal large model modeling training multi-source heterogeneous data is acquired, wherein the multi-source heterogeneous data includes patent data (claim, specification, review history), non-patent literature (PubMed paper, ClinicalTrials.gov test data), market data (medical sales report, medical insurance coverage policy), technical dynamics (academic conference abstract), medical image and chart data set (clinical data), legal regulation library (medical supervision regulations of various countries), medical device registration database (such as FDA EUDAMED);

[0081] S102, identifying the data type of the multi-source heterogeneous data, and fusing and processing the multi-source heterogeneous data based on the data type, wherein for text modal data, a BioBERT model is used to extract a technical feature vector, for image modal data, a VisionTransformer model is used to identify medical image structure features, and for sequence modal data, a CNN combined with a BiLSTM model is used to analyze gene sequence biological features;

[0082] In the embodiment of the application, the process of fusing and processing the multi-source heterogeneous data based on the data type is as follows:

[0083] Text modal: medical patent text -> BioBERT model -> technical feature vector

[0084] Image modal: medical schematic diagram / image -> Vision Transformer model -> spatial feature vector sequence modal: gene sequence -> CNN+BiLSTM model -> biological feature vector

[0085] Fusion mechanism: a cross-modal alignment network is constructed to map text, chemical structure formula and biological sequence to a unified vector space, and feature alignment is realized.

[0086] S103, loading the multi-source heterogeneous data fused and processed, constructing a cross-modal alignment network, mapping text, chemical structure formula and biological sequence to a unified vector space, and realizing feature alignment;

[0087] S104, dynamic knowledge graph construction: acquiring the data vector after feature alignment, constructing a dynamic knowledge graph based on the data vector after feature alignment, Figure 2FIG. 1 shows a dynamic updating mechanism diagram of the dynamic knowledge graph in Embodiment 1 of the present application, the dynamic knowledge graph takes medical technology entities as nodes, and is associated with technical attributes, business attributes, and clinical attributes, and learns the relevance of newly published literature / patents in real time based on a graph neural network GNN model;

[0088] For example, in the construction of the dynamic knowledge graph, medical technology entities are taken as nodes (such as “cardiovascular” and “minimally invasive surgical robot”), and the following attributes are associated: technical attributes (mechanism of action, indications); business attributes (market size, competitor patent layout); and clinical attributes (trial stage, adverse reaction rate). As shown in FIG. 2, the dynamic updating mechanism of the dynamic knowledge graph is to learn the relevance of newly published literature / patents in real time based on a graph neural network (GNN). Figure 2

[0089] S105, medical field large model training and fine-tuning: selecting a base model of a multi-modal large model, wherein ChatGLM3 combined with LLaMA3 is taken as the base model of the multi-modal large model, the multi-modal large model is trained and fine-tuned, a loss function of the multi-modal large model is defined, and a converged multi-modal large model is output.

[0090] In this embodiment, when the multi-modal large model is trained and fine-tuned, the multi-modal large model is fine-tuned based on LoRA / QLoRA fine-tuning technology, and at the same time, the role of “examiner” or “evaluation expert” is simulated through prompt word design during fine-tuning. The training samples of the multi-modal large model during training are medical patent abstracts and specifications (annotated technical features / claims), medical literature materials, medical images and charts, the loss function of the multi-modal large model is represented as:

[0091] L = a1·CrossEntropy + a2·(1 / ClaimCoverage)

[0092] Wherein, a1 and a2 represent two weight coefficients for balancing the importance of two loss terms CrossEntropy and ClaimCoverage, respectively. CrossEntropy controls the accuracy of the language model output (such as fluent and correct language generation); ClaimCoverage controls whether all important medical claims or facts (such as etiology, drugs, and treatment methods) are covered.

[0093] The multi-modal large model also has a vector database and a network search tool, wherein the vector database (such as FAISS) is used to construct a patent paragraph similarity retrieval system and a RAG (retrieval augmented generation) system, and the network search tool is used to obtain medical field information dynamically in real time.

[0094] ​Multi-model collaborative architecture: multi-modal large model realizes multi-dimensional parallel evaluation of patent multi-source data through innovative evaluation layer, value prediction layer and infringement risk detection layer;

[0095] Among them, the innovation evaluation layer uses the Deepseek-r1 generation technical scheme comparison report to locate the difference points;

[0096] The value prediction layer XGBoost integrated model, the input features include:

[0097] V=b1·T novelty +b2·C trial +b3·M growth

[0098] Wherein b1, b2, b3 are adaptive weights, which are dynamically adjusted by reinforcement learning.

[0099] The infringement risk detection layer is based on subgraph matching of knowledge graph, and identifies potential infringement patents.

[0100] In the evaluation of cardiovascular interventional instrument patents, the multi-modal large model after pre-data preparation and large model fine-tuning is used, specifically, the medical patent evaluation method based on multi-modal large model and big data, specifically comprising:

[0101] S10, receiving the patent document and target market information uploaded by the user, wherein the patent document includes the claims, the specification and the drawings;

[0102] Among them, the user uploads the patent "degradable heart stent and its preparation method" (including structure diagram, material composition), and the target market information is the specified target market: the United States, the European Union.

[0103] S20, based on the multi-modal large model, the patent document is analyzed in multiple modes, the technical elements and legal elements in the text, image and structured data are extracted, and the extracted patent multi-source data is aligned;

[0104] In the embodiment of the application, the patent document is analyzed in multiple modes based on the multi-modal large model, comprising:

[0105] Text analysis: extract features such as "magnesium alloy stent", "degradation rate ≤0.02mm / year";

[0106] Image analysis: identify the stent grid structure through graph convolutional neural network;

[0107] Knowledge graph association: link similar patents US20090209982A1 (biodegradable scaffold), and bind clinical literature (JACC: Cardiovascular Interventions Vol.15).

[0108] S30, pre-constructing a dynamic knowledge graph, associating patent-literature-clinical data nodes, using a dynamic knowledge graph technology and based on multi-modal analysis results of patent documents to associate multi-dimensional data;

[0109] S40, obtaining multi-dimensional data and patent multi-source data, multi-modal large model performing multi-dimensional parallel evaluation on patent multi-source data, and generating dynamic weight comprehensive score and risk report based on evaluation results;

[0110] In the embodiment of the application, the multi-modal large model performs multi-dimensional parallel evaluation on patent multi-source data based on multi-dimensional AI evaluation of technical dimension, clinical dimension, registration dimension, market dimension and legal dimension. Table 1 shows the execution results of multi-dimensional AI evaluation.

[0111] Table 1

[0112]

[0113] When generating dynamic weight comprehensive score and risk report based on evaluation results, the large model layer is assigned weights of technology (0.3), clinical (0.25), registration (0.2), legal (0.15) and market (0.1).

[0114] Comprehensive score = 92*0.3 + 90*0.25 + 68*0.2 + 60*0.15 + 70*0.1 = 78.1 / 100

[0115] The risk report includes generating a risk prompt: the scope of claims needs to be narrowed, and animal experiment data of degradation rate needs to be supplemented.

[0116] S50, according to user feedback, real-time updating multi-modal large model weight and knowledge graph node through dynamic optimization mechanism.

[0117] In this embodiment, when the multi-modal large model weight and knowledge graph node are real-time updated through the dynamic optimization mechanism, the user marks "infringement risk assessment is inaccurate", and the following process is executed:

[0118] S501, searching patent database review cases

[0119] S502, fine-tuning the attention mechanism of the legal analysis model

[0120] S503, updating the "degradable stent" infringement case node in the knowledge graph.

[0121] Embodiment 3

[0122] The embodiment of the application provides a medical patent evaluation system based on a multi-modal large model and big data, Figure 3A structural diagram of a medical patent evaluation system based on a multi-modal large model and big data is shown, and the system adopts a five-layer distributed architecture. The medical patent evaluation system based on a multi-modal large model and big data specifically comprises:

[0123] A user interaction layer is used to provide a Web / mini-program / APP multi-terminal portal. The Web / mini-program / APP multi-terminal portal receives patent documents and target market information uploaded by a user and displays a visual result.

[0124] An application layer deploys a medical patent AI evaluation engine including five evaluation dimensions of technology, clinic, registration, market, and law. The medical patent AI evaluation engine performs multi-dimensional parallel evaluation on patent multi-source data based on a multi-modal large model and generates a dynamic weight comprehensive score and a risk report based on the evaluation result.

[0125] When the medical patent AI evaluation engine including the five evaluation dimensions of technology, clinic, registration, market, and law is deployed, the evaluation dimensions are as follows:

[0126] Technical dimension (innovativeness / industry development trend / applicability / supplementary technology dependency / replaceability / maturity)

[0127] Market dimension (commercialization potential / market application / market size / market share / competitor / policy adaptability)

[0128] Clinical value (efficacy / safety / operability / conversion feasibility / effectiveness / application value positioning / supervision fit)

[0129] Registration dimension (regulatory compliance / supervision path matching degree / data property and exclusivity / patent time effectiveness coordination / supervision-patent interaction risk)

[0130] Legal dimension (claim stability / avoidability / dependency / patent infringement determinability / effective period / multi-country application / patent licensing status)

[0131] A large model layer is used to update multi-modal large model weights and knowledge graph nodes in real time through a dynamic optimization mechanism.

[0132] The large model layer includes:

[0133] A multi-model collaborative engine is used to schedule a multi-modal large model dedicated to the medical field. The multi-modal large model can be a technical analysis model and a legal compliance model.

[0134] The medical large model fine-tuning module is based on a multi-modal large model, fine-tuned using fine-tuning technology and medical field data samples, fine-tuned based on a multi-modal large model using fine-tuning technology and medical field data samples to improve medical patent data accuracy and strengthen medical terminology, regulations and policies, and medical image understanding, etc.

[0135] The dynamic optimization mechanism module is used to update the multi-modal large model weight and knowledge graph node in real time according to the evaluation feedback, wherein the dynamic optimization mechanism is: updating the model weight in real time according to the evaluation feedback; building a RAG (retrieval augmented generation) + networking search system, and updating and saving the medical field dynamic information such as the latest medical research results, medical related technology breakthroughs, market pattern changes (such as medical insurance policy adjustment) in the vector database in real time.

[0136] The data processing layer is used to obtain cross-modal data, fuse and process the cross-modal data, and generate a dynamic knowledge graph.

[0137] The data processing layer includes:

[0138] The multi-modal feature fusion engine is used to extract technical elements and legal elements in text, image and structured data, and align the extracted patent multi-source data;

[0139] The knowledge graph construction module is used to associate patent-literature-clinical data nodes, construct a dynamic knowledge graph, and correlate multi-dimensional data using dynamic knowledge graph technology based on multi-modal analysis results of patent documents.

[0140] The big data layer is used to integrate global patent databases, medical literature databases, medical image data sets, medical device registration databases, market analysis reports and legal regulation databases.

[0141] The big data layer can integrate the following six types of data sources:

[0142] Global patent database (such as USPTO, WIPO)

[0143] Medical literature database (PubMed / ClinicalTrials)

[0144] Medical image and chart data set

[0145] Medical device registration database (such as FDA EUDAMED)

[0146] Market analysis report (such as EvaluatePharma)

[0147] Legal regulation database (medical supervision regulations of various countries).

[0148] In the embodiment, the multi-model collaborative engine comprises:

[0149] An innovation evaluation layer for generating a technical scheme comparison report;

[0150] A value prediction layer for predicting patent value based on an XGBoost integrated model;

[0151] An infringement risk detection layer for identifying potential infringing patents based on subgraph matching of a dynamic knowledge graph.

[0152] The innovation evaluation layer uses Deepseek-r1 to generate a technical scheme comparison report and locate difference points.

[0153] The XGBoost integrated model of the value prediction layer, the input features include:

[0154] V = b1·T novelty + b2·C trial + b3·M growth

[0155] Wherein, b1, b2, b3 are adaptive weights, which are dynamically adjusted through reinforcement learning.

[0156] The infringement risk detection layer identifies potential infringing patents based on subgraph matching of a knowledge graph.

[0157] In summary, the present application provides a medical patent evaluation method and system based on multi-modal large models and big data. In the embodiment, multi-modal medical data and big data analysis technology are integrated, and multi-modal large models are used to intelligently evaluate medical patents from multiple dimensions. The patent evaluation is more comprehensive and accurate, and can be analyzed from multiple dimensions, thereby avoiding the problems of single data and one-sided evaluation in the prior art. The automated evaluation process can greatly improve the evaluation efficiency, and is suitable for rapid evaluation of large-scale patents. Especially in the context of rapid development of medical technology, it can keep up with the rapid pace of patent applications, thereby improving the efficiency, accuracy, real-time performance and objectivity of medical patent evaluation.

[0158] In the embodiment, the multi-modal large model can effectively eliminate subjective factors in the artificial review process, ensure the fairness and consistency of the evaluation results, and dynamically update according to the latest patent and market information, so that the evaluation method always remains cutting-edge and timely. The present application is not only suitable for patent evaluation in the medical field, but also can be popularized to other technical fields, and has a wide market application prospect.

[0159] It should be noted that, for the foregoing embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.

[0160] In several embodiments provided in the present application, it should be understood that the disclosed device can be implemented by other ways. For example, the device embodiments described above are only schematic, for example, the division of the above units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or communication connection between the shown or discussed units can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0161] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0162] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the protection scope of the application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application. Although the present application is described in detail with reference to the above embodiments, those skilled in the art can still combine, add or delete or make other adjustments to the features of the embodiments of the present application according to the circumstances without creative labor, so as to obtain different other technical solutions which do not deviate from the concept of the present application in essence. These technical solutions also belong to the scope of the present application.

Claims

1. A medical patent evaluation method based on multimodal large models and big data, characterized in that, The method includes: S10, Receive patent documents and target market information uploaded by the user, wherein the patent documents include claims, specification and drawings; S20, perform multimodal parsing on the patent document based on a multimodal large model, extract technical and legal elements from text, images, and structured data, and perform alignment processing on the extracted multi-source patent data; S30, pre-constructs a dynamic knowledge graph, links patent-literature-clinical data nodes, and uses dynamic knowledge graph technology and multi-modal parsing results of patent documents to link multi-dimensional data; S40 acquires multi-dimensional data and multi-source patent data. The multi-modal large model performs multi-dimensional parallel evaluation of the multi-source patent data and generates a dynamic weighted comprehensive score and risk report based on the evaluation results. S50, based on user feedback, uses a dynamic optimization mechanism to update the weights of multimodal large models and knowledge graph nodes in real time.

2. The medical patent evaluation method based on multimodal large models and big data as described in claim 1, characterized in that: When extracting technical and legal elements from text, images, and structured data, the technical elements include indications, technical problems / solutions, and embodiments / experimental data, while the legal elements include the scope of claims, priority information, and examination history.

3. The medical patent evaluation method based on multimodal large models and big data as described in claim 2, characterized in that: The multimodal large model training method includes: Acquire multi-source heterogeneous data for training multimodal large models. The multi-source heterogeneous data includes patent data, non-patent literature, market data, technology trends, medical image and chart datasets, legal and regulatory databases, and medical device registration databases. The data types of multi-source heterogeneous data are identified, and multi-source heterogeneous data are fused and processed based on data types. Specifically, for text modal data, the BioBERT model is used to extract feature vectors; for image modal data, the VisionTransformer model is used to identify medical image structural features; and for sequence modal data, CNN combined with BiLSTM model is used to parse gene sequence biological features. By loading multi-source heterogeneous data after fusion processing, a cross-modal alignment network is constructed to map text, chemical structural formulas, and biological sequences to a unified vector space to achieve feature alignment. Obtain feature-aligned data vectors, construct dynamic knowledge graphs based on feature-aligned data vectors, with medical technology entities as nodes, associating technical attributes, commercial attributes, and clinical attributes, and learning the relevance of newly published literature / patents in real time based on graph neural network (GNN) models; We select the base model for the multimodal large model, using ChatGLM3 combined with LLaMA3 as the base model. We train and fine-tune the multimodal large model, define the loss function of the multimodal large model, and output the converged multimodal large model.

4. The medical patent evaluation method based on multimodal large models and big data as described in claim 3, characterized in that: When fine-tuning the training of a large multimodal model, LoRA / QLoRA fine-tuning techniques are used. The loss function of the large multimodal model is expressed as: L=a1·CrossEntropy+a2·(1 / ClaimCoverage) Here, a1 and a2 represent two weighting coefficients used to balance the importance of the two loss terms, CrossEntropy and ClaimCoverage, respectively.

5. The medical patent evaluation method based on multimodal large models and big data as described in claim 2, characterized in that: When the multimodal large model performs multi-dimensional parallel evaluation of multi-source patent data, it evaluates the data based on technical, clinical, registration, legal, and market dimensions respectively. When generating a dynamic weighted comprehensive score and risk report based on the evaluation results, the dynamic weighted comprehensive score formula is expressed as: Score=α·Tech+β·Clinical+γ·Regulatory+δ·Legal+ε·Market Here, Score represents a dynamic weighted comprehensive score, with α, β, γ, δ, and ε being the weights for the technology dimension, clinical value dimension, registration dimension, legal dimension, and market dimension, respectively.

6. A medical patent evaluation system based on multimodal large models and big data, used to implement the medical patent evaluation method based on multimodal large models and big data as described in any one of claims 1-5, characterized in that: The medical patent evaluation system based on multimodal large models and big data includes: The user interaction layer is used to provide multiple entry points on Web / Mini Program / APP, and to receive patent documents and target market information uploaded by users through these multiple entry points. At the application layer, a medical patent AI evaluation engine is deployed, which includes five major evaluation dimensions: technology, clinical, registration, market, and law. The medical patent AI evaluation engine performs multi-dimensional parallel evaluation of multi-source patent data based on a multimodal large model, and generates dynamic weighted comprehensive scores and risk reports based on the evaluation results. The large model layer is used to update the weights of the multimodal large model and the knowledge graph nodes in real time through a dynamic optimization mechanism; The data processing layer is used to acquire cross-modal data, fuse and process the cross-modal data, and generate dynamic knowledge graphs. The big data layer is used to integrate global patent databases, medical literature databases, medical image datasets, medical device registration databases, market analysis reports, and legal and regulatory databases.

7. The medical patent evaluation system based on multimodal large models and big data as described in claim 6, characterized in that: The large model layer includes: Multi-model collaboration engine, used for scheduling large multimodal models specifically for the medical field; The Medical Large Model Fine-Tuning Module is based on a multimodal large model and uses fine-tuning techniques and data samples from the medical field for fine-tuning training. The dynamic optimization mechanism module is used to update the weights of the multimodal large model and the knowledge graph nodes in real time based on evaluation feedback.

8. The medical patent evaluation system based on multimodal large models and big data as described in claim 7, characterized in that: The data processing layer includes: A multimodal feature fusion engine is used to extract technical and legal elements from text, images, and structured data, and to align the extracted multi-source patent data. The knowledge graph construction module is used to link patent-document-clinical data nodes, build a dynamic knowledge graph, and use dynamic knowledge graph technology to link multi-dimensional data based on the multimodal parsing results of patent documents.

9. The medical patent evaluation system based on multimodal large models and big data as described in claim 8, characterized in that: The multi-model collaborative engine includes: The innovation assessment layer is used to generate a comparative report of technical solutions; The value prediction layer predicts patent value based on the XGBoost ensemble model. The infringement risk detection layer identifies potential infringing patents based on subgraph matching of a dynamic knowledge graph.

Images