A Deep Learning-Based Automatic Classification Method and System for Medical Device Patents
By constructing a BERT-TextCNN model and using generative data augmentation technology, the problem of mapping medical device patent classification to regulatory categories was solved, achieving efficient and accurate automatic classification and improving the efficiency of regulatory review and industry analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG PHARMA UNIV
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-30
AI Technical Summary
Existing patent classification systems cannot directly map medical device patents to industry regulatory categories. Traditional methods are time-consuming, labor-intensive, and have limited classification accuracy. Models such as BERT struggle to capture highly discriminative local keywords in medical devices and lack labeled data for regulatory categories.
We constructed a high-quality automatically labeled dataset, designed a BERT-TextCNN model that integrates global and local features, performed label calibration by combining a large language model with retrieval enhancement generation techniques, generated synthetic data for categories with insufficient sample numbers, and constructed a BERT-TextCNN fusion model for classification.
It achieves a precise mapping from patent technology classification to industry regulatory standards, improves classification accuracy and robustness, solves the problem of lack of labeled data in vertical fields, and alleviates the problem of long-tail category identification.
Smart Images
Figure CN122309743A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing and artificial intelligence technology, specifically a method and system for automatic classification of medical device patents based on deep learning. Background Technology
[0002] The medical device industry is a strategic high ground for global technological innovation and high-end manufacturing competition. With the rapid iteration of technology, the number of medical device invention patent applications has exploded. However, the existing International Patent Classification (IPC) and Joint Patent Classification (CPC) systems are mainly based on technical principles and are difficult to directly correspond to the "Medical Device Classification Catalog" adopted by regulatory authorities.
[0003] Current patent databases cannot directly inform users which category a patent belongs to in the regulatory catalog, posing a significant challenge to enterprises' infringement risk warnings and regulatory authorities' technical review. Traditional manual classification methods are time-consuming, labor-intensive, and difficult to guarantee consistency. Furthermore, medical device patent texts are typically lengthy, use obscure terminology, and involve multiple disciplines; traditional machine learning algorithms (such as SVM and KNN) struggle to capture deep semantic features when processing such long texts, resulting in limited classification accuracy.
[0004] While pre-trained models like BERT perform well in general text classification, relying solely on BERT makes it difficult to accurately capture highly discriminative local keywords (such as specific materials and parameters) in medical device classification; while CNNs alone lack the ability to understand long texts globally. Furthermore, existing public datasets lack labeled data for regulatory categories, resulting in significant label noise.
[0005] Therefore, there is an urgent need for an automatic classification method that can efficiently and accurately map medical device patents to industry regulatory categories. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for automatic classification of medical device patents based on deep learning. This method achieves accurate mapping of patent text to preset classification categories by constructing a high-quality automatically labeled dataset and designing a BERT-TextCNN model that integrates global and local features.
[0007] The technical solution adopted by this invention to achieve the above objectives is: an automatic classification method for medical device patents based on deep learning, comprising the following steps:
[0008] Step S1: Obtain the text data of the medical device patent and perform text cleaning and standardization preprocessing operations;
[0009] Step S2: Using a large language model combined with retrieval enhancement generation technology, the medical device classification standard is introduced as an external knowledge base to perform label calibration on the preprocessed text data to form a standard dataset.
[0010] Step S3: For categories with insufficient sample size in the standard dataset, use a generative model to generate synthetic data based on the patent text to balance the sample distribution of each category;
[0011] Step S4: Construct a deep learning classification model that integrates global and local semantic features;
[0012] Step S5: Train the deep learning classification model using the enhanced dataset, and input the medical device patent text to be classified into the trained model for classification prediction, and output its corresponding medical device classification standard category.
[0013] Step S2 includes the following steps:
[0014] Step S2-1: Use an embedding model to convert the patent text into a vector representation, perform semantic retrieval in a pre-built medical device classification standard vector knowledge base, and obtain the Top-N knowledge fragments that are semantically related to the current patent text.
[0015] Step S2-2: Combine the patent text with the retrieved knowledge fragments to construct prompt words, and input them into the large language model;
[0016] Step S2-3: Based on the preset role settings, classification rules and reasoning logic, the large language model outputs preliminary classification labels and reasoning basis;
[0017] Step S2-4: Set up conditional branch verification nodes for easily confused categories, and perform secondary verification and correction on the preliminary classification results through preset keyword matching rules or logical judgment rules, and output the calibrated classification labels.
[0018] In steps S2-4, the conditional branch verification node performs a secondary verification on the preliminary classification result based on the preset easily confused category pairs and the corresponding verification rule library; the verification rule library contains at least one verification rule; the verification rule includes keyword-based matching rules, logical expression-based judgment rules, or feature weight-based calculation rules.
[0019] Step S3 includes the following steps:
[0020] Step S3-1: Extract the independent claims portion of the patent text from the patent text in the insufficient sample category;
[0021] Step S3-2: Construct prompt words containing role settings, task descriptions, example guidance, and core semantic preservation constraints, and input them into the generative large language model;
[0022] Step S3-3: Based on prompt words, the generative large language model generates synthetic text with a structure, length, and style that are similar to the patent abstract while maintaining the core technical features and semantics of the original patent.
[0023] Step S3-4: Add the generated synthetic text to the training set of the corresponding category so that the number of samples in each category reaches the preset balance threshold.
[0024] In step S4, the deep learning classification model includes: a first neural network model for extracting global semantic features of the text and a second neural network model for extracting local key features of the text.
[0025] The global semantic features are obtained through the following steps:
[0026] Step a: The first neural network model uses a pre-trained language model as the encoding layer to encode the input text and output a sequence of hidden states;
[0027] Step b: Extract global semantic representation from the hidden state sequence;
[0028] The global semantics are obtained by concatenating the CLS tag vector output by the pre-trained language model with the vector obtained by averaging the hidden states of all valid tokens.
[0029] Step c: Input the global semantic representation into the first fully connected layer for dimension mapping to obtain the global feature vector.
[0030] The local semantic features are obtained through the following steps:
[0031] Step d: The second neural network model uses a multi-size convolutional neural network and takes the hidden state sequence output by the pre-trained language model as input;
[0032] The multi-size convolutional neural network is configured with four groups of one-dimensional convolutional kernels with sizes of 2, 3, 4, and 5, respectively, to capture phrase-level local semantic patterns of two-word, three-word, four-word, and five-word phrases.
[0033] Step e: Set at least two sets of one-dimensional convolution kernels of different sizes, perform convolution operations on the input sequence respectively, and extract N-gram local semantic features of different granularities;
[0034] Step f: Perform max pooling on the output of each group of convolutional kernels to retain the most salient local features;
[0035] Step g: Concatenate the pooled feature vectors from each group to obtain local feature vectors.
[0036] Step S4 includes the following steps:
[0037] Step S4-1: Concatenate and fuse the global feature vector and the local feature vector along the feature dimension to obtain the fused feature vector, thus completing the construction of the feature fusion layer; the feature fusion layer sends the fused feature vector to the feature optimization layer;
[0038] Step S4-2: The feature optimization layer performs layer normalization and random deactivation on the fused feature vector in sequence to accelerate model convergence and prevent overfitting;
[0039] Step S4-3: The classification mapping layer performs the process of inputting the fused feature vector into the multilayer perceptron classifier;
[0040] The multilayer perceptron classifier comprises: an input layer, at least one hidden layer, and a result output layer stacked together in sequence;
[0041] The hidden layer uses the ReLU activation function to perform a non-linear transformation on the input; the output layer uses the Softmax activation function to map the output of the hidden layer to a probability distribution of a preset category.
[0042] Step S4-4: The result output layer receives the probability distribution of each category and determines the final classification result based on the probability distribution.
[0043] A classification system for an automatic classification method of medical device patents based on deep learning includes the following functional modules:
[0044] The data preprocessing module is used to obtain the raw text data of medical device patents and perform text cleaning and standardization.
[0045] The automatic annotation module has a built-in large language model and retrieval enhancement generation engine. It introduces medical device classification standards as an external knowledge base to perform label calibration on preprocessed patent data and generate annotated datasets.
[0046] The data augmentation module has a built-in generative large language model that generates synthetic summaries based on patent claims for categories with insufficient sample size, in order to balance the sample distribution of each category.
[0047] The model building and training module is used to build and train a deep learning classification model that integrates global and local semantic features. The model includes a first neural network submodule that uses a pre-trained language model to extract global semantic features, a second neural network submodule that uses a multi-size convolutional neural network to extract local semantic features, and a feature fusion and classification submodule that hierarchically fuses global and local features and outputs classification results.
[0048] The model inference module encapsulates a pre-trained deep learning classification model and provides a model call interface for receiving the patent text to be classified and outputting its corresponding medical device classification standard category and confidence level.
[0049] The system adopts a B / S architecture, including the following layers connected from top to bottom:
[0050] The presentation layer, deployed on the user terminal, is used to provide a human-computer interaction interface, receive the patent text to be classified input by the user and send it to the business logic layer, and receive the classification results returned by the business logic layer and display them to the user in a visual form.
[0051] The business logic layer, deployed on the application server, communicates with the presentation layer to receive patent text sent by the presentation layer, calls the encapsulated deep learning classification model to perform inference and prediction, and returns the prediction results to the presentation layer.
[0052] The business logic layer includes:
[0053] The first service module is built using the Java Spring Boot framework and is responsible for user authentication, request distribution, logging, and data interaction with the presentation layer.
[0054] The second service module is built using the Python Flask framework. It communicates with the first service module, encapsulates the trained deep learning classification model, and provides a RESTful inference interface to receive patent text and return classification prediction results.
[0055] The data persistence layer, deployed on the database server, connects to the business logic layer and is used to receive and store user information, historical classification records, operation logs, and system configuration data generated by the business logic layer.
[0056] The present invention has the following beneficial effects and advantages:
[0057] 1. This invention achieves a leap from patent technology classification (IPC) to industry regulatory standards (Medical Device Classification Catalog), directly serving regulatory review and industry analysis.
[0058] 2. This invention innovatively utilizes LLM+RAG technology to construct an automatic annotation workflow, solving the problem of a lack of high-quality annotation data in vertical fields and significantly improving annotation accuracy.
[0059] 3. The BERT-TextCNN model proposed in this invention integrates the global semantic understanding capability of RoBERTa with the local feature extraction capability of TextCNN, effectively solving the classification problem of long and terminology-dense medical device patent texts. It achieves classification performance superior to traditional models and single deep learning models on both the test set and the external validation set.
[0060] 4. This invention alleviates the difficulty of identifying long-tail categories and improves the robustness of the model by constructing generative data to expand the sample for regulatory categories with a small number of samples and training it under a unified classification model framework. Attached Figure Description
[0061] Figure 1 This is a flowchart of the method of the present invention;
[0062] Figure 2 This is a schematic diagram of the architecture of the automatic annotation workflow in this invention;
[0063] Figure 3 This is a schematic diagram of the BERT-TextCNN fusion classification model in this invention;
[0064] Figure 4 This is the logical architecture diagram of the system of the present invention. Detailed Implementation
[0065] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0066] Example 1: Implementation of Classification Method
[0067] like Figure 1 As shown, the method flow of the present invention is as follows:
[0068] The first step is data acquisition and cleaning.
[0069] Retrieve medical device invention patents with application dates within a specified range and IPC main classification number A61B from patent databases (such as PatSnap), and obtain their text data, including abstracts and claims. Clean and standardize the original text: remove HTML tags (such as HTML tags...). Invisible characters are removed, English punctuation is standardized to Chinese full-width punctuation, and consecutive spaces are merged to obtain the preprocessed patent text data.
[0070] The acquired raw text data is preprocessed, including:
[0071] 1-1) Remove HTML tags, XML tags, and invisible characters (such as garbled characters and control characters);
[0072] 1-2) Standardize punctuation marks, for example, convert English punctuation marks to Chinese full-width punctuation marks;
[0073] 1-3) Merge consecutive spaces and newlines to standardize the text format;
[0074] 1-4) Convert the text to simplified / traditional Chinese and unify it to simplified Chinese;
[0075] 1-5) Truncate or pad the patent abstract and claims text to a preset length (e.g., 512 characters) to meet the input requirements of subsequent models.
[0076] The second step is to automatically annotate and build the workflow.
[0077] To address the issue of the lack of corresponding regulatory category labels in the "Medical Device Classification Catalog" for the raw data, this step utilizes Large Language Modeling (LLM) combined with Retrieval Augmentation (RAG) technology to construct an automated annotation workflow, such as... Figure 2 As shown. Specifically, it includes the following sub-steps:
[0078] 2-1) Constructing a vector knowledge base for the "Medical Device Classification Catalog"
[0079] We collected the "Medical Device Classification Catalog" document published by the National Medical Products Administration (NMPA), which includes 22 primary product categories and their descriptions, intended uses, typical product examples, etc. The catalog text was segmented into knowledge fragments, and each fragment was converted into a vector representation using an embedding model (e.g., bge-large-zh-v1.5). These vector representations were then stored in a vector database (e.g., Milvus, Faiss) to construct a knowledge base index.
[0080] 2-2) Knowledge retrieval: The patent abstract is vectorized using the bge-large-zh-v1.5 model, and the top-5 relevant knowledge fragments (including product description and intended use) are retrieved from the "Medical Device Classification Catalog" vector library.
[0081] 2-3) Preliminary classification: Input the retrieved knowledge as context into the Qwen big model, and require the model to make a preliminary judgment based on the "distinguishing key points" rule (such as the difference between "active" and "passive").
[0082] 2-4) Conditional Branch Validation: For easily confused category pairs (such as active implantable devices and active surgical instruments, medical diagnostic instruments and medical software), conditional branch validation nodes are set up. The validation rule base contains a variety of pre-defined validation rules, including:
[0083] a) Keyword-based matching rules: For example, if keywords such as "long-term indwelling" or "electrode" appear in the patent abstract, it tends to be classified as an active implantable device;
[0084] b) Judgment rules based on logical expressions: For example, if "signal processing" and "image display" appear simultaneously but "diagnostic decision" does not appear, it may belong to medical software rather than medical diagnostic equipment;
[0085] c) Calculation rules based on feature weights: Weighted scores are applied to multiple relevant features, and secondary discrimination is performed based on the total score.
[0086] The verification node selects the appropriate rules for secondary confirmation based on the preliminary classification results. If the verification results are inconsistent with the preliminary classification, corrections are made.
[0087] 2-5) Result Aggregation: Output the final calibrated labels to form a standard dataset containing the first-level product categories of the "Medical Device Classification Catalog". Experiments show that this workflow achieves an annotation accuracy of over 91.85%.
[0088] The third step is data augmentation.
[0089] For long-tail categories (such as hemodialysis devices) with a small number of samples in the standard dataset, this step employs generative data augmentation methods to balance the sample distribution. Specifically, the independent claims of the patents in this category are extracted, and a prompt-guided generative large language model is constructed to generate synthetic text with a structure, length, and style similar to the patent abstract while maintaining the core semantics. The synthetic text is then added to the training set of the corresponding category to ensure that the number of samples in each category reaches a preset equilibrium threshold.
[0090] The fourth step is to build the BERT-TextCNN model.
[0091] like Figure 3 As shown, the deep learning classification model constructed in this step includes a BERT encoding layer, a dual-path feature extraction layer, and a fusion classification layer:
[0092] 1) BERT Encoding Layer: The input is a text sequence truncated or padded to a length of 512, encoded using the Chinese-RoBERTa-wwm-ext model, and the output is a sequence of hidden states through the BERT encoding layer. ,in, For sequence length, For the hidden layer dimension.
[0093] 2) Dual-path feature extraction:
[0094] Global Path: Extracting the [CLS] vector from BERT output and the average pooling vector of all valid tokens After splicing the two together, we get After mapping through a fully connected layer, the global feature vector is obtained. .
[0095] Local Pathways: The hidden state sequence H is input into the TextCNN network, and four sets of one-dimensional convolutional kernels with sizes of 2, 3, 4, and 5 are used for convolution operations to capture N-gram local semantic features of different granularities. After max pooling, these features are concatenated to obtain the local feature vector. .
[0096] 3) Integration and classification: and By concatenating the features along the feature dimension, a fused feature vector is obtained. The fused feature vectors are processed sequentially by LayerNorm and Dropout. The processed vectors are then input into a Multilayer Perceptron (MLP). The MLP contains a hidden layer (ReLU activation) and an output layer (Softmax activation). The output layer shows the probability distribution of the text belonging to 22 primary product categories, and the final classification result is determined based on the probability distribution.
[0097] The fifth step is model training.
[0098] A hierarchical fine-tuning strategy was employed during training, freezing the low-level parameters of BERT and updating and optimizing only the high-level parameters of BERT, as well as the parameters of TextCNN and the fusion classification layer. The loss function used was the label-smoothed cross-entropy loss function, and the optimizer was AdamW. Experiments showed that the model achieved an accuracy of 0.8153 and a Macro F1 score of 0.7527 on the test set, significantly outperforming XGBoost, pure BERT, or pure TextCNN models.
[0099] Model Inference: The trained model is deployed on the inference server. For a medical device patent to be classified, its abstract text is input into the model. After forward propagation calculation, the model outputs the primary product category of the "Medical Device Classification Catalog" and its confidence level. If the confidence level is lower than a preset threshold (e.g., 0.7), a manual review process can be triggered.
[0100] Example 2: Classification System Implementation
[0101] This embodiment provides a deep learning-based automatic classification system for medical device patents, used to implement the method described in Embodiment 1. The system adopts a B / S architecture, such as... Figure 4 As shown, it includes the following functional modules, which work together to complete the patent classification task:
[0102] The data preprocessing module is used to acquire the original text data of medical device patents and perform text cleaning and standardization.
[0103] The automatic annotation module incorporates a large language model and a retrieval enhancement generation engine, and introduces the "Medical Device Classification Catalog" as an external knowledge base to perform label calibration on the preprocessed patent data and generate annotated datasets.
[0104] The data augmentation module incorporates a generative large language model that generates synthetic abstracts based on patent claims for categories with insufficient sample numbers, thereby balancing the sample distribution across categories.
[0105] The model building and training module is used to build and train the BERT-TextCNN fusion classification model. The model includes a first neural network submodule that uses a pre-trained language model to extract global semantic features, a second neural network submodule that uses a multi-size convolutional neural network to extract local semantic features, and a feature fusion and classification submodule that hierarchically fuses global and local features and outputs classification results.
[0106] The model inference module encapsulates a pre-trained deep learning classification model and provides a RESTful inference interface to receive the patent text to be classified and output its corresponding category in the "Medical Device Classification Catalog" and its confidence level.
[0107] like Figure 4 As shown, this embodiment provides a deep learning-based automatic classification system for medical device patents, used to implement the method described in Embodiment 1. The system adopts a B / S architecture, including a presentation layer, a business logic layer, and a data persistence layer.
[0108] 1) Presentation layer:
[0109] The user interface is built using the Vue.js framework combined with the Element UI component library and deployed on the user's terminal (such as a browser). The presentation layer includes:
[0110] Patent text input area: Supports users to input or paste patent abstract text;
[0111] Submit button: Triggers a category request;
[0112] Results display area: The categories are displayed in card format, and the confidence distribution of each category is shown in an ECharts bar chart. Categories with high confidence are marked in green, and categories with low confidence are marked in yellow or red, to help users make quick decisions.
[0113] 2) Business Logic Layer:
[0114] This Java Spring Boot service is responsible for user login authentication (JWT token), request dispatching, logging, and data interaction with the presentation layer. After receiving the patent text from the frontend, it calls the second service module via a RESTful interface.
[0115] Python Flask Inference Service: This service encapsulates a PyTorch environment, loads a pre-trained BERT-TextCNN model, and provides a RESTful inference interface. The interface receives patent text, calls the model to perform predictions, and returns the class label and confidence score.
[0116] 3) Data persistence layer:
[0117] It uses a MySQL relational database, deployed on a database server, and connects to the business logic layer. It stores user information, historical classification records (including input text, prediction results, confidence levels, operation times, etc.), operation logs, and system configuration parameters.
[0118] The workflow is as follows:
[0119] 1) Users access the system through a browser, enter a patent abstract, and submit it;
[0120] 2) The presentation layer sends the request to the first service module;
[0121] 3) The first service module verifies the user's identity, records logs, and forwards the patent text to the second service module;
[0122] 4) The second service module loads the model, performs inference, and returns the classification results;
[0123] 5) The first service module stores the results in the database and returns them to the presentation layer;
[0124] 6) The presentation layer displays the classification results in a visual format.
[0125] In summary, this invention discloses an automatic classification method and system for medical device patents based on deep learning. This method constructs an automatic annotation workflow by introducing a large language model and retrieval-enhanced generation technology, and uses the *Medical Device Classification Catalog* as an external knowledge base, effectively solving the problem of a lack of high-quality labeled data in vertical domains. For long-tail categories, a generative data augmentation strategy is adopted to balance the sample distribution and improve the model's ability to identify sparse categories. In terms of model design, a BERT-TextCNN fusion architecture is constructed, capturing global semantic features and local key features of the text through dual-path feature extraction, and performing hierarchical processing in the fusion classification layer to achieve accurate classification of medical device patent texts. The classification system built based on this method adopts a B / S architecture, providing a convenient user interface and efficient model inference services through front-end and back-end separation. It can map the patents to be classified to the first-level product categories specified in the *Medical Device Classification Catalog* in real time and visualize the classification confidence. This invention achieves effective alignment between patent technology classification and industry regulatory standards, significantly improving classification accuracy and automation levels. It can be widely applied in multiple fields such as medical device regulatory review, patent navigation, infringement risk warning, and industry analysis, and has significant practical value and broad application prospects.
[0126] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0127] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.
Claims
1. A deep learning-based medical instrument patent automatic classification method, characterized in that, Includes the following steps: Step S1: Obtain the text data of the medical device patent and perform text cleaning and standardization preprocessing operations; Step S2: Using a large language model combined with retrieval enhancement generation technology, the medical device classification standard is introduced as an external knowledge base to perform label calibration on the preprocessed text data to form a standard dataset. Step S3: For categories with insufficient sample size in the standard dataset, use a generative model to generate synthetic data based on the patent text to balance the sample distribution of each category; Step S4: Construct a deep learning classification model that integrates global and local semantic features; Step S5: Train the deep learning classification model using the enhanced dataset, and input the medical device patent text to be classified into the trained model for classification prediction, and output its corresponding medical device classification standard category. 2.The deep learning-based medical instrument patent automatic classification method of claim 1, wherein, Step S2 includes the following steps: Step S2-1: Use an embedding model to convert the patent text into a vector representation, perform semantic retrieval in a pre-built medical device classification standard vector knowledge base, and obtain the Top-N knowledge fragments that are semantically related to the current patent text. Step S2-2: Combine the patent text with the retrieved knowledge fragments to construct prompt words, and input them into the large language model; Step S2-3: Based on the preset role settings, classification rules and reasoning logic, the large language model outputs preliminary classification labels and reasoning basis; Step S2-4: Set up conditional branch verification nodes for easily confused categories, and perform secondary verification and correction on the preliminary classification results through preset keyword matching rules or logical judgment rules, and output the calibrated classification labels. 3.The deep learning-based medical instrument patent automatic classification method of claim 2, wherein, In steps S2-4, the conditional branch verification node performs a secondary verification on the preliminary classification result based on the preset easily confused category pairs and the corresponding verification rule library; the verification rule library contains at least one verification rule; the verification rule includes keyword-based matching rules, logical expression-based judgment rules, or feature weight-based calculation rules. 4.The deep learning-based medical instrument patent automatic classification method of claim 1, wherein, Step S3 includes the following steps: Step S3-1: Extract the independent claims portion of the patent text from the patent text in the insufficient sample category; Step S3-2: Construct prompt words containing role settings, task descriptions, example guidance, and core semantic preservation constraints, and input them into the generative large language model; Step S3-3: Based on prompt words, the generative large language model generates synthetic text with a structure, length, and style that are similar to the patent abstract while maintaining the core technical features and semantics of the original patent. Step S3-4: Add the generated synthetic text to the training set of the corresponding category so that the number of samples in each category reaches the preset balance threshold.
5. The method of claim 1, wherein the method is based on deep learning. In step S4, the deep learning classification model includes: a first neural network model for extracting global semantic features of the text and a second neural network model for extracting local key features of the text.
6. A method for automatic classification of medical device patents based on deep learning according to claim 1 or 5, characterized in that, The global semantic features are obtained through the following steps: Step a: The first neural network model uses a pre-trained language model as the encoding layer to encode the input text and output a sequence of hidden states; Step b: Extract global semantic representation from the hidden state sequence; The global semantics are obtained by concatenating the CLS tag vector output by the pre-trained language model with the vector obtained by averaging the hidden states of all valid tokens. Step c: Input the global semantic representation into the first fully connected layer for dimension mapping to obtain the global feature vector.
7. A method for automatic classification of medical device patents based on deep learning according to claim 1 or 5, characterized in that, The local semantic features are obtained through the following steps: Step d: The second neural network model uses a multi-size convolutional neural network and takes the hidden state sequence output by the pre-trained language model as input; The multi-size convolutional neural network is configured with four groups of one-dimensional convolutional kernels with sizes of 2, 3, 4, and 5, respectively, to capture phrase-level local semantic patterns of two-word, three-word, four-word, and five-word phrases. Step e: Set at least two sets of one-dimensional convolution kernels of different sizes, perform convolution operations on the input sequence respectively, and extract N-gram local semantic features of different granularities; Step f: Perform max pooling on the output of each group of convolutional kernels to retain the most salient local features; Step g: Concatenate the pooled feature vectors from each group to obtain local feature vectors.
8. The method for automatic classification of medical device patents based on deep learning according to claim 1, characterized in that, Step S4 includes the following steps: Step S4-1: Concatenate and fuse the global feature vector and the local feature vector along the feature dimension to obtain the fused feature vector, thus completing the construction of the feature fusion layer; the feature fusion layer sends the fused feature vector to the feature optimization layer; Step S4-2: The feature optimization layer performs layer normalization and random deactivation on the fused feature vector in sequence to accelerate model convergence and prevent overfitting; Step S4-3: The classification mapping layer performs the process of inputting the fused feature vector into the multilayer perceptron classifier; The multilayer perceptron classifier comprises: an input layer, at least one hidden layer, and a result output layer stacked together in sequence; The hidden layer uses the ReLU activation function to perform a non-linear transformation on the input; the output layer uses the Softmax activation function to map the output of the hidden layer to a probability distribution of a preset category. Step S4-4: The result output layer receives the probability distribution of each category and determines the final classification result based on the probability distribution.
9. A classification system for an automatic classification method of medical device patents based on deep learning according to any one of claims 1-8, characterized in that, Includes the following functional modules: The data preprocessing module is used to obtain the raw text data of medical device patents and perform text cleaning and standardization. The automatic annotation module has a built-in large language model and retrieval enhancement generation engine. It introduces medical device classification standards as an external knowledge base to perform label calibration on preprocessed patent data and generate annotated datasets. The data augmentation module has a built-in generative large language model that generates synthetic summaries based on patent claims for categories with insufficient sample size, in order to balance the sample distribution of each category. The model building and training module is used to build and train a deep learning classification model that integrates global and local semantic features. The model includes a first neural network submodule that uses a pre-trained language model to extract global semantic features, a second neural network submodule that uses a multi-size convolutional neural network to extract local semantic features, and a feature fusion and classification submodule that hierarchically fuses global and local features and outputs classification results. The model inference module encapsulates a pre-trained deep learning classification model and provides a model call interface for receiving the patent text to be classified and outputting its corresponding medical device classification standard category and confidence level.
10. The classification system of the automatic classification method for medical device patents based on deep learning according to claim 9, characterized in that, The system adopts a B / S architecture, including the following layers connected from top to bottom: The presentation layer, deployed on the user terminal, is used to provide a human-computer interaction interface, receive the patent text to be classified input by the user and send it to the business logic layer, and receive the classification results returned by the business logic layer and display them to the user in a visual form. The business logic layer, deployed on the application server, communicates with the presentation layer to receive patent text sent by the presentation layer, calls the encapsulated deep learning classification model to perform inference and prediction, and returns the prediction results to the presentation layer. The business logic layer includes: The first service module is built using the Java Spring Boot framework and is responsible for user authentication, request distribution, logging, and data interaction with the presentation layer. The second service module is built using the Python Flask framework. It communicates with the first service module, encapsulates the trained deep learning classification model, and provides a RESTful inference interface to receive patent text and return classification prediction results. The data persistence layer, deployed on the database server, connects to the business logic layer and is used to receive and store user information, historical classification records, operation logs, and system configuration data generated by the business logic layer.