A claim settlement information desensitization collection method, device and system

By combining deep learning and natural language processing technologies with image and graph neural networks, the problems of inaccurate identification and manual verification in existing methods for de-identifying and collecting claims information have been solved, achieving efficient and secure privacy information processing and data protection.

CN117454426BActive Publication Date: 2026-11-10SHANGHAI CHINA HEALTH CONNECT HEALTH TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311549869.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-20
Publication Date
2026-11-10
Estimated Expiration
2043-11-20

AI Technical Summary

Technical Problem

Existing methods for de-identifying and collecting claims data lack the application of deep learning and natural language processing technologies, resulting in inaccurate text recognition and privacy data identification. They also lack scientific and systematic de-identification strategy decisions, and the verification process relies on manual processes, which are inefficient and prone to errors.

Method used

We employ deep learning image and text recognition technology combined with long short-term memory networks for feature extraction and sequence recognition, use a pre-trained model with a Transformers architecture for semantic analysis, support vector machines for risk assessment, image processing methods for privacy information masking, graph neural networks for cross-document association analysis, and combine automated scripts with manual review for verification.

Benefits of technology

It achieves high-precision privacy information labeling and desensitization, ensuring data security and consistency, improving processing efficiency, reducing the risk of human error, and enhancing the protection of customer privacy and regulatory compliance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117454426B_ABST
    Figure CN117454426B_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of data recognition, in particular to a claim settlement information desensitization collection method, device and system, comprising the following steps: based on original insurance claim settlement information, using deep learning image text recognition technology for feature extraction, and combining long short-term memory network for sequence feature recognition to generate text recognition data set.In the present application, through deep learning image text recognition and long short-term memory network, the characteristics of insurance claim settlement information are accurately extracted, and further using Transformers pre-training model realizes high-quality semantic analysis and privacy information labeling, ensures accurate identification of privacy data, and the application of support vector machine provides a scientific and reliable basis for privacy information risk assessment, combined with OpenCV image processing, realizes real-time and efficient privacy data coding, introduces graph neural network technology to simplify cross-document privacy information association, provides comprehensive privacy protection of insurance claim settlement information, which provides multi-dimensional protection for overall privacy protection and ensures data security.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data recognition technology, and in particular to a method, apparatus and system for de-identifying and collecting claims data information. Background Technology

[0002] The field of data identification technology encompasses various techniques and methods for identifying, extracting, and processing different types of data. This can include a variety of data types such as text, images, and audio. A key focus in data identification is how to effectively identify, classify, analyze, and process data to extract useful information or insights. Data anonymization is one important application area, addressing how to collect sensitive data while protecting individual privacy.

[0003] Claims data anonymization is a specific data collection process commonly used in the claims process of insurance businesses. In this context, personal information, health data, and other sensitive information must be collected to assess the legality and extent of insurance claims. However, this information needs to be anonymized to ensure that personal privacy is not violated. Data anonymization refers to replacing sensitive information in the original data with non-sensitive data, such as replacing real names with codes or detailed addresses with geographic coordinates, to protect the privacy of the identified individuals. The main purpose is to protect the privacy of the insured when collecting claims-related information. Anonymization methods ensure that sensitive information is not leaked or misused during data collection and transmission. This helps maintain customer trust, comply with relevant privacy regulations, and reduce the potential risk of data breaches.

[0004] In practical application, existing methods for anonymizing and collecting claims data do not fully utilize advanced technologies such as deep learning and natural language processing, resulting in inaccurate text recognition and privacy data identification. Furthermore, existing methods lack scientific and systematic evaluation methods when making anonymization strategy decisions, easily leading to privacy data leakage or over-anonymization. In cross-document privacy information correlation analysis, advanced technologies such as graph neural networks are not utilized, resulting in room for improvement in the accuracy and efficiency of correlation analysis. Simultaneously, existing verification processes rely too heavily on manual processes, lacking automated verification methods, which increases processing time and introduces errors due to human error. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a method, device and system for de-identifying and collecting claims data information.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for de-identifying and collecting claims data, comprising the following steps:

[0007] S1: Based on the original insurance claim data, deep learning image text recognition technology is used for feature extraction, and long short-term memory network is combined to identify sequence features to generate a text recognition dataset;

[0008] S2: Based on the text recognition dataset, natural language processing algorithms are used to perform semantic analysis through a pre-trained model based on the Transformers architecture, and privacy information is labeled to generate privacy information labeled data;

[0009] S3: Based on the privacy information annotation data, a support vector machine is used to conduct a risk assessment of the privacy information, determine the de-identification level, and generate de-identification strategy decision data;

[0010] S4: Based on the desensitization strategy decision data, an image processing method is used, and image blurring and masking techniques in the OpenCV library are used to perform privacy information decryption and generate preliminary desensitized data;

[0011] S5: Based on the preliminary de-identified data, graph neural network technology is used to link and analyze the information of the same person in different documents to generate cross-document de-identified related data;

[0012] S6: Based on the aforementioned cross-document de-identified associated data, data consistency verification technology is adopted, combined with automated scripts and manual review processes, to verify the de-identification effect and generate a de-identified insurance claim data set;

[0013] The text recognition dataset specifically includes a collection of text content and its location in an image; the privacy information annotation data includes the text, category, and location information of the privacy information in the document; the desensitization strategy decision data specifically refers to the desensitization level and processing method of each piece of privacy information; the preliminary desensitization data is processed claim data, in which privacy information is blurred or covered; and the cross-document desensitization association data includes the association and desensitization status of privacy information of the same individual in multiple claim data.

[0014] As a further aspect of the present invention, based on the original insurance claim data, the steps of using deep learning image text recognition technology for feature extraction and combining it with a long short-term memory network to identify sequence features and generate a text recognition dataset are as follows:

[0015] S101: Based on the original insurance claim data images, a deep convolutional neural network algorithm is used to extract features and generate a preliminary feature map.

[0016] S102: Based on the preliminary feature map, the region proposal network algorithm is used to locate the text region and perform bounding box regression adjustment to generate the text region location result;

[0017] S103: Based on the text region positioning results, optical character recognition technology is used to accurately identify the text content, classify the characters, and generate text content recognition results;

[0018] S104: Based on the text content recognition results, a long short-term memory network is used to optimize the recognition of sequence features and process contextual information to generate a text recognition dataset;

[0019] The deep convolutional neural network specifically uses VGG or ResNet models for deep feature extraction of images. The region proposal network specifically uses RPN in Faster R-CNN to propose image regions containing text. The optical character recognition technology specifically uses the Tesseract OCR framework to recognize and convert text information in images. The long short-term memory network specifically uses the BiLSTM model to handle the temporal dependency problem of text data.

[0020] As a further aspect of the present invention, based on the text recognition dataset, a natural language processing algorithm is used to perform semantic analysis through a pre-trained model based on the Transformers architecture to annotate privacy information and generate privacy information labeled data. The specific steps are as follows:

[0021] S201: Based on the text recognition dataset, a pre-trained model based on the Transformers architecture is used to perform semantic encoding and generate the encoding results, thus generating text semantic encoding results;

[0022] S202: Based on the text semantic encoding result, named entity recognition technology is used to identify and classify entities, and entity classification results are generated to generate entity recognition results;

[0023] S203: Based on the entity recognition results, a text classification algorithm is used to classify privacy levels and generate privacy level classification results.

[0024] S204: Based on the privacy level classification results, apply data desensitization technology to annotate sensitive information and generate privacy information annotation data;

[0025] The pre-trained model of the Transformers architecture specifically uses BERT, GPT, or RoBERTa models for deep semantic understanding of text. The named entity recognition technology includes applying the BiLSTM-CRF model to identify entities in text that include person names and place names. The text classification algorithm specifically refers to using support vector machines or deep neural network algorithms to classify text data. The data desensitization technology specifically applies data masking, data camouflage, or differential privacy methods to process the identified sensitive information.

[0026] As a further aspect of the present invention, the steps of using a support vector machine to assess the risk of privacy information based on the privacy information annotation data, determining the de-identification level, and generating de-identification strategy decision data are as follows:

[0027] S301: Based on the privacy information labeled data, data cleaning and feature encoding algorithms are used to perform data normalization processing, and feature format is converted to generate preprocessed risk assessment data;

[0028] S302: Based on the preprocessed risk assessment data, principal component analysis is used to reduce dimensionality and extract key features to generate filtered key feature data;

[0029] S303: Based on the selected key feature data, the support vector machine algorithm is used to train the model, including using kernel trick and soft margin maximization, to generate a trained SVM model;

[0030] S304: Based on the trained SVM model, perform a risk assessment to determine the sensitivity of the data, and accordingly determine the level of data anonymization, generating data for data anonymization strategy decisions;

[0031] The feature encoding specifically involves converting text data into numerical data that can be processed by machine learning algorithms. The principal component analysis method uses orthogonal transformation to convert correlated variables into a set of linearly uncorrelated variables. The kernel trick specifically maps data to a high-dimensional space. The soft margin maximization allows some data points to be within the margin region or misclassified.

[0032] As a further aspect of the present invention, based on the desensitization strategy decision data, the steps of using image processing methods and employing image blurring and masking techniques from the OpenCV library to perform privacy information decryption and generate preliminary desensitized data are as follows:

[0033] S401: Based on the desensitization strategy decision data, image segmentation technology is used to identify and locate the desensitization area, and generate masking area location data;

[0034] S402: Based on the masking region positioning data, Gaussian blur or mean blur method is used to blur the sensitive region to generate blurred image data;

[0035] S403: Based on the blurred image data, image masking technology is used to cover sensitive information by adding layered masking layers to generate masked image data;

[0036] S404: Based on the masked image data, perform image quality assessment to ensure the desensitization effect while maintaining the clarity of non-sensitive areas, and generate preliminary desensitization data;

[0037] The image segmentation techniques include thresholding and region growing, used to distinguish different objects in an image. The Gaussian blurring specifically involves convolving the image with a Gaussian function. The image masking technique specifically refers to using a binary image to locally mask the target image.

[0038] As a further aspect of the present invention, based on the preliminary de-identified data, the specific steps of using graph neural network technology to link and analyze information about the same person in different documents to generate cross-document de-identified related data are as follows:

[0039] S501: Based on the preliminary desensitized data, a graph neural network is used to extract node embedding features and generate an individual feature representation set;

[0040] S502: Based on the individual feature representation set, a graph convolutional network is used to aggregate neighbor information and generate an enhanced individual connectivity graph;

[0041] S503: Based on the enhanced individual connection graph, a graph attention network is used to perform correlation weight analysis and generate correlation weighted analysis results;

[0042] S504: Based on the aforementioned correlation weighted analysis results, a graph isomorphic network is used to establish information links based on people and generate cross-document de-identified correlation data;

[0043] The individual feature representation set is specifically a vectorized representation of multiple entities in a document, including text content and structural information.

[0044] As a further aspect of the present invention, based on the aforementioned cross-document de-identified associated data, data consistency verification technology is employed, combined with automated scripts and manual review processes, to verify the de-identification effect. The specific steps for generating the de-identified insurance claim data set are as follows:

[0045] S601: Based on the aforementioned cross-document de-identified associated data, data hash matching is used to perform preliminary consistency verification and generate preliminary verification results;

[0046] S602: Based on the preliminary verification results, a comparative analysis technique is used to detect differences and generate a detailed consistency verification report;

[0047] S603: Based on the refined consistency verification report, use an automated script to perform database rule verification and generate automated verification results;

[0048] S604: Based on the automated verification results, a manual review process is adopted to confirm the effect and generate a de-identified insurance claim data set.

[0049] A claims data information anonymization collection device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the claims data information anonymization collection method described above.

[0050] A claims data information anonymization collection system, the claims data information anonymization collection system being used to execute the above-mentioned claims data information anonymization collection method, including a feature extraction module, a semantic recognition module, a data preprocessing module, an anonymization strategy decision module, an image processing module, a correlation analysis module, and a result verification module;

[0051] The data preprocessing module, based on privacy-labeled data, performs data cleaning and feature encoding, and uses principal component analysis to extract key features, generating filtered key feature data.

[0052] The desensitization strategy decision module generates desensitization strategy decision data based on the screened key feature data and using the support vector machine algorithm for model training, risk assessment, and desensitization strategy decision data.

[0053] The image processing module uses desensitization strategy decision data, employs image segmentation technology to locate the desensitization area, and applies Gaussian blur or mean blur for processing. It also uses image masking technology to cover privacy information and generates preliminary desensitization data after image quality assessment.

[0054] The correlation analysis module, based on the preliminary de-identified data, uses graph neural networks for feature extraction, graph convolutional networks for information aggregation, graph attention networks for weight analysis, and graph isomorphic networks to establish information links, generating cross-document de-identified correlation data.

[0055] The result verification module verifies the validity of the de-identification results based on cross-document de-identification association data, uses machine learning classification algorithms for data verification, and uses evaluation index algorithms for performance evaluation to generate de-identified data.

[0056] As a further aspect of the present invention, the feature extraction module includes a feature map generation submodule, a text region localization submodule, a text content recognition submodule, and a sequence feature optimization submodule;

[0057] The semantic recognition module includes a semantic encoding submodule, an entity recognition and classification submodule, a privacy level classification submodule, and a privacy information annotation submodule;

[0058] The data preprocessing module includes a data normalization submodule, a feature transformation submodule, a dimensionality reduction submodule, and a key feature extraction submodule.

[0059] The desensitization strategy decision module includes a model training submodule, a kernel trick application submodule, a soft margin maximization submodule, and a risk assessment submodule.

[0060] The image processing module includes a region localization submodule, a blur processing submodule, a masking application submodule, and a quality assessment submodule.

[0061] The correlation analysis module includes a feature extraction submodule, an information aggregation submodule, a weight analysis submodule, and an information linking submodule;

[0062] The result verification module includes a preliminary verification submodule, a difference detection submodule, a database verification submodule, and a manual review submodule.

[0063] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0064] This invention employs deep learning image-text recognition technology to accurately extract features from original insurance claim data. Combined with a Long Short-Term Memory (LSTM) network, it can accurately identify continuous sequence features. Furthermore, a pre-trained model based on the Transformers architecture is utilized to achieve high-quality semantic analysis and privacy information annotation, ensuring accurate identification of privacy data. The application of Support Vector Machines brings scientific rigor and reliability to privacy information risk assessment, and combined with OpenCV image processing, it provides real-time and efficient privacy data decryption functionality. The introduction of Graph Neural Network technology makes cross-document privacy information association simple and efficient, providing a higher level of protection for the overall privacy of insurance claim data. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0066] Figure 2 This is a detailed flowchart of step S1 of the present invention;

[0067] Figure 3 This is a detailed flowchart of step S2 of the present invention;

[0068] Figure 4 This is a detailed flowchart of step S3 of the present invention;

[0069] Figure 5 This is a detailed flowchart of step S4 of the present invention;

[0070] Figure 6 This is a detailed flowchart of step S5 of the present invention;

[0071] Figure 7 This is a detailed flowchart of step S6 of the present invention;

[0072] Figure 8 This is a system flowchart of the present invention;

[0073] Figure 9 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0075] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0076] Example 1

[0077] Please see Figure 1 This invention provides a technical solution: a method for de-identifying and collecting claims data, comprising the following steps:

[0078] S1: Based on the original insurance claim data, deep learning image text recognition technology is used for feature extraction, and long short-term memory network is combined to identify sequence features to generate a text recognition dataset;

[0079] S2: Based on a text recognition dataset, natural language processing algorithms are used to perform semantic analysis through a pre-trained model based on the Transformers architecture, and privacy information is labeled to generate privacy information labeled data;

[0080] S3: Based on privacy-annotated data, support vector machines are used to assess the risk of privacy information, determine the level of de-identification, and generate de-identification strategy decision data;

[0081] S4: Based on the desensitization strategy decision data, image processing methods are used, and image blurring and masking techniques in the OpenCV library are used to perform privacy information decryption and generate preliminary desensitized data;

[0082] S5: Based on the initial de-identified data, graph neural network technology is used to link and analyze the information of the same person in different documents to generate cross-document de-identified related data;

[0083] S6: Based on cross-document de-identified associated data, data consistency verification technology is adopted, combined with automated scripts and manual review processes to verify the de-identification effect and generate a de-identified insurance claim data set;

[0084] The text recognition dataset specifically includes a collection of text content and its location in an image; the privacy information annotation data includes the text, category, and location information of privacy information in a document; the desensitization strategy decision data specifically refers to the desensitization level and processing method of each piece of privacy information; the preliminary desensitization data is the processed claims data, in which privacy information is blurred or covered; and the cross-document desensitization association data includes the association and desensitization status of privacy information of the same individual in multiple claims data.

[0085] Automating the processing of large volumes of claims data improves efficiency and reduces costs. High-precision deep learning and natural language processing technologies ensure accurate annotation and de-identification of privacy information, effectively protecting sensitive data. Multi-layered de-identification strategies and data association technologies enhance data security, making it difficult to recover sensitive information even across multiple documents. Data consistency verification ensures the accuracy and consistency of de-identified data, reducing the risk of data quality issues. This approach helps companies comply with privacy and data security regulations, mitigating legal risks. Most importantly, through effective privacy protection, companies can enhance customer trust in their data processing methods, increasing customer loyalty.

[0086] Please see Figure 2 Based on original insurance claim data, deep learning image-text recognition technology is used for feature extraction, and a long short-term memory network is combined to identify sequence features, generating a text recognition dataset. The specific steps are as follows:

[0087] S101: Based on the original insurance claim data images, a deep convolutional neural network algorithm is used to extract features and generate a preliminary feature map.

[0088] S102: Based on the preliminary feature map, the region proposal network algorithm is used to locate the text region and perform bounding box regression adjustment to generate the text region location result;

[0089] S103: Based on the text region localization results, optical character recognition technology is used to accurately identify the text content, classify the characters, and generate text content recognition results;

[0090] S104: Based on the text content recognition results, a long short-term memory network is used to optimize the recognition of sequence features and process contextual information to generate a text recognition dataset;

[0091] Deep convolutional neural networks specifically use VGG or ResNet models for deep feature extraction of images; region proposal networks specifically use RPN in Faster R-CNN to propose image regions containing text; optical character recognition technology specifically uses the Tesseract OCR framework to recognize and convert text information in images; and long short-term memory networks specifically use BiLSTM models to handle the temporal dependency problem of text data.

[0092] In step S101, images of the original insurance claim data are imported into the system and stored in digital image format. Next, a deep convolutional neural network (typically VGG or ResNet models) is used to extract features from the images. This process involves multiple convolutional and pooling layers to capture high-level features of the images. The result of feature extraction generates a preliminary feature map, a data structure containing various image features.

[0093] In S102, using the preliminary feature maps, a region proposal network (RPN), typically employed in Faster R-CNN, is used to detect potential regions that may contain text. These region proposals are suggestions for image regions containing text. Boundary regression adjustment is performed on these region proposals to more accurately locate text regions, thereby reducing errors and ensuring the accuracy of text region localization.

[0094] In S103, optical character recognition technology (such as the Tesseract OCR framework) is used to identify each located text region. This involves extracting text content from the image and classifying characters, converting the text content in the image into text data for further processing.

[0095] In S104, based on the results of text content recognition, a Long Short-Term Memory network (such as a BiLSTM model) is used to handle the temporal dependencies of the text data. This includes recognizing sequence features in the text, such as dates, amounts, and names. Contextual information processing is performed to enhance the semantic understanding of the text content. A text recognition dataset is generated, which includes a set of text content and its locations in images, for use in subsequent privacy information annotation and other processing steps.

[0096] Please see Figure 3 Based on a text recognition dataset, natural language processing algorithms are used to perform semantic analysis and annotate privacy information using a pre-trained model based on the Transformers architecture. The specific steps for generating privacy-annotated data are as follows:

[0097] S201: Based on the text recognition dataset, a pre-trained model based on the Transformers architecture is used to perform semantic encoding and generate the encoding results, thus generating text semantic encoding results;

[0098] S202: Based on the text semantic encoding results, named entity recognition technology is used to identify and classify entities, generate entity classification results, and generate entity recognition results;

[0099] S203: Based on the entity recognition results, a text classification algorithm is used to classify privacy levels and generate privacy level classification results.

[0100] S204: Based on the privacy level classification results, apply data desensitization technology to annotate sensitive information and generate privacy information labeled data;

[0101] The pre-trained models of the Transformers architecture specifically use BERT, GPT, or RoBERTa models for deep semantic understanding of text. Named entity recognition technology includes applying the BiLSTM-CRF model to identify entities in text that include names of people and places. Text classification algorithms specifically refer to using support vector machines or deep neural network algorithms to classify text data. Data anonymization technology specifically applies data masking, data camouflage, or differential privacy methods to process the identified sensitive information.

[0102] In S201, a pre-trained model based on the Transformers architecture (such as BERT, GPT, or RoBERTa) is used to perform deep semantic encoding on the text recognition dataset. This means converting the text into high-dimensional semantic vectors, capturing the semantic information of the text, and enabling further processing by a computer. The encoding result of each sample is used as the text semantic code.

[0103] In S202, the model identifies various entities in the text and classifies them into multiple entity categories, such as names of people and places. This generates entity recognition results, which include the location and category information of the entities.

[0104] In S203, combining the text semantic encoding results and entity recognition results, a text classification algorithm, typically a support vector machine or deep neural network, is used to classify the text data. The purpose of classification is to categorize text samples into multiple privacy levels, such as low-risk, medium-risk, or high-risk levels. A privacy level classification result is generated, which includes the privacy level information of the text sample.

[0105] In S204, based on the privacy level classification results, data anonymization techniques, such as data masking, data masquerading, or differential privacy methods, are applied to process the identified privacy information. This includes replacing, deleting, or encrypting sensitive information to ensure privacy protection. Privacy-annotated data is generated, containing anonymized text information, facilitating further processing or sharing without disclosing sensitive information.

[0106] Please see Figure 4 The specific steps for generating desensitization strategy decision data, based on privacy-annotated data, using support vector machines to assess privacy risks, determine the desensitization level, and generate desensitization strategy decision data are as follows:

[0107] S301: Based on privacy-labeled data, data cleaning and feature encoding algorithms are used to standardize the data and convert the feature format to generate preprocessed risk assessment data;

[0108] S302: Based on the preprocessed risk assessment data, principal component analysis is used to reduce dimensionality and extract key features to generate filtered key feature data;

[0109] S303: Based on the selected key feature data, the support vector machine algorithm is used to train the model, including using kernel tricks and soft margin maximization, to generate a trained SVM model;

[0110] S304: Based on the trained SVM model, perform a risk assessment, determine the sensitivity of the data, and decide on the level of data anonymization accordingly, generating data for data anonymization strategy decisions;

[0111] Feature encoding specifically converts text data into numerical data that can be processed by machine learning algorithms. Principal component analysis transforms correlated variables into a set of linearly uncorrelated variables through orthogonal transformation. Kernel trick specifically maps data to a high-dimensional space. Soft margin maximization allows some data points to be in the margin region or misclassified.

[0112] In S301, data cleaning is performed on the privacy-annotated data to ensure data quality. Erroneous or anomalous data will be deleted or corrected. Feature encoding algorithms are used to convert text data into numerical data that can be processed by machine learning algorithms. This includes using bag-of-words models, TF-IDF, or word embedding methods to extract features from the text data and generate preprocessed risk assessment data, including transformed feature vectors and related labels.

[0113] In S302, dimensionality reduction techniques such as Principal Component Analysis (PCA) are used to reduce the dimensionality of the preprocessed data, thereby reducing data complexity. This helps extract the most important features from the data, reduce noise and redundant information, and generate filtered key feature data, which are typically those features that contribute significantly to explaining the variance of the data.

[0114] In S303, the Support Vector Machine (SVM) algorithm is used to train the model on the selected key feature data. Model training involves selecting an appropriate kernel technique, such as a linear kernel, multinomial kernel, or Gaussian kernel, to map the data to a high-dimensional space. A soft margin parameter needs to be set to determine the tolerance of the SVM model, allowing some data points to fall within the margin region or be misclassified, thus improving the model's generalization ability. After training is complete, a trained SVM model is generated, which is used for risk assessment and desensitization level decision-making.

[0115] In S304, a risk assessment is performed based on a pre-trained SVM model, analyzing the input data to determine its sensitivity. This includes calculating the distance between data points and the decision boundary to determine the risk level. Based on the risk assessment results, a desensitization level is determined, such as low risk, medium risk, or high risk. Desensitization strategy decision data is generated, containing suggestions on how to desensitize different data points to ensure that sensitive information is protected.

[0116] Please see Figure 5 Based on the desensitization strategy decision data, and using image processing methods, including image blurring and masking techniques from the OpenCV library, the specific steps for generating preliminary desensitized data are as follows:

[0117] S401: Based on the desensitization strategy decision data, image segmentation technology is used to identify and locate the desensitization area, and generate masking area location data;

[0118] S402: Based on the masked region positioning data, Gaussian blur or mean blur method is used to blur the sensitive region to generate blurred image data;

[0119] S403: Based on the blurred image data, image masking technology is used to cover sensitive information by adding layered masking layers to generate masked image data;

[0120] S404: Based on the masked image data, perform image quality assessment to ensure the desensitization effect while maintaining the clarity of non-sensitive areas, and generate preliminary desensitization data;

[0121] Image segmentation techniques include thresholding and region growing, used to distinguish different objects in an image; Gaussian blurring specifically involves convolving an image with a Gaussian function; and image masking techniques specifically refer to using a binary image to locally occlude a target image.

[0122] In S401, image segmentation techniques are used to identify and locate desensitized regions. Here are some commonly used image segmentation methods:

[0123] Thresholding segmentation: Using a threshold to divide an image into different regions, for example, dividing pixels in an image into foreground and background.

[0124] import cv2

[0125] #Read image

[0126] image=cv2.imread('input_image.jpg',0)

[0127] # Use threshold segmentation

[0128] ret,binary_image=cv2.threshold(image,threshold_value,max_value,cv2.THRESH_BINARY)

[0129] Region growing: Starting from a seed pixel, adjacent pixels are merged into a region based on the similarity between pixels.

[0130] import cv2

[0131] #Read image

[0132] image=cv2.imread('input_image.jpg',0)

[0133] #Regional Growth

[0134] seed_point = (x, y)

[0135] connected_regions=cv2.floodFill(image,mask,seed_point,newVal)

[0136] Generate masking area location data, indicating which parts need to be desensitized.

[0137] In S402, based on the masked area location data, Gaussian blurring or mean blurring methods are used to blur the privacy area. This helps to hide sensitive information.

[0138] import cv2

[0139] #Read image

[0140] image=cv2.imread('input_image.jpg')

[0141] #Gaussian Blur

[0142] blurred_image=cv2.GaussianBlur(image,(ksize_x,ksize_y),sigmaX)

[0143] #Mean Fuzzy

[0144] blurred_image=cv2.blur(image,(ksize_x,ksize_y))

[0145] In S403, image masking technology is used to add a masking layer to cover privacy information. A binary image is created, the areas that need to be desensitized are marked as 1, and then this image is multiplied by the original image to achieve the masking effect.

[0146] import cv2

[0147] import numpy as np

[0148] #Read image

[0149] image=cv2.imread('input_image.jpg')

[0150] # Create a masking image

[0151] mask=np.zeros(image.shape[:2],dtype=np.uint8)

[0152] #Set the areas that need to be desensitized

[0153] cv2.rectangle(mask,(x1,y1),(x2,y2),1,-1)

[0154] #Apply Mask

[0155] masked_image=cv2.bitwise_and(image,image,mask=mask)

[0156] In S404, image quality assessment is performed to ensure effective desensitization and maintain the clarity of non-privacy areas. Image quality assessment metrics such as PSNR (Peak Signal-to-Noise Ratio) or SSIM (Structural Similarity Index) can be used to evaluate image quality.

[0157] import cv2

[0158] from skimage.metrics import structural_similarity as ssim

[0159] #Read the original image and the anonymized image

[0160] original_image=cv2.imread('original_image.jpg')

[0161] masked_image=cv2.imread('masked_image.jpg')

[0162] #Calculate SSIM

[0163] ssim_score=ssim(original_image,masked_image)

[0164] #Calculate PSNR

[0165] psnr=cv2.PSNR(original_image,masked_image)

[0166] Please see Figure 6 Based on preliminary anonymized data, graph neural network technology is used to link and analyze information about the same person in different documents, generating cross-document anonymized related data. The specific steps are as follows:

[0167] S501: Based on the preliminary desensitized data, a graph neural network is used to extract node embedding features and generate an individual feature representation set;

[0168] S502: Based on the individual feature representation set, a graph convolutional network is used to aggregate neighbor information and generate an enhanced individual connectivity graph;

[0169] S503: Based on the enhanced individual connectivity graph, a graph attention network is used to perform correlation weight analysis and generate correlation weighted analysis results;

[0170] S504: Based on the results of the correlation weighted analysis, a graph isomorphic network is used to establish information links based on people and generate cross-document de-identified correlation data;

[0171] The individual feature representation set is specifically the vectorized representation of multiple individuals in a document, including text content and structural information.

[0172] In S501, a graph neural network is used to extract node embedding features and generate an individual feature representation set. To this end, individual information in the document is converted into vector representations, and word embedding techniques are used to convert text content and structural information into numerical features. A node is created for each individual, and the node embedding algorithm of the graph neural network is used to extract the node's feature embeddings, generating an individual feature representation set.

[0173] In S502, neighbor information is aggregated based on the individual feature representation set, and an enhanced individual connectivity graph is generated. This involves constructing a graph data structure where each node represents an individual in a document, and edges represent the relationships between them. Then, graph convolutional network technology is used to aggregate the neighbor information of the nodes to obtain richer information and generate the enhanced individual connectivity graph.

[0174] In S503, correlation weight analysis is performed and correlation weighted analysis results are generated. This involves using methods such as graph attention networks to analyze correlation weights, emphasizing nodes with high correlation. By calculating the correlation weights between nodes, correlation weighted analysis results are generated, which include correlation weight information between nodes, reflecting the connections between people in different documents.

[0175] In S504, information links based on persons are established based on the results of relevance weighted analysis, and cross-document anonymized relational data is generated. This includes linking nodes representing the same person and matching them based on relevance weights. The final generated cross-document anonymized relational data includes the linked person information and the relevance weights between them, which can be used for further analysis, such as cross-document person relationship analysis and privacy anonymization processing.

[0176] Please see Figure 7 Based on cross-document de-identified associated data, data consistency verification technology is used, combined with automated scripts and manual review processes, to verify the de-identification effect and generate a de-identified insurance claims data set. The specific steps are as follows:

[0177] S601: Based on cross-document de-identified associated data, data hash matching is used to perform preliminary consistency verification and generate preliminary verification results;

[0178] S602: Based on the preliminary verification results, comparative analysis technology is used to detect differences and generate a detailed consistency verification report;

[0179] S603: Based on the refined consistency verification report, use automated scripts to perform database rule verification and generate automated verification results;

[0180] S604: Based on the automated proofreading results, a manual review process is adopted to confirm the effect and generate a de-identified insurance claims data set.

[0181] In S601, cross-document de-identified associated data is used for preliminary data consistency verification, employing data hash matching to generate preliminary verification results. The operation is as follows:

[0182] A data hash value is generated to represent each document or dataset. An initial validation is performed, comparing the data hash values ​​to verify data consistency across different documents. Preliminary validation results are generated, indicating which documents or datasets passed the consistency check and which failed.

[0183] In S602, based on the preliminary verification results, comparative analysis technology is used to detect differences and generate a detailed consistency verification report. The operation is as follows:

[0184] For documents or datasets that fail the initial validation, use comparative analysis tools to compare their content and detect differences. Generate a detailed consistency validation report containing detailed information on the differences, facilitating further processing.

[0185] In S603, based on the refined consistency verification report, an automated script is used to perform database rule verification and generate automated verification results. The operation is as follows:

[0186] Using automated scripts, documents or datasets are checked against predetermined database rules based on a refined consistency verification report. Automated verification results are generated, indicating which documents or datasets conform to the rules and which do not.

[0187] In S604, a manual review process is performed to confirm the effectiveness of the anonymization and generate an anonymized insurance claim data set. The operation is as follows:

[0188] Professional staff conduct manual reviews to examine documents or datasets that do not meet the rules, ensuring that the anonymization effect meets the requirements. Based on the results of the manual review, an anonymized insurance claims dataset is generated, which includes documents or datasets that have passed the proofreading and review, and confirms that the anonymization effect has been achieved.

[0189] A claims data information anonymization collection device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the claims data information anonymization collection method described above.

[0190] Please see Figure 8A claims data information anonymization collection system is used to execute the above-mentioned claims data information anonymization collection method, including a feature extraction module, a semantic recognition module, a data preprocessing module, an anonymization strategy decision module, an image processing module, a correlation analysis module, and a result verification module;

[0191] The feature extraction module is based on the original insurance claim data images. It uses a deep convolutional neural network algorithm to extract features and uses optical character recognition technology and long short-term memory network to perform text recognition and optimization, generating a text recognition dataset.

[0192] The semantic recognition module receives the text recognition dataset, uses a pre-trained model based on the Transformers architecture and named entity recognition technology to perform semantic encoding and entity authentication, adopts a text classification algorithm to classify privacy levels, and applies data anonymization technology to label the data to obtain privacy information labeled data.

[0193] The data preprocessing module uses privacy-labeled data, cleans and encodes the data, and uses principal component analysis to extract key features, generating filtered key feature data.

[0194] The desensitization strategy decision module is based on the screened key feature data and uses the support vector machine algorithm for model training, risk assessment, and generates desensitization strategy decision data.

[0195] The image processing module uses desensitization strategy decision data, employs image segmentation technology to locate desensitized areas, and applies Gaussian blur or mean blur for processing. It also uses image masking technology to cover privacy information and generates preliminary desensitized data after image quality assessment.

[0196] The correlation analysis module, based on the initial de-identified data, uses graph neural networks for feature extraction, graph convolutional networks for information aggregation, graph attention networks for weight analysis, and graph isomorphic networks to establish information links, generating cross-document de-identified correlation data.

[0197] The results verification module verifies the validity of the de-identification results based on cross-document de-identification association data, uses machine learning classification algorithms for data verification, and employs evaluation index algorithms for performance evaluation to generate de-identified data.

[0198] Enhanced data security is achieved through the efficient and accurate labeling of private information using deep convolutional neural networks and text recognition technology, effectively preventing data leaks. Automated feature extraction and information recognition reduce manual intervention, improving processing efficiency and lowering costs. Data preprocessing ensures data consistency and accuracy, making the anonymized data more trustworthy. The system also provides personalized anonymization strategies, using machine learning models to develop customized strategies for different data types, balancing privacy protection and data usability. Image processing and quality assessment guarantee the quality of anonymized images, maintaining data usability. Cross-document anonymization and data correlation facilitate comprehensive analysis of claims data, while result verification and performance evaluation improve data quality and system effectiveness.

[0199] Please see Figure 9 The feature extraction module includes a feature map generation submodule, a text region localization submodule, a text content recognition submodule, and a sequence feature optimization submodule;

[0200] The semantic recognition module includes a semantic encoding submodule, an entity recognition and classification submodule, a privacy level classification submodule, and a privacy information annotation submodule;

[0201] The data preprocessing module includes a data normalization submodule, a feature transformation submodule, a dimensionality reduction submodule, and a key feature extraction submodule;

[0202] The desensitization strategy decision-making module includes a model training submodule, a kernel trick application submodule, a soft margin maximization submodule, and a risk assessment submodule;

[0203] The image processing module includes a region localization submodule, a blur processing submodule, a masking application submodule, and a quality assessment submodule;

[0204] The correlation analysis module includes a feature extraction submodule, an information aggregation submodule, a weight analysis submodule, and an information linking submodule;

[0205] The results verification module includes a preliminary verification submodule, a difference detection submodule, a database verification submodule, and a manual review submodule.

[0206] The system uses a feature map generation submodule to extract features from original insurance claim data images using a deep convolutional neural network algorithm. The text region localization submodule uses image processing techniques to locate text regions, providing accurate text recognition for the text content recognition submodule. The text content recognition submodule utilizes optical character recognition (OCR) technology and a long short-term memory network to identify and optimize text regions, generating a text recognition dataset. The sequence feature optimization submodule further processes and optimizes the text recognition data, improving the accuracy and quality of text recognition.

[0207] The semantic recognition module receives the text recognition dataset and processes it further. The semantic encoding submodule uses a pre-trained model based on the Transformers architecture to semantically encode the text recognition data, capturing the semantic information of the text. Subsequently, the entity recognition and classification submodule uses named entity recognition technology to identify and classify entities in the text. The privacy level classification submodule uses a text classification algorithm to classify privacy levels and determine the information that needs to be anonymized. Finally, the privacy information annotation submodule annotates the privacy information in the text, preparing for subsequent anonymization processing.

[0208] The data preprocessing module further processes the labeled privacy information. The data normalization submodule normalizes the labeled data to ensure consistency and accuracy. The feature transformation submodule encodes and transforms features, preparing for subsequent data analysis and model training. The dimensionality reduction submodule uses principal component analysis to reduce data dimensionality, decreasing complexity and improving processing efficiency. The key feature extraction submodule extracts key features from the processed data, providing crucial information for subsequent model training and desensitization strategy decisions.

[0209] The desensitization strategy decision module is responsible for formulating the desensitization strategy. The model training submodule trains the model using a support vector machine algorithm based on the selected key feature data to determine the desensitization strategy. The kernel trick application submodule improves the model's performance and generalization ability, while the soft margin maximization submodule ensures the model can handle imperfect data. The risk assessment submodule evaluates the model's performance and determines the risks and effectiveness of the desensitization strategy.

[0210] The image processing module is responsible for handling the desensitization of images. The region localization submodule uses desensitization strategy decision data and image segmentation techniques to locate the desensitized regions. The blurring submodule applies techniques such as Gaussian blur or mean blur to blur the image information in the desensitized regions. The masking application submodule uses image masking techniques to cover privacy information, ensuring that it is not visible. The quality assessment submodule evaluates the quality of the desensitized image to ensure that the desensitization process does not affect the image's usability.

[0211] The correlation analysis module handles the correlations between different documents. The feature extraction submodule extracts features from the initially anonymized data for subsequent information correlation analysis. The information aggregation submodule uses graph convolutional networks to aggregate relevant information and establish connections between them. The weight analysis submodule uses graph attention networks to perform weight analysis and determine the importance of information correlations. The information linking submodule uses graph isomorphic networks to establish information links and generate cross-document anonymized correlation data.

[0212] The results verification module validates the anonymization results. The preliminary verification submodule performs initial verification of cross-document anonymized related data to ensure data consistency and validity. The difference detection submodule uses machine learning classification algorithms to verify the data, detecting differences and anomalies. The database verification submodule verifies the anonymized data against the database to ensure consistency between the anonymized results and the original data. The manual review submodule performs manual review to ensure the accuracy and compliance of the anonymized results.

[0213] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for anonymizing and collecting claims data, characterized in that, Includes the following steps: Based on the original insurance claims data, deep learning image text recognition technology is used for feature extraction, and long short-term memory network is combined to recognize sequence features, generating a text recognition dataset; Based on the text recognition dataset, natural language processing algorithms are used to perform semantic analysis through a pre-trained model based on the Transformers architecture, and privacy information is labeled to generate privacy information labeled data. Based on the privacy information labeled data, a support vector machine is used to assess the risk of privacy information, determine the level of de-identification, and generate de-identification strategy decision data; Based on the desensitization strategy decision data, an image processing method is used, employing image blurring and masking techniques from the OpenCV library, to perform privacy information decryption and generate preliminary desensitized data; Based on the preliminary de-identified data, graph neural network technology is used to link and analyze the information of the same person in different documents to generate cross-document de-identified related data. Based on the aforementioned cross-document de-identified associated data, data consistency verification technology is used, combined with automated scripts and manual review processes, to verify the de-identification effect and generate a de-identified insurance claim data set; The text recognition dataset specifically includes a collection of text content and its location in an image; the privacy information annotation data includes the text, category, and location information of the privacy information in the document; the desensitization strategy decision data specifically refers to the desensitization level and processing method of each piece of privacy information; the preliminary desensitization data is processed claim data, in which privacy information is blurred or covered; and the cross-document desensitization association data includes the association and desensitization status of privacy information of the same individual in multiple claim data.

2. The method for anonymizing and collecting claims data according to claim 1, characterized in that, Based on original insurance claim data, the following steps are taken to generate a text recognition dataset: Deep learning image-text recognition technology is used for feature extraction, and a long short-term memory network is combined to identify sequence features. Based on the original insurance claim data images, a deep convolutional neural network algorithm is used to extract features and generate a preliminary feature map. Based on the preliminary feature map, a region proposal network algorithm is used to locate the text region and perform bounding box regression adjustment to generate the text region location result. Based on the text region localization results, optical character recognition technology is used to accurately identify the text content, classify the characters, and generate text content recognition results. Based on the text content recognition results, a long short-term memory network is used to optimize the recognition of sequence features and process contextual information to generate a text recognition dataset. The deep convolutional neural network specifically uses VGG or ResNet models for deep feature extraction of images. The region proposal network specifically uses RPN in Faster R-CNN to propose image regions containing text. The optical character recognition technology specifically uses the Tesseract OCR framework to recognize and convert text information in images. The long short-term memory network specifically uses the BiLSTM model to handle the temporal dependency problem of text data.

3. The method for de-identifying and collecting claims data information according to claim 1, characterized in that, Based on the aforementioned text recognition dataset, the specific steps for generating privacy-annotated data, using natural language processing algorithms and a pre-trained model based on the Transformers architecture to perform semantic analysis and annotate privacy information, are as follows: Based on the text recognition dataset, a pre-trained model based on the Transformers architecture is used to perform semantic encoding and generate the encoding results, thus generating text semantic encoding results. Based on the text semantic encoding results, named entity recognition technology is used to identify and classify entities, generate entity classification results, and generate entity recognition results. Based on the entity recognition results, a text classification algorithm is used to classify privacy levels and generate privacy level classification results. Based on the privacy level classification results, data anonymization technology is applied to annotate sensitive information and generate privacy information annotation data; The pre-trained model of the Transformers architecture specifically uses BERT, GPT, or RoBERTa models for deep semantic understanding of text. The named entity recognition technology includes applying the BiLSTM-CRF model to identify entities in text that include person names and place names. The text classification algorithm specifically refers to using support vector machines or deep neural network algorithms to classify text data. The data desensitization technology specifically applies data masking, data camouflage, or differential privacy methods to process the identified sensitive information.

4. The method for de-identifying and collecting claims data information according to claim 1, characterized in that, Based on the privacy information labeled data, the steps of using support vector machines to assess the risk of privacy information, determine the de-identification level, and generate de-identification strategy decision data are as follows: Based on the privacy information labeled data, data cleaning and feature encoding algorithms are used to perform data standardization processing and convert feature formats to generate preprocessed risk assessment data. Based on the preprocessed risk assessment data, principal component analysis is used to reduce dimensionality and extract key features to generate filtered key feature data. Based on the selected key feature data, the support vector machine algorithm is used to train the model, including using kernel trick and soft margin maximization, to generate a trained SVM model. Based on the trained SVM model, a risk assessment is performed to determine the sensitivity of the data, and the level of data anonymization is determined accordingly, generating data for data anonymization strategy decisions. The feature encoding specifically involves converting text data into numerical data that can be processed by machine learning algorithms. The principal component analysis method uses orthogonal transformation to convert correlated variables into a set of linearly uncorrelated variables. The kernel trick specifically maps data to a high-dimensional space. The soft margin maximization allows some data points to be within the margin region or misclassified.

5. The method for de-identifying and collecting claims data information according to claim 1, characterized in that, Based on the aforementioned desensitization strategy decision data, the steps for generating preliminary desensitized data by employing image processing methods and utilizing image blurring and masking techniques from the OpenCV library to perform privacy information decryption are as follows: Based on the desensitization strategy decision data, image segmentation technology is used to identify and locate the desensitization area, and generate masking area location data. Based on the masked area positioning data, Gaussian blur or mean blur method is used to blur the sensitive area to generate blurred image data. Based on the blurred image data, an image masking technique is used to cover sensitive information by adding a layered masking layer to generate masked image data. Based on the masked image data, image quality assessment is performed to ensure the desensitization effect while maintaining the clarity of non-sensitive areas, and preliminary desensitization data is generated. The image segmentation techniques include thresholding and region growing, used to distinguish different objects in an image. The Gaussian blurring specifically involves convolving the image with a Gaussian function. The image masking technique specifically refers to using a binary image to locally mask the target image.

6. The method for anonymizing and collecting claims data according to claim 1, characterized in that, Based on the preliminary de-identified data, the specific steps for generating cross-document de-identified related data by using graph neural network technology to link and analyze information about the same person in different documents are as follows: Based on the preliminary desensitized data, a graph neural network is used to extract node embedding features and generate an individual feature representation set. Based on the individual feature representation set, a graph convolutional network is used to aggregate neighbor information and generate an enhanced individual connectivity graph. Based on the enhanced individual connectivity graph, a graph attention network is used to perform correlation weight analysis and generate correlation weighted analysis results. Based on the aforementioned correlation weighted analysis results, a graph isomorphic network is used to establish information links based on people and generate cross-document anonymized correlation data. The individual feature representation set is specifically a vectorized representation of multiple entities in a document, including text content and structural information.

7. The method for anonymizing and collecting claims data according to claim 1, characterized in that, Based on the aforementioned cross-document de-identified associated data, using data consistency verification technology, combined with automated scripts and manual review processes, the steps for verifying the de-identification effect and generating a de-identified insurance claim data set are as follows: Based on the aforementioned cross-document de-identified associated data, data hash matching is used to perform preliminary consistency verification and generate preliminary verification results; Based on the preliminary verification results, a comparative analysis technique was used to detect differences and generate a detailed consistency verification report. Based on the refined consistency verification report, an automated script is used to perform database rule verification and generate automated verification results. Based on the automated verification results, a manual review process is adopted to confirm the effectiveness and generate a de-identified insurance claims data set.

8. A device for de-identifying and collecting claims data, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the claims data information desensitization collection method according to any one of claims 1 to 7.

9. A claims data information anonymization collection system, characterized in that, The claims data information anonymization collection system is used to execute the claims data information anonymization collection method according to any one of claims 1-7. The claims data information anonymization collection system includes a feature extraction module, a semantic recognition module, a data preprocessing module, an anonymization strategy decision module, an image processing module, a correlation analysis module, and a result verification module. The feature extraction module is based on the original insurance claim data images, uses a deep convolutional neural network algorithm to extract features, and uses optical character recognition technology and long short-term memory network to perform text recognition and optimization, generating a text recognition dataset. The semantic recognition module receives a text recognition dataset, uses a pre-trained model based on the Transformers architecture and named entity recognition technology for semantic encoding and entity authentication, employs a text classification algorithm to classify privacy levels, and applies data anonymization technology for labeling to obtain privacy information labeled data. The data preprocessing module, based on privacy-labeled data, performs data cleaning and feature encoding, and uses principal component analysis to extract key features, generating filtered key feature data. The desensitization strategy decision module generates desensitization strategy decision data based on the screened key feature data and using the support vector machine algorithm for model training, risk assessment, and desensitization strategy decision data. The image processing module uses desensitization strategy decision data, employs image segmentation technology to locate the desensitization area, and applies Gaussian blur or mean blur for processing. It also uses image masking technology to cover privacy information and generates preliminary desensitization data after image quality assessment. The correlation analysis module, based on the preliminary de-identified data, uses graph neural networks for feature extraction, graph convolutional networks for information aggregation, graph attention networks for weight analysis, and graph isomorphic networks to establish information links, generating cross-document de-identified correlation data. The result verification module verifies the validity of the de-identification results based on cross-document de-identification association data, uses machine learning classification algorithms for data verification, and uses evaluation index algorithms for performance evaluation to generate de-identified data.

10. The claims data information anonymization collection system according to claim 9, characterized in that, The feature extraction module includes a feature map generation submodule, a text region localization submodule, a text content recognition submodule, and a sequence feature optimization submodule; The semantic recognition module includes a semantic encoding submodule, an entity recognition and classification submodule, a privacy level classification submodule, and a privacy information annotation submodule; The data preprocessing module includes a data normalization submodule, a feature transformation submodule, a dimensionality reduction submodule, and a key feature extraction submodule. The desensitization strategy decision-making module includes a model training submodule, a kernel trick application submodule, a soft margin maximization submodule, and a risk assessment submodule. The image processing module includes a region localization submodule, a blur processing submodule, a masking application submodule, and a quality assessment submodule. The correlation analysis module includes a feature extraction submodule, an information aggregation submodule, a weight analysis submodule, and an information linking submodule; The result verification module includes a preliminary verification submodule, a difference detection submodule, a database verification submodule, and a manual review submodule.

Citation Information

Patent Citations

  • Form information extraction system and method with privacy protection

    CN113221181A

  • Claim settlement data processing method and device, electronic equipment and readable medium

    CN113792570A