A method, device, equipment and medium for dynamically pushing claim documents

Abstract

The present application relates to the field of artificial intelligence technology, which can be applied to medical health and financial technology business system platforms, and discloses a kind of methods for dynamic pushing of claim documents, comprising: collecting multi-modal data, performing multi-dimensional feature extraction on the multi-modal data, and generating multi-modal document features;The multi-modal data and multi-modal document features are used to jointly train the claim document demand prediction model;The multi-modal document features are input into the trained claim document demand prediction model for prediction, and the claim document demand list is output;According to the current claim link, user portrait data and claim document demand list, a personalized claim document list is generated, and the personalized claim document list is pushed to the client.The present application uses multi-dimensional data for collection, extracts multi-dimensional features, jointly trains the prediction model, and outputs the claim document demand list.Combining the claim link with the user portrait to generate a personalized list, the efficiency of claim settlement is improved, and the cost is optimized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for dynamically pushing claims documents. Background Technology

[0002] The current document processing workflow in the insurance claims industry suffers from numerous technical bottlenecks. These combined issues not only severely restrict claims efficiency but also significantly impact service quality. On one hand, the traditional manual document processing model is highly reliant on manpower. Users must manually upload claims documents, and claims personnel then sequentially complete the verification, classification, and review processes. This entire process is time-consuming and labor-intensive, and highly susceptible to delays or errors due to human negligence. On the other hand, the problem of data silos is prominent. Hospital information systems and insurance company claims systems operate independently, with inconsistent data standards, making seamless integration impossible. Although some companies have attempted to introduce OCR technology to automate document recognition, general OCR models lack adaptability to specialized documents such as medical invoices. When faced with the varying invoice formats across different hospitals nationwide, recognition accuracy drops significantly.

[0003] The lack of flexibility in existing intelligent claims systems makes them ill-suited for complex and ever-changing claims scenarios. These systems are mostly driven by rule engines, requiring the pre-configuration of numerous rigid rules. The types of claims documents are diverse, and the required document combinations vary greatly from case to case, directly leading to high system maintenance costs and difficulties in iterative updates. Meanwhile, the low accuracy of document delivery is also a key drawback. Existing systems generally use a static document list delivery model, unable to dynamically adjust the content based on the user's actual claims progress and case characteristics. Users often have to repeatedly submit supplementary materials due to missing or incorrect documents, resulting in an average case processing time extension of 3-5 days.

[0004] Furthermore, the technology modules are severely fragmented; computer vision, natural language processing, knowledge graphs, and other related technologies often operate independently, lacking the collaborative integration of multi-dimensional machine learning algorithms. The combination of these problems ultimately results in the current inefficiency of insurance claims document processing.

[0005] In the healthcare sector, core technical challenges lie in the insufficient coordination of medical data flow and analysis, as well as weak privacy and security protection. On one hand, data silos are prominent, with inconsistent data standards between hospital information systems and external systems, lacking unified interaction interfaces and interoperability protocols. This hinders the seamless transfer of critical medical data such as medical bills and discharge summaries, failing to provide efficient data support for cross-domain operations like claims processing. On the other hand, specialized medical document analysis technologies lack adaptability. General recognition models struggle to handle complex scenarios such as varying bill formats, specialized medical terminology, and handwritten prescriptions across different hospitals. Furthermore, the disconnect between bill recognition and medical semantic analysis prevents the accurate extraction of core clinical information like past medical history, hindering the effective value transformation of medical data. In addition, the application of privacy-preserving computing technologies is inadequate. During cross-institutional sharing and joint analysis of medical data, it fails to meet the stringent protection requirements of regulations such as the Data Security Law for sensitive patient information, while also posing a risk of privacy breaches, thus restricting the compliant flow and in-depth utilization of medical data.

[0006] In the fintech business sector, technical issues primarily manifest in the insufficient adaptability of intelligent claims systems to various scenarios, inadequate technical collaboration efficiency, and insufficient compliance and risk control levels. Firstly, the intelligent systems lack flexibility. Traditional rule-based architectures require numerous pre-defined hard rules, making it difficult to adapt to complex scenarios with diverse claims document types and varying case combinations. This results in high system maintenance costs and difficulties in iteration. Simultaneously, dynamic technical support for document push is insufficient; static push modes cannot match user claims progress with case characteristics, leading to issues such as missing or incorrect document transmission, extending business processing cycles. Secondly, technical modules are severely fragmented. Key technologies such as computer vision and natural language processing operate independently, lacking the collaborative integration of multi-dimensional algorithms. This prevents the formation of end-to-end technical support from document recognition and information extraction to claims decision-making, hindering the improvement of automation and intelligence levels. Finally, the data compliance and security technology system is incomplete. When introducing external medical data for risk analysis, the application maturity of security technologies such as privacy computing is insufficient. This not only fails to address the industry pain point of "data usable but not visible" but also faces compliance risks such as cross-border data flow and sensitive information protection, affecting the security and credibility of fintech services. Summary of the Invention

[0007] The main objective of this invention is to provide a method, apparatus, device, and storage medium for dynamic push of claims documents, aiming to solve the technical problems of high reliance on manual claims processing and low accuracy of document push.

[0008] To achieve the above objectives, the present invention provides a method for dynamically pushing claims documents, comprising: Collect multimodal data, extract multidimensional features from the multimodal data, and generate multimodal document features; The document demand prediction model is jointly trained using the multimodal data and multimodal document features. The multimodal document features are input into the trained document demand prediction model for prediction, and a document demand list is output. A personalized document list is generated based on the current claims process, user profile data, and document requirement list, and then pushed to the client.

[0009] Furthermore, to achieve the above objectives, the present invention provides a dynamic push device for claims documents, comprising: The data processing module is used to collect multimodal data, extract multi-dimensional features from the multimodal data, and generate multimodal document features; The model training module is used to jointly train the document demand prediction model using the multimodal data and multimodal document features; The prediction list module is used to input the multimodal document features into the trained document demand prediction model for prediction and output a document demand list. The list push module is used to generate a personalized list of documents based on the current claims process, user profile data, and document requirement list, and push the personalized list of documents to the client.

[0010] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a claims document dynamic push program stored in the memory and executable on the processor, wherein when the claims document dynamic push program is executed by the processor, it implements the steps of the claims document dynamic push method as described above.

[0011] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a claims document dynamic push program, wherein the claims document dynamic push program, when executed by a processor, implements the steps of the claims document dynamic push method as described above.

[0012] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business system platforms such as healthcare and fintech. It discloses a method for dynamically pushing claims documents, comprising: collecting multimodal data; extracting multi-dimensional features from the multimodal data to generate multimodal document features; jointly training a document demand prediction model using the multimodal data and the multimodal document features; inputting the multimodal document features into the trained document demand prediction model for prediction, and outputting a document demand list; generating a personalized document list based on the current claims process, user profile data, and the document demand list, and pushing the personalized document list to the client. This invention, through collecting multi-dimensional data, extracting multi-dimensional features, jointly training a prediction model, and outputting a document demand list, combined with the claims process and user profile to generate a personalized list for push, achieves improved claims efficiency and cost optimization. Attached Figure Description

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for a method of dynamically pushing claims documents according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the method for dynamically pushing claims documents according to the present invention. Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the claims document dynamic push device of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0015] The method for dynamically pushing claims documents provided in this embodiment of the invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server terminal via a network. The server terminal can collect multimodal data from the user terminal, extract multi-dimensional features from the multimodal data, and generate multimodal document features; it then jointly trains a document demand prediction model using the multimodal data and multimodal document features; the multimodal document features are input into the trained document demand prediction model for prediction, and a document demand list is output; a personalized document list is generated based on the current claims process, user profile data, and the document demand list, and this personalized document list is pushed to the client. This invention uses multi-dimensional data collection, extracts multi-dimensional features, jointly trains a prediction model, and outputs a document demand list. Combining the claims process with user profiles to generate and push personalized lists improves claims efficiency and optimizes costs. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server terminal can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0016] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the method for dynamically pushing claims documents provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0017] like Figure 2 As shown, the method for dynamically pushing claims documents proposed in this invention includes the following steps: S100. Collect multimodal data, extract multidimensional features from the multimodal data, and generate multimodal document features; S200. Jointly train the document demand prediction model using the multimodal data and multimodal document features; S300. Input the multimodal document features into the trained document demand prediction model for prediction and output a document demand list. S400: Generate a personalized document list based on the current claims process, user profile data, and document requirement list, and push the personalized document list to the client.

[0018] In this embodiment, multimodal data collection encompasses document images (such as medical invoices, diagnostic certificates, and ID cards) captured and encrypted by users in real-time via mobile applications, verifiable medical institution data obtained using blockchain technology, and historical claims data containing fields such as case type, document list, and processing time. During the collection process, all sensitive data is processed using privacy-preserving computation techniques. Secure computation operators within a relevant technical framework are used to classify and categorize the data, enabling joint analysis without the original data leaving the local machine. For example, medical data is converted into ciphertext format that only authorized nodes can participate in the computation through secure encryption algorithms.

[0019] Multi-dimensional feature extraction employs corresponding technologies for different types of data: For document images, convolutional neural networks (CNNs) are used to extract visual features, which can automatically identify key fields such as serial numbers, names, and amounts on medical invoices, as well as medical entity terms in non-standard documents such as discharge summaries, supporting high-accuracy recognition of more than 100 types of medical vouchers nationwide; with the help of natural language processing (NLP) technology, the BERT model (a model that can normalize medical terms) is used to map medical terms in medical record texts to standard ICD codes (International Classification of Diseases, used to unify the coding standards for medical information such as diseases), and a knowledge graph of diseases-drugs-treatment items is constructed; at the same time, basic information such as insurance type, cause of claim, and level of medical institution in claim cases is analyzed, and converted into feature vectors through an embedding layer, and finally fused to form multimodal document features.

[0020] The document demand prediction model is trained based on multimodal document features. It employs a graph neural network (GNN, a neural network capable of handling structured relational data) combined with an attention mechanism. Claims cases are abstracted into a heterogeneous graph structure containing entities such as patients, medical institutions, and document types, as well as the relationships between these entities. Neighborhood information is aggregated through a message passing mechanism, and the importance weight of each document is dynamically calculated. The training process uses historical claims data (including case features and the final set of documents used) as the training set. The model parameters are optimized with the goal of minimizing document recommendation error. Furthermore, a federated learning framework (a technical framework that enables joint training from multiple data sources without sharing the original data) is introduced to support cross-institutional collaborative training.

[0021] By inputting multimodal document features into the trained model, a document requirement list can be output. The system combines the user's current claims process (submission, review, payment, etc.) with user profile data (such as age) to generate a personalized document list. For example, for car insurance cases, driver's licenses and traffic accident liability certificates are prioritized, while for health insurance cases, medical invoices and expense details are pushed. Subsequently, through the API interface, using a multi-armed gambling machine algorithm (an algorithm that can dynamically optimize the timing and channels of push notifications), the list is pushed to the client through multiple channels such as the App, SMS, and email. For example, in-app messages are prioritized for younger users, while SMS reminders are added for older users.

[0022] In the fintech sector, this solution deeply empowers insurance claims processing. Traditional insurance claims document processing relies on manual labor, resulting in low efficiency, high costs, and a high risk of errors. This solution automates and improves document processing through multimodal data collection and multi-dimensional feature extraction, significantly shortening the claims cycle, reducing insurance company operating costs, enhancing claims risk control capabilities, effectively identifying document anomalies and suspected fraud cases, and ensuring the compliant and stable operation of insurance business. Furthermore, the application of privacy-preserving computing technology addresses data security and compliance issues in cross-institutional data collaboration, providing a secure and feasible path for data sharing between financial institutions and other relevant institutions, and driving deep innovation in fintech within the insurance sector.

[0023] In the healthcare field, this solution builds an efficient bridge between medical data and insurance claims. Through precise analysis of multimodal medical data such as medical invoices and medical records, it standardizes medical terminology and structures medical information, providing data support for medical insurance settlement and medical quality assessment. Simultaneously, personalized document delivery reduces the document preparation costs for patients during the claims process, improving their medical treatment and claims experience. Furthermore, the disease-drug-treatment knowledge graph constructed within the solution can provide valuable references for medical research, drug development, and clinical decision-making, promoting the digital and intelligent development of the healthcare field and contributing to the advancement of precision medicine.

[0024] In one embodiment, step S100 includes: S101. Obtain the document image data uploaded by the user on the client; S102. Collect verifiable voucher data and historical claims data; S103. Preprocess the document image data, verifiable voucher data and historical claims data using privacy computing technology; S104. Use a convolutional neural network to extract features from the preprocessed document image data to obtain visual features; S105. Natural language processing technology is used to perform semantic analysis on the preprocessed verifiable credential data to obtain text semantic features; S106. Extract context features from the preprocessed historical claims data to obtain case context features; S107. The visual features, textual semantic features, and case context features are fused to generate multimodal document features.

[0025] In this embodiment, the acquisition of document image data uploaded by users on the client side is mainly achieved through a mobile application. Users are allowed to take real-time photos of relevant documents such as medical invoices, diagnostic certificates, and ID cards, which are then encrypted and transmitted to the system. This image data is the core foundational material for claims review and is directly linked to key information in the user's claims case. Verifiable credential data is collected from medical institutions using blockchain technology, ensuring the authenticity and immutability of the data source and providing reliable data support for claims. The collection of historical claims case data covers multiple fields such as case type, required document list, and processing time. This past data contains rich claims patterns and can provide a reference for current case processing.

[0026] During the data preprocessing stage, privacy-preserving computation technology is comprehensively employed to securely process the collected document image data, verifiable voucher data, and historical claims data. The MPC SQL (Secure Multi-Party Computation Structured Query Language) operator within the "Hidden Language" framework is used to classify and categorize the data. All sensitive data, especially medical-related privacy data, is converted into ciphertext using secure encryption algorithms, allowing only authorized nodes to participate in computation. This ensures that the original data does not leave the local machine, enabling cross-institutional and cross-source joint data analysis while guaranteeing data security and compliance.

[0027] The feature extraction stage employs targeted techniques to handle different types of data: For preprocessed document image data, a convolutional neural network (CNN) is used to extract visual features. This model can accurately identify key fields such as serial numbers, names, and amounts on medical invoices. For non-standard documents such as discharge summaries, it can also extract medical entity terms such as disease names and surgical records through a semantic parsing model, supporting high-accuracy recognition of over 100 types of medical vouchers nationwide. For preprocessed verifiable voucher data, semantic analysis is conducted using natural language processing (NLP) technology. The BERT model is used to normalize medical terms to standard ICD encoding, thereby mining the core semantics behind the text and forming textual semantic features. For preprocessed historical claims data, basic information such as insurance type, cause of claim, and level of medical institution is analyzed. The embedding layer is used to convert this discrete information into feature vectors that can be used for algorithm calculation, i.e., case context features.

[0028] Finally, the extracted visual features, textual semantic features, and case context features are organically integrated to break down the barriers between different types of data and generate comprehensive and three-dimensional multimodal document features, providing solid feature support for subsequent document demand prediction, accurate delivery, and risk prevention and control.

[0029] In the fintech sector, this series of processes has brought revolutionary optimization to insurance claims processing. Problems such as fragmented data collection, insufficient privacy protection, and limited feature extraction in traditional insurance claims have been effectively solved. The secure integration of multi-source data and the accurate extraction of multi-dimensional features have automated and intelligentized claims document processing. This not only significantly shortens the processing cycle of claims cases and reduces the error rate and operating costs of insurance companies due to manual operations, but also improves the effectiveness of claims risk control through precise feature analysis. It can more accurately identify risks such as fraudulent claims and document forgery, ensuring the sound development of the insurance business. Furthermore, it provides a secure and compliant solution for cross-institutional financial data collaboration.

[0030] In the healthcare field, this process builds an efficient bridge connecting medical data with external application scenarios. Through in-depth analysis and feature extraction of medical document images and verifiable medical data, it achieves the standardization and structuring of medical information, providing precise data support for medical insurance settlement and medical expense review. The disease, diagnosis, and drug-related information contained in the multimodal document features can also provide valuable references for medical research, clinical pathway optimization, and drug development, contributing to the digital transformation of the healthcare field and the advancement of precision medicine. At the same time, strict privacy protection measures ensure the security of patient medical data and enhance patients' trust in medical data sharing.

[0031] In one embodiment, step S200 includes: S201. Pre-construct a single-certificate demand prediction model based on graph neural networks and attention mechanisms; S202. Filter historical claims data from multimodal data to generate full-process claims data; S203. Use multi-label classification to label the entire claims process data to obtain a set of valid documents, and sort the set of valid documents by priority to generate a pre-training dataset. S204. Convert the multimodal document features into multidimensional document vectors; S205. Using a federated learning framework, the document demand prediction model is jointly trained based on the pre-trained dataset and the multi-dimensional document vector.

[0032] In this embodiment, the pre-construction of the document demand prediction model is supported by graph neural networks (GNNs, a type of neural network model adept at handling data related to relationships between entities) and attention mechanisms. The model abstracts key elements in the claims scenario into a heterogeneous graph structure, where patients, medical institutions, and document types are all treated as independent nodes. The relationships between nodes (such as the patient-medical institution relationship and the document type-case compatibility relationship) are presented as edges. The integration of the attention mechanism allows the model to dynamically focus on key information, aggregating neighborhood information of nodes through a message passing mechanism, and accurately calculating the importance weights of various documents in specific cases, laying the architectural foundation for subsequent accurate predictions.

[0033] Historical claims data from the multimodal dataset undergoes targeted screening to remove invalid, duplicate, or incomplete records. The dataset retains complete information including case type, cause of claim, level of medical institution treated, required documentation list, processing steps, and final settlement result, forming a dataset covering the entire claims process. This step aims to extract high-quality data with reference value to ensure the effectiveness of subsequent model training.

[0034] A multi-label classification method was used to label the filtered claims process data. Based on the actual processing results of each case, all valid documents required for each case were identified, forming a set of valid documents. Subsequently, the set of valid documents was prioritized based on factors such as the role of documents in claims review and the timeliness requirements for submission. For example, core documents (such as medical invoices and diagnostic certificates) had higher priority than auxiliary documents (such as travel certificates). Finally, a pre-training dataset with both completeness and order was generated to provide standardized training samples for the model.

[0035] Multimodal document features encompass visual features, textual semantic features, and case context features. By using techniques such as embedding layers, these different dimensions and types of features are transformed into multi-dimensional document vectors with a unified dimension that can be directly processed by the model. This achieves the structuring and standardization of features, breaks down the barriers between different types of features, and enables the model to efficiently extract key patterns from the fused information.

[0036] The joint training phase of the model is conducted using a federated learning framework. Each participating party (such as different insurance companies or medical institutions) retains the original data locally, transmitting only the multi-dimensional document vectors and training parameters based on the pre-trained dataset to the joint training node via an encrypted channel. This ensures that the original data does not leave the local machine and that privacy and security are guaranteed, enabling collaborative training across institutions and data sources, continuously optimizing model parameters, and improving the model's accuracy in predicting document requirements in different scenarios.

[0037] In the fintech sector, this series of processes provides core support for the intelligent upgrade of insurance claims processing. Through precise document requirement prediction models, insurance companies can dynamically match the document requirements of different cases, significantly reducing instances of users omitting or misreporting documents, shortening the claims processing cycle, and lowering operational costs caused by repeated communication and document resubmission. The application of the federated learning framework addresses the pain points of data silos and privacy protection among financial institutions, enabling the mining and sharing of cross-institutional data value. Simultaneously, it improves the accuracy of claims risk control, effectively identifying risks such as fraudulent documents and malicious insurance fraud, ensuring the stable operation of insurance business, and promoting the deep application of fintech in areas such as risk control and service optimization.

[0038] In the healthcare sector, this process has built an efficient bridge connecting medical data and claims services. The model's accurate prediction of medical documentation needs, based on standardized medical terminology and structured medical data, not only simplifies the process of preparing claims materials for patients and improves their healthcare experience, but also provides data support for medical insurance settlement and medical expense review. The deep fusion and feature extraction of multimodal data can help build a more comprehensive disease-drug-treatment knowledge graph, providing valuable references for clinical decision-making, medical quality assessment, and medical research. Simultaneously, the combination of privacy-preserving computing technology and federated learning ensures the security and compliance of medical data during sharing and use, promotes cross-institutional collaborative applications of healthcare data, and drives the digital and intelligent development of the healthcare sector.

[0039] In one embodiment, S400 includes: S411. Obtain the user's current claim process stage; S412. Generate a basic document list and a special document list based on the document requirement list and the current claims process; S413. Obtain user information and generate user profile data based on the user information; S414. Based on the user profile data, generate a personalized document list according to the basic document list and the special document list. S415. Select a push strategy based on the user profile data; S416. Push the personalized document list to the client according to the push strategy.

[0040] In this embodiment, obtaining the user's current claim process involves the system tracking the user's progress in the claim process in real time, clarifying the specific stage the user is in, such as material submission, review and verification, and payment settlement. Different stages correspond to different document requirements, which is the basic premise for generating an accurate document list.

[0041] Based on the document requirement list and the current claims process, a basic document list and a special document list are generated. The basic document list contains essential and universally applicable documents for all claims stages, such as identity documents and copies of insurance policies, ensuring basic compliance of the claims process. The special document list is customized for specific claims stages and case types. For example, in the document submission stage, car insurance cases require supplementary documents such as driver's licenses and traffic accident liability statements, while health insurance cases require specific submissions such as medical invoices and expense details, making the document list both comprehensive and targeted.

[0042] Obtaining user information requires collecting basic user data through compliant channels, including age, occupation, contact preferences, past claims records, and insurance product types. This data is then integrated and analyzed to generate user profile data. User profile data can accurately depict user characteristics and needs, such as differentiating communication habits between young and elderly users, and document familiarity between frequent claimants and first-time claimants.

[0043] Based on user profile data, the basic document list and the special document list are personalized to generate a customized document list. For example, for users making their first claim, supplementary information such as document filling instructions and upload guidelines will be added to the list; for corporate users with frequent claims, the process of repeatedly submitting basic documents is simplified, and special documents directly related to the current case are presented first; for users who seek medical treatment in other places, additional prompts for documents related to out-of-town medical treatment registration are pushed to ensure that the document list is completely tailored to the individual user's situation.

[0044] The core of selecting a push notification strategy based on user profile data is matching users' behavioral habits and preferences. For example, for younger users who are more accustomed to using mobile applications, the push notification strategy prioritizes in-app message pushes with a quick access point for uploading documents; for older users, considering their higher adaptability to receiving SMS messages, the push notification strategy will primarily focus on SMS reminders, along with simple and clear operation instructions; for business professionals, email pushes can be selected to facilitate their easy access to their document lists; and for users who prefer instant communication, reminders can be enhanced by combining them with app pop-ups and other methods.

[0045] Based on the selected push strategy, a personalized list of documents is accurately pushed to the client via API. During the push process, secure encryption is ensured for information transmission, and useful information such as document upload links and deadline reminders are included. If the user fails to view or upload the documents in a timely manner, the system will issue appropriate reminders according to the push strategy to ensure the smooth progress of the claims process.

[0046] In the fintech sector, this process significantly optimizes the experience and efficiency of insurance claims services. By precisely matching users' claims processes with their individual needs, it reduces the probability of users missing or missubmitting documents, shortens the claims cycle, and lowers the operational costs for insurance companies caused by repeated communication and document resubmission. Personalized push strategies allow different types of users to conveniently receive and process documents, improving user satisfaction with insurance services. It also helps insurance companies optimize service resource allocation, focus on high-value service segments, and promote the deep development of fintech in the areas of refined and intelligent insurance services.

[0047] In the healthcare sector, this process builds an efficient bridge between medical services and claims services. It generates personalized document lists based on patients' medical scenarios and individual characteristics, simplifying the process of preparing claims materials. This is particularly beneficial for elderly patients and those seeking medical treatment in other locations, alleviating their post-treatment claims burdens. Precise push strategies ensure patients receive the required documents promptly, preventing delays in medical insurance settlement or commercial insurance claims due to document issues, thus improving the closed-loop experience of healthcare services. Simultaneously, optimizing document push and services through user profile data provides a reference for collaborative services between healthcare institutions and financial institutions, promoting deep integration and collaborative innovation between the healthcare and financial sectors.

[0048] In one embodiment, step S400 includes: S421. Obtain user historical response data and scenario characteristics; S422. Calculate the optimal push time based on the user's historical response data and scenario characteristics; S423. Push the personalized document list to the client according to the optimal push time, and obtain user response data in real time; S424. Update the user's historical response data based on the user response data.

[0049] In this embodiment, obtaining user historical response data and scenario characteristics is a fundamental prerequisite for accurate push notifications. User historical response data covers the user's past behavior records after receiving document push notifications, including the viewing time of different push channels (App, SMS, email), the time taken to complete document upload, and whether a secondary reminder is needed. This data directly reflects the user's response habits. Scenario characteristics focus on key information in the current claims scenario, such as the type of claims (car insurance, health insurance, etc.), the current stage of the claims process (submission of materials, review, payment), the level of the medical institution, and the time interval between the time of the incident and the current situation. These characteristics directly affect the user's attention to push notifications and their processing priority.

[0050] The optimal push notification time is calculated based on historical user response data and scenario characteristics. The core principle is to use algorithms to uncover the matching relationship between user behavior patterns and scenario needs. Utilizing optimization algorithms such as multi-armed gambling machines, and combining user response efficiency at different times (e.g., 9 AM and 7 PM on weekdays, and all day on weekends), the system analyzes users' high-frequency activity periods. Simultaneously, it dynamically adjusts the push notification time based on scenario characteristics. For example, for health insurance policyholders recently discharged from the hospital, a document checklist is pushed within 24 hours of discharge, aligning with their need to organize post-operative claim materials. For car insurance policyholders involved in accidents, push notifications are pushed within 1-3 hours after the accident is resolved, leveraging peak user attention to post-accident matters to improve response rates. Ultimately, the optimal push notification time is determined to maximize the probability of timely user response.

[0051] Based on the calculated optimal push time, the personalized document list is accurately pushed to the client through the preset push channels. During the push process, the system maintains real-time data monitoring and continuously acquires user response data, including whether the user viewed the push information, the viewing time, the time when the document upload started, the completion status of the document upload, and whether the user requested a re-upload guide, ensuring that user feedback dynamics can be captured in a timely manner.

[0052] Based on real-time user response data, the system dynamically updates historical user response data. If a user quickly views and completes document upload within the optimal push time, the system records the time point's suitability for the scenario, strengthening the time recommendation weight for that scenario. If a user does not respond in time or requires a secondary reminder, the system updates relevant data such as response timeliness and reminder needs, providing the latest basis for adjusting the optimal push time and optimizing the push strategy, forming a data-driven closed-loop optimization mechanism.

[0053] In the fintech sector, this process has significantly improved the precision and operational efficiency of insurance claims services. By precisely timing the push notifications, it has greatly increased user response rates and processing efficiency for document pushes, reducing issues such as missing or delayed document uploads caused by inappropriate timing, and shortening the overall processing cycle for claims. Simultaneously, continuous updates and strategy optimization based on user behavior data have reduced unnecessary secondary reminder costs, allowing insurance companies to focus their service resources on high-value aspects, improving user satisfaction and business competitiveness, and driving more precise user reach and service delivery through fintech in insurance service scenarios.

[0054] In the healthcare sector, this process builds a bridge for claims services that better meet patients' needs. It optimizes push notification times based on patients' post-operative recovery and accident handling status, taking into account their daily habits. This avoids disruptions caused by push notifications during critical treatment periods or rest times, while ensuring patients receive timely guidance when it's convenient for them to handle claims, reducing delays in medical insurance settlements or commercial insurance claims due to untimely document preparation. Furthermore, by continuously updating user response data and optimizing push notification timing strategies, it better adapts to the lifestyles and habits of different patients (such as elderly patients and working professionals), improving the closed-loop experience of healthcare services and facilitating deep collaboration between healthcare and financial services.

[0055] In one embodiment, after step S400, the method further includes: S501. Obtain document image data uploaded by the user based on the personalized document list; S502. Perform multi-dimensional quality verification on the document image data; S503. If the document image data passes the document quality verification, then mark the personalized document list upload as complete. S504. If the document image data fails the document quality verification, a document quality verification report is generated. S505. The document quality verification report is sent back to the client, and a re-upload reminder is triggered.

[0056] In this embodiment, after receiving the personalized document list, the user uploads the corresponding document image data through the client. This data covers medical invoices, diagnostic certificates, ID documents and other vouchers related to claims. The upload process uses encrypted transmission to ensure data security during transmission and avoid information leakage.

[0057] The system performs multi-dimensional quality checks on uploaded document images, ensuring comprehensive and accurate verification. Firstly, it performs a completeness check, verifying whether the uploaded documents cover all required vouchers in the personalized list and whether any key documents are missing. Secondly, it performs clarity and standardization checks, using image recognition technology to detect whether document images are blurry, obstructed, or have intact edges and corners. It also verifies whether key information on the documents (such as serial number, name, amount, and treatment items) is clearly legible and whether the format conforms to standards. For medical documents, it also incorporates Natural Language Processing (NLP) technology to verify the completeness and standardization of medical terminology, ensuring that the document information can support subsequent claims review.

[0058] If the document image data passes the quality verification in all dimensions, the system will automatically mark the personalized document list as uploaded successfully and synchronize the relevant data to the subsequent claims review process, thus promoting the orderly progress of the claims process.

[0059] If the document image data fails the quality verification, the system will immediately generate a detailed document quality verification failure report. The report clearly indicates the specific reasons for the failure, such as "the medical invoice image is blurry and the amount field cannot be recognized," "a scanned copy of the original diagnostic certificate is missing," or "the corner of the ID card photo is missing and key information is incomplete," allowing users to clearly understand the problem.

[0060] Subsequently, the system will send the failed report to the client via previously designated push channels (such as app messages, SMS, email, etc.) and trigger a re-upload reminder. The reminder message will include specific rectification guidelines and suggested re-upload timelines, allowing users to quickly correct issues, supplement and improve documents based on the report, and re-upload, ensuring that the claims process is not excessively delayed due to document quality issues.

[0061] In the fintech business, this series of processes further ensures the accuracy and efficiency of insurance claims. Multi-dimensional quality checks pre-filter invalid and substandard documentation, avoiding rejections due to document issues, reducing repetitive workload for claims personnel, and shortening the overall claims cycle. Clearly defined failure reports and rectification guidelines reduce communication costs between users and insurance companies, improve user experience, and provide high-quality data support for subsequent intelligent claims processing. This helps insurance companies optimize operational processes, reduce operating costs, and enhance the professionalism and reliability of claims services.

[0062] In the healthcare sector, this process bridges the gap between medical data and claims services, ensuring quality control. Through rigorous quality verification of medical document images, the authenticity, completeness, and standardization of medical data are guaranteed, providing a reliable data source for medical insurance settlements and commercial insurance claims, and preventing disputes caused by incorrect or incomplete document information. For patients, clear feedback reports and re-upload instructions reduce the difficulty of preparing claim materials, minimize the hassle of repeatedly going to supplement documents, and improve the post-treatment claims experience. Simultaneously, high-quality medical document data also provides reliable support for data analysis and scientific research statistics in the healthcare field, promoting the coordinated development of healthcare and financial services.

[0063] In one embodiment, step S501 includes: S5011. Based on the document image data, multimodal document features, historical claims data, and preset multi-level risk control strategies, perform real-time risk scoring on the cases corresponding to the personalized document list; S5012. If the risk score is lower than the preset threshold, it is determined to be a low-risk case and enters the claims process. S5013. If the risk score is higher than the preset threshold, it is determined to be a high-risk case and suspected fraud cases are intercepted.

[0064] In this embodiment, after generating and pushing a personalized list of documents and the user uploads document images, the system initiates a real-time risk scoring process. The core of this process is to integrate multi-source data with preset risk control strategies for comprehensive judgment. The data relied upon includes user-uploaded document image data (valid credentials verified through multi-dimensional quality checks, including key information such as invoices and diagnostic certificates), multimodal document features formed by fusing visual features, textual semantic features, and case context features, as well as historical claims data covering past case types, processing results, and risk records. Simultaneously, the system incorporates preset multi-level risk control strategies. These strategies include rules across multiple dimensions, such as "invoice amount does not match treatment items," "abnormal frequency of visits," and "conflicts between document information and historical cases." Custom configuration is supported based on different insurance product types and business scenarios to ensure the adaptability of risk control standards.

[0065] During the risk scoring process, the system conducts in-depth analysis of the aforementioned data using multi-dimensional machine learning algorithms. Utilizing a risk scoring model within a federated learning framework, and combining key information from multimodal document features (such as standardized medical terminology results and case context feature vectors), it verifies each case against multi-level risk control strategies. For example, it uses natural language processing technology to analyze the consistency between medical record text and invoice treatment items, and compares the frequency of claims and claim amounts in the current case against historical claims data to determine if they are within a reasonable range. Ultimately, it outputs a quantitative risk score, intuitively reflecting the degree of fraud suspicion in the case.

[0066] If the risk score is below a preset threshold, the case is classified as low-risk. These cases do not show obvious signs of fraud, but may have incomplete information or require further verification. The system will automatically generate clear material supplementation prompts and send them to the user through previous push channels, informing them of the specific materials that need to be supplemented (such as supplementary medical record details, expense details, etc.), guiding the user to complete the information and continue to advance the claims process. This ensures the accuracy of the claims process without affecting the processing efficiency of normal cases.

[0067] If the risk score exceeds a preset threshold, the case is classified as high-risk. The system will immediately activate an interception mechanism, marking suspected fraud cases and automatically transferring them to the manual review channel. Simultaneously, the system will lock the subsequent automated processing flow to prevent erroneous claims. The manual review stage will focus on verifying anomalies in the risk score, such as falsified documentation, fabricated medical procedures, and duplicate claims. Professional reviewers will determine the nature of the case; cases confirmed as fraud will be processed according to regulations, while misjudged high-risk cases will be unblocked and the normal claims process will resume.

[0068] In the fintech business, traditional insurance claims fraud identification relies on manual screening, which is inefficient and prone to omissions. This real-time risk scoring mechanism, through multi-source data fusion and intelligent algorithm analysis, significantly improves the accuracy of fraud identification, effectively reducing insurance companies' losses from erroneous claims. The customizable configuration function of multi-level risk control strategies adapts to the risk control needs of different insurance products (such as auto insurance and health insurance), enhancing the flexibility and practicality of fintech in insurance risk control. Simultaneously, the rapid guidance of low-risk cases and the precise interception of high-risk cases optimize resource allocation in the claims process, allowing insurance companies to focus more on high-value risk verification and service improvement, driving the intelligent upgrade of fintech in insurance risk control.

[0069] In the healthcare sector, cross-verification of medical documents, treatment information, and historical data can effectively identify violations such as fabricated medical expenses and false treatment records, ensuring the safe and compliant use of medical insurance funds and commercial insurance funds. For patients seeking normal medical care, the rapid processing of low-risk cases and clear guidelines for supplementing materials reduce unnecessary procedural obstacles and improve the claims experience. Intercepting high-risk violations also helps regulate the medical service market and promote the healthy development of the healthcare sector. Furthermore, in-depth analysis of medical data during risk scoring can provide data support for risk warnings and standardized treatment practices in the healthcare sector, driving collaborative risk control and innovative development between healthcare and fintech.

[0070] In one embodiment, a dynamic claims document push device is provided, which corresponds one-to-one with the dynamic claims document push method described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the claims document dynamic push device of the present invention. The modules include a context encoding module 10, a historical feature management module 20, a distribution prediction module 30, a speech waveform point generation module 40, a waveform feature encoding module 50, and a speech synthesis control module 60. Detailed descriptions of each functional module are as follows: Data processing module 10 is used to collect multimodal data, extract multi-dimensional features from the multimodal data, and generate multimodal document features; The model training module 20 is used to jointly train the document demand prediction model using the multimodal data and multimodal document features; Prediction list module 30 is used to input the multimodal document features into the trained document demand prediction model for prediction and output a document demand list. The list push module 40 is used to generate a personalized list of documents based on the current claims process, user profile data, and document requirement list, and push the personalized list of documents to the client.

[0071] In one embodiment, the data processing module 10 includes: Obtain document image data uploaded by the user on the client; Collect verifiable voucher data and historical claims data; Privacy-preserving computation techniques are used to preprocess the document image data, verifiable credential data, and historical claims data. A convolutional neural network is used to extract features from the preprocessed document image data to obtain visual features; Natural language processing techniques are used to perform semantic analysis on the preprocessed verifiable credential data to obtain textual semantic features; Contextual features are extracted from the preprocessed historical claims data to obtain case contextual features; The visual features, textual semantic features, and case context features are fused to generate multimodal document features.

[0072] In one embodiment, the model training module 20 includes: A document demand prediction model is pre-built based on graph neural networks and attention mechanisms; Historical claims data is filtered from multimodal data to generate full-process claims data; The entire claims process data is labeled using multi-label classification to obtain a set of valid documents, and the set of valid documents is prioritized to generate a pre-training dataset. The multimodal document features are converted into multidimensional document vectors; The document demand prediction model is jointly trained using a federated learning framework based on the pre-trained dataset and multi-dimensional document vectors.

[0073] In one embodiment, the list push module 40 includes: Get the user's current claim process stage; Generate a basic document list and a special document list based on the document requirement list and the current claims process; Obtain user information and generate user profile data based on the user information; Based on the user profile data, a personalized document list is generated according to the basic document list and the special document list. Select a push strategy based on the user profile data; The personalized document list is pushed to the client according to the push strategy.

[0074] In one embodiment, the list push module 40 includes: Obtain user historical response data and scenario characteristics; The optimal push time is calculated based on the user's historical response data and scenario characteristics; The personalized document list is pushed to the client according to the optimal push time, and user response data is obtained in real time. Update the user's historical response data based on the user response data.

[0075] In one embodiment, the document verification module further includes: Obtain document image data uploaded by the user based on a personalized document list; Perform multidimensional quality verification on the document image data; If the document image data passes the document quality verification, then the personalized document list is marked as uploaded successfully. If the document image data fails the document quality verification, a document quality verification report will be generated. The document quality verification report is sent back to the client, triggering a re-upload reminder.

[0076] In one embodiment, the risk control module includes: Based on the document image data, multimodal document features, historical claims data, and preset multi-level risk control strategies, real-time risk scoring is performed on the cases corresponding to the personalized document list; If the risk score is lower than the preset threshold, it is determined to be a low-risk case and enters the claims process; If the risk score is higher than a preset threshold, it is determined to be a high-risk case, and suspected fraud cases are blocked.

[0077] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements the server-side functions or steps of a method for dynamically pushing claims documents.

[0078] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements a method for dynamically pushing claims documents, fulfilling user-side functions or steps. In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Collect multimodal data, extract multidimensional features from the multimodal data, and generate multimodal document features; The document demand prediction model is jointly trained using the multimodal data and multimodal document features. The multimodal document features are input into the trained document demand prediction model for prediction, and a document demand list is output. A personalized document list is generated based on the current claims process, user profile data, and document requirement list, and then pushed to the client.

[0079] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Collect multimodal data, extract multidimensional features from the multimodal data, and generate multimodal document features; The document demand prediction model is jointly trained using the multimodal data and multimodal document features. The multimodal document features are input into the trained document demand prediction model for prediction, and a document demand list is output. A personalized document list is generated based on the current claims process, user profile data, and document requirement list, and then pushed to the client.

[0080] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0083] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for dynamically pushing claims documents, characterized in that, Includes the following steps: Collect multimodal data, extract multidimensional features from the multimodal data, and generate multimodal document features; The document demand prediction model is jointly trained using the multimodal data and multimodal document features. The multimodal document features are input into the trained document demand prediction model for prediction, and a document demand list is output. A personalized document list is generated based on the current claims process, user profile data, and document requirement list, and then pushed to the client.

2. The method for dynamically pushing claims documents as described in claim 1, characterized in that, The process of collecting multimodal data and extracting multidimensional features from the multimodal data to generate multimodal document features includes: Obtain document image data uploaded by the user on the client; Collect verifiable voucher data and historical claims data; Privacy-preserving computation techniques are used to preprocess the document image data, verifiable credential data, and historical claims data. A convolutional neural network is used to extract features from the preprocessed document image data to obtain visual features; Natural language processing techniques are used to perform semantic analysis on the preprocessed verifiable credential data to obtain textual semantic features; Contextual features are extracted from the preprocessed historical claims data to obtain case contextual features; The visual features, textual semantic features, and case context features are fused to generate multimodal document features.

3. The method for dynamically pushing claims documents as described in claim 1, characterized in that, The joint training of the document demand prediction model using the multimodal data and multimodal document features includes: A document demand prediction model is pre-built based on graph neural networks and attention mechanisms; Historical claims data is filtered from multimodal data to generate full-process claims data; The entire claims process data is labeled using multi-label classification to obtain a set of valid documents, and the set of valid documents is prioritized to generate a pre-training dataset. The multimodal document features are converted into multidimensional document vectors; The document demand prediction model is jointly trained using a federated learning framework based on the pre-trained dataset and multi-dimensional document vectors.

4. The method for dynamically pushing claims documents as described in claim 1, characterized in that, The process of generating a personalized document list based on the current claims process, user profile data, and document requirement list, and then pushing the personalized document list to the client, includes: Get the user's current claim process stage; Generate a basic document list and a special document list based on the document requirement list and the current claims process; Obtain user information and generate user profile data based on the user information; Based on the user profile data, a personalized document list is generated according to the basic document list and the special document list. Select a push strategy based on the user profile data; The personalized document list is pushed to the client according to the push strategy.

5. The method for dynamically pushing claims documents as described in claim 1, characterized in that, The process of generating a personalized document list based on the current claims process, user profile data, and document requirement list, and then pushing the personalized document list to the client, includes: Obtain user historical response data and scenario characteristics; The optimal push time is calculated based on the user's historical response data and scenario characteristics; The personalized document list is pushed to the client according to the optimal push time, and user response data is obtained in real time. Update the user's historical response data based on the user response data.

6. The method for dynamically pushing claims documents as described in claim 1, characterized in that, After generating a personalized document list based on the current claims process, user profile data, and document requirement list, and pushing the personalized document list to the client, the process further includes: Obtain document image data uploaded by the user based on a personalized document list; Perform multidimensional quality verification on the document image data; If the document image data passes the document quality verification, then the personalized document list is marked as uploaded successfully. If the document image data fails the document quality verification, a document quality verification report will be generated. The document quality verification report is sent back to the client, triggering a re-upload reminder.

7. The method for dynamically pushing claims documents as described in claim 6, characterized in that, The process of obtaining document image data uploaded by the user based on a personalized document list also includes: Based on the document image data, multimodal document features, historical claims data, and preset multi-level risk control strategies, real-time risk scoring is performed on the cases corresponding to the personalized document list; If the risk score is lower than the preset threshold, it is determined to be a low-risk case and enters the claims process; If the risk score is higher than a preset threshold, it is determined to be a high-risk case, and suspected fraud cases are blocked.

8. A dynamic push device for claims documents, characterized in that, The dynamic push device for claims documents includes: The data processing module is used to collect multimodal data, extract multi-dimensional features from the multimodal data, and generate multimodal document features; The model training module is used to jointly train the document demand prediction model using the multimodal data and multimodal document features; The prediction list module is used to input the multimodal document features into the trained document demand prediction model for prediction and output a document demand list. The list push module is used to generate a personalized list of documents based on the current claims process, user profile data, and document requirement list, and push the personalized list of documents to the client.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a claims document dynamic push program stored in the memory and executable on the processor. When the claims document dynamic push program is executed by the processor, it implements the steps of the claims document dynamic push method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a dynamic push program for claims documents, which, when executed by a processor, implements the steps of the method for dynamic push of claims documents as described in any one of claims 1-7.

Images