Intelligent claim settlement material dynamic guidance method based on diagnosis driving

Abstract

This invention proposes a diagnostic-driven intelligent method for dynamic guidance of claim materials. The method includes: first, receiving images of claim materials uploaded by the user, performing quality checks, and automatically classifying the images using a trained image classification model; then, extracting key structured information from the classified claim material images using a combination of OCR technology and a multimodal large model; next, using this key structured information as query conditions to obtain a list of required materials corresponding to the current diagnosis, comparing and generating targeted guidance information to guide the user to complete the missing materials in real time; finally, packaging the key structured information, targeted guidance information, and classified claim material images into a claim pre-application package for confirmation and submission. This solves the problems of delayed material missing detection, lack of targeted guidance, and reliance on manual review in the traditional claim process, achieving real-time, accurate missing material detection and dynamic guidance during the claim material upload stage.

Description

Technical Field

[0001] This invention relates to the field of computer information technology, and in particular to a diagnostic-driven intelligent claims material dynamic guidance method. Background Technology

[0002] In insurance claims, customers are required to submit various supporting documents. In traditional claims processes, customers are often unclear about the specific documents needed and simply upload a generic checklist, easily overlooking crucial files. When reviewers discover missing documents, the claims are returned and the customer is asked to provide the missing information, leading to longer processing times and a poor user experience.

[0003] Existing technologies include several systems that assist in claims processing. For example, Chinese patent CN117011872A discloses a "claims data processing method," which includes: preprocessing images uploaded by users; classifying the images using a classification model; calculating and matching the classified image information with the applicant's information, policy information, and claims information, and performing integrity checks; determining the image information that needs to be supplemented, and notifying the user via an outbound call device. However, this technical solution has the following drawbacks: The current technology merely performs a generalized and static matching of categorized materials with policy and claims information to check the completeness of materials, without dynamically constructing differentiated verification rules based on specific disease diagnoses, treatment methods, and other factors. For example, the key supporting documents required for pneumonia claims differ significantly from those for fracture claims. Current technology cannot adaptively adjust the list of required materials based on diagnostic information, still relying on fixed templates for completeness judgment. This cannot avoid omissions or duplicate submissions due to mismatches between material requirements and actual conditions.

[0004] Furthermore, this solution notifies users via outbound calling after submission, which is a post-event guidance process. Users need to wait for manual review before they can submit supplementary materials, resulting in low efficiency. In addition, the outbound notification is a one-time voice interaction, and users cannot review the list of required supplementary materials at any time after the call ends. The outbound notification also cannot carry a structured link to the supplementary materials list or upload entry, so users still need to understand and memorize the supplementary materials requirements themselves, resulting in low information transmission efficiency. Moreover, the outbound notification lacks linkage with the system's front-end upload interface, failing to achieve a seamless experience of clicking the notification to jump to the supplementary materials page. This leads to a disconnect between the supplementary materials process and the main claims process, resulting in a poor user experience. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a diagnostic-driven dynamic guidance method for claims materials, comprising the following steps: S1: Receive the images of claim materials uploaded by the user, perform quality inspection on the claim material images, and then use a trained image classification model to automatically classify the claim material images into preset material types.

[0006] Furthermore, the quality inspection includes checking the clarity, brightness, and tilt of the claims material images, and prompting unqualified images to be retaken; the image classification model adopts a low-cost, self-evolving training scheme based on transfer learning and active learning, thereby obtaining a high-performance classification model with only a small amount of manual annotation; the preset material types include at least "hospitalization invoice", "discharge summary", "diagnosis certificate" and "expense list".

[0007] Furthermore, the specific training steps for the image classification model are as follows: 1) A lightweight convolutional neural network pre-trained on a general image dataset is selected as the base model. The shallow feature extraction layer is frozen, and only the classification layer at the end is retrained. This strategy utilizes the existing general feature extraction capabilities of the pre-trained model to reduce the need for manual annotation from hundreds of thousands to thousands. 2) In terms of training data construction, we integrate two mechanisms: weakly supervised automatic labeling and active learning. We extract images uploaded by users through the specified material type entry from historical claims business logs and automatically assign them the corresponding material type as "weak labels", thereby obtaining tens of thousands to hundreds of thousands of noisy training weak label data at no cost. 3) The following active learning iterative strategy is adopted: first, the initial image classification model is trained with the seed training set, and then the image classification model is used to predict the automatically labeled weakly labeled data to obtain the confidence score; 4) High-confidence samples with confidence levels higher than α are automatically added to the seed training set, while suspected erroneous samples with confidence levels lower than β are submitted for manual review; the reviewed and corrected samples are also added to the seed training set to retrain the image classification model; where α and β are both values ​​greater than 0 and less than 1, and α > β; preferably, α ≥ 0.9 and β ≤ 0.6.

[0008] S2: For classified claim material images, key structured information is extracted from the classified claim material images by combining OCR technology with multimodal large model.

[0009] The key structured information includes at least diagnostic information, invoice number, date of visit, and medical expenses.

[0010] Furthermore, this step designs a hierarchical, complementary, and confidence-based intelligent extraction strategy to balance extraction efficiency and accuracy. The specific process is as follows: 1) Utilize a mature OCR engine to quickly and fully recognize categorized claim material images, directly extracting structured fields such as invoice number, consultation date, and medical expenses, and attempting to locate diagnostic information; 2) Determine whether the diagnostic information extracted by the OCR engine meets the following conditions: the confidence level is greater than or equal to the preset confidence threshold, and the current claim material image belongs to a structured document; if both conditions are met, it is determined to be a high-confidence scenario. At this time, the multimodal large model is no longer called, and the diagnostic information and other structured fields extracted by the OCR engine are directly used as the key structured information, and then proceed directly to step 4); This can avoid unnecessary calls to the large model API, reduce operating costs and response latency; 3) If the diagnostic information extracted by the OCR engine is empty, or the confidence level is lower than the preset confidence threshold, or the document being processed is unstructured, the multimodal large model will be automatically triggered and the multimodal large model will be guided by carefully designed prompts to output key structured information that includes at least the diagnostic information, invoice number, consultation date, and medical expenses. 4) Preset the authority priority of diagnostic information sources. When different sources extract the same information and the results are inconsistent, the source with higher priority will be automatically adopted, and the low-priority result will be marked as "pending review" for subsequent manual review. For diagnostic information, priority will be given to extracting it from materials of the "discharge summary" and "diagnosis certificate" types.

[0011] S3: Using the diagnostic information extracted from the key structured information as query conditions, access the preset diagnostic-material mapping knowledge base, obtain the list of required materials corresponding to the current diagnosis, and generate a list of missing materials by comparison.

[0012] Furthermore, the diagnosis-materials mapping knowledge base contains a mapping relationship between diagnostic information and the types of materials required for claims. The core data table includes fields such as "Diagnosis Information," "Basic Material List," "Applicable Conditions," "Condition Matching Value," "Additional Materials," and "Excluded Materials," and supports dynamic adjustment based on diagnostic conditions. The "Basic Material List" field contains the general materials typically required under the diagnosis; the "Applicable Conditions" field specifies the case context in which the mapping record takes effect; the "Condition Matching Value" field specifies the specific value of the condition corresponding to the case context; the "Additional Materials" field contains materials that need to be added when the applicable conditions are met; and the "Excluded Materials" field contains materials that need to be deleted from the "Basic Material List" when the applicable conditions are met.

[0013] Specifically, after obtaining the user's key structured information and case context, all relevant mapping records are queried using the diagnostic information as the primary key. Then, the corresponding condition records are matched according to the case context, and the final list of required materials is dynamically generated. The list of required materials = basic material list + additional materials corresponding to condition matching - excluded materials when matching conditions. Then, the uploaded material types are compared with the list of required materials to generate a list of missing materials.

[0014] The case context refers to a set of auxiliary attributes related to the current claim case, excluding diagnostic information, used to refine the list of required materials. The case context includes at least the type of medical visit, whether surgery is involved, and the department visited, and is obtained in the following two ways: ① Front-end data collection: basic information filled in by the user at the claim application portal, or obtained through selected options; ② Inference from extracted structured information.

[0015] Furthermore, the missing materials in the missing materials list are graded and marked: a basic importance level is predefined for each material type, including mandatory, critical, and recommended. The mandatory missing materials directly result in failure to meet the basic requirements for filing a case, requiring the user to submit them again. The critical missing materials result in the inability to determine core responsibility or amount, but are not an absolute necessity for filing a case. The recommended missing materials are used to assist in the review process, do not affect filing a case, but help speed up the review; the user can choose to submit them again. A set of dynamic upgrade and downgrade rules based on diagnosis and conditions is also built-in. Then, the missing materials list is traversed, and for each missing material, its basic importance level is first obtained, then the dynamic upgrade and downgrade rules are applied to adjust it to obtain the final level. Finally, missing materials with the final level of mandatory or critical are marked as mandatory, and the remaining missing materials are marked as recommended.

[0016] Furthermore, the dynamic upgrade and downgrade rules are a set of built-in decision rules based on diagnostic information and case context, used to dynamically adjust the importance level of material types. The general logic is: if the diagnostic information belongs to a preset diagnostic coding range or the case context meets the value of a preset applicable condition field, then the importance level of the material type is adjusted from the original level to the new level.

[0017] Through the above-mentioned conditional queries and dynamic hierarchical marking, the importance of materials in different diagnostic scenarios can be accurately distinguished, avoiding unreasonable mandatory requirements on users, while ensuring that core claims evidence is not omitted.

[0018] S4: Based on the list of missing materials, generate targeted guidance information to guide users to complete the missing materials in real time.

[0019] The targeted guidance information includes: the type and reason for the missing materials, sample photos, and key points for shooting. This targeted guidance information is displayed in real time through the APP interface or mini-program, and provides an "Shoot Now" entry. After the user uploads the missing materials, steps S1-S3 are repeated until the user has submitted all the necessary materials.

[0020] Furthermore, the generation of targeted guidance information is based on a combination of missing material type, missing reason, and user profile in the generation and rendering process. Specifically, a guidance content material library is pre-built, storing the following structured elements for each material type: 1. Missing material type and reason text template; 2. Standard shooting example image; 3. Shooting key point prompt text list. The missing material list is traversed to obtain each missing material item. The corresponding missing material type and reason text template is retrieved from the guidance content material library. The variables in the missing material type and reason text template are replaced with actual values, and assembled into a guidance card object containing the missing material type, reason text, standard shooting example image URL, shooting key points, and an "Shoot Now" action button. After receiving the guidance card object, the front-end APP or mini-program displays the guidance card objects to the user in a list using a dynamic rendering engine. The "Shoot Now" button on each guidance card object is pre-associated with the material type to be uploaded. After the user clicks it, the camera is opened directly and the upload is automatic. Once the upload is complete, the closed-loop detection process of steps S1-S3 is re-executed.

[0021] Through the above solution, the present invention achieves a seamless connection from "detecting missing data" to "explaining the reasons" and then to "guiding resubmission", transforming complex review rules into intuitive guidance that is easy for users to understand, and significantly improving resubmission efficiency and user experience.

[0022] S5: Arrange the classified claim materials images in a preset order, and package them together with the extracted key structured information and targeted guidance information to generate a claim pre-application package. After the user confirms, submit the claim pre-application package to the claim review end for subsequent verification.

[0023] Once all the necessary materials are complete, a structured pre-claim package is generated for user confirmation. After the user confirms that everything is correct, the pre-claim package is submitted to the claims review end, forming a unified, archiveable, and auditable application document.

[0024] Furthermore, the method for constructing and optimizing the diagnostic-material mapping knowledge base includes: By mining common diagnoses and corresponding material combinations from historical claims data and combining them with expert rules to generate initial mapping relationships, an initial diagnosis-material mapping knowledge base is constructed. Subsequently, feedback data from the manual review process is collected, and the mapping rules are periodically trained and optimized using the feedback data to achieve adaptive optimization of the diagnosis-material mapping knowledge base.

[0025] The beneficial effects of the intelligent claims material dynamic guidance method based on diagnosis-driven method of the present invention are as follows: (1) Based on diagnosis-driven missing detection, it can accurately determine the required materials according to different diseases, avoid omissions or redundancies caused by general lists, improve the one-time pass rate of materials, and avoid omissions or repeated supplementary materials due to mismatch between material requirements and actual situation; (2) Instant feedback on missing materials during the user upload process, without waiting for manual review, solve the problem before submission, shorten the average processing time of cases; and there is no need to understand and memorize the supplementary material requirements on your own. The card notification is linked with the front-end upload interface, associated with the type of supplementary materials, and provides real-time shooting and uploading function, realizing seamless connection from detection to explanation to guidance, and improving user experience; (3) Combining multimodal large model and self-optimizing knowledge base, it can continuously learn new diagnosis-material relationships, and reduce the workload of manual review. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a diagnostic-driven intelligent claims material dynamic guidance method according to the present invention. Detailed Implementation

[0027] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0028] like Figure 1 As shown, this application provides a diagnostic-driven intelligent claims material dynamic guidance method, including the following steps: S1: Receive the images of claim materials uploaded by the user, perform quality inspection on each image, and then use a trained image classification model to automatically classify the claim material images into preset material types.

[0029] Furthermore, the quality inspection includes checking the clarity, brightness, and tilt of the claims material images, and prompting unqualified images to be retaken; the image classification model adopts a low-cost, self-evolving training scheme based on transfer learning and active learning, thereby obtaining a high-performance classification model with only a small amount of manual annotation; the preset material types include at least "hospitalization invoice", "discharge summary", "diagnosis certificate" and "expense list".

[0030] Furthermore, the specific training steps for the image classification model are as follows: 1) Select a lightweight convolutional neural network (MobileNetV3 or ResNet50) pre-trained on the general image dataset ImageNet as the base model, freeze the shallow feature extraction layer, and only retrain the final classification layer. This strategy utilizes the existing general feature extraction capabilities of the pre-trained model to reduce the need for manual annotation from hundreds of thousands to thousands. 2) In terms of training data construction, a combination of weakly supervised automatic annotation and active learning mechanisms is used. On the one hand, images uploaded by users through specified material type entry points are extracted from historical claims business logs, and the corresponding material type is automatically assigned as a "weak label," thereby obtaining tens of thousands to hundreds of thousands of noisy weakly labeled training data at no cost. On the other hand, only 200-300 typical images for each material type (approximately 2000-3000 images in total) need to be manually labeled as a high-quality seed training set. 3) The following active learning iterative strategy is adopted: first, the initial image classification model is trained with the seed training set, and then the image classification model is used to predict the automatically labeled weakly labeled data to obtain the confidence score; 4) High-confidence samples with a confidence level higher than 0.95 are automatically added to the seed training set, while suspected erroneous samples with a confidence level lower than 0.6 are submitted for manual review; the reviewed and corrected samples are also added to the seed training set to retrain the image classification model.

[0031] S2: For classified claim material images, key structured information is extracted from the classified claim material images by combining OCR technology with multimodal large model.

[0032] Furthermore, this step designs a hierarchical, complementary, and confidence-based intelligent extraction strategy to balance extraction efficiency and accuracy. The specific process is as follows: 1) The mature OCR engine PaddleOCR is used to quickly and fully recognize the classified claims material images, directly extracting structured fields such as "invoice number", "date of visit", "medical expenses", "surgical record" and "anesthesia method", and attempting to locate diagnostic information; 2) Determine whether the diagnostic information extracted by the OCR engine meets the following conditions: the confidence level is greater than or equal to the preset confidence threshold of 0.7, and the current claim material image belongs to a structured document; if both conditions are met, it is determined to be a high-confidence scenario. At this time, the multimodal large model is no longer called, and the diagnostic information and other structured fields extracted by the OCR engine are directly used as the key structured information, and then directly proceed to step 4). This branch processing can avoid unnecessary calls to the large model API, reduce operating costs and response latency; 3) If the diagnostic information extracted by the OCR engine is empty, or the confidence level is lower than the preset confidence threshold, or the document being processed is unstructured, the multimodal large model Qwen-VL will be automatically triggered, and the multimodal large model Qwen-VL will be guided by carefully designed prompts to output key structured information in JSON format, including at least diagnostic information, invoice number, consultation date, and medical expenses. 4) The authority of diagnostic information sources is prioritized. When different sources extract the same information but yield inconsistent results, the result from the higher-priority source is automatically adopted, and the lower-priority result is marked as "pending review" for subsequent manual review. Specifically, the priority relationship is: results extracted by the multimodal large model from "Discharge Summary" > results extracted by the multimodal large model from "Diagnosis Certificate" > results extracted by the OCR engine from "Admission Record" > others. Through this hierarchical calling and arbitration mechanism, the frequent calls to the multimodal large model API are reduced (only enabled when processing critical unstructured materials) while ensuring extraction accuracy, achieving an optimal balance between cost and performance.

[0033] S3: Using the diagnostic information extracted from the key structured information as query conditions, access the preset diagnostic-material mapping knowledge base, obtain the list of required materials corresponding to the current diagnosis, and generate a list of missing materials by comparison.

[0034] In this embodiment, the ICD-10 diagnostic code is used as the diagnostic information; Furthermore, the diagnosis-materials mapping knowledge base contains a mapping relationship between ICD-10 diagnosis codes and the types of materials required for claims. The core data table includes fields such as "ICD-10 diagnosis code," "basic materials list," "applicable conditions (whether hospitalized or surgical)," "condition matching value," and "additional materials," and supports dynamic adjustment based on diagnosis conditions (hospitalized / outpatient, surgical / non-surgical). The basic materials list field contains the commonly required materials under the diagnosis; the applicable conditions field specifies the case context in which the mapping record takes effect; the condition matching value field specifies the specific value of the condition corresponding to the case context; the additional materials field contains materials that need to be added when the applicable conditions are met; and the excluded materials field contains materials that need to be deleted from the basic materials list when the applicable conditions are met.

[0035] In this embodiment, for the diagnosis of pneumonia, when the "whether hospitalized" condition is true, the basic materials list is [diagnosis certificate, discharge summary], and the additional materials are [hospitalization invoice, expense list]; when the "whether hospitalized" condition is false, the basic materials list is [diagnosis certificate], the additional materials are [outpatient invoice, medical record], and the excluded material is [discharge summary].

[0036] Specifically, after obtaining the user's key structured information and case context, all relevant mapping records are queried using the diagnostic information as the primary key. Then, the corresponding condition records are matched according to the case context, and the final list of required materials is dynamically generated. The list of required materials = basic materials list + additional materials corresponding to condition matching - excluded materials when matching conditions (if there are no additional materials or excluded materials, the corresponding items are ignored). Then, the uploaded material types are compared with the list of required materials to generate a list of missing materials.

[0037] The case context refers to a set of auxiliary attributes related to the current claim case, excluding diagnostic information, used to refine the list of required materials. The case context includes at least the type of treatment (inpatient / outpatient / emergency), whether surgery is involved (yes / no), and the department visited, which are obtained through the following two methods: ① Front-end data collection: basic information filled in by the user at the claim application portal (such as selecting "inpatient claim" or "outpatient claim"), or by checking options such as "whether surgery is required"; ② Inference from extracted structured information. In this embodiment, the process of inferring from the extracted structured information is as follows: if step S2 extracts the keyword "surgical record" or the field "anesthesia method" from the "discharge summary", then "whether surgery is involved" is automatically marked as "yes"; if the extracted key structured information contains "invoice number" but the material type does not include "discharge summary", then the default "visit type" is inpatient.

[0038] Furthermore, the missing materials in the missing materials list are graded and marked: a basic importance level is predefined for each material type, including mandatory, critical, and recommended. The mandatory missing materials directly result in failure to meet the basic requirements for filing a case, requiring the user to submit them again. The critical missing materials result in the inability to determine core responsibility or amount, but are not an absolute necessity for filing a case. The recommended missing materials are used to assist in the review process, do not affect filing a case, but help speed up the review; the user can choose to submit them again. A set of dynamic upgrade and downgrade rules based on diagnosis and conditions is also built-in. Then, the missing materials list is traversed, and for each missing material, its basic importance level is first obtained, then the dynamic upgrade and downgrade rules are applied to adjust it to obtain the final level. Finally, missing materials with the final level of mandatory or critical are marked as mandatory, and the remaining missing materials are marked as recommended.

[0039] Furthermore, the dynamic upgrade and downgrade rules are a set of built-in decision rules based on diagnostic information and case context, used to dynamically adjust the importance level of the material type. The general logic is as follows: if the diagnostic information belongs to the preset diagnostic coding range (injury and poisoning category S00-T98 in ICD-10 diagnostic coding) or the case context meets the value of the preset applicable condition field (whether hospitalization is true, whether surgery is true), then the importance level of the material type will be adjusted from the original level to the new level.

[0040] By using the aforementioned conditional queries and dynamic hierarchical marking, the importance of materials in different diagnostic scenarios can be accurately distinguished, avoiding unreasonable mandatory requirements on users, while ensuring that core claims evidence is not omitted.

[0041] S4: Based on the list of missing materials, generate targeted guidance information to guide users to complete the missing materials in real time.

[0042] Furthermore, the targeted guidance information includes: the type and reason for the missing material, sample shooting images, and key shooting tips.

[0043] Targeted guidance information is displayed in real time through the app interface or mini-program, and an "Instant Photo" entry is provided. After the user uploads additional materials, steps S1-S3 are repeated until the user has submitted all the required materials.

[0044] Furthermore, the generation of the targeted guidance information is based on a combination of missing material type, missing reason, and user profile in the generation and rendering process. Specifically, a guidance content material library is pre-built, storing the following structured elements for each material type: 1. Missing material type and reason text template ("Based on your diagnosis [{diagnosis_name}], a [diagnosis certificate] is required to confirm the ICD-10 diagnostic code, which is a necessary basis for claim filing"); 2. Standard shooting example images (URLs of 2-3 qualified shooting images from different angles for each material); 3. List of shooting key points prompts text. The missing material list is traversed to obtain each missing material item. The corresponding missing material type and reason text template is retrieved from the guide content material library. The variables in the missing material type and reason text template are replaced with actual values, and assembled into a guide card object containing the missing material type and reason, the URL of the standard shooting example image, shooting points, and the "Shoot Now" action button. After the front-end APP or mini-program receives the guide card object, it displays the guide card objects to the user in a list through the dynamic rendering engine. The "Shoot Now" button on each guide card object is pre-associated with the material type to be uploaded. After the user clicks it, the camera is opened directly and the upload is automatically completed. Once the upload is complete, the closed-loop detection process of steps S1-S3 is re-executed.

[0045] In this embodiment, the missing material type and reason text in the guide card object is "Based on your diagnosis of [pneumonia], you need to supplement [diagnosis certificate] to confirm the ICD-10 diagnosis code, which is a necessary basis for filing a claim", and the shooting points are "Please ensure that the hospital's official seal is clearly visible and all four corners are fully visible in the shot". The guide card object adopts JSON format.

[0046] Through the above solution, the present invention achieves a seamless connection from "detecting missing data" to "explaining the reasons" and then to "guiding resubmission", transforming complex review rules into intuitive guidance that is easy for users to understand, and significantly improving resubmission efficiency and user experience.

[0047] S5: Arrange the classified claim materials images in a preset order, and package them together with the extracted key structured information and targeted guidance information to generate a claim pre-application package. After the user confirms, submit the claim pre-application package to the claim review end for subsequent verification.

[0048] Once all necessary materials are complete, the categorized claim materials images are arranged in a preset order: invoice, discharge summary, diagnosis certificate, expense list, and identity document. These materials are then packaged together with extracted key structured information and targeted guidance information to generate a structured claim pre-application package for user confirmation. After the user confirms that everything is correct, the claim pre-application package is submitted to the insurance institution's server, which is used to process claims and for reviewers to verify materials and process claims. This server is the claim review server.

[0049] Subsequently, insurance companies can use the aforementioned pre-claim application package for claims processing, risk control, and financial disbursement, which facilitates traceability, archiving, and compliance checks, reduces repeated communication and document returns at the claims processing end, and significantly shortens the overall claims and review process cycle.

[0050] Furthermore, the method for constructing and optimizing the diagnostic-material mapping knowledge base includes: By mining common diagnoses and corresponding material combinations from historical claims data, and combining them with expert rules (clinical pathways, medical insurance regulations) to generate initial mapping relationships, an initial diagnosis-material mapping knowledge base is constructed. Subsequently, feedback data from the manual review process is collected, and the mapping rules are periodically trained and optimized using the feedback data to achieve adaptive optimization of the diagnosis-material mapping knowledge base.

[0051] The present invention has been described by the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

[0052] The contents of this invention not described in detail are existing technologies known to those skilled in the art.

Claims

1. A diagnostic-driven intelligent claims material dynamic guidance method, characterized in that, Includes the following steps: S1: Receive the images of claim materials uploaded by the user, perform quality inspection on each image, and then use the trained image classification model to automatically classify the claim material images into preset material types; The quality inspection includes checking the clarity, brightness, and tilt of the claims material images, and prompting unqualified images to be retaken; the image classification model adopts a low-cost, self-evolving training scheme based on transfer learning and active learning, thereby obtaining a high-performance classification model with only a small amount of manual annotation; the preset material types include at least "hospitalization invoice", "discharge summary", "diagnosis certificate" and "item list"; S2: For classified claim material images, key structured information is extracted from the classified claim material images by combining OCR technology with multimodal large model; Once the user's key structured information and case context are obtained, all relevant mapping records are queried using the diagnostic information as the primary key. Then, the corresponding condition records are matched according to the case context, and the final list of required materials is dynamically generated. The list of required materials = basic list of materials + additional materials corresponding to condition matching - materials excluded when matching conditions. Then, the uploaded material types are compared with the list of required materials to generate a list of missing materials; The case context refers to a set of auxiliary attributes related to the current claim case, excluding diagnostic information, used to refine the list of required materials. The case context includes at least "type of treatment", "whether surgery is involved", and "department of treatment". The case context is obtained in the following two ways: ① front-end data collection: basic information filled in by the user at the claim application portal, or obtained through selected options; ② inferred from extracted structured information. S3: Use the diagnostic information in the extracted key structured information as the query condition, access the preset diagnosis-material mapping knowledge base, obtain the list of necessary materials corresponding to the current diagnosis, and compare and generate a list of missing materials. S4: Based on the list of missing materials, generate targeted guidance information to guide users to complete the missing materials in real time; The targeted guidance information includes: the type and reason for the missing materials, sample photos, and tips for taking photos; after the user uploads the missing materials, steps S1-S3 are repeated until the user has uploaded all the necessary materials. S5: Arrange the classified claim materials images in a preset order, and package them together with the extracted key structured information and targeted guidance information to generate a claim pre-application package. After the user confirms, submit the claim pre-application package to the claim review end for subsequent verification. Once all the necessary materials are complete, a structured pre-claim package is generated for user confirmation. After the user confirms that everything is correct, the pre-claim package is submitted to the claims review end, forming a unified, archiveable, and auditable application document.

2. The method according to claim 1, characterized in that, The specific training steps for the image classification model described in step S1 are as follows: 1) Select a lightweight convolutional neural network pre-trained on a general image dataset as the base model, freeze the shallow feature extraction layer, and only retrain the classification layer at the end. 2) In terms of training data construction, we integrate two mechanisms: weakly supervised automatic labeling and active learning. We extract images uploaded by users through the specified material type entry from historical claims business logs and automatically assign them the corresponding material type as "weak labels", thereby obtaining tens of thousands to hundreds of thousands of noisy training weak label data at no cost. 3) The following active learning iterative strategy is adopted: first, the initial image classification model is trained with the seed training set, and then the image classification model is used to predict the automatically labeled weakly labeled data to obtain the confidence score; 4) High-confidence samples with confidence levels higher than α are automatically added to the seed training set, while suspected erroneous samples with confidence levels lower than β are submitted for manual review; the reviewed and corrected samples are also added to the seed training set to retrain the image classification model; where α and β are both values ​​greater than 0 and less than 1, and α > β.

3. The method according to claim 1, characterized in that, The specific process of step S2 is as follows: 1) Utilize a mature OCR engine to quickly and fully recognize categorized claim material images, directly extracting structured fields such as invoice number, consultation date, and medical expenses, and attempting to locate diagnostic information; 2) Determine whether the diagnostic information extracted by the OCR engine meets the following conditions: the confidence level is greater than or equal to the preset confidence threshold, and the current claim material image belongs to a structured document; if both conditions are met, it is determined to be a high-confidence scenario. At this time, the multimodal large model is no longer called, and the diagnostic information and other structured fields extracted by the OCR engine are directly used as the key structured information, and then proceed directly to step 4). This can avoid unnecessary calls to the large model API, reduce system operating costs and response latency. 3) If the diagnostic information extracted by the OCR engine is empty, or the confidence level is lower than the preset confidence threshold, or the document being processed is unstructured, the multimodal large model will be automatically triggered and the multimodal large model will be guided by carefully designed prompts to output key structured information that includes at least the diagnostic information, invoice number, consultation date, and medical expenses. 4) Preset the authority priority of diagnostic information sources. When different sources extract the same information and the results are inconsistent, the result from the source with higher priority will be automatically adopted, and the result from the lower priority source will be marked as "pending review" for subsequent manual review. For diagnostic information, priority will be given to extracting it from materials of the type "discharge summary" and "diagnosis certificate".

4. The method according to claim 1, characterized in that, The diagnosis-material mapping knowledge base mentioned in step S3 contains the mapping relationship between diagnosis information and the types of materials required for claims. The core data table includes fields such as "Diagnosis Information", "Basic Material List", "Applicable Conditions", "Condition Matching Value", "Additional Materials", and "Excluded Materials", and supports dynamic adjustment based on diagnosis conditions. Among them, the Basic Material List field is the general materials that are usually required under this diagnosis; the Applicable Conditions field is used to specify the case context in which this mapping record takes effect; the Condition Matching Value field is the specific value of the condition corresponding to the case context; the Additional Materials field is the materials that need to be added when the Applicable Conditions are met; and the Excluded Materials field is the materials that need to be deleted from the Basic Material List when the Applicable Conditions are met.

5. The method according to claim 1, characterized in that, Step S3 further performs hierarchical labeling on the missing materials in the missing materials list: a basic importance level is predefined for each material type, including mandatory, critical, and recommended; among them, the missing mandatory materials directly result in failure to meet the basic requirements for case filing, requiring the user to submit them again; the missing critical materials result in the inability to verify the core responsibility or amount, but are not an absolute necessity for case filing; the missing recommended materials are used to assist in the review, do not affect case filing, but help speed up the review, and the user can choose to submit them again; at the same time, a set of dynamic upgrade and downgrade rules based on diagnosis and conditions are built-in; then, the missing materials list is traversed, and for each missing material, its basic importance level is first obtained, and then the dynamic upgrade and downgrade rules are applied to adjust it to obtain the final level; finally, the missing materials with the final level of mandatory or critical are marked as mandatory level, and the remaining missing materials are marked as recommended level.

6. The method according to claim 5, characterized in that, The dynamic upgrade and downgrade rules are a set of built-in decision rules based on diagnostic information and case context, used to dynamically adjust the importance level of material types. The general logic is: if the diagnostic information belongs to the preset diagnostic coding range or the case context meets the value of the preset applicable condition field, then the importance level of the material type is adjusted from the original level to the new level.

7. The method according to claim 1, characterized in that, The generation of targeted guidance information in step S4 is based on a combination of missing material type, missing reason, and user profile generation and rendering process. Specifically, a guidance content material library is pre-built, storing the following structured elements for each material type:

1. Missing material type and reason text template; 2. Standard shooting example image; 3. Shooting key point prompt text list; then, the missing material list is traversed to obtain each missing material item, and the corresponding missing material type and reason text template is obtained from the guidance content material library. The variables in the missing material type and reason text template are replaced with actual values, and assembled into a guidance card object containing the missing material type, reason text, standard shooting example image URL, shooting key points, and "Shoot Now" action button; after the front-end APP or mini-program receives the guidance card object, it displays the guidance card object to the user in a list through the dynamic rendering engine. The "Shoot Now" button on each guidance card object is pre-associated with the material type to be uploaded. After the user clicks it, the camera is opened directly and the upload is automatic. Once the upload is complete, the closed-loop detection process of steps S1-S3 is re-executed.

8. The method according to claim 1, characterized in that, The method for constructing and optimizing the diagnosis-material mapping knowledge base includes: mining material combinations corresponding to common diagnoses from historical claims data, generating initial mapping relationships by combining expert rules, and constructing an initial diagnosis-material mapping knowledge base; subsequently collecting feedback data from the manual review process, and periodically using the feedback data to train and optimize the mapping rules to achieve adaptive optimization of the diagnosis-material mapping knowledge base.

Images