Intelligent insurance underwriting method

Abstract

The invention discloses an intelligent insurance underwriting method. Insurance requirements, health notification information and multi-mode physical examination data of a user are obtained through a dialogue interaction interface; extracting text physiological indexes by adopting a named entity recognition technology, analyzing medical images through a visual language model to generate pathological descriptions, and mapping the pathological descriptions into standard medical term codes; performing preliminary screening based on a decision tree rule engine, and calculating a deviation degree of health indexes and terms in combination with a disease correlation model to screen candidate products; constructing a hierarchical cue word template to drive the large language model to execute clause semantic analysis, and outputting a structured underwriting conclusion; and generating a product recommendation sequence by using a weighted scoring model, or triggering a difference comparison description. Through multi-modal data fusion, dynamic deviation calculation and an anti-factual reasoning mechanism, the problems that unstructured data is difficult to process and clause matching is rigid in traditional underwriting are solved, the underwriting accuracy is effectively improved, the decision-making efficiency is remarkably improved, real-time synchronous updating of supervision clauses is supported, and the compliance risk is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the field of insurance underwriting, and in particular to a method for intelligent insurance underwriting. Background Art

[0002] Intelligent underwriting is crucial for insurance purchases. Currently, there are numerous insurance products with complex terms and conditions, and the underwriting process is cumbersome. These products rely primarily on the review and judgment of insurance personnel, or through methods like condition identification and questionnaire analysis. This often takes a lengthy review process, which reduces insurance purchase efficiency. Rapidly analyzing user situations and recommending suitable insurance products is a key issue in insurance underwriting. Complicated review methods make it difficult to accurately and intelligently analyze user situations. For example, user-provided medical examination information with images is not easily analyzed. Leveraging the power of large language models, we are designing new intelligent underwriting methods that can more quickly and effectively analyze user needs, health conditions, and other aspects, and provide suitable insurance products.

[0003] Therefore, a more efficient method of intelligent insurance underwriting is needed. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method for intelligent insurance underwriting, which can quickly analyze the user's needs and health conditions, and recommend suitable insurance products based on insurance terms.

[0005] A method for intelligent insurance underwriting, characterized by comprising the following steps:

[0006] (a) User information collection: Obtaining the user's insurance demand description, health information questionnaire data, and physical examination report input through the dialogue interface. The physical examination report contains structured text and unstructured medical images;

[0007] (b) multimodal information processing: extracting disease keywords from the health information questionnaire data, performing multimodal analysis on the physical examination report document, and generating a health risk feature vector containing text indicators and image features;

[0008] (c) Insurance terms matching: User feature data is input into a pre-trained insurance terms analysis model, which performs preliminary screening based on a decision tree rule engine and outputs a set of products that meet insurance eligibility.

[0009] (d) Intelligent underwriting decision-making: Build a hierarchical prompt template to drive the large language model to perform clause semantic analysis, generating a three-part structured analysis result consisting of [underwriting conclusion] - [matching clause] - [exclusion reason];

[0010] (e) Comprehensive recommendation generation: An approximate ideal solution ranking algorithm is used to calculate the matching score of each product and generate an underwriting recommendation report that includes a recommended ranking, risk warnings, and alternative solutions.

[0011] The multimodal analysis in step (b) includes:

[0012] Use named entity recognition technology to extract key physiological indicators from the physical examination report text;

[0013] Generate pathology description text from medical imaging data through visual language models;

[0014] Establish an associated data structure of <indicator name>-<detection value>-<image evidence>.

[0015] The decision tree rule engine of step (c) implements a two-stage validation:

[0016] (c1) Generate an initial screening exclusion list based on hard condition filtering of age and occupation;

[0017] (c2) The remaining products are calculated using a disease association model to determine the degree of deviation between the user's health indicators and the terms and conditions. Products with a degree of deviation less than or equal to a preset threshold are retained as candidate products. The degree of deviation is calculated using the formula:

[0018]

[0019] Among them, xi is the user's health index value, yi is the clause threshold, and αi is the disease risk weight coefficient.

[0020] The layered prompt template of step (d) includes:

[0021] Role definition layer: Set the large language model as the interactive identity of the underwriting expert;

[0022] Dynamic data layer: inject user feature fields and target insurance clause keywords;

[0023] Verification constraint layer: embeds counterfactual reasoning instructions and requires the output of a structured response containing a clause number reference.

[0024] The matching score in step (e) is calculated using the following formula:

[0025]

[0026] Among them, Si is the matching score of the i-th product, wj is the dynamic weight of the j-th evaluation dimension, dij+ represents the Euclidean distance from the ideal solution, and dij- represents the Euclidean distance from the negative ideal solution.

[0027] After step (e), the method further comprises the steps of:

[0028] (f) Enhanced traceability: Underwriting reports will include traceability links to the original clauses in the insurance regulatory database;

[0029] (g) Dynamic update mechanism: When the terms of the regulatory database change, the large language model is triggered to perform incremental training and update the decision threshold.

[0030] Beneficial effects:

[0031] The present invention obtains the user's insurance needs, health disclosure information and multimodal physical examination data (including text and medical images) through a dialogue interaction interface; uses named entity recognition technology to extract text physiological indicators, uses visual language models to parse medical images to generate pathological descriptions, and maps them into standard medical terminology codes; based on the decision tree rule engine, the user's age, occupation and other hard conditions are preliminarily screened, and the deviation between health indicators and terms is calculated in combination with the disease correlation model, and candidate products with deviations that meet the requirements are screened; a hierarchical prompt template is constructed to drive the large language model to perform semantic analysis of terms and output structured underwriting conclusions; a weighted scoring model is used to generate product recommendation rankings, and a difference comparison description is triggered when the difference in scores between the best and second-best products reaches a threshold. The present invention solves the problems of difficult unstructured data processing and rigid terms matching in traditional underwriting through multimodal data fusion, dynamic deviation calculation and counterfactual reasoning mechanism, significantly improving underwriting accuracy and decision-making efficiency, and supporting real-time synchronous updates of regulatory terms, significantly reducing compliance risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A method flow chart of an exemplary embodiment of the present invention;

[0033] Figure 2 Module workflow diagram of an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0034] like Figure 1 As shown, in an exemplary embodiment of this case, a method for intelligent insurance underwriting includes the steps of:

[0035] (a) User information collection: Obtaining the user's insurance demand description, health information questionnaire data, and physical examination report input through the dialogue interface. The physical examination report contains structured text and unstructured medical images;

[0036] (b) multimodal information processing: extracting disease keywords from the health information questionnaire data, performing multimodal analysis on the physical examination report document, and generating a health risk feature vector containing text indicators and image features;

[0037] (c) Insurance terms matching: User feature data is input into a pre-trained insurance terms analysis model, which performs preliminary screening based on a decision tree rule engine and outputs a set of products that meet insurance eligibility.

[0038] (d) Intelligent underwriting decision-making: Build a hierarchical prompt template to drive the large language model to perform clause semantic analysis, generating a three-part structured analysis result consisting of [underwriting conclusion] - [matching clause] - [exclusion reason];

[0039] (e) Comprehensive recommendation generation: An approximate ideal solution ranking algorithm is used to calculate the matching score of each product and generate an underwriting recommendation report that includes a recommended ranking, risk warnings, and alternative solutions.

[0040] Specifically, a large language model is preset, and the large language model is trained using insurance underwriting data. The large language model is used to conduct inquiry interactions with users to obtain inquiry data for analysis, and the inquiry data is identified and classified. The inquiry data includes insurance data, health disclosure information, and physical examination reports. The inquiry data are input into a preset insurance data analysis module, a health disclosure structured sub-module, and a physical examination report analysis module respectively; the large language model is used to analyze the inquiry data to generate recommendations for insurance products. Based on the user input and the recommendations for insurance products, the large language model is used to generate a comprehensive underwriting situation. The comprehensive underwriting situation includes recommended insurance products and additional information, and the additional information includes disclaimers.

[0041] The insurance data analysis module is used for clause text segmentation and question-answering. The segmentation utilizes a paragraph segmentation method to divide the complete clause into subsections to obtain individual clause points. The divided clause documents are constructed into an insurance clause knowledge base, which is used as input to train microcalls to answer questions about insurance clauses. The insurance data analysis module uses the Llama2 model and includes a conversation history parsing submodule that uses contextual understanding technology to extract user attribute features.

[0042] The suitability judgment is performed in the insurance data analysis module, and the suitability judgment is used to establish an insurance product database and an insurance terms rule set, perform preliminary screening through the rule engine, and generate a large language model prompt for the insurance products that pass the screening.

[0043] In the above applicability judgment, the user's insurance eligibility is determined through a double verification mechanism, including:

[0044] Build a rule engine pre-screening unit: Build a decision tree model based on the insurance clause rule set, perform Boolean logic verification on the user's age, occupation category, and medical history, and generate an initial screening exclusion list;

[0045] Construct a semantic understanding verification unit: Use a large language model to perform contextual semantic analysis on the disclaimer clauses of products that have passed the initial screening, and use a disease correlation calculation model to determine the deviation between the user's health indicators and the clause requirements. When the deviation is ≤ the preset threshold, mark it as an insurable candidate product. The deviation calculation uses the formula:

[0046]

[0047] Among them, xi is the user's health index value, yi is the clause threshold, and αi is the disease risk weight coefficient, which is determined based on clinical data statistics.

[0048] If the user's blood pressure is 150 / 95 mmHg, a critical illness insurance policy requires blood pressure ≤ 140 / 90

[0049] The disease risk weight coefficient α was set to 0.7 (determined based on clinical data statistics);

[0050] Application formula:

[0051]

[0052] The deviation threshold is set to 10%, and the product is determined to be a candidate.

[0053] This solution avoids the absolute exclusion inherent in traditional Boolean logic, increasing the number of potentially insurable products. A dynamic weighting mechanism improves underwriting approval rates for chronic disease patients. This solution optimizes the matching efficiency of unstructured medical images with policy terms through a dynamic deviation calculation model, breaking through the technical bottleneck of traditional underwriting data processing.

[0054] A health information knowledge base is constructed in the health notification structured submodule, and the health information knowledge base includes matching information of health information and diseases. When a user inquires about a related disease, text matching is performed. If the match is successful, health notification is triggered.

[0055] Intelligent underwriting is performed in the health notification structured submodule. The intelligent underwriting is used to construct a prompt template that matches the insurance terms, input the user's comprehensive information into the large language model through the natural language processing interface, and obtain the underwriting feasibility analysis of each product and the detailed description of the terms that are not met.

[0056] The prompt template in the intelligent underwriting adopts a hierarchical structure design:

[0057] The first layer contains role-defining instructions: "As a professional insurance underwriting assistant, you need to conduct an insurance feasibility analysis based on the user's objective information";

[0058] The second layer embeds dynamic parameter placeholders, including user feature fields and insurance clause key items;

[0059] The third layer sets the output format specification, requiring a three-part structured response of [underwriting conclusion]-[matching terms]-[reasons for exclusion].

[0060] The prompt template includes a dynamic element injection mechanism, including:

[0061] Context-sensitive instruction layer: automatically selects the strength of professional term explanation based on user characteristics, and inserts a simplified explanation module when it is detected that the user is a non-professional;

[0062] Structured comparison layer: Generates a comparison matrix template containing [insurance product]-[key terms]-[user matching points], forcing the large language model to perform item-by-item comparison analysis;

[0063] Counterfactual reasoning constraint layer: embeds verification instructions in the prompt to determine whether the predetermined indicator exceeds the threshold.

[0064] Template structure (example):

[0065] {

[0066] "role":"As a senior underwriter, you must strictly analyze the feasibility of insurance according to the following terms",

[0067] "dynamic_data":{

[0068] "user_info": "Age 35, blood pressure 150 / 95 mmHg, lung nodule 3 cm",

[0069] "clause":"Case 3.2 of Insurance Clause A: Systolic blood pressure ≤ 140 mmHg and no history of malignant tumor"

[0070] },

[0071] "constraints":[

[0072] "The output must include the clause number, format: [Conclusion]-[Clause]-[Reason]",

[0073] "Assuming the user's blood pressure drops to 130 mmHg, please re-evaluate the conclusion." ]

[0075] }

[0076] Using layered prompts can effectively improve user experience and increase conversion rates.

[0077] The physical examination report analysis module uses named entity recognition technology to extract key health indicators and extract and analyze text on the physical examination report. The physical examination report analysis module includes a text detection model OCR for extracting text information from images. If the physical examination report contains pathological images, the pathological images are saved and the multimodal understanding model LLAVA is used to perform intelligent analysis on the pathological images to obtain relevant pathological information. The text information and the pathological information constitute the multimodal information of the physical examination report.

[0078] The use of multimodal information fusion technology in the medical examination report analysis module can improve underwriting accuracy, effectively reduce claims disputes, and lower operating costs. Specifically, it includes:

[0079] Conversation history feature extraction: Use a temporal attention network to capture key statements in the conversation history and build a user risk feature vector;

[0080] Physical examination report parsing engine: converts unstructured test data into standard SNOMEDCT coding through medical ontology mapping;

[0081] Evidence chain construction unit: Create a triple knowledge graph containing <data source>-<extraction time>-<confidence level> to provide traceable basis for underwriting decisions.

[0082] For example, a user uploads a medical examination report containing text and a chest X-ray:

[0083] Text parsing:

[0084] Use the pre-trained BERT-NER model to fine-tune the medical examination report text and extract key indicators in the text (such as "blood pressure: 150 / 95 mmHg").

[0085] Output structured data: {Indicator: Blood Pressure, Value: 150 / 95, Unit: mmHg, Confidence: 0.92}.

[0086] Image analysis:

[0087] The ViT-LSTM visual language model is used to analyze X-rays and generate a description text: "A 3 cm ground-glass nodule is visible in the left lower lobe of the lung."

[0088] The SNOMED CT code maps to: 74400008 | Pulmonary nodule (disorder).

[0089] This solves the problem of low utilization of unstructured data and improves the completeness of medical examination report analysis. By linking medical images with text indicators, the accuracy of underwriting risk identification is improved.

[0090] The large language model generates a comprehensive recommendation based on the comprehensive underwriting situation. The comprehensive recommendation uses a multi-dimensional evaluation algorithm to integrate the underwriting results of various insurance products, triggers the large language model to perform comparative analysis through a secondary prompt generation strategy, and outputs the specific exclusion basis for the recommended product ranking and rejected products; feedback generation is executed after the comprehensive recommendation, and the feedback generation is used to contextually associate the recommendation results with the conversation history to generate a personalized recommendation report containing insurance suggestions, risk warnings and alternative plans.

[0091] The approximate ideal solution sorting algorithm is used to calculate the matching score of each product. The matching score is calculated using the following formula:

[0092]

[0093] Among them, Si is the matching score of the i-th product, wj is the dynamic weight of the j-th evaluation dimension, dij+ represents the Euclidean distance from the ideal solution, and dij- represents the Euclidean distance from the negative ideal solution.

[0094] Large language model output (example):

[0095] [Partial Coverage] - [Clause A3.2] - [Blood pressure exceeds the normal limit by 10%, but the lung nodules have been ruled out as malignant]

[0096] [Counterfactual analysis]: If blood pressure is ≤140mmHg, it can be upgraded to full coverage

[0097] The present invention is applied to serve various insurance products and users. It needs to input user context query information and health information in sequence to determine whether it is applicable to the relevant insurance products. If not, it explains which regulations are not met.

[0098] Another possible prompt template is: You are playing the role of an insurance sales assistant. Below is the user's insurance information, as well as health information and physical examination report information, such as "***". Please judge the following insurance products based on the user's actual information. Please give reasons for whether the user can be insured. The insurance requirements are as follows: "***".

[0099] For comprehensive recommendations: We obtain the insurance results for each product and then use the large language model to comprehensively identify the best product and provide reasons why other products are unsuitable. Based on this information, we provide the user with insurance-related feedback. This is equivalent to evaluating the underwriting of each product. The large language model draws conclusions and reasons based on the user's actual situation before providing a conclusion. Possible prompt templates: Please combine the following insurance scenarios: "***" to provide the user's insurance conclusion and provide detailed explanations.

[0100] Enhanced traceability: The underwriting report is marked with a traceability link to the original clause in the insurance regulatory database;

[0101] Dynamic update mechanism: When the terms of the regulatory database change, the large language model is triggered to perform incremental training and update the decision threshold.

[0102] When the regulatory database detects a change in any clause threshold, the system automatically captures updated data and triggers model fine-tuning. For example, when the regulatory database detects a clause change (such as the relaxation of insurance restrictions for hepatitis B carriers), the following process is automatically triggered:

[0103] Capture the original text of the change terms and extract key thresholds (such as the ALT test value is adjusted from 80U / L to 100U / L).

[0104] Based on incremental data (1,000 historical underwriting cases), the large language model is fine-tuned with Lora, and the Yi parameter in the deviation calculation formula is updated.

[0105] Verification after update: 87% of the 120 users (ALT 85-99U / L) who were originally denied insurance were eligible for coverage.

[0106] The response time for terms updates is significantly shortened compared to manual processing. This avoids customer churn caused by policy delays, expands the potential customer base, and improves user satisfaction.

[0107] The present invention will be further illustrated below in conjunction with another specific embodiment. The embodiment is implemented based on the technical solution of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0108] As attached Figure 2 As shown, the intelligent insurance underwriting method of the present invention realizes multi-dimensional analysis of user information and decision optimization through a modular process, which specifically includes the following technical modules and their interactive relationships:

[0109] User inquiry analysis module: receives user-entered insurance demand description, health disclosure questionnaire data and physical examination report (including structured text and unstructured medical images) through a dialogue interaction interface.

[0110] Technical implementation: Use a large language model (such as LLaMA2) to identify and classify user questions, determine whether they belong to one of the categories of "insurance terms inquiry", "health disclosure review", "physical examination report analysis" or "comprehensive underwriting request", and trigger the corresponding downstream module.

[0111] Data flow: The user's original input data is broken down into structured fields and unstructured data, and transmitted to the insurance terms data processing and analysis module, health disclosure information analysis module, and physical examination report analysis module respectively.

[0112] Insurance clause data processing and analysis module: performs semantic analysis and dynamic matching of insurance clauses.

[0113] Technical Implementation: Based on a pre-trained insurance clause knowledge base (built through paragraph segmentation and named entity recognition), a decision tree rule engine is used to perform two-stage verification:

[0114] Initial screening stage: Generate an exclusion list based on hard conditions such as user age and occupation;

[0115] Semantic verification stage: Calculate the deviation between the user's health indicators and the clause threshold through the disease association model to screen the candidate product set.

[0116] Data output: Generate preliminary screening results including clause number, matching threshold and exclusion reason, and push them to the comprehensive assessment and recommendation module.

[0117] Health disclosure information analysis module: structured processing of user health disclosure questionnaire data.

[0118] Technical Implementation: Named Entity Recognition (NER) technology is used to extract disease keywords (e.g., hypertension, diabetes) and map them to standard medical terminology (SNOMED CT codes) in the health information knowledge base. If health notification conditions are triggered (e.g., a user declares a past medical history), a risk warning label is generated.

[0119] Data output: Output health risk feature vector (format: <disease name>-<risk level>-<associated clause number>) as input parameter for underwriting decision.

[0120] Physical examination report analysis module: multimodal analysis of user physical examination reports (including text and medical images).

[0121] Technical implementation:

[0122] Text parsing: Use the BERT-NER model to extract key physiological indicators and convert them into structured data;

[0123] Image parsing: Generate pathology description text through a visual language model (ViT-LSTM) and map it to standard medical codes;

[0124] Data fusion: Construct a triple association structure of <indicator name>-<detection value>-<image evidence> to form a traceable chain of evidence.

[0125] Data output: Output health risk feature vectors containing text indicators and image features, and push them to the comprehensive assessment and recommendation module.

[0126] Comprehensive assessment and recommendation module: Integrates the analysis results of multiple modules to generate the final underwriting conclusion.

[0127] Technical implementation:

[0128] Historical conversation records: Capturing key statements in the user's conversation history through a temporal attention network to enhance contextual understanding;

[0129] Matching degree calculation: The TOPSIS algorithm is used to calculate the matching degree score of each product and dynamically assign weights wj;

[0130] Counterfactual reasoning: embedding hypothetical instructions to drive large language models to generate comparative analysis of differences.

[0131] Data output: Generate a structured underwriting report that includes recommended rankings, risk warnings, and alternatives, and link it to the original text in the regulatory database through a clause traceability link.

[0132] After the insurance-related questions raised by the user are preliminarily analyzed by the user inquiry analysis module, the system processes the user needs in parallel through the insurance terms data processing and analysis module, the health disclosure information analysis module and the physical examination report analysis module. The insurance terms data processing and analysis module outputs the terms matching results, the health disclosure information analysis module generates the health risk assessment conclusion, and the physical examination report analysis module extracts key health indicators and forms the analysis results. The analysis data of each module (including the terms analysis results, health disclosure assessment results and physical examination report analysis results) and the recorded historical conversation information are jointly input into the comprehensive assessment and recommendation module. Finally, the module integrates multi-source information, generates a comprehensive underwriting conclusion, and recommends suitable insurance products to users, realizing intelligent decision-making throughout the entire process.

[0133] The implementation process involves a user asking an insurance-related question. The User Inquiry Analysis Module, including the User Question Intent Recognition Module, leverages the language understanding capabilities of the large language model to identify four scenarios: inquiries related to insurance terms, health disclosure information, insurance information and a physical examination report, or other inquiries. Based on the large language model's judgment, inquiries related to insurance terms are answered using the aforementioned modules. Other inquiries are answered by combining the results of the three modules. The User Inquiry Analysis Module then proceeds to one of the three modules: the Insurance Terms Data Processing and Analysis Module, the Health Disclosure Structured Submodule, or the Physical Examination Report Analysis Module. Other inquiries are answered by combining the answers of all three modules.

[0134] Entering the Insurance Terms Data Processing and Analysis module, which includes the Terms Text Segmentation and Terms Question and Answer module, uses paragraph segmentation to divide the complete regulations into subsections, identifying individual clause points. This segmented document is then used to construct an insurance terms knowledge base. The Terms Question and Answer module includes a large language model. After fine-tuning the large language model, it is trained and fine-tuned using the insurance terms knowledge base to create the insurance terms question and answer large language model. This model then responds to clause-related content and provides answers to relevant questions. Select the appropriate module based on your needs.

[0135] If you enter the health notification structured sub-module, including the construction of the health information knowledge base, diseases with health problems will be added to the health notification library. When the user asks about related diseases, text matching will be performed to determine whether the health notification is touched. If it is in the health notification, the user will be reminded that there may be health problems with the insurance. If not, the user will be returned to the point that your situation does not match the health notification and you can take out insurance.

[0136] Entering the medical report analysis module involves extracting text from the medical examination conclusion and performing large language model analysis. If the medical report is in document format, the medical examination information can be directly retrieved. If it is in image format, the text detection model Optical Character Recognition (OCR) is used to extract the text from the medical examination results. The pathology images are then analyzed using the multimodal analysis model LLAVA. Ultimately, the text and image analysis information is used as the user's medical examination information. This combination of data modalities provides a more comprehensive understanding of the user's health status. The large language model, trained on content related to insurance-related diseases, can analyze and determine the user's medical examination conditions. Taking the user's textual medical examination information as input, the large language model determines whether the medical examination meets the insurance requirements. If not, it determines which medical conditions are disqualified and provides the user with a result. Introducing multimodal information from images and text enriches the content of health information and provides a more comprehensive analysis of the user's actual health status than the text report alone.

[0137] If the intent is other than the above, the system will sequentially enter the insurance terms data processing and analysis module, the health notification structured submodule, and the physical examination report analysis module, and return the comprehensive results to the user.

[0138] The previously mentioned large language model fine-tuning module includes data input and a language model. The data consists of insurance underwriting data, including information about covered diseases and other related information provided by relevant insurance policies. The language model can be trained using the LLaMA series large language model, using the fine-tuning data from the large language model to enable the large language model to analyze relevant underwriting situations. The preferred LLaMA2 model offers improved answering performance and is easy to train and fine-tune.

[0139] This embodiment uses a comprehensive assessment and recommendation module to generate comprehensive results. This involves recording the user's historical conversation information and, based on analysis from a large language model, providing an insurance application report and corresponding insurance recommendations. Specifically, the user is fed with the specific products they can insure, along with their historical multi-round conversation questions and answers. The large language model prompts are then designed to recommend the most suitable insurance products based on the user's specific circumstances. The comprehensive assessment combines the user's insurance needs, personal circumstances, and inquiries to provide comprehensive recommendations for eligible products, along with additional information such as liability exclusions.

[0140] In the process of communicating with users, by classifying intent in different modules and dividing different processes, we can provide more accurate analysis and answers, reduce the number of inaccurate AI answers, and record historical conversation information with users to better explore user needs and actual situations. This is also an important part of the underwriting process. Through intent understanding and multiple rounds of conversation, we can improve the user experience during the underwriting process and record it. We can understand the user's needs, help users better understand related products, and provide recommendations.

[0141] The present invention obtains the user's insurance needs, health disclosure information and multimodal physical examination data (including text and medical images) through a dialogue interaction interface; uses named entity recognition technology to extract text physiological indicators, uses visual language models to parse medical images to generate pathological descriptions, and maps them into standard medical terminology codes; based on the decision tree rule engine, the user's age, occupation and other hard conditions are preliminarily screened, and the deviation between health indicators and terms is calculated in combination with the disease correlation model, and candidate products with deviations that meet the requirements are screened; a hierarchical prompt template is constructed to drive the large language model to perform semantic analysis of terms and output structured underwriting conclusions; a weighted scoring model is used to generate product recommendation rankings, and a difference comparison description is triggered when the difference in scores between the best and second-best products reaches a threshold. The present invention solves the problems of difficult unstructured data processing and rigid terms matching in traditional underwriting through multimodal data fusion, dynamic deviation calculation and counterfactual reasoning mechanism, significantly improving underwriting accuracy and decision-making efficiency, and supporting real-time synchronous updates of regulatory terms, significantly reducing compliance risks.

[0142] The present invention may also have many other implementation methods. The above embodiments do not limit the present invention in any way. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention. Any other improvements and applications made to the above embodiments in an equivalent transformation manner shall fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for intelligent insurance underwriting, characterized in that: The following steps are involved: (a) User information collection: Obtaining the user's insurance demand description, health information questionnaire data, and physical examination report input through the dialogue interface. The physical examination report contains structured text and unstructured medical images; (b) multimodal information processing: extracting disease keywords from the health information questionnaire data, performing multimodal analysis on the physical examination report document, and generating a health risk feature vector containing text indicators and image features; (c) Insurance terms matching: User feature data is input into a pre-trained insurance terms analysis model, which performs preliminary screening based on a decision tree rule engine and outputs a set of products that meet insurance eligibility. (d) Intelligent underwriting decision-making: Build a hierarchical prompt template to drive the large language model to perform clause semantic analysis, generating a three-part structured analysis result consisting of [underwriting conclusion] - [matching clause] - [exclusion reason]; (e) Comprehensive recommendation generation: An approximate ideal solution ranking algorithm is used to calculate the matching score of each product and generate an underwriting recommendation report that includes a recommended ranking, risk warnings, and alternative solutions.

2. The intelligent insurance underwriting method according to claim 1, characterized in that: The multimodal analysis in step (b) includes: Use named entity recognition technology to extract key physiological indicators from the physical examination report text; Generate pathology description text from medical imaging data through visual language models; Establish an associated data structure of <indicator name>-<detection value>-<image evidence>.

3. The intelligent insurance underwriting method according to claim 2, characterized in that: The decision tree rule engine of step (c) implements a two-stage validation: (c1) Generate an initial screening exclusion list based on hard condition filtering of age and occupation; (c2) The remaining products are calculated using a disease association model to determine the degree of deviation between the user's health indicators and the terms and conditions. Products with a degree of deviation less than or equal to a preset threshold are retained as candidate products. The degree of deviation is calculated using the formula: Among them, xi is the user's health index value, yi is the clause threshold, and αi is the disease risk weight coefficient.

4. The intelligent insurance underwriting method according to claim 3, characterized in that: The layered prompt template of step (d) includes: Role definition layer: Set the large language model as the interactive identity of the underwriting expert; Dynamic data layer: inject user feature fields and target insurance clause keywords; Verification constraint layer: embeds counterfactual reasoning instructions and requires the output of a structured response containing a clause number reference.

5. The intelligent insurance underwriting method according to claim 4, characterized in that: The matching score in step (e) is calculated using the following formula: Among them, Si is the matching score of the i-th product, wj is the dynamic weight of the j-th evaluation dimension, dij+ represents the Euclidean distance from the ideal solution, and dij- represents the Euclidean distance from the negative ideal solution.

6. The intelligent insurance underwriting method according to claim 1, characterized in that: After step (e), the method further comprises the steps of: (f) Enhanced traceability: Underwriting reports will include traceability links to the original clauses in the insurance regulatory database; (g) Dynamic update mechanism: When the terms of the regulatory database change, the large language model is triggered to perform incremental training and update the decision threshold.

Images