A digital insurance intelligent underwriting and claim settlement system

Abstract

The present application relates to the technical field of insurance technology, in particular to a digital insurance intelligent underwriting and claim settlement system, comprising: a data analysis module, the data analysis module comprising an insurance vertical large model unit and an OCR intelligent identification unit; a dynamic risk rating module, which is in real-time linkage with the data analysis module; a pricing and approval module, which is connected with the dynamic risk rating module; a whole-process closed-loop anti-fraud module, which is connected with the data analysis module, the dynamic risk rating module and the pricing and approval module respectively, and is used for data monitoring of the whole process of insurance, underwriting and claim settlement, the whole-process closed-loop anti-fraud module comprising a time sequence comparison submodule, a behavior recognition submodule and a false claim feature submodule. Through the cooperation of the insurance vertical large model and the OCR identification technology, the system can automatically organize various scattered and disordered data such as medical records, physical examination reports and insurance information into a clean and unified format.

Description

Technical Field

[0001] This invention relates to the field of insurance technology, specifically to a digital insurance intelligent underwriting and claims settlement system. Background Technology

[0002] Underwriting and claims settlement are core aspects of risk management and compliance assurance in insurance business, directly impacting the operating efficiency and industry reputation of insurance institutions. Currently, with the continuous growth in national insurance demand and the expanding scale of insurance business, the types and quantities of application and claims materials are complex and numerous, placing stringent demands on the efficiency and accuracy of risk control audits.

[0003] Currently, most insurance institutions still use a combination of manual and basic fixed-system underwriting and claims processing. This requires staff to manually organize and verify various insurance and medical application materials, completing information verification, risk assessment, and limit determination item by item. The manual review model is labor-intensive and cumbersome, significantly reducing business processing efficiency and easily affected by factors such as human fatigue, subjective judgment, and information omissions, leading to review errors and significant deviations in risk control results. According to authorization announcement number CN116245490A, a digital insurance intelligent underwriting and claims processing system is disclosed, relating to the field of insurance claims technology. The system includes an acquisition module, a first processing module, a second processing module, and a claims processing module. This invention can identify and verify claims-related images uploaded by users and automatically process claims based on the identification and verification results. By adopting an automated underwriting and claims processing method, insurance companies can conduct more scientific risk control in insurance application and claims processing, preventing malicious insurance fraud, controlling repair costs more reasonably, and allowing users to apply for insurance and process claims faster, thus improving the user experience.

[0004] Currently, existing insurance underwriting and claims processing technologies have limited data processing capabilities, supporting only structured data processing and unable to intelligently analyze various unstructured materials. They struggle to achieve multi-dimensional data cleaning and integration, cannot break down industry data silos, and cannot provide comprehensive and effective risk assessment data support. Traditional underwriting, claims processing, risk control, and pricing models are rigid, unable to combine multi-dimensional core risk factors to build a user risk system and generate dynamic risk scores, making it difficult to achieve differentiated review and pricing, and resulting in low accuracy in risk matching. Furthermore, existing technologies lack a fully intelligent closed-loop anti-fraud mechanism, lack the ability to dynamically monitor and intelligently identify fraudulent behavior, and cannot effectively intercept various insurance and claims fraud risks, resulting in significant vulnerabilities in risk control and protection. Summary of the Invention

[0005] The purpose of this invention is to provide a digital insurance intelligent underwriting and claims settlement system to solve the problems mentioned in the background art.

[0006] To address the above problems, the present invention provides a technical solution: a digital insurance intelligent underwriting and claims settlement system, comprising:

[0007] The data parsing module includes an insurance vertical large model unit and an OCR intelligent recognition unit. The OCR intelligent recognition unit is used to parse unstructured and structured data in medical records, physical examination reports, and insurance information, and output initial text data. The insurance vertical large model unit receives the initial text data, performs multi-dimensional data cleaning, fusion, and precise extraction, and generates standardized user data records.

[0008] The dynamic risk rating module works in real time with the data parsing module. The dynamic risk rating module receives the standardized user data records and extracts the user's age (A), occupational category code (J), past medical history vector (H), and insured amount (C). The dynamic risk rating module calculates the dynamic risk score (R) based on the following formula:

[0009]

[0010] in, For age-related risk functions, This is a mapping value for occupational risk level. Let be the index value for the i-th type of past medical history. The preset weighting coefficients are n, where n is the total number of medical history types; the dynamic risk rating module also generates risk profile data based on the dynamic risk score R;

[0011] The pricing and assessment module, connected to the dynamic risk rating module, is used to generate differentiated underwriting pricing results and claims amount assessment results based on the risk profile data.

[0012] The full-process closed-loop anti-fraud module is connected to the data analysis module, the dynamic risk rating module, and the pricing and verification module, respectively, and is used to monitor data throughout the entire process of insurance application, underwriting, and claims settlement. The full-process closed-loop anti-fraud module includes a time-series comparison submodule, a behavior recognition submodule, and a fraudulent claims feature submodule.

[0013] Preferably, the full-process closed-loop anti-fraud module further includes a duplicate insurance detection unit, which is used to obtain the historical insurance record sequence of the same insured within a time window T, and calculate the duplicate insurance risk index M as follows:

[0014]

[0015] in, This is an indicator function showing the similarity between the Kth historical insurance record and the current insurance request in terms of the insured object and the scope of insurance liability, and its value is 0 or 1; The time interval between the k-th historical insurance record and the current request is in days. is the time decay coefficient, and m is the total number of historical insurance records within the time window T.

[0016] Preferably, the time-series comparison submodule includes a claims time-series feature extraction unit, which is used to obtain the claims request sequence of the same insured at consecutive time points, extract the time interval sequence of adjacent claims requests, and calculate the time-series anomaly degree S based on the following formula:

[0017]

[0018] in, The total number of claims. Let be the time interval between the p-th and p+1-th claim requests. The average of the time intervals, Standard deviation For time points Adjustment factors for corresponding holidays or non-working days.

[0019] Preferably, the behavior recognition submodule includes a behavior sequence encoder, which is used to convert user operations during the insurance application, underwriting, and claims processes into a behavior vector sequence, and calculate the behavior deviation degree D. The formula for calculating the behavior deviation degree D is:

[0020]

[0021] in, This represents the actual value of the vector for the q-th row. Let Q be the expected value of the q-th behavior vector learned from historical normal behavior samples, where Q is the total number of behavior vectors. This is the preset maximum allowable deviation value.

[0022] Preferably, the fraudulent claim feature submodule includes a claim material feature analysis unit, which is used to extract a set of key fields from medical records and physical examination reports, and calculate the fraudulent claim feature score F based on the following formula:

[0023]

[0024] in, This refers to the number of key fields in the claim materials that are inconsistent with the insurance information. The total number of key fields. Let j be the encoded length of the j-th abnormal text segment. This represents the total number of abnormal text fragments. This is the normalized value of the difference between the diagnosis time and the insurance waiting period. Preset weights.

[0025] Preferably, in the data parsing module, the insurance vertical large model unit further includes an entity relationship extraction subunit. The entity relationship extraction subunit is used to extract disease names, diagnosis and treatment operations, and drug names from unstructured text, and establish temporal and causal relationships between entities to generate structured relationship graph data. The standardized user data record contains the structured relationship graph data.

[0026] Preferably, the dynamic risk rating module further includes a dynamic factor update unit, which is used to receive newly added standardized user data records at preset time intervals and recalculate the age risk function. The segmented base parameter and the occupational risk level mapping value in the text The system uses the occupational risk matrix and the medical history weight coefficients in the medical history vector H to update the parameters and feeds them back to the dynamic risk rating module.

[0027] Preferably, the pricing and assessment module includes a coverage adjustment submodule, which receives the dynamic risk score R and calculates the basic coverage adjustment coefficient Adj.

[0028]

[0029] in, As a benchmark risk score, As an adjustment factor, the sum insured adjustment submodule determines the final approved upper limit of the compensation amount based on the product of the basic sum insured adjustment coefficient Adj and the user's insured amount C.

[0030] Preferably, the full-process closed-loop anti-fraud module further includes a real-time early warning submodule. The real-time early warning submodule obtains the time series anomaly degree S from the time series comparison submodule, the behavior deviation degree D from the behavior recognition submodule, and the fraudulent claim feature score F from the fraudulent claim feature submodule. When any one of the indicators S, D, and F exceeds the corresponding preset threshold, an early warning signal containing the specific anomaly type is generated, and the early warning signal is sent to the pricing and verification module to trigger the manual review process.

[0031] Preferably, the data parsing module, the dynamic risk rating module, the pricing and verification module, and the full-process closed-loop anti-fraud module interact with each other via an asynchronous message queue; the asynchronous message queue includes at least one first queue for storing the standardized user data records, at least one second queue for storing the dynamic risk score R, and at least one third queue for storing the warning signal.

[0032] The beneficial effects of this invention are reflected in:

[0033] 1. This invention combines a large vertical insurance model with OCR recognition technology. The system can automatically organize various scattered and messy data such as medical records, physical examination reports, and insurance information into a clean and uniform format. The system will combine key factors such as the user's age, occupation, past illnesses, and desired insurance amount to calculate a dynamic risk score. Each person's underwriting result and claim amount are determined based on their actual situation, rather than the previous one-size-fits-all fixed model.

[0034] 2. The system of this invention has a dedicated anti-fraud mechanism that runs smoothly from insurance application to underwriting to claims settlement. It automatically compares the user's insurance application and claims records at different time points to see if there are any abnormal time intervals. At the same time, it analyzes whether the operation behavior is abnormal and also identifies any traces of forgery in the claims materials. As long as suspicious situations such as duplicate insurance, insurance application with pre-existing conditions, or fraudulent claims are found, the system can immediately detect them and issue an early warning.

[0035] 3. The data analysis, risk rating, pricing verification and anti-fraud monitoring parts of the entire system of this invention are communicated through an asynchronous message queue. Each module can run independently and without interference. Even if a certain link is slower, it will not slow down the response speed of the entire system. At the same time, the parameters used in the dynamic risk scoring can be automatically updated regularly based on the new data, making the risk assessment closer to the actual situation. Attached Figure Description

[0036] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0037] Figure 1 This is a schematic diagram of the main process of intelligent underwriting and claims settlement of the present invention;

[0038] Figure 2 A schematic diagram of the invention's closed-loop anti-fraud monitoring process;

[0039] Figure 3 This is a schematic diagram of the dynamic parameter update and asynchronous message queue interaction process for the invention. Detailed Implementation

[0040] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0041] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0042] Example 1

[0043] like Figures 1-3 As shown, a digital insurance intelligent underwriting and claims system includes:

[0044] The data parsing module includes an insurance vertical large model unit and an OCR intelligent recognition unit. The OCR intelligent recognition unit is used to parse unstructured and structured data in medical records, physical examination reports, and insurance information, and outputs initial text data. The insurance vertical large model unit receives the initial text data, performs multi-dimensional data cleaning, fusion, and precise extraction, and generates standardized user data records.

[0045] The dynamic risk rating module works in real time with the data analysis module. It receives standardized user data records and extracts the user's age (A), occupational category code (J), past medical history vector (H), and insured amount (C). The dynamic risk rating module then calculates the dynamic risk score (R) based on the following formula:

[0046]

[0047] in, For age-related risk functions, This is a mapping value for occupational risk level. Let be the index value for the i-th type of past medical history. The preset weighting coefficients are used, where n is the total number of medical history types; the dynamic risk rating module also generates risk profile data based on the dynamic risk score R.

[0048] The pricing and assessment module, connected to the dynamic risk rating module, is used to generate differentiated underwriting pricing results and claims amount assessment results based on risk profile data.

[0049] The end-to-end closed-loop anti-fraud module is connected to the data analysis module, dynamic risk rating module, and pricing and verification module, respectively. It is used to monitor data throughout the entire process of insurance application, underwriting, and claims settlement. The end-to-end closed-loop anti-fraud module includes a time-series comparison sub-module, a behavior recognition sub-module, and a fraudulent claims feature sub-module.

[0050] The data analysis module, dynamic risk rating module, and end-to-end closed-loop anti-fraud module form a collaborative closed loop. Each time the data analysis module generates a standardized user data record, it simultaneously triggers the risk score calculation of the dynamic risk rating module and the anomaly detection of the end-to-end closed-loop anti-fraud module. When the warning signal output by the end-to-end closed-loop anti-fraud module exceeds the preset severity level, the dynamic risk rating module automatically uses the anomaly factors contained in the warning signal as additional input to recalculate the dynamic risk score R. The updated risk score is then fed back to the pricing and assessment module to adjust the generated underwriting results. Through data collaboration among the OCR recognition, risk rating, and anti-fraud monitoring modules, the system no longer operates in isolation. The user data processed by OCR is simultaneously fed to both risk rating and anti-fraud. The score calculated by risk rating serves as a reference baseline for anti-fraud, and once anti-fraud detects anomalies, it can trigger a recalculation of the risk rating.

[0051] Furthermore, the end-to-end closed-loop anti-fraud module also includes a duplicate insurance detection unit. This unit is used to obtain the historical insurance record sequence of the same insured within a time window T and calculate the duplicate insurance risk index M as follows:

[0052]

[0053] in, This is an indicator function showing the similarity between the Kth historical insurance record and the current insurance request in terms of the insured object and the scope of insurance liability, and its value is 0 or 1; The time interval between the k-th historical insurance record and the current request is in days. is the time decay coefficient, and m is the total number of historical insurance records within the time window T.

[0054] Furthermore, the time-series comparison submodule includes a claims time-series feature extraction unit. This unit is used to obtain the claims request sequence of the same insured at consecutive time points, extract the time interval sequence of adjacent claims requests, and calculate the time-series anomaly degree S based on the following formula:

[0055]

[0056] in, The total number of claims. Let be the time interval between the p-th and p+1-th claim requests. The average of the time intervals, Standard deviation For time points Adjustment factors for corresponding holidays or non-working days.

[0057] Furthermore, the behavior recognition submodule includes a behavior sequence encoder, which is used to convert user actions during the insurance application, underwriting, and claims processes into a sequence of behavior vectors, and calculate the behavior deviation degree D. The formula for calculating the behavior deviation degree D is:

[0058]

[0059] in, This represents the actual value of the vector for the q-th row. Let Q be the expected value of the q-th behavior vector learned from historical normal behavior samples, where Q is the total number of behavior vectors. This is the preset maximum allowable deviation value.

[0060] Furthermore, the fraudulent claim feature submodule includes a claim material feature analysis unit. This unit extracts a set of key fields from medical records and physical examination reports, and calculates the fraudulent claim feature score F based on the following formula:

[0061]

[0062] in, This refers to the number of key fields in the claim materials that are inconsistent with the insurance information. The total number of key fields. Let j be the encoded length of the j-th abnormal text segment. This represents the total number of abnormal text fragments. This is the normalized value of the difference between the diagnosis time and the insurance waiting period. Preset weights.

[0063] Furthermore, in the data parsing module, the insurance vertical large model unit also includes an entity relationship extraction subunit. The entity relationship extraction subunit is used to extract disease names, diagnosis and treatment operations, and drug names from unstructured text, and establish temporal and causal relationships between entities to generate structured relationship graph data. Standardized user data records contain structured relationship graph data.

[0064] Furthermore, the dynamic risk rating module also includes a dynamic factor update unit, which receives newly added standardized user data records at preset time intervals and recalculates the age risk function. The segmented base parameter and occupational risk level mapping value in The system uses the occupational risk matrix and the medical history weight coefficients in the past medical history vector H, and feeds the updated parameters back to the dynamic risk rating module.

[0065] The dynamic factor update unit updates the medical history weight coefficients based on the following formula. :

[0066]

[0067] in, The weighting coefficient for the i-th type of medical history before the update. The updated weighting coefficients, The learning rate is preset and 0 ≤ ≤0.1, where K is the total number of standardized user data records added within the preset time period. The predicted dynamic risk score for the kth newly added record. The risk score is adjusted after the actual incident corresponding to the newly added record k. This represents the index value for the i-th type of past medical history in the k-th newly added record.

[0068] Furthermore, the pricing and assessment module includes a coverage adjustment submodule, which receives a dynamic risk score R and calculates the basic coverage adjustment coefficient Adj.

[0069]

[0070] in, As a benchmark risk score, As an adjustment factor, the sum assured adjustment submodule determines the final approved upper limit of the compensation amount based on the product of the basic sum assured adjustment coefficient Adj and the user's insured amount C.

[0071] Furthermore, the full-process closed-loop anti-fraud module also includes a real-time early warning sub-module. The real-time early warning sub-module obtains the time series anomaly degree S from the time series comparison sub-module, the behavior deviation degree D from the behavior recognition sub-module, and the fraudulent claim feature score F from the fraudulent claim feature sub-module. When any of the indicators S, D, and F exceeds the corresponding preset threshold, an early warning signal containing the specific anomaly type is generated and sent to the pricing and verification module to trigger the manual review process.

[0072] The real-time early warning submodule also calculates the comprehensive early warning index W as follows:

[0073]

[0074] in, A preset threshold is set for the time series anomaly degree. Preset a threshold for behavioral deviation. Preset thresholds for scoring characteristics of fraudulent claims. Preset weighting coefficients and satisfying When W > 1, the real-time early warning submodule generates a multi-factor joint early warning signal.

[0075] Furthermore, the data parsing module, dynamic risk rating module, pricing and verification module, and the full-process closed-loop anti-fraud module interact with each other through an asynchronous message queue; the asynchronous message queue includes at least one first queue for storing standardized user data records, at least one second queue for storing dynamic risk scores R, and at least one third queue for storing early warning signals.

[0076] Example 2

[0077] The differentiated underwriting and claims settlement solutions based on dynamic risk rating are as follows:

[0078] A 35-year-old male insured, an occupational worker (high-risk occupation code J=3), purchased critical illness insurance with a coverage amount of C=500,000 yuan. His pre-existing medical history includes hypertension (corresponding to...). =0.6), no history of diabetes ( =0);

[0079] The steps are as follows:

[0080] A1. The system uses OCR to recognize the user's physical examination report and medical record. The insurance vertical big model extracts standardized user data records, including age A=35, occupation code J=3, medical history vector H=[0.6,0], and insurance amount C=500,000 yuan. At the same time, the entity relationship extraction sub-unit identifies the time-series causal triple "hypertension - taking nifedipine - blood pressure control is good" from the medical record and stores it in the structured relationship graph;

[0081] A2. Dynamic Risk Rating Module, Dynamic Risk Score R, with weighting coefficients. =0.25, =0.20, =0.35, =0.20, age risk function =0.3, occupational risk mapping =0.8, substitute into the formula:

[0082]

[0083]

[0084] The generated risk profile data has an R value of approximately 1.23, indicating a medium-to-high risk range.

[0085] A3. Pricing and Assessment Module: Coverage Adjustment Factor, Setting Benchmark Risk Score =1.0, adjustment factor =2.0:

[0086]

[0087] The final approved maximum compensation amount = Adj × C = 0.6137 × 500,000 yuan = 306,850 yuan, while the basic premium will increase by approximately 38.6%;

[0088] A4. One week later, the system added K=200 new claim data entries, which included the insured's actual hospitalization records. =1.45, higher than the predicted value =1.23, dynamic factor update unit, set learning rate =0.05, update the weighting coefficient for history of hypertension. :

[0089]

[0090] Enable online learning of the model.

[0091] Example 3

[0092] The time-series early warning scheme based on duplicate insurance applications and abnormal behavior is as follows:

[0093] The same insured, Mr. Zhang, submitted the same critical illness insurance application to three different insurance companies within 30 days (time window T=30 days), with completely identical insured objects and coverage. The system monitored this through a closed-loop anti-fraud module.

[0094] The steps are as follows:

[0095] B1. The duplicate insurance detection unit acquires the historical insurance record sequence, m=3, with a duplicate insurance risk index M, and a time decay coefficient. =0.1, time interval The sameness indicator functions for 5 days, 12 days, and 20 days respectively. =1:

[0096]

[0097]

[0098] Exceeding the preset threshold =1.2, triggering a duplicate insurance warning;

[0099] B2. The time-series comparison submodule extracts Zhang's four past claim requests, L=4, and the time interval sequence. =[3,8,2], mean =4.33 days, standard deviation =3.21 days, non-working adjustment factor All are 1, no holidays:

[0100]

[0101]

[0102] Exceeding the threshold When the value is 0.6, a timing anomaly warning is generated;

[0103] B3. The behavior recognition submodule encodes Zhang's insurance application sequence, including rapid product switching, nighttime operations, and multiple modifications to medical history, into a behavior vector. In line with historical expectations For comparison, assuming the total number of behavior vectors Q=10, the calculated value is D=0.83. =0.7;

[0104] B4. Real-time Early Warning Submodule, set =0.4, =0.3, =0.3, threshold =0.6, =0.7, =0.5, F is temporarily set to 0.2, no false material warning triggered:

[0105]

[0106]

[0107] W=0.983<1. Although the multi-factor joint early warning threshold has not been reached, S and D have each triggered single-indicator early warnings. The system still generates an early warning signal containing the type of "duplicate insurance + time sequence abnormality" and sends it to the pricing and verification module to trigger manual review.

[0108] Example 4

[0109] The solution based on identifying fraudulent claim materials and using asynchronous message queues is as follows:

[0110] In the pneumonia hospitalization claim materials submitted by the insured, Mr. Li, the interval between the diagnosis date and the effective date of the insurance was 3 days (the waiting period is usually 30 days), and there were many inconsistencies between the key fields in the medical record and the insurance information.

[0111] The steps are as follows:

[0112] C1, the fraudulent claims feature submodule, sets the total number of key fields. =15, number of inconsistent fields =4, total number of abnormal text fragments =6, Length of abnormal fragment encoding The average is 1.2, and the normalized value of the difference between diagnosis time and insurance waiting period is... =0.85 (extremely abnormal), set weights =0.3, =0.4, =0.3:

[0113]

[0114]

[0115] F=0.815 exceeds the preset threshold. =0.5, triggering a false claim warning, the system automatically blocks the claim request;

[0116] C2. The data parsing module sends Li's standardized user data record to the first queue. The dynamic risk rating module calculates R=2.35 (high risk) and sends the result to the second queue. The closed-loop anti-fraud module sends the warning signal (type code: FRAUD_CLAIM_005) to the third queue. The pricing and verification module consumes the R value from the second queue and the warning signal from the third queue, and finally outputs the verification result of "rejection of claim + blacklisting". The three queues operate decoupled to ensure high system throughput.

[0117] Finally, it should be noted that the above embodiments 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A digital insurance intelligent underwriting and claims settlement system, characterized in that: Include: The data parsing module includes an insurance vertical large model unit and an OCR intelligent recognition unit. The OCR intelligent recognition unit is used to parse unstructured and structured data in medical records, physical examination reports, and insurance information, and output initial text data. The insurance vertical large model unit receives the initial text data, performs multi-dimensional data cleaning, fusion, and precise extraction, and generates standardized user data records. The dynamic risk rating module works in real time with the data parsing module. The dynamic risk rating module receives the standardized user data records and extracts the user's age (A), occupational category code (J), past medical history vector (H), and insured amount (C). The dynamic risk rating module calculates the dynamic risk score (R) based on the following formula: in, For age-related risk functions, This is a mapping value for occupational risk level. Let be the index value for the i-th type of past medical history. The preset weighting coefficients are n, where n is the total number of medical history types; the dynamic risk rating module also generates risk profile data based on the dynamic risk score R; The pricing and assessment module, connected to the dynamic risk rating module, is used to generate differentiated underwriting pricing results and claims amount assessment results based on the risk profile data. The full-process closed-loop anti-fraud module is connected to the data analysis module, the dynamic risk rating module, and the pricing and verification module, respectively, and is used to monitor data throughout the entire process of insurance application, underwriting, and claims settlement. The full-process closed-loop anti-fraud module includes a time-series comparison submodule, a behavior recognition submodule, and a fraudulent claims feature submodule.

2. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The closed-loop anti-fraud module also includes a duplicate insurance detection unit. This unit is used to obtain the historical insurance record sequence of the same insured within a time window T and calculate the duplicate insurance risk index M as follows: in, This is an indicator function showing the similarity between the Kth historical insurance record and the current insurance request in terms of the insured object and the scope of insurance liability, and its value is 0 or 1; The time interval between the k-th historical insurance record and the current request is in days. is the time decay coefficient, and m is the total number of historical insurance records within the time window T.

3. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The time-series comparison submodule includes a claims time-series feature extraction unit. This unit is used to obtain the claims request sequence of the same insured at consecutive time points, extract the time interval sequence of adjacent claims requests, and calculate the time-series anomaly degree S based on the following formula: in, The total number of claims. Let be the time interval between the p-th and p+1-th claim requests. The average of the time intervals. Standard deviation, For time points Adjustment factors for corresponding holidays or non-working days.

4. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The behavior recognition submodule includes a behavior sequence encoder, which is used to convert user actions during the insurance application, underwriting, and claims processes into a behavior vector sequence and calculate the behavior deviation degree D. The formula for calculating the behavior deviation degree D is as follows: in, This represents the actual value of the vector for the q-th row. Let Q be the expected value of the q-th behavior vector learned from historical normal behavior samples, where Q is the total number of behavior vectors. This is the preset maximum allowable deviation value.

5. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The fraudulent claim feature submodule includes a claim material feature analysis unit, which extracts a set of key fields from medical records and physical examination reports, and calculates the fraudulent claim feature score F based on the following formula: in, This refers to the number of key fields in the claim materials that are inconsistent with the insurance information. This represents the total number of key fields. Let j be the encoded length of the j-th abnormal text segment. This represents the total number of abnormal text fragments. This is the normalized value of the difference between the diagnosis time and the insurance waiting period. Preset weights.

6. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: In the data parsing module, the insurance vertical large model unit also includes an entity relationship extraction subunit. The entity relationship extraction subunit is used to extract disease names, diagnosis and treatment operations, and drug names from unstructured text, and establish temporal and causal relationships between entities to generate structured relationship graph data. The standardized user data record contains the structured relationship graph data.

7. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The dynamic risk rating module also includes a dynamic factor update unit, which receives newly added standardized user data records at preset time intervals and recalculates the age risk function. The segmented base parameter and the occupational risk level mapping value in the text The system uses the occupational risk matrix and the medical history weight coefficients in the medical history vector H to update the parameters and feeds them back to the dynamic risk rating module.

8. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The pricing and assessment module includes a coverage adjustment submodule, which receives the dynamic risk score R and calculates the basic coverage adjustment coefficient Adj. in, As a benchmark risk score, As an adjustment factor, the sum insured adjustment submodule determines the final approved upper limit of the compensation amount based on the product of the basic sum insured adjustment coefficient Adj and the user's insured amount C.

9. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The full-process closed-loop anti-fraud module also includes a real-time early warning submodule. The real-time early warning submodule obtains the time series anomaly degree S from the time series comparison submodule, the behavior deviation degree D from the behavior recognition submodule, and the fraudulent claim feature score F from the fraudulent claim feature submodule. When any of the indicators S, D, and F exceeds the corresponding preset threshold, an early warning signal containing the specific anomaly type is generated and the early warning signal is sent to the pricing and verification module to trigger the manual review process.

10. The digital insurance intelligent underwriting and claims settlement system according to claim 1, characterized in that: The data parsing module, the dynamic risk rating module, the pricing and verification module, and the full-process closed-loop anti-fraud module interact with each other via an asynchronous message queue; the asynchronous message queue includes at least one first queue for storing the standardized user data records, at least one second queue for storing the dynamic risk score R, and at least one third queue for storing the warning signal.

Images