Internet health insurance underwriting method and device based on disease characteristics

Abstract

The invention relates to the technical field of insurance and intelligent data processing, and discloses an internet health insurance underwriting method and device based on disease characteristics, and the method comprises the following steps: S1, obtaining to-be-underwritten historical insurance data and current insurance application data; s2, constructing a health insurance special word vector library based on the historical insurance data; s3, inputting the feature vector into a pre-trained hierarchical disease recognition model, and obtaining an output disease severity score; s4, generating an underwriting risk prompt according to the disease severity score, and sending the underwriting risk prompt to an underwriting operation terminal; and S5, collecting an artificial underwriting decision result for the underwriting risk prompt, and taking the artificial underwriting decision result as a new training sample to carry out incremental updating on the hierarchical disease recognition model. According to the method, the non-standard disease description analysis precision is improved by fusing the claim co-occurrence word vector and the vertical domain attention mechanism.

Description

Technical Field

[0001] This invention relates to the fields of insurance and intelligent data processing technology, specifically to an internet-based health insurance underwriting method and apparatus based on disease characteristics. Background Technology

[0002] The rapid development of internet-based health insurance has placed higher demands on the automation and precision of the underwriting process. Traditional health insurance underwriting systems mainly rely on rule engines, using preset "if-then" logic to determine whether the disease name disclosed by the insured triggers the underwriting control rules.

[0003] However, existing technologies have limitations when processing non-standardized disease descriptions. Diagnostic texts submitted by policyholders often contain non-standard medical abbreviations, colloquial expressions, or complex descriptions of accompanying symptoms. Rule engines struggle to accurately identify the underlying risk through simple keyword matching, easily leading to misjudgments or omissions. Furthermore, existing underwriting models are mostly built on general natural language processing algorithms, which typically focus only on the semantic similarity of text, ignoring the actual risk relevance of medical terminology in insurance claims scenarios. For example, some diseases with vastly different literal descriptions may have similar claim risk characteristics, and general models struggle to capture this implicit business logic.

[0004] On the other hand, existing underwriting risk assessment models are typically statically deployed. With advancements in medical diagnostic technology and the iteration of insurance products, the risk attributes of diseases and claims rules change dynamically. After running online for a period, the discriminative ability of static models gradually lags behind new business scenarios. While models can be updated by retraining with full data, this method is computationally expensive and time-consuming, and also faces the problem of "catastrophic forgetting," where the model easily loses its memory of historically stable underwriting logic when learning new rules, leading to fluctuations in system performance. Therefore, how to construct an underwriting method that can deeply integrate insurance claims characteristics for accurate risk identification and adapt to business changes with low-cost incremental updates has become a problem to be solved in the current technological field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an internet-based health insurance underwriting method and apparatus based on disease characteristics. This solves the problems of low recognition accuracy in existing health insurance underwriting technologies when processing complex disease descriptions due to the lack of correlation with insurance vertical domain features, and poor system stability when the model adapts to new risk rule updates due to high full retraining costs and the tendency to forget historical knowledge.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an internet health insurance underwriting method based on disease characteristics, comprising the following steps: S1. Obtain historical insurance data to be underwritten and current insurance application data, wherein the current insurance application data includes disease diagnosis text; S2. Construct a health insurance-specific word vector library based on the historical insurance data, and convert the disease diagnosis text into feature vectors; S3. Input the feature vector into a pre-trained hierarchical disease recognition model to obtain the output disease severity score; the hierarchical disease recognition model includes a physiological system classification layer, a specific disease recognition layer, and a severity assessment layer connected in sequence. S4. Generate an underwriting risk warning based on the severity score of the disease, and send the underwriting risk warning to the underwriting terminal; S5. Collect the results of manual underwriting decisions in response to the underwriting risk warning, and use the results of manual underwriting decisions as new training samples to incrementally update the hierarchical disease identification model.

[0007] Preferably, when constructing a dedicated word vector library for health insurance based on the historical insurance data, the specific steps include: Calculate the health insurance domain similarity between two medical terms; the calculation process of the health insurance domain similarity includes: calculating the cosine similarity of the two medical terms in the vector space, and calculating the co-occurrence correlation of the two medical terms appearing simultaneously in historical claims cases; after weighting the co-occurrence correlation of claims through a preset domain adjustment coefficient, multiply it with the cosine similarity to obtain the health insurance domain similarity.

[0008] Preferably, the physiological system classification layer uses a random forest algorithm to construct the classification model. The classification splitting criterion for the decision tree nodes in the random forest algorithm is determined based on the health insurance weighted Gini coefficient. The calculation logic of the health insurance weighted Gini coefficient is as follows: calculate the original Gini impurity of the current node sample set and calculate the risk importance of the candidate feature. The risk importance is determined based on the contribution of the candidate feature to the historical claim amount in the historical insurance data. The original Gini impurity is combined with the risk importance weighted by the risk sensitivity parameter to obtain the health insurance weighted Gini coefficient, and the feature that minimizes the coefficient is selected as the splitting feature.

[0009] Preferably, the specific disease identification layer adopts a deep learning network based on the Transformer encoder architecture, and the deep learning network is configured with a health insurance vertical domain attention mechanism; the calculation logic of the health insurance vertical domain attention mechanism is as follows: a medical term mask matrix is ​​introduced in the standard dot product attention calculation process; the medical term mask matrix is ​​used to mark the positions in the input sequence that belong to the predefined medical entity library; when calculating the attention weight, preference weights are superimposed on the positions marked by the medical term mask matrix.

[0010] Preferably, the logic for the severity assessment layer to output the disease severity score is as follows: The scores for the identified diseases are calculated in four dimensions: disease progression stage, treatment complexity, prognosis, and insurance impact. The scores for the four dimensions are weighted and summed, and then multiplied by a time decay factor determined based on the time of disease diagnosis to obtain the final disease severity score.

[0011] In one specific embodiment, the dimension score of the insurance impact dimension is determined based on the claims risk coefficient, which is obtained by calculating the ratio of the average claims amount of the disease in the historical insurance data to the average premium income of the corresponding insurance product, and combining the claim frequency weight of the disease in the insured population.

[0012] Preferably, the process of generating underwriting risk warnings specifically includes: The disease severity score is compared with a preset risk threshold to determine the corresponding risk level identifier; the predicted confidence level output by the hierarchical disease identification model is obtained; the risk level identifier, the predicted confidence level, and the underwriting recommendations based on the disease severity score are encapsulated into a JSON data packet as the underwriting risk warning.

[0013] Preferably, the process of collecting the manual underwriting decision results in response to the underwriting risk warning specifically includes: The system receives the final underwriting conclusion fed back through the underwriting operation terminal, which includes normal coverage, rejection, or conditional coverage; it associates and tags the final underwriting conclusion with the disease diagnosis text in the current insurance application data to construct a new training sample containing the correct label, and stores the new training sample in the sample database; it monitors the number of the new training samples in the sample database in real time, and triggers the incremental update process when the number reaches a preset update threshold.

[0014] Preferably, the process of incrementally updating the hierarchical disease identification model specifically includes: An elastic weight consolidation algorithm is used to construct a total loss function, which consists of a new task loss term for the new training samples and a regularization term for protecting old knowledge. The calculation logic of the regularization term is as follows: the importance weight of each model parameter of the hierarchical disease identification model in the historical task is calculated based on the Fisher information matrix; in the process of minimizing the total loss function, the update magnitude of the corresponding model parameters is constrained by the importance weight to retain the hierarchical disease identification model's ability to remember historical underwriting rules.

[0015] A second aspect of the present invention provides an internet-based health insurance underwriting device based on disease characteristics, comprising: The data acquisition module is used to acquire historical insurance data to be underwritten and current insurance application data, wherein the current insurance application data includes disease diagnosis text; The feature construction module, connected to the data acquisition module, is used to construct a health insurance-specific word vector library based on the historical insurance data and convert the disease diagnosis text into feature vectors. The disease identification module, connected to the feature construction module, is used to input the feature vector into a pre-trained hierarchical disease identification model to obtain the output disease severity score; the hierarchical disease identification model includes a physiological system classification layer, a specific disease identification layer, and a severity assessment layer connected in sequence. The risk warning module, connected to the disease identification module, is used to generate underwriting risk warnings based on the severity score of the disease and send the underwriting risk warnings to the underwriting operation terminal. The model optimization module is connected to the risk warning module and the disease identification module respectively. It is used to collect the results of manual underwriting decisions in response to the underwriting risk warning, and use the results of manual underwriting decisions as new training samples to incrementally update the hierarchical disease identification model.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it is able to implement an internet health insurance underwriting method based on disease characteristics.

[0017] This invention provides a method and apparatus for internet-based health insurance underwriting based on disease characteristics. It has the following beneficial effects: 1. By constructing a dedicated word vector library for health insurance that integrates co-occurrence correlations in claims processing, and combining it with a vertical domain attention mechanism in deep learning networks, this invention can capture the implicit associations of medical terms in insurance business scenarios. This approach not only considers literal semantic similarity but also quantifies the co-occurrence relationships of terms in historical claims cases, enabling the model to accurately identify disease features with significant literal differences but similar risk attributes, thus improving the parsing accuracy of non-standard or complex disease descriptions.

[0018] 2. This invention introduces a weighted Gini coefficient based on the contribution of claim amount and a claim risk coefficient based on payout leverage in the model construction. Unlike traditional algorithms that only pursue classification accuracy, this technical solution directly embeds economic risk indicators into the feature selection and node splitting process. This makes the model more sensitive to high payout risk features, thus outputting a severity score that better aligns with actuarial logic and more intuitively reflects the potential impact of insurance applications on the payout ratio.

[0019] 3. This invention employs an incremental update mechanism based on an elastic weight consolidation algorithm, utilizing the results of manual underwriting decisions to perform closed-loop optimization of the model. By constraining the update magnitude of key parameters through the Fisher information matrix, this mechanism effectively suppresses catastrophic forgetting of historical knowledge when absorbing newly emerging disease risk characteristics or underwriting rules, ensuring that the underwriting system maintains the stability and continuity of historical underwriting standards while adapting to business changes. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram of the device architecture of the present invention; Figure 3 This is a schematic diagram of the electronic device described in this application. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figure 1 , Figure 1 This is a flowchart illustrating an internet health insurance underwriting method based on disease characteristics according to an embodiment of the present invention. The present invention provides an internet health insurance underwriting method based on disease characteristics, the method comprising the following steps: Step S100: Obtain historical insurance data and current insurance application data to be underwritten. The historical insurance data is stored in the core insurance business database and includes past policyholders' claim records, underwriting logs, and corresponding disease diagnosis conclusions. The current insurance application data is collected in real time through the front-end interactive interface and includes the disease diagnosis text submitted by the policyholder. In a preferred embodiment, the current insurance application data also includes physiological monitoring indicator data collected through a wearable device interface.

[0023] Step S200: Construct a dedicated word vector library for health insurance based on the historical insurance data, and convert the disease diagnosis text into feature vectors. Step S200 first performs word segmentation on the historical insurance data to establish a dedicated dictionary containing disease diagnosis terms, medical anatomy terms, and underwriting professional terms. Based on this, the health insurance domain similarity between any two terms is calculated, and then a word vector representation is generated. For the disease diagnosis text, the generated word vectors are used for spatial mapping to obtain the corresponding text feature vector. If physiological monitoring indicator data exists, this step also includes calculating a dynamic health score, normalizing the score value, and incorporating it into the feature vector.

[0024] Step S300: Input the feature vector into the pre-trained hierarchical disease recognition model to obtain the output disease severity score. The hierarchical disease recognition model is logically composed of a physiological system classification layer, a specific disease recognition layer, and a severity assessment layer connected sequentially.

[0025] The physiological system classification layer performs coarse-grained partitioning of the input feature vector to determine the physiological system category to which the disease belongs. This layer employs an improved random forest algorithm, whose node splitting criterion is determined based on the health insurance weighted Gini coefficient to prioritize the differentiation of high-claims-risk features.

[0026] The specific disease identification layer, based on physiological system classification, performs fine-grained semantic parsing of feature vectors to identify specific disease name entities. This layer employs a deep learning network based on a Transformer encoder architecture and introduces a health insurance vertical domain attention mechanism during the computation process, enhancing the feature extraction weights of key medical entities through a medical terminology mask matrix.

[0027] The severity assessment layer performs multi-dimensional quantitative analysis on the identified diseases and outputs a disease severity score. This score integrates four dimensions: disease progression stage, treatment complexity, prognosis, and insurance impact, and incorporates a time decay factor. The calculation of the insurance impact dimension involves a claims risk coefficient, which is directly related to the ratio of the historical average claims amount to the average premium income for the disease.

[0028] Step S400: Generate an underwriting risk warning based on the disease severity score and send the warning to the underwriting terminal. Step S400 compares the disease severity score with a preset risk threshold range to determine the corresponding risk level. Simultaneously, it obtains the prediction confidence level output by the model. The system encapsulates the risk level indicator, prediction confidence level, and corresponding underwriting recommendations into a standardized JSON data packet, sends it via network transmission protocol to the terminal used by underwriters, and displays it visually on the interface.

[0029] Step S500: Collect the manual underwriting decision results for the underwriting risk warning, and use the manual underwriting decision results as new training samples to incrementally update the hierarchical disease recognition model. The underwriting terminal records the final underwriting, rejection, or conditional underwriting conclusion issued by the underwriter, and associates the conclusion with the original disease diagnosis text to form new training samples with labels.

[0030] When the number of newly added training samples reaches a preset threshold, the incremental update process of the model is triggered. This process uses the elastic weight consolidation algorithm to update the model parameters by constructing a total loss function that includes a new task loss term and a regularization term. The regularization term is calculated based on the Fisher information matrix and is used to constrain the magnitude of changes in model parameters during the update process, ensuring that the model retains its memory of historical underwriting rules while learning the features of new samples.

[0031] Reference Figure 1 In step S100, the acquisition of multi-source heterogeneous data is performed through the data interface module, which establishes communication connections with the database of the core insurance business system and the front-end insurance application server, respectively.

[0032] For historical insurance data awaiting underwriting, the system extracts structured fields and unstructured text from the core database using SQL queries. Structured fields include the policyholder's age, gender, previous claim amounts, claim dates, insurance type, and underwriting conclusion. Unstructured text includes descriptions of past claims, OCR-recognized scanned medical invoices, and electronic medical records issued by hospitals. After extraction, this data undergoes deduplication and cleaning, and is stored in a distributed file system as the foundational corpus for subsequent word vector library construction and model training.

[0033] For current insurance application data, the system receives the disease diagnosis text entered by the policyholder on the front-end interface via an API interface. This text typically consists of the chief complaint, present medical history, and the name of the disease diagnosed by the doctor. Simultaneously with receiving the text data, the system checks for authorized wearable device data sources. If such sources exist, it synchronously acquires physiological monitoring indicator data for a preset monitoring period, including but not limited to resting heart rate, sleep duration, blood oxygen saturation, and daily steps.

[0034] To transform physiological monitoring indicators of different dimensions into unified features usable for risk assessment, step S100 further includes calculating a dynamic health score. This calculation process first preprocesses the acquired raw monitoring data to remove outliers caused by improper device use, and then uses a dynamic health score model to quantify physiological health status. Dynamic Health Score The calculation formula is as follows: ; In the formula, This indicates the total number of categories of physiological monitoring indicators accessed; For indexing indicators; For the first The average measured value of each physiological monitoring indicator during the monitoring period; For the first The risk weight coefficients for each indicator are preset. These coefficients are set according to the correlation between different physiological indicators and health insurance payout risks. For example, the weight of cardiovascular-related indicators is set higher than that of exercise steps. This is a normalization function used to map measurements of different physical units to a dimensionless interval of [0,1]. It is usually achieved using the minimum-maximum normalization method.

[0035] In the formula Indicates the first The time consistency factor of an indicator is used to characterize the numerical stability of that indicator within the monitoring period. This factor is negatively correlated with the variance of the monitoring data; that is, the smaller the data fluctuation, the closer the consistency factor is to 1, indicating that the physiological indicator is in a stable state. Conversely, if the data fluctuates sharply in the short term, the consistency factor value decreases, thereby lowering the overall health score. By introducing the time consistency factor, this method can distinguish between the long-term stable level of physiological indicators and short-term occasional fluctuations, preventing risk assessment bias caused by abnormalities in a single measurement. The final calculated... The numerical value, as an independent numerical feature dimension, is concatenated with the text feature vector generated in subsequent steps.

[0036] Reference Figure 1In step S200, to enable the computer to understand the specific meaning of medical text in insurance business, a dedicated vector space needs to be constructed. This process specifically includes three sub-steps: establishing a health insurance-specific word segmentation dictionary, calculating domain similarity, and generating feature vectors.

[0037] First, the system establishes a specialized health insurance terminology dictionary, which consists of three independent terminology databases: the first database is a disease diagnosis terminology database, containing all disease names and codes under the ICD-10 standard; the second database is a medical terminology database, containing names of human anatomical locations (such as "mitral valve" and "coronary artery") and descriptions of pathological changes (such as "hyperplasia" and "calcification"); the third database is a health insurance specialized terminology database, containing expressions specific to underwriting rules and claims reports (such as "pre-existing conditions," "exclusions," and "claims during the waiting period"). This dictionary is then used to segment disease diagnosis texts in historical insurance data and current insurance application data, removing stop words to obtain standardized word sequences.

[0038] In constructing the word vector space, this invention does not solely rely on semantic co-occurrence within the text context, but instead introduces co-occurrence information based on claims data. For any two medical / insurance terms in the dictionary... and Calculate their similarity in the health insurance field. The computational logic aims to bridge the gap in vector space between terms that are semantically distant but frequently appear in claims cases.

[0039] Similarity in the health insurance sector The calculation formula is as follows: ; In the formula, and Each represents a term and Original word vector representation in high-dimensional space The cosine similarity between the two is used to characterize the general semantic closeness. The preset domain adjustment coefficient, with a value range between 0.1 and 0.3, is used to adjust the weight of the impact of claims relevance on the overall similarity.

[0040] In the formula The formula for calculating the co-occurrence correlation of two terms in historical claims records is as follows: ; In the formula, This indicates that the historical database contains terms simultaneously. and The number of claims processed; and Each represents a term containing and The total number of claims. This formula shows that if two terms frequently appear in the same claim (e.g., "hypertension" and "stroke"), even if their semantic connection is weak in a general corpus, their... The value will also be higher, thus increasing The value allows the trained word vectors to better reflect the correlation between disease complications and payout risks in terms of spatial distribution.

[0041] Finally, based on the calculated health insurance-specific word vectors, the disease diagnosis text obtained in step S100 is converted into a numerical feature vector. Specifically, the specific operation involves weighting or concatenating the specific word vectors corresponding to each word obtained after word segmentation of the disease diagnosis text to form a fixed-dimensional text feature vector. If a dynamic health score was calculated in step S100... The score is then appended as a scalar to the end of the text feature vector to form the joint feature vector of the final input model.

[0042] Reference Figure 1 The hierarchical disease identification model in step S300 is designed with a three-layer progressive structure, aiming to gradually analyze the disease risk characteristics in the insurance application from macro-system positioning to micro-risk quantification.

[0043] Physiological system classification layers: This layer is used to roughly classify the input joint feature vector into physiological system categories such as cardiovascular, respiratory, and digestive systems. To improve the model's sensitivity to high-payout-risk diseases, this layer employs an improved random forest algorithm. When constructing the decision tree, the selection of node splitting features no longer relies solely on traditional information gain or Gini impurity reduction, but instead introduces the concept of "risk importance."

[0044] Specifically, the classification splitting criterion for decision tree nodes is based on the health insurance-weighted Gini coefficient. The calculation formula is as follows: ; In the formula, This represents the set of data samples for the current node; The standard Gini impurity is used to measure the degree of disorder in the distribution of categories in a set; This is a risk sensitivity parameter used to adjust the influence of risk weights; Candidate features The importance of the risks.

[0045] Importance of risk The calculation is based on the contribution of this feature to the claim amount in historical data: ; In the formula, To determine features Does it belong to a set? ; This represents the total number of historical claims. For the first The actual amount of compensation paid in each case; For the first A set of characteristics for each case; This is the indicator function. This formula assigns higher importance values ​​to features that frequently appear in high-payout cases. During the splitting process, the algorithm iterates through all candidate features and calculates the corresponding... The model selects the feature that minimizes this coefficient as the optimal split point. This means that the model tends to prioritize using features that are strongly correlated with high payout risk for classification, thereby separating high-risk samples at shallow nodes of the tree.

[0046] Specific disease identification layer: After identifying the physiological system, the data proceeds to the second layer for fine-grained identification of specific disease entities. This layer employs a deep learning network based on a Transformer encoder architecture, utilizing its self-attention mechanism to capture long-range semantic dependencies. To enhance the model's perception of key medical terms, a health insurance vertical domain attention mechanism is integrated into the network. .

[0047] This mechanism incorporates a medical terminology mask matrix when calculating the dot product attention of the query and the key. The calculation formula is as follows: ; In the formula, This means that each component of the input vector is mapped to the interval (0, 1), and the sum of all components is 1; For matrix and The result of multiplication; This is the set of value vectors generated after the input sequence undergoes a linear transformation (projection). Attention mechanisms in the health insurance vertical sector; , , The projection matrix of the input sequence; This is the scaling factor; This is the terminology preference coefficient. This is a matrix of the same size as the attention map, used to mark the positions in the input sequence that belong to a predefined medical entity library (such as the dictionary in step S200). If the input sequence's... The word and the first If all the words are medical entities, then Set to 1 if it is set to 1, otherwise set to 0. This is achieved by adding... In this aspect, the model artificially increases the attention score between medical entities, forcing the network to pay more attention to the association between key descriptions such as "hypertension," "accompanied by," and "kidney disease," while ignoring general conjunctions or adverbs that have no practical meaning, thereby improving the accuracy of understanding complex disease descriptions.

[0048] Severity assessment layer: After identifying the specific disease name, the third-layer model performs multi-dimensional risk quantification and outputs the final disease severity score. This score not only reflects the severity of the medical condition but also incorporates the risk of payouts from an insurance business perspective.

[0049] The scoring is based on four dimensions: disease progression stage (e.g., early, middle, late stage), treatment complexity (e.g., conservative treatment, surgery, long-term medication), prognosis (e.g., cure rate, recurrence rate), and insurance impact. Each dimension is quantified with a score from 0 to 10, denoted as... The final scoring formula is as follows: ; In the formula, The loop variable represents the currently being calculated [number of iterations]. One evaluation dimension; Score the severity of the disease; These are the weighting coefficients for each dimension; As a time decay factor, based on the time of disease diagnosis. The calculation is used to reflect the diminishing effect of the impact of old medical history on current risk.

[0050] The fourth dimension, "Insurance Impact Level," is scored based on the claims risk coefficient. Determined. This coefficient is calculated by examining the payout leverage (average claim amount) for this disease in historical data. with average premium income (ratio) and combined with the risk frequency weight get: ; In the formula, For claims risk coefficient; This represents the average claim amount. Average premium income; Risk frequency weighting.

[0051] Frequency weight The score is determined by the percentage of historical claims for the disease and a time adjustment factor. This design ensures that the score not only reflects the biological harm of the disease itself, but also directly reflects the financial pressure it may put on the insurance fund pool.

[0052] Reference Figure 1 In step S400, the system transforms the quantitative results output in step S300 into decision support information that underwriters can directly refer to. This process mainly includes risk level determination, confidence correlation, suggestion matching, and data encapsulation and transmission.

[0053] The system executes risk level determination logic. Preset risk threshold ranges are stored in the system configuration parameters, typically including low-risk thresholds. and high risk threshold The system will use the disease severity score calculated in step S300. Compare the values ​​with these two thresholds: like The system will identify the risk level as "low risk", and the corresponding standard underwriting process will be automatically approved. like The system identifies the risk level as "medium risk" and prompts that manual review or conditional underwriting should be considered. like The system will identify the risk level as "high risk" and suggest rejection or postponement of the application.

[0054] The system obtains the prediction confidence of the hierarchical disease identification model output. This confidence level is derived from the maximum value of the Softmax probability distribution output by the specific disease identification layer (Transformer encoder), or the inverse of the variance of the regression prediction from the severity assessment layer. It characterizes the degree of certainty the model has regarding the current judgment result. If the prediction confidence level is lower than a preset confidence threshold (e.g., 0.75), even if the severity score is low, the system will forcibly raise the risk level label by one level to prevent the risk of missed coverage due to model uncertainty.

[0055] Subsequently, based on the disease name entity and risk level identifier, the system matches the corresponding underwriting recommendation measures in a pre-set underwriting recommendation rule base. This rule base contains structured processing logic, such as: "If the identified entity is 'thyroid nodule' and the risk level is 'medium risk', then the matching recommendation is 'retrieve ultrasound reports within the last 6 months' and 'coverage with exclusions'."

[0056] Finally, the system encapsulates the above information into a standardized JSON format data packet. The structure of this data packet is as follows: it includes a unique insurance application ID field, a calculated numeric type... Fields include enumerated risk level identifier fields, floating-point predicted confidence level fields, and text list-type underwriting recommendation fields.

[0057] The JSON data packet is sent to the underwriting terminal via an encrypted network transmission protocol. Upon receiving the data packet, the underwriting terminal parses and renders the front-end interface. The interface displays risk information in a visually prominent manner; for example, for high-risk levels, the disease name and score are highlighted in red, and a pop-up window displays underwriting recommendations, requiring underwriters to confirm each recommendation before submitting their final conclusion. This interactive mechanism ensures that the risk characteristics calculated by the model are transmitted seamlessly to the decision-making process, achieving an effective combination of algorithmic assistance and human experience.

[0058] Reference Figure 1 Step S500 constructs a reverse data loop from the application end to the model end, aiming to continuously calibrate the model's decision boundary using high-confidence human underwriting experience. Step S500 specifically covers three core technical aspects: the collection and standardization of decision data, the maintenance of the incremental sample library, and the updating of model parameters based on the Elastic Weight Consolidation (EWC) algorithm.

[0059] The system collects in real-time the final decisions made by underwriters in response to the underwriting risk warnings through an interactive interface provided by the underwriting terminal. These final underwriting conclusions are typically discrete category labels, including "normal underwriting" (corresponding to low risk), "rejection" (corresponding to high risk), and "conditional underwriting" (corresponding to medium risk and specific exclusions). The system uses these manually determined conclusions as Ground Truth labels and associates them with the disease diagnosis text in the original current insurance application data to construct standardized training samples. The samples are then stored in a separate incremental sample database. The system monitors the cumulative number of samples in this database in real time. ,when When the preset update threshold is reached (e.g., 1000 or 5000 records), the system suspends the background inference service copy and triggers the offline incremental update process.

[0060] In the incremental update process, to prevent the model from forgetting existing medical logic when learning new disease features or underwriting rules (i.e., the catastrophic forgetting problem), this invention does not employ a direct fine-tuning strategy, but instead uses an elastic weight consolidation algorithm to constrain the parameter update trajectory. The core of this algorithm lies in constructing a total loss function that includes regularization constraints. .

[0061] The total loss function The calculation formula is defined as follows: ; In the formula, The model parameter vector that is currently being updated. The first in A specific numerical value; This represents the vector of model parameters that needs to be updated. The original loss function calculated for newly collected incremental samples is usually the cross-entropy loss function, which is used to guide the model to fit the new manual underwriting rules. This is the regularization strength coefficient, used to balance the weight between learning new knowledge and retaining old memories; The model was determined at the end of the last training task. The optimal values ​​for each parameter.

[0062] The second term in the formula is the regularization term, where It is the first one on the diagonal of the Fisher Information Matrix. The element represents the first element. Each parameter has a weight representing its importance to the historical task. The calculation logic is based on historical data (or sampling of the previous round of data), and its value approximates the loss function in terms of parameters. The second derivative at a given point reflects the curvature of the parameter in the parameter space.

[0063] Specifically, if A larger value indicates that the first This parameter lies in the steep region of the loss function; a small change in this parameter can lead to a significant increase in the loss of past tasks, meaning that this parameter is crucial for maintaining historical underwriting rules. In this case, the regularization term incurs a large penalty value, forcing the optimization algorithm to be cautious when updating this parameter and limiting its deviation. The magnitude. Conversely, if A smaller value indicates that the parameter is in a flat region and has little impact on historical tasks; the model then allows the parameter to adjust according to... The gradient guidance was adjusted significantly to adapt to the new data characteristics.

[0064] By minimizing the above total loss function The system uses stochastic gradient descent (SGD) or the Adam optimizer to iteratively update the model parameters. After the update is complete, the system will update the new parameters. The existing model parameters are replaced, and a new Fisher information matrix is ​​recalculated for use in the next update. This mechanism ensures that the hierarchical disease identification model can dynamically absorb newly emerging disease risk perceptions as insurance business data accumulates, while maintaining the stability of the system's memory of basic medical principles and existing core business rules.

[0065] Reference Figure 2 , Figure 2 This is an architectural diagram of an internet health insurance underwriting device based on disease characteristics, according to an embodiment of the present invention. This embodiment provides an internet health insurance underwriting device based on disease characteristics. Logically, the device consists of a data acquisition module, a feature construction module, a disease identification module, a risk warning module, and a model optimization module. Each module interacts and operates collaboratively through a system bus or internal communication protocol.

[0066] The data acquisition module, serving as the device's input interface, is equipped with a standardized database connector (such as a JDBC / ODBC bridge) and a RESTful API listening interface. This module establishes bidirectional communication with the insurance company's core business system, executing SQL queries to extract historical insurance data awaiting underwriting, including structured records of past claims and unstructured medical record text and images. Simultaneously, the module receives current insurance application data in real-time via the API interface, parsing the chief complaint text, description of present illness, and diagnostic conclusions contained within. In scenarios with wearable device access permissions, the data acquisition module is also equipped with a streaming data acquisition unit to buffer and clean high-frequency physiological monitoring data from external sensors, integrating it into time-series segments for transmission to the next-level module.

[0067] The feature construction module is signal-connected to the data acquisition module and integrates a natural language processing engine and vector computation unit. The feature construction module's storage space contains a pre-built word segmentation dictionary specifically for health insurance and a high-dimensional word vector lookup table. During runtime, the feature construction module first calls a word segmentation algorithm to clean and standardize the disease diagnosis text. Then, it activates the vector computation unit to perform similarity calculations based on claims co-occurrence relevance, mapping discrete text words into continuous numerical feature vectors. If the input data includes physiological monitoring indicators, the feature construction module is also responsible for calling a dynamic health scoring algorithm to calculate normalized score values ​​and fusing them into the text feature vector through vector concatenation, forming a joint feature vector output.

[0068] The disease identification module is the core computing unit of this device, used to host and execute a pre-trained hierarchical disease identification model. Logically, this module is divided into three cascaded processing subunits: a physiological system classification subunit, a specific disease identification subunit, and a severity assessment subunit. The physiological system classification subunit runs an improved random forest algorithm, quickly determining the major category of a disease based on the health insurance weighted Gini coefficient; the specific disease identification subunit runs a deep neural network based on the Transformer architecture, using a built-in medical terminology mask matrix to perform vertical domain attention calculations and accurately extract specific disease entity features; the severity assessment subunit is responsible for executing multi-dimensional scoring logic, combining the degree of insurance impact and time decay factors to output the final disease severity score.

[0069] The risk warning module is used for logical judgment and information encapsulation in underwriting decisions. Internally, this module stores risk threshold configuration files and compares the received disease severity score with preset low-risk and high-risk thresholds using a comparator circuit or logic program to generate a corresponding risk level identifier. Simultaneously, this module is equipped with a JSON serialization engine, which encapsulates the risk level, prediction confidence level, and matched underwriting recommendations into a data packet conforming to transmission standards. This data packet is then sent to the underwriting terminal via a network interface card, driving the terminal screen for visual rendering.

[0070] The model optimization module establishes data pathways with both the risk warning module and the disease identification module, forming a closed-loop feedback mechanism. This module monitors human decision-making instructions from the underwriting terminal, parses the conclusions of "normal underwriting," "rejection," or "conditional underwriting" into supervisory labels, and stores them in an incremental sample buffer paired with the original text data. An integrated parameter update controller monitors the sample backlog in the buffer in real time. Once a trigger threshold is reached, the controller immediately initiates a training process based on the elastic weight consolidation algorithm. It calculates the importance weights of historical parameters using the Fisher information matrix, and while constraining parameter offsets, fine-tunes the weights of the hierarchical disease identification model using the loss gradient calculated from new samples. The updated parameters are then written back to the disease identification module.

[0071] The electronic device described below can be referenced in relation to the internet health insurance underwriting method based on disease characteristics described above.

[0072] Please see Figure 3 The present invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it is able to perform the above method.

Claims

1. A method for underwriting internet-based health insurance based on disease characteristics, characterized in that, Includes the following steps: S1. Obtain historical insurance data to be underwritten and current insurance application data, wherein the current insurance application data includes disease diagnosis text; S2. Construct a health insurance-specific word vector library based on the historical insurance data, and convert the disease diagnosis text into feature vectors; S3. Input the feature vector into a pre-trained hierarchical disease recognition model to obtain the output disease severity score; the hierarchical disease recognition model includes a physiological system classification layer, a specific disease recognition layer, and a severity assessment layer connected in sequence. S4. Generate an underwriting risk warning based on the severity score of the disease, and send the underwriting risk warning to the underwriting terminal; S5. Collect the results of manual underwriting decisions in response to the underwriting risk warning, and use the results of manual underwriting decisions as new training samples to incrementally update the hierarchical disease identification model.

2. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S2, the construction of a dedicated word vector library for health insurance based on the historical insurance data specifically includes: Calculate the health insurance similarity between two medical terms; The calculation logic for the similarity in the health insurance field is as follows: calculate the cosine similarity of two medical terms in the vector space, and then calculate the co-occurrence correlation of the two medical terms in historical claims cases. The health insurance domain similarity is obtained by weighting the claim co-occurrence correlation using a preset domain adjustment coefficient and then multiplying it with the cosine similarity.

3. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S3, the physiological system classification layer specifically includes: A classification model is constructed using the random forest algorithm, wherein the classification splitting criterion of the decision tree nodes in the random forest algorithm is determined based on the health insurance weighted Gini coefficient; The calculation logic for the weighted Gini coefficient of health insurance is as follows: calculate the original Gini impurity of the current node sample set, and calculate the risk importance of the candidate feature; the risk importance is determined based on the contribution of the candidate feature to the historical claims amount in the historical insurance data. The original Gini impurity is combined with the risk importance weighted by the risk sensitivity parameter to obtain the health insurance weighted Gini coefficient, and the feature that minimizes this coefficient is selected as the splitting feature.

4. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S3, the specific disease identification layer specifically includes: A deep learning network based on the Transformer encoder architecture is adopted, and the deep learning network is configured with a health insurance vertical domain attention mechanism; The computational logic of the attention mechanism in the health insurance vertical domain is as follows: a medical terminology mask matrix is ​​introduced during the standard dot product attention calculation process; the medical terminology mask matrix is ​​used to mark the positions in the input sequence that belong to a predefined medical entity library; When calculating attention weights, preference weights are superimposed on the positions of the medical term mask matrix markers.

5. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S3, the severity assessment layer specifically includes: Calculate the dimensional scores of the identified diseases in four dimensions: disease progression stage, treatment complexity, prognosis, and insurance impact. The weighted sum of the scores for the four dimensions is multiplied by a time decay factor determined based on the time of disease diagnosis to obtain the final disease severity score. The dimension score for the insurance impact dimension is determined based on the claims risk coefficient, which is calculated by taking into account the ratio of the average claims amount for the disease in the historical insurance data to the average premium income of the corresponding insurance product, and combining the claim frequency weight of the disease in the insured population.

6. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S4, generating the underwriting risk warning specifically includes: The disease severity score is compared with a preset risk threshold to determine the corresponding risk level identifier; Obtain the prediction confidence score output by the hierarchical disease identification model; The risk level identifier, the predicted confidence level, and the underwriting recommendations based on the disease severity score are encapsulated into a JSON data packet as the underwriting risk warning.

7. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S5, the collection of manual underwriting decision results for the underwriting risk warning specifically includes: Receive the final underwriting conclusion fed back through the underwriting operation terminal, the final underwriting conclusion including normal coverage, rejection, or conditional coverage; The final underwriting conclusion is associated with the disease diagnosis text in the current insurance application data to construct a new training sample containing the correct label, and the new training sample is stored in the sample database. The number of newly added training samples in the sample database is monitored in real time, and an incremental update process is triggered when the number reaches a preset update threshold.

8. The method for internet health insurance underwriting based on disease characteristics according to claim 1, characterized in that, In step S5, the incremental update of the hierarchical disease identification model specifically includes: An elastic weight consolidation algorithm is used to construct a total loss function, which consists of a new task loss term for the new training samples and a regularization term for protecting old knowledge. The calculation logic for the regularization term is as follows: calculate the importance weight of each model parameter of the hierarchical disease identification model in the historical task based on the Fisher information matrix; In the process of minimizing the total loss function, the update magnitude of the model parameters corresponding to the importance weight constraints is used to preserve the hierarchical disease identification model's ability to remember historical underwriting rules.

9. A disease-characteristic-based internet health insurance underwriting device, applied to the disease-characteristic-based internet health insurance underwriting method described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire historical insurance data to be underwritten and current insurance application data, wherein the current insurance application data includes disease diagnosis text; The feature construction module, connected to the data acquisition module, is used to construct a health insurance-specific word vector library based on the historical insurance data and convert the disease diagnosis text into feature vectors. The disease identification module, connected to the feature construction module, is used to input the feature vector into a pre-trained hierarchical disease identification model to obtain the output disease severity score; the hierarchical disease identification model includes a physiological system classification layer, a specific disease identification layer, and a severity assessment layer connected in sequence. The risk warning module, connected to the disease identification module, is used to generate underwriting risk warnings based on the severity score of the disease and send the underwriting risk warnings to the underwriting operation terminal. The model optimization module is connected to the risk warning module and the disease identification module respectively. It is used to collect the results of manual underwriting decisions in response to the underwriting risk warning, and use the results of manual underwriting decisions as new training samples to incrementally update the hierarchical disease identification model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements an internet health insurance underwriting method based on disease characteristics as described in any one of claims 1-8.

Images