Dynamic data-driven risk control method for health insurance claim

Abstract

The application discloses a dynamic data-driven risk control method for health insurance claim settlement, comprising: collecting and preprocessing claim text data to generate a text corpus; screening normal claim settlement samples based on compliance labels, latent Dirichlet distribution topic modeling, and constructing a topic probability feature library; K nearest neighbor algorithm comparison to determine the K nearest normal claim settlement samples; calculating an abnormal score, comparing with a preset threshold, and identifying and warning abnormalities; dynamically collecting and updating data, regularly adjusting parameters, and forming a risk control closed loop. The application realizes automatic abnormal detection and risk warning for the whole process of health insurance claim settlement, and improves the intelligent and accurate level of claim settlement risk control.

Description

Technical Field

[0001] This invention relates to the field of intelligent risk control technology, and in particular to a dynamic data-driven risk control method for health insurance claims. Background Technology

[0002] Against the backdrop of rapid growth in health insurance claims, insurance companies face multiple challenges, including a large volume of claims, complex data types, and strong hidden risks. Existing claims risk control technologies primarily rely on rule engines and static feature analysis, often depending on expert-defined judgment rules or a limited amount of structured data features. While these technologies achieved initial screening and detection of some abnormal claims in the early stages, they suffer from significant shortcomings when dealing with unstructured data such as text-based medical documents, medical records, and claims explanations with complex semantic relationships. These shortcomings include weak modeling capabilities, poor knowledge transferability, and difficulty in dynamically adapting to business changes. Traditional risk control methods typically lack deep semantic mining and feature modeling of claims applications, failing to effectively identify high-risk behavioral patterns hidden within claims materials. Complex risks such as repeated claims, malicious collusion, or disguised excessive medical reimbursement are difficult to detect in real time. Existing methods often employ single static thresholds and judgment mechanisms dominated by human experience, which easily leads to false positives and false negatives. This increases the manual review burden on claims departments and negatively impacts customer experience and business compliance.

[0003] With the continuous development of data-driven risk control and artificial intelligence technologies, some claims risk identification methods based on statistical learning and machine learning are gradually being applied to the health insurance field. However, current mainstream solutions mostly focus on risk identification based on structured features, lacking multi-dimensional modeling and dynamic anomaly detection mechanisms for unstructured text semantics. As a result, the models are difficult to adapt to the rapid changes in high-frequency and new types of abnormal applications in actual claims scenarios.

[0004] Therefore, how to provide dynamic data-driven risk control methods for health insurance claims is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a dynamic data-driven risk control method for health insurance claims. This invention fully combines latent Dirichlet distribution topic modeling and the K-nearest neighbor algorithm to perform deep semantic feature mining and intelligent risk identification on text data in the health insurance claims process. It describes in detail the anomaly detection and adaptive risk control process based on topic probability distribution, which has the advantages of accurate risk identification, strong dynamic adaptability and high degree of risk control automation.

[0006] The dynamic data-driven risk control method for health insurance claims according to embodiments of the present invention includes:

[0007] Collect text data generated during the health insurance claims process, including historical claims text data and claims text data to be analyzed, and preprocess the text data to generate a text corpus;

[0008] Based on the compliance review tags of historical claims text data, normal claims samples are screened, and topic modeling is performed on the text corpus through latent Dirichlet distribution to obtain the topic probability distribution vector corresponding to each claims text. A topic probability distribution feature library for normal claims samples is constructed, and corresponding topic probability distribution feature vectors are generated for the claims text data to be analyzed.

[0009] The K-nearest neighbor algorithm is used to calculate the nearest neighbor distance between the feature vector of the topic probability distribution of the claim text data to be analyzed and the normal claim samples in the topic probability distribution feature library, and to determine the K nearest claim samples to the claim text data to be analyzed.

[0010] Calculate the anomaly score of the claim text data to be analyzed based on the feature vector of the topic probability distribution of K claim samples, preset the claim anomaly score threshold, compare the anomaly score with the claim anomaly score threshold, identify abnormal claim applications and trigger risk warnings;

[0011] The system dynamically collects new claims data and updates the text corpus, and regularly updates the relevant parameters of the potential Dirichlet distribution and the K-nearest neighbor algorithm to achieve a dynamic data-driven closed loop for health insurance claims risk control.

[0012] Optionally, the historical claims text data specifically includes medical diagnostic reports, medical records, claims instructions, and invoice item descriptions corresponding to closed health insurance claims, and the claims text data to be analyzed specifically includes medical diagnostic reports, medical records, claims instructions, and invoice item descriptions corresponding to current health insurance claims that require risk identification.

[0013] Optionally, the preprocessing of the text data specifically includes word segmentation, stop word removal, and semantic standardization.

[0014] Optionally, the compliance review tags based on historical claims text data are used to filter normal claims samples. The text corpus is then modeled using a latent Dirichlet distribution to obtain the topic probability distribution vector corresponding to each claims text. A topic probability distribution feature library for normal claims samples is constructed. For the claims text data to be analyzed, a corresponding topic probability distribution feature vector is generated, including:

[0015] The collected historical claims text data is filtered according to compliance review tags to obtain normal claims samples, and the normal claims samples and the claims text data to be analyzed are included in the text corpus.

[0016] Based on the aforementioned text corpus, we statistically analyzed the specialized terms related to disease names, treatment items, and insurance terms in the health insurance field, and constructed a field-specific glossary.

[0017] In the initialization process of the topic-word distribution of the latent Dirichlet distribution, a fixed prior probability value of topic is set for each word in the domain-specific vocabulary, and the domain-specific words are assigned to each topic based on the prior probability value of topic. The domain-specific vocabulary is used to impose prior constraints on the topic generation process of the latent Dirichlet distribution.

[0018] After completing the prior constraints on topic-word distribution, topic modeling is performed on the text corpus based on the latent Dirichlet distribution. Combining the content complexity of the text corpus with the sparsity of topic distribution, the perplexity index is used to dynamically adjust the number of topics k.

[0019] Based on the dynamically adjusted number of topics k, for each claim text data in the text corpus, topic inference is performed on the claim text data through the latent Dirichlet distribution. The posterior probability value of the claim text data assigned to each topic is statistically calculated, and the posterior probability values ​​corresponding to all topics are arranged in order according to the topic number to form the topic probability distribution feature vector of the claim text data.

[0020] Based on the tag information obtained through compliance review in historical claims text data, the topic probability distribution of normal claims samples is used as the optimization objective. A semi-supervised regularization term is designed. In the process of topic modeling based on the potential Dirichlet distribution, the tag information is used to guide parameter updates. The topic probability distribution of normal claims samples remains stable and meets the reasonable requirements of the health insurance business field.

[0021] For all normal claim samples, the topic probability distribution is inferred one by one to obtain the posterior probability value of each normal claim sample on all topics. The posterior probability values ​​of the same normal claim sample on each topic are arranged in the order of topic number to form the topic probability distribution feature vector of normal claim samples. The feature vectors are summarized according to the claim sample number to establish the topic probability distribution feature library of normal claim samples. The topic probability distribution feature library contains the topic probability distribution feature vector corresponding to each normal claim sample and the total number of all normal claim samples.

[0022] For the claims text data to be analyzed, the latent Dirichlet distribution method based on semi-supervised regularization to complete topic modeling is adopted. The topic probability distribution inference is performed on each claims text data to be analyzed, and the topic probability distribution feature vector of each claims text data to be analyzed is generated as the input feature of the K nearest neighbor algorithm.

[0023] Optionally, the step of using the K-nearest neighbor algorithm to calculate the nearest neighbor distance between the topic probability distribution feature vector of the claims text data to be analyzed and the normal claims samples in the topic probability distribution feature library, and determining the K nearest claims samples to the claims text data to be analyzed, includes:

[0024] For each claim text data to be analyzed, extract the theme probability distribution feature vector generated by potential Dirichlet distribution theme modeling;

[0025] Obtain the theme probability distribution feature vector of all normal claims samples from the theme probability distribution feature library of normal claims samples, and obtain the theme probability distribution feature vector of historical abnormal claims samples from the claims samples marked as abnormal by the compliance review label in historical claims text data;

[0026] Based on the topic probability distribution characteristics of normal claims samples and historical abnormal claims samples, the nearest neighbor discrimination accuracy is evaluated on the sample sets of normal claims samples and historical abnormal claims samples for three distance measurement functions: Euclidean distance, Manhattan distance and cosine distance. The distance measurement function with the highest accuracy is selected as the distance calculation method of the K-nearest neighbor algorithm.

[0027] Based on the ability of each theme component to distinguish anomalies in historical risk control tasks, weights are assigned to each theme component of the theme probability distribution feature vector to form a set of weighted coefficients.

[0028] Based on the thematic probability distribution characteristics of normal claims samples, the degree of fluctuation of each thematic component is statistically analyzed, and a dynamic offset correction factor is set for thematic components with large fluctuations.

[0029] Using a defined distance metric function, combined with a set of weighted coefficients and a dynamic offset correction factor, the weighted distance between the topic probability distribution feature vector of the claims text data to be analyzed and the topic probability distribution feature vector of each normal claims sample is calculated.

[0030] Sort all weighted distances in ascending order and select the K nearest normal claim samples to form a preliminary candidate nearest neighbor sample set;

[0031] A consistency test is performed on the preliminary candidate nearest neighbor sample set. The weighted distance between each normal claim sample and the other normal claim samples in the preliminary candidate nearest neighbor sample set is calculated. If the weighted distance between a normal claim sample and other normal claim samples is greater than a preset distance deviation threshold, the normal claim sample is removed from the preliminary candidate nearest neighbor sample set. Only normal claim samples whose weighted distances are all within the threshold range are retained. Finally, the K nearest normal claim samples to the claim text data to be analyzed are determined.

[0032] Optionally, the step of calculating the anomaly score of the claim text data to be analyzed based on the feature vector of the topic probability distribution of K claim samples, setting a claim anomaly score threshold, comparing the anomaly score with the claim anomaly score threshold, identifying abnormal claim applications and triggering risk warnings includes:

[0033] For each piece of claim text data to be analyzed, obtain the topic probability distribution feature vectors of the K nearest normal claim samples;

[0034] For each nearest normal claim sample, a defined weighted distance metric is used to calculate the weighted distance between the topic probability distribution feature vector of the claim text data to be analyzed and the topic probability distribution feature vector of each normal claim sample, resulting in K weighted distance values.

[0035] Perform statistical analysis on the K weighted distance values, calculate the mean and standard deviation of the weighted distance, and at the same time calculate the average weighted distance between the claim text data to be analyzed and the K nearest normal claim samples;

[0036] Based on the calculated K weighted distance values, the average of all weighted distance values ​​is calculated, and the average of the K weighted distances is used as the anomaly score of the claim text data to be analyzed.

[0037] Set a threshold for abnormal claims scores. The threshold for abnormal claims scores can be determined based on empirical statistics of the weighted distance distribution of historical claims samples.

[0038] The abnormal score of the claim text data to be analyzed is compared with the claim abnormal score threshold. If the abnormal score is greater than the claim abnormal score threshold, it is determined to be an abnormal claim application, triggering the risk warning process.

[0039] For all claim text data that is determined to be abnormal, record the abnormality score and weighted distance statistical characteristics, generate a risk warning report and store it.

[0040] Optionally, the dynamic collection of new claims data and updating of the text corpus, and the periodic updating of relevant parameters of the latent Dirichlet distribution and the K-nearest neighbor algorithm, to achieve a dynamic data-driven closed loop for health insurance claims risk control, includes:

[0041] Dynamically collect newly added claim text data in health insurance claims business, including medical diagnosis reports, medical records, claim instructions and invoice item descriptions, and incorporate the newly added claim text data into a text corpus;

[0042] All claims text data in the text corpus are processed according to the preprocessing workflow, including word segmentation, stop word removal, and semantic standardization.

[0043] Regularly update the compliance review tags of the claims text data in the text corpus, and include newly added normal claims samples and abnormal claims samples into the corresponding datasets. In a timely manner, supplement the topic probability distribution feature library of normal claims samples and the feature library of historical abnormal claims samples.

[0044] For the updated text corpus, the potential Dirichlet distribution topic model is re-performed periodically, the topic probability distribution is inferred for all claims text data, the topic distribution parameters are optimized and the topic probability distribution feature vector of each sample is updated;

[0045] The topic component weighting coefficients, dynamic offset correction factors, and distance metric function parameters involved in the K-nearest neighbor algorithm are periodically re-evaluated and optimized.

[0046] Based on dynamically updated text corpora, potential Dirichlet distribution topic modeling results, and K-nearest neighbor algorithm parameters, we continuously conduct claims risk detection, abnormal application identification, and risk warning.

[0047] The beneficial effects of this invention are:

[0048] This invention enhances the ability of health insurance claims risk control systems to mine deep semantic features and detect anomalies in textual data by integrating latent Dirichlet distribution topic modeling with the K-nearest neighbor algorithm. Compared with traditional risk control methods that rely on rules and structured features, this invention automatically extracts topic probability distribution features from unstructured texts such as medical diagnosis reports, medical records, and claims instructions, achieving accurate characterization and discrimination of implicit risk patterns in claims content. The K-nearest neighbor algorithm overcomes the problems of distance failure and outlier interference in high-dimensional feature spaces through adaptive selection of distance metrics, introduction of topic component weighting, and dynamic offset correction mechanisms, thereby enhancing the sensitivity and accuracy of discrimination against complex claims anomalies.

[0049] In practical applications, this invention dynamically collects new claims data and updates thematic modeling and anomaly detection parameters in real time, enabling the risk control model to self-learn and achieve closed-loop optimization. It supports flexible adjustments to risk control thresholds and strategies, continuously adapting to the rapid changes in health insurance claims. The system can automatically identify diverse and concealed high-risk claims, reducing the burden of manual review and the error rate, and improving overall claims service efficiency and risk management capabilities. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of the dynamic data-driven risk control method for health insurance claims proposed in this invention;

[0052] Figure 2 This diagram illustrates the potential Dirichlet distribution topic modeling and claims text feature extraction for the dynamic data-driven risk control method for health insurance claims proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1 and Figure 2 Dynamic data-driven risk control methods for health insurance claims include:

[0055] Collect text data generated during the health insurance claims process, including historical claims text data and claims text data to be analyzed, and preprocess the text data to generate a text corpus;

[0056] Based on the compliance review tags of historical claims text data, normal claims samples are screened, and topic modeling is performed on the text corpus through latent Dirichlet distribution to obtain the topic probability distribution vector corresponding to each claims text. A topic probability distribution feature library for normal claims samples is constructed, and corresponding topic probability distribution feature vectors are generated for the claims text data to be analyzed.

[0057] The K-nearest neighbor algorithm is used to calculate the nearest neighbor distance between the feature vector of the topic probability distribution of the claim text data to be analyzed and the normal claim samples in the topic probability distribution feature library, and to determine the K nearest claim samples to the claim text data to be analyzed.

[0058] Calculate the anomaly score of the claim text data to be analyzed based on the feature vector of the topic probability distribution of K claim samples, preset the claim anomaly score threshold, compare the anomaly score with the claim anomaly score threshold, identify abnormal claim applications and trigger risk warnings;

[0059] The system dynamically collects new claims data and updates the text corpus, and regularly updates the relevant parameters of the potential Dirichlet distribution and the K-nearest neighbor algorithm to achieve a dynamic data-driven closed loop for health insurance claims risk control.

[0060] In this embodiment, the historical claims text data specifically includes medical diagnostic reports, medical records, claims explanations, and invoice item descriptions corresponding to closed health insurance claims, and the claims text data to be analyzed specifically includes medical diagnostic reports, medical records, claims explanations, and invoice item descriptions corresponding to current health insurance claims that require risk identification.

[0061] In this embodiment, the preprocessing of text data specifically includes word segmentation, stop word removal, and semantic standardization.

[0062] In this embodiment, the compliance review tags based on historical claims text data are used to screen normal claims samples. The text corpus is then modeled using a latent Dirichlet distribution to obtain the topic probability distribution vector corresponding to each claims text. A topic probability distribution feature library for normal claims samples is constructed. For the claims text data to be analyzed, a corresponding topic probability distribution feature vector is generated, including:

[0063] The collected historical claims text data is filtered according to compliance review tags to obtain normal claims samples, and the normal claims samples and the claims text data to be analyzed are included in the text corpus.

[0064] Based on the aforementioned text corpus, we statistically analyzed the specialized terms related to disease names, treatment items, and insurance terms in the health insurance field, and constructed a field-specific glossary.

[0065] In the initialization process of the topic-word distribution of the latent Dirichlet distribution, a fixed prior probability value of topic is set for each word in the domain-specific vocabulary, and the domain-specific words are assigned to each topic based on the prior probability value of topic. The domain-specific vocabulary is used to impose prior constraints on the topic generation process of the latent Dirichlet distribution.

[0066] After completing the prior constraints on topic-word distribution, topic modeling is performed on the text corpus based on the latent Dirichlet distribution. Combining the content complexity of the text corpus with the sparsity of topic distribution, the perplexity index is used to dynamically adjust the number of topics k.

[0067] Based on the dynamically adjusted number of topics k, for each claim text data in the text corpus, topic inference is performed on the claim text data through the latent Dirichlet distribution. The posterior probability value of the claim text data assigned to each topic is statistically calculated, and the posterior probability values ​​corresponding to all topics are arranged in order according to the topic number to form the topic probability distribution feature vector of the claim text data.

[0068] Based on the tag information obtained through compliance review in historical claims text data, the topic probability distribution of normal claims samples is used as the optimization objective. A semi-supervised regularization term is designed. In the process of topic modeling based on the potential Dirichlet distribution, the tag information is used to guide parameter updates. The topic probability distribution of normal claims samples remains stable and meets the reasonable requirements of the health insurance business field.

[0069] For all normal claim samples, the topic probability distribution is inferred one by one to obtain the posterior probability value of each normal claim sample on all topics. The posterior probability values ​​of the same normal claim sample on each topic are arranged in the order of topic number to form the topic probability distribution feature vector of normal claim samples. The feature vectors are summarized according to the claim sample number to establish the topic probability distribution feature library of normal claim samples. The topic probability distribution feature library contains the topic probability distribution feature vector corresponding to each normal claim sample and the total number of all normal claim samples.

[0070] For the claims text data to be analyzed, the latent Dirichlet distribution method based on semi-supervised regularization to complete topic modeling is adopted. The topic probability distribution inference is performed on each claims text data to be analyzed, and the topic probability distribution feature vector of each claims text data to be analyzed is generated as the input feature of the K nearest neighbor algorithm.

[0071] In this embodiment, the step of using the K-nearest neighbor algorithm to calculate the nearest neighbor distance between the topic probability distribution feature vector of the claims text data to be analyzed and the normal claims samples in the topic probability distribution feature library, and to determine the K nearest claims samples to the claims text data to be analyzed, includes:

[0072] For each claim text data to be analyzed, extract the theme probability distribution feature vector generated by potential Dirichlet distribution theme modeling;

[0073] Obtain the theme probability distribution feature vector of all normal claims samples from the theme probability distribution feature library of normal claims samples, and obtain the theme probability distribution feature vector of historical abnormal claims samples from the claims samples marked as abnormal by the compliance review label in historical claims text data;

[0074] Based on the topic probability distribution characteristics of normal claims samples and historical abnormal claims samples, the nearest neighbor discrimination accuracy is evaluated on the sample sets of normal claims samples and historical abnormal claims samples for three distance measurement functions: Euclidean distance, Manhattan distance and cosine distance. The distance measurement function with the highest accuracy is selected as the distance calculation method of the K-nearest neighbor algorithm.

[0075] Based on the ability of each theme component to distinguish anomalies in historical risk control tasks, weights are assigned to each theme component of the theme probability distribution feature vector to form a set of weighted coefficients.

[0076] Based on the thematic probability distribution characteristics of normal claims samples, the degree of fluctuation of each thematic component is statistically analyzed, and a dynamic offset correction factor is set for thematic components with large fluctuations.

[0077] Using a defined distance metric function, combined with a set of weighted coefficients and a dynamic offset correction factor, the weighted distance is calculated between the topic probability distribution feature vector of the claims text data to be analyzed and the topic probability distribution feature vector of each normal claims sample:

[0078]

[0079] Where, d qj and d nj D respectively q and D n The component value of the j-th topic, δ j dist is the dynamic offset correction factor for the j-th topic component. w，δ (D q D n ) represents the weighted distance between the claims text data to be analyzed and the nth normal claims sample, k is the number of topics, and w j D is the weighting coefficient for the j-th topic component. q Let D be the feature vector of the topic probability distribution of the claims text data to be analyzed. n Let be the feature vector of the thematic probability distribution of the nth normal claim sample;

[0080] Sort all weighted distances in ascending order and select the K nearest normal claim samples to form a preliminary candidate nearest neighbor sample set;

[0081] A consistency test is performed on the preliminary candidate nearest neighbor sample set. The weighted distance between each normal claim sample and the other normal claim samples in the preliminary candidate nearest neighbor sample set is calculated. If the weighted distance between a normal claim sample and other normal claim samples is greater than a preset distance deviation threshold, the normal claim sample is removed from the preliminary candidate nearest neighbor sample set. Only normal claim samples whose weighted distances are all within the threshold range are retained. Finally, the K nearest normal claim samples to the claim text data to be analyzed are determined.

[0082] In this embodiment, the step of calculating the anomaly score of the claim text data to be analyzed based on the feature vector of the topic probability distribution of K claim samples, setting a claim anomaly score threshold, comparing the anomaly score with the claim anomaly score threshold, identifying abnormal claim applications and triggering risk warnings includes:

[0083] For each piece of claim text data to be analyzed, obtain the topic probability distribution feature vectors of the K nearest normal claim samples;

[0084] For each nearest normal claim sample, a defined weighted distance metric is used to calculate the weighted distance between the topic probability distribution feature vector of the claim text data to be analyzed and the topic probability distribution feature vector of each normal claim sample, resulting in K weighted distance values.

[0085] Perform statistical analysis on the K weighted distance values, calculate the mean and standard deviation of the weighted distance, and at the same time calculate the average weighted distance between the claim text data to be analyzed and the K nearest normal claim samples;

[0086] Based on the calculated K weighted distance values, the average of all weighted distance values ​​is calculated, and the average of the K weighted distances is used as the anomaly score of the claim text data to be analyzed.

[0087] Set a threshold for abnormal claims scores. The threshold for abnormal claims scores can be determined based on empirical statistics of the weighted distance distribution of historical claims samples.

[0088] The abnormal score of the claim text data to be analyzed is compared with the claim abnormal score threshold. If the abnormal score is greater than the claim abnormal score threshold, it is determined to be an abnormal claim application, triggering the risk warning process.

[0089] For all claim text data that is determined to be abnormal, record the abnormality score and weighted distance statistical characteristics, generate a risk warning report and store it.

[0090] In this embodiment, the dynamic collection of newly added claims data and updating of the text corpus, and the periodic updating of relevant parameters of the latent Dirichlet distribution and the K-nearest neighbor algorithm, to achieve a dynamic data-driven closed loop for health insurance claims risk control, include:

[0091] Dynamically collect newly added claim text data in health insurance claims business, including medical diagnosis reports, medical records, claim instructions and invoice item descriptions, and incorporate the newly added claim text data into a text corpus;

[0092] All claims text data in the text corpus are processed according to the preprocessing workflow, including word segmentation, stop word removal, and semantic standardization.

[0093] Regularly update the compliance review tags of the claims text data in the text corpus, and include newly added normal claims samples and abnormal claims samples into the corresponding datasets. In a timely manner, supplement the topic probability distribution feature library of normal claims samples and the feature library of historical abnormal claims samples.

[0094] For the updated text corpus, the potential Dirichlet distribution topic model is re-performed periodically, the topic probability distribution is inferred for all claims text data, the topic distribution parameters are optimized and the topic probability distribution feature vector of each sample is updated;

[0095] The topic component weighting coefficients, dynamic offset correction factors, and distance metric function parameters involved in the K-nearest neighbor algorithm are periodically re-evaluated and optimized.

[0096] Based on dynamically updated text corpora, potential Dirichlet distribution topic modeling results, and K-nearest neighbor algorithm parameters, we continuously conduct claims risk detection, abnormal application identification, and risk warning.

[0097] Example 1:

[0098] To verify the feasibility of this invention in practice, it was applied to the claims center of an insurance company. The total number of claims each month was constantly increasing, and the insurance company needed to manually review a large amount of textual materials, such as diagnostic certificates, hospital records, claims instructions, and medical invoices. Faced with complex and diverse unstructured text data, human reviewers needed to invest a significant amount of time in initial screening and multiple rounds of review, which easily led to untimely identification of high-risk claims, missed double claims, and delays in reviewing some abnormal cases. The claims center reviewed an average of about 730 claims per month, with an average review period of about 2.5 days per claim, and a manual review rate of 14.2%. Statistics from January to June of the year showed that 41 high-risk claims were identified through a combination of automatic detection and manual review, but the accuracy rate of manual review in identifying high-risk cases was only 70.7%, the false positive rate for anomaly detection was 3.2%, the missed detection rate was 2.4%, the claims complaint rate was high, and customer feedback on claims processing timeliness and experience was poor.

[0099] In January 2025, the insurance company officially launched the health insurance claims dynamic data-driven risk control system of this invention. All newly submitted claims text data are automatically collected and stored in the system. The system performs word segmentation, stop word removal, and standardization on the text materials, and uniformly incorporates them into topic modeling analysis. Based on the latent Dirichlet distribution method, the system generates a topic probability distribution feature vector for each claims text and performs multi-dimensional comparison with the feature library of historical normal claims samples. The K-nearest neighbor algorithm introduces feature weighting, dynamic correction, and anomaly removal strategies in distance calculation, making it more sensitive to high-risk topics. The anomaly score of each claims application is automatically compared with a preset claim anomaly score threshold. High-risk cases are immediately alerted and pushed to the manual review team, achieving a combination of automatic screening and key manual verification. The system supports periodic dynamic collection of new claims data and automatic updates to the text corpus, topic modeling parameters, and K-nearest neighbor discrimination parameters.

[0100] Table 1 Comparison of the effects of the health insurance claims risk control system before and after its application

[0101]

[0102] As shown in Table 1, the method of this invention has brought significant improvements in efficiency and risk control in actual health insurance claims processing. After the system went live, the total number of claims increased from 5,120 to 5,380, but the average review time was significantly shortened from 2.5 days to 1.1 days, indicating a marked improvement in claims processing speed and enhancing the customer's claims experience. The number of automatically detected high-risk cases increased from 41 to 54, demonstrating the system's stronger ability and wider coverage in identifying high-risk cases. The accuracy rate of high-risk cases also improved significantly, from 70.7% to 88.9%, indicating that the intelligent risk control method reduced the probability of misjudgments and omissions.

[0103] The manual review rate decreased from 14.2% to 2.7%, significantly enhancing the accuracy and intelligence of the system's automatic screening and reducing the workload of manual reviewers. The false alarm rate for anomaly detection decreased from 3.2% to 0.8%, and the false negative rate also decreased from 2.4% to 0.3%, indicating that the method of this invention not only improves the sensitivity of risk warnings but also reduces the occurrence of false alarms and false negatives. The system has achieved remarkable results in improving the efficiency of health insurance claims processing, risk control capabilities, and customer service experience, fully demonstrating the practical value of data-driven intelligent risk control in actual health insurance claims.

[0104] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic data-driven fraud detection method for health insurance claims, characterized in that, The method comprises the following steps: Collecting text data generated in the health insurance claim process, including historical claim text data and to-be-analyzed claim text data, and preprocessing the text data to generate a text corpus; Based on the compliance audit label of the historical claim text data, screening normal claim samples, modeling the text corpus based on the latent Dirichlet distribution to obtain the topic probability distribution vector corresponding to each claim text, and constructing a topic probability distribution feature library of normal claim samples, and generating a corresponding topic probability distribution feature vector for the to-be-analyzed claim text data; Using K-Nearest Neighbor algorithm to calculate the nearest neighbor distance between the topic probability distribution feature vector of the to-be-analyzed claim text data and the normal claim samples in the topic probability distribution feature library, and determining the K nearest claim samples of the to-be-analyzed claim text data; According to the topic probability distribution feature vector of the K claim samples, the abnormal score of the to-be-analyzed claim text data is calculated, a preset claim abnormal score threshold is set, the abnormal score and the claim abnormal score threshold are compared, the claim abnormal application is identified, and the risk early warning is triggered; Dynamically collecting new claim data and updating the text corpus, regularly updating the related parameters of the latent Dirichlet distribution and the K-Nearest Neighbor algorithm, and realizing the dynamic data-driven health insurance claim risk control closed loop.

2. The dynamic data driven fraud detection method for health insurance claims according to claim 1, wherein, The historical claim text data specifically includes the medical diagnosis report, medical record, claim explanation and invoice item description corresponding to the closed health insurance claim application, and the to-be-analyzed claim text data specifically includes the medical diagnosis report, medical record, claim explanation and invoice item description corresponding to the health insurance claim application which needs to be risk-identified at present.

3. The method of claim 1, wherein the method further comprises: The preprocessing of the text data specifically includes word segmentation, stop word removal and semantic standardization processing of the text data.

4. The method of claim 1, wherein the method further comprises: The compliance audit label based on the historical claim text data, screening normal claim samples, modeling the text corpus based on the latent Dirichlet distribution to obtain the topic probability distribution vector corresponding to each claim text, and constructing a topic probability distribution feature library of normal claim samples, and generating a corresponding topic probability distribution feature vector for the to-be-analyzed claim text data, comprises: According to the compliance audit label, the normal claim samples are obtained by screening the collected historical claim text data, and the normal claim samples and the to-be-analyzed claim text data are included in the text corpus; Based on the text corpus, the disease name, diagnosis and treatment item and insurance clause related special vocabulary in the health insurance field are counted to construct a field special vocabulary table; In the topic-word distribution initialization process of the latent Dirichlet distribution, a fixed topic prior probability value is set for each vocabulary in the field special vocabulary table, and the field special vocabulary is assigned to each topic based on the topic prior probability value, and the topic generation process of the latent Dirichlet distribution is constrained in advance by using the field special vocabulary table; After the prior constraint of the topic-word distribution is completed, the text corpus is modeled based on the latent Dirichlet distribution, the number of topics k is dynamically adjusted by using the perplexity index in combination with the content complexity of the text corpus and the sparsity degree of the topic distribution; Based on the dynamically adjusted number of topics k, for each claim text data in the text corpus, the latent Dirichlet distribution is used to infer the topics of the claim text data, the posterior probability value of the claim text data assigned to each topic is counted, and the posterior probability values of all topics are arranged in order according to the topic number to form the topic probability distribution feature vector of the claim text data; Based on the label information obtained through compliance review in the historical claim text data, the topic probability distribution of the normal claim sample is taken as the optimization target, and a semi-supervised regularization term is designed to guide the parameter update using the label information during the topic modeling based on the latent Dirichlet distribution, so that the topic probability distribution of the normal claim sample remains stable and meets the rationality requirements of the health insurance business field; For all normal claim samples, the topic probability distribution is inferred one by one to obtain the posterior probability value of each normal claim sample on all topics, and the posterior probability values of the same normal claim sample on each topic are arranged in order according to the topic number to form the topic probability distribution feature vector of the normal claim sample. The topic probability distribution feature vectors of the normal claim samples are summarized according to the claim sample number to establish a topic probability distribution feature library of the normal claim samples, which contains the topic probability distribution feature vector corresponding to each normal claim sample and the total number of all normal claim samples. For the to-be-analyzed claim text data, the latent Dirichlet distribution method based on semi-supervised regularization is used to complete topic modeling, and the topic probability distribution of each to-be-analyzed claim text data is inferred to generate the topic probability distribution feature vector of each to-be-analyzed claim text data as the input feature of the K-nearest neighbor algorithm.

5. The method of claim 1, wherein the method further comprises: The K-nearest neighbor algorithm is used to calculate the neighbor distance between the topic probability distribution feature vector of the to-be-analyzed claim text data and the normal claim samples in the topic probability distribution feature library, and the K nearest claim samples to the to-be-analyzed claim text data are determined, which includes: For each to-be-analyzed claim text data, the topic probability distribution feature vector generated by the latent Dirichlet distribution topic modeling is extracted; From the topic probability distribution feature library of the normal claim samples, the topic probability distribution feature vectors of all normal claim samples are obtained, and from the claim samples labeled as abnormal through compliance review in the historical claim text data, the topic probability distribution feature vectors of the historical abnormal claim samples are obtained; Based on the topic probability distribution features of the normal claim samples and the historical abnormal claim samples, the neighbor discrimination accuracy is evaluated on the sample set of the normal claim samples and the historical abnormal claim samples for three distance measurement functions of Euclidean distance, Manhattan distance and cosine distance, and the distance measurement function with the highest accuracy is selected as the distance calculation method of the K-nearest neighbor algorithm; According to the discrimination ability of each topic component in the historical risk control task, a weight is assigned to each topic component of the topic probability distribution feature vector to form a set of weighting coefficients; For the topic probability distribution features of the normal claim samples, the fluctuation degree of each topic component is counted, and a dynamic offset correction factor is set for the topic component with large fluctuation. The subject probability distribution feature vector of the to-be-analyzed claim text data and the subject probability distribution feature vector of each normal claim sample are calculated for weighted distance by using a determined distance measurement function, in combination with a set of weighting coefficients and a dynamic offset correction factor; All weighted distances are sorted in ascending order, and the K nearest normal claim samples are selected to form a preliminary candidate neighbor sample set; The consistency of the preliminary candidate neighbor sample set is tested, the weighted distance between each normal claim sample and the remaining normal claim samples in the preliminary candidate neighbor sample set is calculated, and if the weighted distance between a certain normal claim sample and other normal claim samples is greater than a preset distance deviation threshold, the normal claim sample is removed from the preliminary candidate neighbor sample set, only the normal claim samples with weighted distances within the threshold range are retained, and finally the K nearest normal claim samples to the to-be-analyzed claim text data are determined.

6. The dynamic data driven fraud detection method for health insurance claims according to claim 1, wherein, The abnormal score of the to-be-analyzed claim text data is calculated according to the subject probability distribution feature vectors of the K claim samples, a preset claim abnormal score threshold is set, the abnormal score and the claim abnormal score threshold are compared, the claim abnormal application is identified, and the risk early warning is triggered, including: For each to-be-analyzed claim text data, the subject probability distribution feature vectors of the determined K nearest normal claim samples are obtained; For each nearest normal claim sample, the weighted distance between the subject probability distribution feature vector of the to-be-analyzed claim text data and the subject probability distribution feature vector of each normal claim sample is calculated by using a determined weighted distance measurement method, and K weighted distance values are obtained; The K weighted distance values are statistically analyzed, the mean and standard deviation of the weighted distance are calculated, and the average weighted distance between the to-be-analyzed claim text data and the K nearest normal claim samples is calculated; The average of all weighted distance values is calculated according to the K calculated weighted distance values, and the average of the K weighted distances is taken as the abnormal score of the to-be-analyzed claim text data; The claim abnormal score threshold is set, and the claim abnormal score threshold can be determined according to the empirical statistics of the historical claim sample weighted distance distribution; The abnormal score of the to-be-analyzed claim text data is compared with the claim abnormal score threshold, if the abnormal score is greater than the claim abnormal score threshold, it is determined as an abnormal claim application, and the risk early warning process is triggered; For all abnormal claim text data, the abnormal score and the weighted distance statistical characteristics are recorded, a risk early warning report is generated and stored.

7. The method of claim 1, wherein the method further comprises: The newly added claim text data is dynamically collected and the text corpus is updated, the related parameters of the latent Dirichlet distribution and the K nearest neighbor algorithm are updated regularly, and a dynamic data-driven health insurance claim risk control closed loop is realized, including: The newly added claim text data in the health insurance claim business is dynamically collected, including medical diagnosis reports, medical records, claim explanations and invoice item descriptions, and the newly added claim text data is included in the text corpus; All claim text data in the text corpus is processed by word segmentation, stop word removal and semantic standardization according to the pretreatment process; Periodically update the compliance audit label of the claim text data in the text corpus, and add the newly added normal claim samples and abnormal claim samples into the corresponding data sets respectively, and timely supplement the theme probability distribution feature library of normal claim samples and the feature library of historical abnormal claim samples; For the updated text corpus, periodically re-perform latent Dirichlet allocation topic modeling, infer the theme probability distribution of all claim text data, optimize the theme distribution parameters and update the theme probability distribution feature vector of each sample; Periodically re-evaluate and optimize the theme component weighting coefficient, dynamic offset correction factor and distance metric function parameter involved in the K nearest neighbor algorithm; Based on the dynamically updated text corpus, the results of latent Dirichlet allocation topic modeling and the parameters of K nearest neighbor algorithm, continuously carry out claim risk detection, abnormal application identification and risk warning.

Images