Healthcare abnormality recognition method based on background structure, medium and device

Abstract

This invention discloses a method, medium, and device for identifying medical insurance anomalies based on background structure. The method includes: acquiring medical invoice data for T consecutive accounting periods in a regulatory region, and aggregating the data to obtain the total regional expenditure C. t The amount reimbursed separately by each institution x i,t For any medical institution i to be verified, construct a background mask and calculate the total background reimbursement amount X. −i,t A constrained regression equation is established; based on time series samples, robust parameter estimation mechanisms such as iterative weighted least squares are used to stably solve for the individual contribution coefficient 'a'. i The invention calculates the set of individual contribution coefficients for all institutions within a regulatory region, constructs a statistical baseline, and calculates a robust risk score. Finally, it uses the time-series variation characteristics of the robust risk score and / or individual contribution coefficients to perform multidimensional anomaly identification and risk warning. This invention effectively solves the model underdeterminacy problem under high-dimensional small sample conditions and can accurately separate the abnormal behavior of a single institution from the overall regional fluctuations.

Description

Technical Field

[0001] This invention relates to the field of intelligent supervision of medical insurance funds, specifically to a method, medium, and device for identifying medical insurance anomalies based on background structure. Background Technology

[0002] Currently, the supervision of medical insurance funds mainly relies on manual auditing, expert rule engines, or basic statistical analysis (same-year and month-on-month comparisons). With the development of data science, statistical models such as regression analysis are beginning to be applied to anomaly detection, identifying risks by analyzing the correlation between medical institution reimbursement amounts and total regional expenditures.

[0003] However, in practical applications, existing technologies suffer from the following significant drawbacks: the curse of dimensionality and underdeterminacy (core pain points): In regional supervision, limited by data statistical standards, only monthly summary data is often available. When the number of medical institutions N (e.g., 100) is much larger than the observable time series sample size T (e.g., 31 months), traditional all-variable regression models suffer from severe underdeterminacy (N+1>T), failing to stably estimate the contribution coefficient of each institution. Furthermore, they are heavily influenced by background noise, and the total regional medical insurance expenditure is significantly affected by seasonal epidemics, policy adjustments, and public health emergencies. Traditional models struggle to accurately isolate atypical abnormal behaviors of individual medical institutions from the drastically fluctuating "background noise." Summary of the Invention

[0004] In view of the above problems, the present invention provides a technical solution for medical insurance anomaly identification based on background structure, in order to solve the technical problem that existing technologies are unable to accurately separate abnormal behavior of a single medical institution from the overall fluctuations of the regulatory area.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for identifying medical insurance anomalies based on background structure, the method comprising:

[0006] S1: Obtain medical invoice data for T consecutive accounting periods within the regulatory region, extract the medical insurance fund payment amount from the medical invoice data, and aggregate them to obtain the total medical insurance reimbursement expenditure C within the regulatory region. t The individual reimbursement amount for each medical institution within the regulated region x i,t Where T is the number of accounting cycles, i is the medical institution index, and t is the accounting cycle index;

[0007] S2: For any medical institution i to be verified within the regulatory area, construct a background mask, treating all medical institutions other than medical institution i as a whole, and calculate the total background reimbursement amount corresponding to medical institution i. The calculation formula is as follows: X -i,t =C t -x i,t And establish a restricted regression equation:

[0008] ;

[0009] in, For the total medical insurance reimbursement expenditure in the regulatory region during the t-th accounting period, X represents the independent reimbursement amount for medical institution i in the t-th accounting period. -i,t For the total amount of background reimbursement, Let be the individual contribution coefficient of medical institution i, used to characterize the degree to which the medical insurance reimbursement behavior of this medical institution is driven by the demand for medical treatment in the external region. Here, c represents the background contribution coefficient, and c represents the environmental constant offset. Let be the regression residual for the t-th accounting period;

[0010] S3: Time series samples based on T accounting periods {(x i,1 ,X -i,1 ,C1),(x i,2 ,X -i,2 ,C2),...,(x i,T ,X -i,T C T A robust parameter estimation mechanism is used to estimate the individual contribution coefficients in the restricted regression equation. To achieve a stable solution, the robust parameter estimation mechanism includes:

[0011] S31: Perform iterative weighted least squares estimation on the time series samples, and assign adaptive weights w to each accounting period t. t ;

[0012] S32: Based on the regression residuals corresponding to each accounting period t Size dynamically adjusts weight w t For outlier sample points with large residuals, their weights should be reduced.

[0013] S33: Through multiple rounds of iteration, the individual contribution coefficient is increased. The estimated value remains stable even with reduced weights for outlier sample points with large residuals, thus obtaining the individual contribution coefficient that reflects the stable behavioral characteristics of the target medical institution across multiple accounting periods. ;

[0014] S4: Calculate the set of individual contribution coefficients {a1, a2, ..., a...} for all N medical institutions within the regulatory area. N Based on the set of individual contribution coefficients, a statistical baseline is constructed, and a robust risk score for each medical institution is calculated.

[0015] S5: Based on the time series variation characteristics of the robust risk score and / or individual contribution coefficient, perform multidimensional anomaly determination on the target medical institution and trigger corresponding risk warnings. The time series variation characteristics include the trend variation characteristics and stability variation characteristics of the individual contribution coefficient in multiple consecutive accounting periods.

[0016] Furthermore, in step S4, constructing the statistical baseline specifically includes:

[0017] Calculate the median Med(a) and absolute median deviation MAD(a) of the set of individual contribution coefficients;

[0018] Calculate robust sizing estimates The calculation formula is as follows: Where k is a preset scale calibration factor used to adjust the absolute median deviation. Convert it into a robust measure of dispersion with dimensions consistent with the standard deviation under a normal distribution;

[0019] Calculate robust risk score The calculation formula is as follows:

[0020] .

[0021] Furthermore, step S5 includes static anomaly determination and dynamic anomaly determination;

[0022] The static anomaly determination includes: based on the robust risk score, when the robust risk score of the target medical institution in a certain accounting period exceeds the preset static threshold, it is determined that there is a static anomaly and an early warning is triggered;

[0023] The dynamic anomaly determination includes: evaluating the individual contribution coefficient sequence of the target medical institution over multiple consecutive accounting periods based on the time series change characteristics of the individual contribution coefficient; when the change characteristics of the individual contribution coefficient sequence meet a preset dynamic anomaly pattern, a dynamic anomaly is determined to exist and an early warning is triggered; the dynamic anomaly pattern includes at least one of the following:

[0024] The individual contribution coefficient sequence shows a statistically significant monotonic trend, and its deviation from the center of the population distribution increases as the trend develops;

[0025] The fluctuation range of the individual contribution coefficient sequence between consecutive accounting periods exceeds the dynamic fluctuation threshold determined by the concurrent fluctuation level of the medical institutions in the regulatory region.

[0026] The individual contribution coefficient remains in the abnormal state range determined by the group distribution for more than the preset number of durations.

[0027] Furthermore, in step S3, after solving for the individual contribution coefficient... Then, the following steps are also included:

[0028] Obtain the static attribute vector Z of the target medical institution i. i The static attribute vector includes one or more of the following: medical institution level, approved number of beds, and number of key departments;

[0029] Constructing a mapping model between medical institution attributes and contribution coefficients, specifically including: based on the static attribute vector set {Z1, Z2, ..., Z...} of all N medical institutions within the regulatory area. N} and its corresponding set of individual contribution coefficients {a1, a2, ..., a N Train a regression model f(·) to obtain the structural contribution determined by the inherent attributes of the medical institution. The calculation formula is as follows: ;

[0030] The individual contribution coefficient a i Decomposed into structural contribution coefficients and behavioral contribution coefficient ,in:

[0031] ;

[0032] In step S5, a multidimensional anomaly assessment is performed on the target medical institution, and corresponding risk warnings are triggered, including:

[0033] Combination and Perform multidimensional anomaly detection:

[0034] When the first condition is met, it is determined to be a risk of distorted medical practice. The first condition includes: the robustness risk score exceeds a preset risk threshold, and simultaneously meets the following conditions: >τ b and >λ· , where τ b λ is the threshold for the behavioral contribution coefficient, and λ is the contribution ratio coefficient.

[0035] When the second condition is met, it is determined to be a risk of mismatch between scale and positioning. The second condition includes: the robustness risk score exceeds a preset risk threshold, and simultaneously meets the following conditions: >τ s and >λ· And based on the static attribute vector Z i The calculated change in the medical institution size positioning index over M consecutive accounting periods is within the preset error range, where τ sThe threshold for the structural contribution coefficient.

[0036] Furthermore, in step S2, constructing the background mask includes:

[0037] Acquire patient transfer data among medical institutions within the regulatory region over a consecutive T accounting periods;

[0038] For any medical institution i to be verified and the accounting period t, a directed weighted network is constructed based on the flow data; wherein, the nodes of the directed weighted network are medical institutions, and if a patient flows from medical institution u to medical institution v within the period t, there exists an edge from u to v, with the weight being the number of patients transferred.

[0039] Starting with medical institution i, in the directed weighted network, all medical institutions that have direct or indirect patient transfer relationships with medical institution i are selected to form the dynamic background institution set B of medical institution i in period t. i,t Among them, the path length between selected medical institutions is less than the preset length;

[0040] Calculate the total amount of reimbursement under dynamic background The calculation formula is as follows:

[0041] ;

[0042] Where j represents the dynamic background mechanism set B i,t Index of Chinese medical institutions This represents the independent reimbursement amount for medical institution j in the dynamic background institution set during the t-th accounting period;

[0043] In the restricted regression equation, the total amount of reimbursement under the dynamic background is used. Replace the original background reimbursement total amount x i,t Perform the calculation.

[0044] Furthermore, the process of identifying multidimensional anomalies in the target medical institution and triggering corresponding risk warnings includes the following steps:

[0045] Calculate the behavioral correlation w between any two medical institutions p and q based on the time series of historical individual contribution coefficients or the business similarity between medical institutions. pq This forms a matrix of relationships among medical institutions;

[0046] For the current accounting period t, calculate the change in the individual contribution coefficient of each medical institution pair (p, q). and And calculate the correlation change score of medical institutions. The calculation formula is as follows: ;in, This represents the individual contribution coefficient of medical institution p within the current accounting period t. This represents the individual contribution coefficient of medical institution p in the previous accounting period t-1. This represents the individual contribution coefficient of medical institution q within the current accounting period t. This represents the individual contribution coefficient of medical institution q in the previous accounting period t-1;

[0047] If a medical institution pair (p, q) that meets the third condition is identified, an association anomaly alert is triggered. The third condition includes the behavioral association degree w between medical institutions p and q. pq The score for the association change of (p,q) is higher than the preset association threshold. The absolute value is higher than the preset correlation change threshold, and its sign is negative.

[0048] Furthermore, in step S4, a statistical baseline is constructed based on the set of individual contribution coefficients, and the robust risk score for each medical institution is calculated, including:

[0049] Based on preset association rules, the direct association strength ρ between any two medical institutions p and q within the regulatory area is calculated. pq , where 0≤ρ pq ≤1, ρ pp = 1, and construct an N-row, N-column correlation strength matrix R;

[0050] For any medical institution i, calculate its individual abnormality intensity. The calculation formula is as follows:

[0051] ;

[0052] in, and These are sets of individual contribution coefficients based on historical cycles, {a i The calculated benchmark mean and benchmark standard deviation;

[0053] Calculate the native robust risk score of healthcare institution i The native robust risk score By analyzing the individual abnormal intensity of medical institution i The result is obtained after smoothing and standardization.

[0054] Based on the aforementioned correlation strength matrix R, the network infection risk bonus of medical institution i is calculated. The calculation formula is:

[0055] ;

[0056] in, and The preset attenuation coefficient, >0, <1; This represents the additional attenuation factor for secondary transmission, 0 < <1, Let be the shortest path length from medical institution j to medical institution k;

[0057] Calculate the comprehensive robust risk score R of medical institution i i The calculation formula is as follows:

[0058] .

[0059] Furthermore, the method also includes:

[0060] Obtain a list of known medical insurance policy shock events that occur within T consecutive accounting periods, and generate a policy shock identifier vector D for each accounting period t. t , where D t It is a multi-dimensional vector, where each dimension corresponds to a specific policy shock event. If the policy shock event continues to be effective in period t or later, the corresponding dimension value is 1; otherwise, it is 0.

[0061] The policy impact identification vector D t Introducing the restricted regression equation as a control variable yields the extended restricted regression equation:

[0062] ;

[0063] Where G is the policy shock identifier vector D t The coefficient vector corresponding to the dimension is used to capture the impact of various policy shocks on the total medical insurance reimbursement expenditure C in the regulated area. t The average systemic impact, It is the transpose of G;

[0064] In step S3, based on the time series samples and the extended restricted regression equation, a robust parameter estimation mechanism is used to simultaneously estimate the individual contribution coefficient 'a'. i The policy shock coefficient vector G is then used to obtain a stable solution.

[0065] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical insurance anomaly identification method based on background structure as described in the first aspect of the present invention.

[0066] In a third aspect, the present invention provides an electronic device having a computer program stored thereon, including a processor and a storage medium, wherein the computer program is stored on the storage medium, and when executed by the processor, the computer program implements the medical insurance anomaly identification method based on background structure as described in the first aspect of the present invention.

[0067] Unlike existing technologies, the above technical solution involves a method, medium, and device for identifying medical insurance anomalies based on background structure. The method includes: acquiring medical invoice data for T consecutive accounting periods within a regulatory region, and aggregating it to obtain the total regional expenditure C. t The amount reimbursed separately by each institution x i,t For any medical institution i to be verified, construct a background mask and calculate the total background reimbursement amount X. -i,t A constrained regression equation is established; based on time series samples, robust parameter estimation mechanisms such as iterative weighted least squares are used to stably solve for the individual contribution coefficient 'a'. i The invention calculates the set of individual contribution coefficients for all institutions within the regulatory region, constructs a statistical baseline, and calculates a robust risk score. Finally, it performs multidimensional anomaly identification and risk warning based on the time series variation characteristics of the robust risk score and / or individual contribution coefficients. This invention effectively solves the model underdeterminacy problem under high-dimensional, small-sample conditions through "individual-background" comparative modeling, and can accurately separate the abnormal behavior of a single institution from the overall regional fluctuations, improving the accuracy and robustness of medical insurance anomaly identification.

[0068] The above description of the invention is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical solution of the present invention and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of the present invention easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of the present invention. Attached Figure Description

[0069] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on the present invention.

[0070] In the accompanying drawings of the instruction manual:

[0071] Figure 1 This is a flowchart of the medical insurance anomaly identification method based on background structure as described in the first exemplary embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of an electronic device according to an exemplary embodiment of the present invention;

[0073] The reference numerals used in the above figures are explained as follows:

[0074] 10. Electronic equipment; 101. Processor; 102. Storage medium. Detailed Implementation

[0075] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that can be achieved by this invention, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this invention, and are therefore only examples, and should not be used to limit the scope of protection of this invention.

[0076] In the first aspect, such as Figure 1 As shown, this application provides a method for identifying medical insurance anomalies based on background structure, the method comprising:

[0077] S1: Obtain medical invoice data for T consecutive accounting periods within the regulatory region, extract the medical insurance fund payment amount from the medical invoice data, and aggregate them to obtain the total medical insurance reimbursement expenditure C within the regulatory region. t The individual reimbursement amount for each medical institution within the regulated region x i,t Where T is the number of accounting cycles, i is the medical institution index, and t is the accounting cycle index;

[0078] S2: For any medical institution i to be verified within the regulatory area, construct a background mask, treating all medical institutions other than medical institution i as a whole, and calculate the total background reimbursement amount X corresponding to medical institution i. -i,t The calculation formula is as follows: X -i,t =C t -x i,t And establish a restricted regression equation:

[0079] ;

[0080] in, For the total medical insurance reimbursement expenditure in the regulatory region during the t-th accounting period, X represents the independent reimbursement amount for medical institution i in the t-th accounting period. -i,t For the total amount of background reimbursement, Let be the individual contribution coefficient of medical institution i, used to characterize the degree to which the medical insurance reimbursement behavior of this medical institution is driven by the demand for medical treatment in the external region. Here, c represents the background contribution coefficient, and c represents the environmental constant offset. Let be the regression residual for the t-th accounting period;

[0081] S3: Time series samples based on T accounting periods {(x i,1 ,X -i,1 ,C1),(x i,2 ,X -i,2,C2),...,(x i,T ,X -i,T C T A robust parameter estimation mechanism is used to estimate the individual contribution coefficients in the restricted regression equation. To achieve a stable solution, the robust parameter estimation mechanism includes:

[0082] S31: Perform iterative weighted least squares estimation on the time series samples, and assign adaptive weights w to each accounting period t. t ;

[0083] S32: Based on the regression residuals corresponding to each accounting period t Size dynamically adjusts weight w t For outlier sample points with large residuals, their weights should be reduced.

[0084] S33: Through multiple rounds of iteration, the individual contribution coefficient is increased. The estimated value remains stable even with reduced weights for outlier sample points with large residuals, thus obtaining the individual contribution coefficient that reflects the stable behavioral characteristics of the target medical institution across multiple accounting periods. ;

[0085] S4: Calculate the set of individual contribution coefficients {a1, a2, ..., a...} for all N medical institutions within the regulatory area. N Based on the set of individual contribution coefficients, a statistical baseline is constructed, and a robust risk score for each medical institution is calculated.

[0086] S5: Based on the time series variation characteristics of the robust risk score and / or individual contribution coefficient, perform multidimensional anomaly determination on the target medical institution and trigger corresponding risk warnings. The time series variation characteristics include the trend variation characteristics and stability variation characteristics of the individual contribution coefficient in multiple consecutive accounting periods.

[0087] In this embodiment, medical invoice data refers to the full data of various invoices such as outpatient medical invoices and inpatient medical invoices generated by medical institutions within the regulatory area during the accounting period. It includes core fields such as invoicing unit, invoice type, settlement time, and detailed cost composition, and is the basic data source for extracting the payment amount of the medical insurance fund.

[0088] The amount paid by the medical insurance pooled fund refers to the portion of medical expenses paid by the medical insurance pooled fund for insured individuals. It is a core indicator for measuring the scale of medical insurance reimbursement for medical institutions and the consumption of regional medical insurance funds, and is accurately extracted from the expense composition details of medical bills.

[0089] The accounting cycle refers to the time statistical unit for the supervision of medical insurance funds. It is preferably a continuous time period (such as a natural month), denoted as T. It is the basis for constructing time series samples and ensures the temporal continuity and comparability of data.

[0090] Background masking refers to the regional reference system constructed by the medical institution to be inspected, which treats all medical institutions in the region except for that medical institution as a whole, thereby separating the behavior of a single institution from the behavior of the region as a whole, and realizing the binary deconstruction of "individual-background".

[0091] The restricted regression equation differs from the low-dimensional regression model of the traditional full-scale multiple regression equation. It only includes parameters such as the independent reimbursement amount, background reimbursement amount, and constant term of the medical institutions to be verified, which satisfies the parameter identifiability condition and solves the problem of parameter unsolvability caused by the number of medical institutions being much larger than the number of time samples.

[0092] The individual contribution coefficient refers to the degree to which the medical insurance reimbursement behavior of the medical institution under investigation is driven by the medical needs of the external region. It is a core indicator for measuring the consistency between the institution's reimbursement behavior and the overall regional situation. Its value is highly correlated with the compliance and rationality of the institution's behavior.

[0093] Robust parameter estimation mechanisms refer to parameter solving methods based on iterative weighted least squares. By assigning adaptive weights to time series samples and dynamically adjusting them, the interference of extreme outliers on parameter estimation is suppressed, ensuring the stability and statistical reliability of individual contribution coefficients.

[0094] Robust risk score is a quantitative risk indicator calculated based on the population distribution characteristics of individual contribution coefficients of all medical institutions in a region. It is used to measure the degree to which the individual contribution coefficient of a single institution deviates from the normal level of the region and is the core quantitative basis for anomaly judgment.

[0095] Multidimensional anomaly detection refers to an anomaly identification method that combines horizontal group distribution deviation characteristics (robust risk score) and vertical time series evolution characteristics (trend and stability changes in individual contribution coefficients). It differs from traditional single-dimensional anomaly detection and improves the accuracy and comprehensiveness of identification.

[0096] This embodiment revolves around the logic of "data preprocessing - individual background deconstruction - robust parameter estimation - risk score calculation - multidimensional anomaly detection", realizing the entire process from raw medical insurance data to anomaly warning. The specific steps and principles are as follows:

[0097] In step S1, firstly, all medical invoice data for T consecutive accounting periods within the regulatory region is acquired. The amount paid by the medical insurance fund is then precisely extracted from the detailed breakdown of invoice expenses. This is because the medical insurance fund payment amount is a core indicator directly reflecting the consumption of the medical insurance fund and the reimbursement behavior of medical institutions, excluding interference from non-medical insurance fund-related expenses such as out-of-pocket payments. Subsequently, the data is aggregated according to two dimensions: "region as a whole" and "single institution," yielding the total regional medical insurance reimbursement expenditure C. t The amount reimbursed independently by each medical institution x i,t The aggregation process strictly adheres to the consistency of the accounting cycle and the statistical subject, ensuring the comparability of data in both time and space dimensions, and laying a data foundation for subsequent modeling. Here, T is the number of accounting cycles, i is the index of medical institutions, and t is the index of accounting cycles.

[0098] In step S2, the principles of background mask construction and constrained regression equation establishment are as follows: For any medical institution i to be verified within the region, constructing a background mask is the core step in this method to solve the curse of dimensionality and underdetermined problems. This is achieved through formula X. -i,t =C t -x i,t The calculation of the total reimbursement amount transforms the original N-dimensional multiple regression problem into a low-dimensional problem of "1 (the institution to be verified vs. N-1 (the rest of the institutions as a whole)", effectively avoiding the problem of unpredictable parameters caused by the number of medical institutions N being much larger than the time sample size T.

[0099] Based on this, a restricted regression equation is constructed. , where b i is the background contribution coefficient, which characterizes the stability of the overall medical treatment structure in the region and generally approaches 1; c is the environmental constant offset, which is used to absorb global noise such as changes in medical insurance policies and seasonal adjustments. The regression residuals represent the impact of random factors within the accounting period. This equation decomposes the total regional medical insurance reimbursement expenditure into two parts: the individual contribution of the institutions to be verified and the overall contribution of the regional background. This achieves a preliminary separation of individual behavior from background noise, where the individual contribution coefficient 'a' is... i This is a core parameter to be evaluated, and its value directly reflects the degree to which an institution's reimbursement behavior is driven by external medical needs. Under normal compliance conditions, a i If the medical institution exhibits abnormal behavior such as insurance fraud or excessive medical treatment, close to the population median level, a i The value will be significantly smaller, reflecting a decrease in the consistency between its behavior and the overall region.

[0100] In step S3, the individual contribution coefficient 'a' of the time series samples from T accounting periods to the constrained regression equation is calculated. iInstead of using the traditional ordinary least squares method, a robust parameter estimation mechanism based on iterative weighted least squares is introduced. This is because medical insurance fund data is easily affected by unstructured factors such as seasonal disease outbreaks, public health emergencies, and centralized settlement. Ordinary least squares is easily dominated by a few extreme samples, leading to distorted parameter estimation. This mechanism achieves robustness through three steps:

[0101] First, an adaptive weight w is assigned to each accounting period t. t This serves as the influence coefficient of the sample in parameter estimation; subsequently, based on the regression residuals of each period... Size dynamically adjusts weight w t Outlier sample points with larger residuals have lower weights to suppress the interference of extreme samples; finally, through multiple iterations, the individual contribution coefficient 'a' is optimized. i The estimated value of a tends to stabilize as the weight of outlier samples decreases, and the final value of a is obtained. i It can accurately reflect the long-term stable behavioral characteristics of the target medical institution over multiple accounting periods, rather than being affected by short-term abnormal fluctuations, thus ensuring the statistical stability and comparability of the parameters.

[0102] In step S4, the set of individual contribution coefficients {a1, a2, ..., a...} of all N medical institutions within the regulatory area is obtained. N Following this, constructing a statistical baseline is a crucial step in transforming individual coefficients into risk scores. The statistical baseline is built upon the distribution characteristics of coefficient groups across all institutions within the region, rather than a single fixed threshold, thus adapting to dynamic changes in the regional healthcare structure. Using this statistical baseline as a reference, robust risk scores are calculated for each healthcare institution, quantifying the degree to which individual institutions deviate from the regional normal level and providing comparable and quantifiable indicators for subsequent anomaly assessment.

[0103] In step S5, anomaly detection does not solely rely on robustness risk scores. Instead, it combines robustness risk scores (horizontal deviation characteristics of group distribution) and the time-series variation characteristics of individual contribution coefficients (vertical time-series evolution characteristics) for multi-dimensional assessment. The time-series variation characteristics specifically include trend-based changes (such as continuous increases or decreases) and stability-based changes (such as fluctuation amplitude or whether the institution remains in an abnormal range for an extended period). Horizontal assessment focuses on the degree of deviation between the institution and the overall region within a single accounting period, while vertical assessment focuses on the evolution of the institution's behavioral patterns within consecutive accounting periods. The combination of these two approaches enables comprehensive identification of abnormal behavior. Finally, based on the assessment results, corresponding risk warnings are triggered, and the warning results are output to the medical insurance supervision platform, providing accurate evidence for manual audits and special investigations.

[0104] The method involved in this embodiment transforms the original unsolvable N-dimensional multivariate regression problem into a low-dimensional identifiable parameter estimation problem through a "1 vs N-1" individual-context deconstruction approach. Even under extremely limited conditions where the number of medical institutions N is much larger than the time sample size T (e.g., there are 100 medical institutions in the region, but only 31 months of monthly summary data can be obtained), it can still achieve stable quantitative modeling of the medical insurance reimbursement behavior of a single medical institution. This completely solves the core pain points of unsolvable parameters and unstable results in traditional all-variable regression models, and fills the technical gap in medical insurance anomaly identification modeling under limited sample size scenarios.

[0105] By absorbing global noise such as medical insurance policy adjustments and seasonal epidemics through the environmental constant shift c in the restricted regression equation, and combining it with a robust parameter estimation mechanism to suppress the interference of extreme samples, a dual anti-interference mechanism of "model structure stripping + parameter solution suppression" is achieved. This can accurately extract the real behavioral characteristics of a single medical institution from the drastically fluctuating regional medical insurance total expenditure, avoiding the problem of behavioral characteristic distortion caused by background noise interference in traditional models, and ensuring the accuracy of institutional behavior assessment.

[0106] Unlike traditional methods that focus solely on single-dimensional anomaly detection based on reimbursement amount or absolute value of regression residuals, this method combines horizontal deviations in population distribution with vertical time-series evolution for multi-dimensional detection. It can identify both explicit anomalies that deviate significantly from the overall regional distribution within a single period and implicit anomalies that capture trend changes and fluctuations in behavioral patterns over continuous periods. This effectively reduces the probability of false alarms caused by occasional peaks in medical visits or short-term policy adjustments, while also enhancing the ability to identify hidden and persistent abnormal behaviors, enabling medical insurance supervision to cover more types of abnormal scenarios.

[0107] This invention uses the individual contribution coefficient as the core indicator. Through the construction of statistical baselines and the calculation of robust risk scores, it achieves a quantitative assessment of the degree of abnormality in medical insurance reimbursement of medical institutions. It replaces the subjective judgment methods of traditional manual auditing and expert rule engines, and provides a scientific, objective and quantifiable assessment tool for medical insurance supervision. It reduces the bias caused by subjective judgment and enables the supervision of medical insurance funds to shift from "experience-driven" to "data-driven".

[0108] The method involved in this embodiment simplifies the modeling process into a low-dimensional parameter estimation problem for a single medical institution, eliminating the need for multivariate regression calculations for all N medical institutions. The overall computational complexity is significantly lower than that of traditional full-scale modeling methods. The system can support parallel computing and rolling updates for a massive number of medical institutions within a region, meeting the application needs of medical insurance fund supervision for long-term, continuous, and real-time risk monitoring of a large number of medical institutions. It has good engineering feasibility and scenario adaptability.

[0109] In some embodiments, in step S4, constructing the statistical baseline specifically includes:

[0110] Calculate the median Med(a) and absolute median deviation MAD(a) of the set of individual contribution coefficients;

[0111] Calculate robust sizing estimates The calculation formula is as follows: Where k is a preset scale calibration factor used to adjust the absolute median deviation. Convert it into a robust measure of dispersion with dimensions consistent with the standard deviation under a normal distribution;

[0112] Calculate robust risk score The calculation formula is as follows:

[0113] .

[0114] In this embodiment, the median Med(a) refers to the set of individual contribution coefficients {a1, a2, ..., a3} of all N medical institutions in the region. N The values ​​in the middle position after being arranged in order of size are robust statistics that reflect the location of the group's data center. They are not sensitive to extreme outliers and can truly reflect the normal individual contribution coefficient levels of most medical institutions in the region.

[0115] The absolute median deviation (MAD(a)) is the median of the absolute deviations of each value in the set of individual contribution coefficients from the median Med(a). It is a robust statistic that reflects the dispersion of population data. Compared with the traditional standard deviation, it is less affected by extreme outliers and can accurately characterize the normal dispersion range of individual contribution coefficients of medical institutions in a region.

[0116] The scaling factor k is a preset constant coefficient. Its core function is to convert the absolute median deviation (MAD(a)) into a robust dispersion measure with the same dimensions as the standard deviation under a normal distribution. This makes the robust risk score calculated based on MAD(a) have standardized dimensions and interpretability, which facilitates comparisons across institutions and accounting periods. Preferably, the scaling factor k is 1.4826, which is suitable for the characteristics of the coefficient group that is approximately normally distributed.

[0117] Robust scalar estimate It refers to the robust dispersion index obtained by multiplying the scale calibration factor k and the absolute median deviation MAD(a), which replaces the traditional standard deviation and serves as the scale for measuring the deviation of an individual contribution coefficient from the normal range of the median. It has the characteristic of resisting outlier interference.

[0118] The above scheme is based on robust statistical principles, abandoning traditional mean and standard deviation statistics. Instead, it uses the median and absolute median deviation to construct a statistical baseline, and then obtains a robust risk score through standardized calculation. The specific steps and principles are as follows:

[0119] The statistical baseline is composed of the median Med(a) of the set of individual contribution coefficients and the robust scaling estimate. Together, they constitute the robust scaling estimate, with the median Med(a) serving as the central benchmark. The median was chosen as the benchmark for the discrete range, rather than the mean, because the mean is easily affected by the individual contribution coefficients of a few extremely anomalous institutions, leading to distortion of the central benchmark. For example, if a few institutions in the region exhibit severe anomalies, their contribution coefficients may be significantly affected. i Significantly small values ​​will lower the overall mean, making it unable to reflect the level of most normal institutions; while the median is not sensitive to extreme values. Regardless of changes in extreme outliers, the median always reflects the average level of the group, ensuring the authenticity of the central benchmark. Robust scaling estimates are obtained by selecting the absolute median deviation (MAD(a)) and combining it with the scaling calibration factor k. Instead of directly using the standard deviation, MAD(a) is used because the standard deviation is also sensitive to extreme values. Extreme outliers can significantly amplify the standard deviation, leading to distortion of the discrete range benchmark. MAD(a), calculated using the "median of deviations," statistically shields the interference of extreme values. Furthermore, it uses k to calibrate the dimensions, making... It retains robustness while also possessing dimensional explanatory power consistent with standard deviation, thus achieving a balance between robustness and standardization.

[0120] Robust Risk Score i Essentially, it refers to the standardized deviation of the individual coefficient, that is, the individual contribution coefficient 'a' of a single medical institution. i The absolute deviation from the regional center benchmark Med(a), divided by the robust scaling estimate ( This method standardizes the degree of deviation. The core logic of this calculation is as follows: the discrete range of the individual contribution coefficients of most normal institutions within the region around Med(a) is taken as the "normal interval," S... i The magnitude of S directly represents the degree to which the institution deviates from the normal range. i The larger the value, the greater the deviation of the institution's individual contribution coefficient from the regional normal level, and the greater the risk of abnormal medical insurance reimbursement behavior; S i The smaller the value, the higher the consistency between the institution's behavior and the overall regional behavior, and the lower the risk of anomalies. Furthermore, since Med(a) and MAD(a) are both robust statistics, the S value calculated based on them... i It also inherits the characteristic of resisting outlier interference, and the risk score will not be distorted due to the existence of a few extremely abnormal institutions in the region.

[0121] This embodiment refines the construction of statistical baselines and the calculation of robust risk scores, further enhancing the robustness, standardization, and engineering practicality of the medical insurance anomaly identification method. Compared with traditional statistical baseline construction methods, the specific beneficial effects are as follows:

[0122] By constructing a statistical baseline using the median Med(a) and the absolute median deviation MAD(a), the distortion effect of a few extreme abnormal medical institutions on the baseline is shielded from the statistical foundation. Even if the individual contribution coefficients of individual institutions in the region deviate significantly from the normal level, the statistical baseline can still truly reflect the behavioral characteristics of most normal institutions. This ensures that the calculation basis of the robust risk score is an objective and true regional normal level, avoids the abnormal identification bias caused by the distortion of the baseline, and makes the abnormal detection results more reliable.

[0123] The MAD(a) is converted into a robust scaling estimate consistent with the standard deviation dimension using a scaling calibration factor k. This makes the robust risk score S i It has become a standardized, dimensionless indicator, overcoming the limitations imposed by differences in medical structures across different regions and accounting periods. The S of the same medical institution in different accounting periods... i It can be directly compared to reflect changes in the degree of abnormality in behavior; S of different institutions in the same accounting period i It can be directly compared, reflecting its relative degree of abnormality; at the same time, the standardized S i This allows medical insurance regulatory authorities to more easily set uniform thresholds for anomaly detection without having to repeatedly adjust the thresholds based on different regions or periods, thus reducing the complexity of regulatory operations.

[0124] Median Med(a), median absolute deviation MAD(a), and robust risk score S i The calculations are all based on basic statistical operations, without the need for complex algorithms and large amounts of computing resources. Compared with traditional complex scoring methods based on machine learning and deep learning, this method has higher computational efficiency and can support real-time or near-real-time calculation of robust risk scores for a large number of medical institutions in a region. It perfectly adapts to the engineering needs of online supervision and rolling monitoring of medical insurance funds, ensuring both scoring accuracy and computational efficiency.

[0125] In some embodiments, step S5 includes static anomaly determination and dynamic anomaly determination;

[0126] The static anomaly determination includes: based on the robust risk score, when the robust risk score of the target medical institution in a certain accounting period exceeds the preset static threshold, it is determined that there is a static anomaly and an early warning is triggered;

[0127] The dynamic anomaly determination includes: evaluating the individual contribution coefficient sequence of the target medical institution over multiple consecutive accounting periods based on the time series change characteristics of the individual contribution coefficient; when the change characteristics of the individual contribution coefficient sequence meet a preset dynamic anomaly pattern, a dynamic anomaly is determined to exist and an early warning is triggered; the dynamic anomaly pattern includes at least one of the following:

[0128] The individual contribution coefficient sequence shows a statistically significant monotonic trend, and its deviation from the center of the population distribution increases as the trend develops;

[0129] The fluctuation range of the individual contribution coefficient sequence between consecutive accounting periods exceeds the dynamic fluctuation threshold determined by the concurrent fluctuation level of the medical institutions in the regulatory region.

[0130] The individual contribution coefficient remains in the abnormal state range determined by the group distribution for more than the preset number of durations.

[0131] In this embodiment, static anomaly determination refers to anomaly determination based on a comparison between the robust risk score of the medical institution and a preset static threshold within a single accounting period. It focuses on the horizontal abnormal deviation at a single point in time and is an abnormal identification of the institution's short-term behavior.

[0132] The static threshold refers to the critical value of robust risk score preset by the medical insurance regulatory department based on the characteristics of regional medical structure and the requirements of medical insurance fund supervision. It is the core basis for static anomaly judgment. When an institution's robust risk score exceeds the threshold, it indicates that its behavior in the accounting period deviates from the overall regional level to a short-term abnormal level.

[0133] Dynamic anomaly determination refers to the determination of anomalies based on the time evolution characteristics of the individual contribution coefficient sequence of medical institutions, using multiple consecutive accounting periods as time units. It focuses on long-term time-series behavioral pattern anomalies and is the identification of anomalies in the continuous and covert behavior of institutions.

[0134] Dynamic anomaly patterns refer to the abnormal evolution characteristics of a pre-defined sequence of individual contribution coefficients of an organization. They are the core basis for judging dynamic anomalies and include three core patterns: monotonic change trend, abnormal fluctuation, and continuous anomaly, which correspond to different types of continuous and covert abnormal behaviors, respectively.

[0135] A statistically significant monotonic trend refers to an institution’s individual contribution coefficient showing a continuous upward or downward trend over multiple consecutive accounting periods. This trend is statistically significant and is not a short-term change caused by random fluctuations. Furthermore, the degree to which it deviates from the center of the population distribution increases as the trend develops.

[0136] The dynamic fluctuation threshold is a critical value calculated based on the concurrent fluctuation level of the individual contribution coefficient sequence of all medical institutions in the region to be regulated. It reflects the normal fluctuation range of institutional behavior in the region. When the fluctuation amplitude of a single institution exceeds this threshold, it indicates that the stability of its behavior is significantly lower than the normal level of the region.

[0137] An abnormal state interval refers to an abnormal value range defined by the group distribution characteristics of individual contribution coefficients within a region. It is determined by a statistical baseline and a preset ratio. When an institution's individual contribution coefficients remain within this interval, it indicates that its behavior has deviated from the normal level of the region for a long period of time.

[0138] In static anomaly assessment, the distribution of individual contribution coefficients of all medical institutions in the region within the same accounting period is used as a reference. A robust risk score quantifies the degree of deviation of a single institution, which is then compared with a preset static threshold. The static threshold is set considering factors such as the strictness of regional medical insurance supervision, historical anomaly identification data, and characteristics of the medical structure; it represents a reasonable upper limit for the robust risk score of normal institutions within the region. When an institution's S in a certain accounting period... i When the static threshold is exceeded, it indicates that the institution's medical insurance reimbursement behavior during that period deviates significantly from the behavior characteristics of most normal institutions in the region, posing a risk of short-term abnormal behavior. This triggers a static anomaly warning. It should be noted that the static anomaly assessment is only a risk signal and does not directly confirm abnormal behavior by the institution, as deviations in a single period may be caused by occasional factors and require comprehensive judgment in conjunction with the dynamic anomaly assessment results.

[0139] The basis for dynamic anomaly determination is the sequence of individual contribution coefficients {a1, a2, ..., a...} of medical institutions over multiple consecutive accounting periods. N Its core logic is long-term behavioral pattern evolution analysis. By identifying abnormal evolutionary characteristics of individual contribution coefficient sequences, it captures persistent and covert abnormal behaviors of organizations, avoiding misjudgments caused by accidental factors, and identifying progressive anomalies that cannot be detected by static anomaly assessment. Dynamic anomaly assessment revolves around three preset dynamic anomaly modes, each corresponding to different behavioral anomaly logic:

[0140] Monotonic trend pattern: The individual contribution coefficient sequence shows a statistically significant and continuous upward or downward trend, and the degree of deviation from the center of the group distribution gradually increases. The business logic corresponding to this pattern is that the institution's medical insurance reimbursement behavior has undergone a systemic change, rather than random fluctuations. For example, if an institution engages in continuous over-treatment or fictitious treatment, its individual contribution coefficient will remain low, and the degree of deviation will continue to widen. This gradual change may not exceed the threshold in a single-cycle static judgment, but the long-term trend has already reflected obvious abnormal risks.

[0141] Abnormal fluctuation pattern: The fluctuation range of the individual contribution coefficient sequence exceeds the dynamic fluctuation threshold between consecutive accounting periods. The dynamic fluctuation threshold is calculated from the fluctuation level of all institutions in the region during the same period and reflects the normal fluctuation range of institutional behavior in the region. The business logic corresponding to this pattern is: the medical insurance reimbursement behavior of institutions lacks stability and may have abnormal centralized settlement, splitting of charging items, etc., which leads to significant fluctuations in their individual contribution coefficients in different periods, forming a significant difference from the stable behavior characteristics of normal institutions in the region.

[0142] Persistent Abnormality Pattern: When an individual contribution coefficient remains within an abnormal range determined by the group distribution for a period exceeding a preset duration, the corresponding business logic is that the institution's medical insurance reimbursement behavior deviates from the regional normal level for an extended period. This is not due to accidental factors but rather to persistent abnormal behavior. Even if its single-cycle robustness risk score does not significantly exceed the static threshold, the prolonged abnormality reflects a high compliance risk. When an institution's individual contribution coefficient sequence satisfies any of the above dynamic abnormality patterns, it is determined to have dynamic abnormality risk, triggering a dynamic abnormality warning.

[0143] The method described in this embodiment does not treat static and dynamic anomaly detection as independent entities, but rather as two core dimensions for multi-dimensional anomaly detection, enabling comprehensive judgment and early warning triggering. Static anomaly warnings are short-term risk signals, indicating abnormal behavior within a single cycle; dynamic anomaly warnings are long-term risk signals, indicating systemic and persistent anomalies in the institution's behavioral patterns. When an institution triggers only a static anomaly warning, regulatory authorities can classify it as a general monitoring target for short-term monitoring; when an institution triggers only a dynamic anomaly warning, regulatory authorities can classify it as a key monitoring target for in-depth behavioral analysis; when an institution triggers both static and dynamic anomaly warnings simultaneously, regulatory authorities can classify it as a high-risk entity, immediately initiating special investigations and manual audits to achieve precise allocation of regulatory resources.

[0144] Static anomaly detection can quickly identify explicit anomalies that deviate significantly from the overall regional situation within a single period, such as a sudden surge in fraudulent hospitalizations or over-treatment at a medical institution within a certain period. Dynamic anomaly detection can accurately capture latent anomalies in the evolution of behavioral patterns over continuous periods, such as progressive over-treatment or long-term splitting of charges by an institution. These behaviors may not reach the static threshold within a single period, but their long-term trends are clearly abnormal. The combination of these two approaches achieves comprehensive coverage of various abnormal behaviors, ensuring that medical insurance supervision does not overlook either explicit short-term anomalies or hidden long-term anomalies.

[0145] By analyzing the trends and stability of individual contribution coefficient sequences through dynamic anomaly detection, abnormal changes in an institution's behavioral patterns can be identified before the institution's abnormal behavior develops into serious violations or causes significant losses to the medical insurance fund, thus enabling early warning of abnormal risks. For example, if a medical institution's individual contribution coefficient begins to show a continuous downward trend and the degree of deviation gradually increases, even if the static anomaly threshold has not been reached, dynamic anomaly detection can issue an early warning. Regulatory authorities can then intervene in a timely manner to monitor and remind relevant departments, preventing further deterioration of abnormal behavior, achieving proactive prevention and control of medical insurance fund risks, and reducing fund losses.

[0146] Static anomaly detection is based on robust risk scores, while dynamic anomaly detection is based on the time evolution characteristics of individual contribution coefficient sequences. Both are objective judgments based on quantitative data, and all judgment indicators can be traced and verified from the original medical insurance data, avoiding the subjective arbitrariness of traditional manual judgments. When an anomaly warning is triggered, regulators can accurately locate the abnormal behavior characteristics of the institution by tracing the calculation process of individual contribution coefficients, changes in robust risk scores, and the evolution trend of individual contribution coefficient sequences. This provides clear clues for manual review and evidence collection analysis, improving the operability and efficiency of medical insurance supervision.

[0147] In some embodiments, in step S3, after solving for the individual contribution coefficient... Then, the following steps are also included:

[0148] Obtain the static attribute vector Z of the target medical institution i. i The static attribute vector includes one or more of the following: medical institution level, approved number of beds, and number of key departments;

[0149] Constructing a mapping model between medical institution attributes and contribution coefficients, specifically including: based on the static attribute vector set {Z1, Z2, ..., Z...} of all N medical institutions within the regulatory area. N} and its corresponding set of individual contribution coefficients {a1, a2, ..., a N Train a regression model f(·) to obtain the structural contribution determined by the inherent attributes of the medical institution. The calculation formula is as follows: ;

[0150] The individual contribution coefficient a i Decomposed into structural contribution coefficients and behavioral contribution coefficient ,in:

[0151] ;

[0152] In step S5, a multidimensional anomaly assessment is performed on the target medical institution, and corresponding risk warnings are triggered, including:

[0153] Combination and Perform multidimensional anomaly detection:

[0154] When the first condition is met, it is determined to be a risk of distorted medical practice. The first condition includes: the robustness risk score exceeds a preset risk threshold, and simultaneously meets the following conditions: >τ b and >λ· , where τ b λ is the threshold for the behavioral contribution coefficient, and λ is the contribution ratio coefficient.

[0155] When the second condition is met, it is determined to be a risk of mismatch between scale and positioning. The second condition includes: the robustness risk score exceeds a preset risk threshold, and simultaneously meets the following conditions: >τ s and >λ· And based on the static attribute vector Z i The calculated change in the medical institution size positioning index over M consecutive accounting periods is within the preset error range, where τ s The threshold for the structural contribution coefficient.

[0156] In this embodiment, static attribute vector It refers to a multi-dimensional vector that characterizes the inherent attributes of the medical institution to be verified. Its core includes indicators that are highly correlated with the scale, positioning and service capacity of the medical institution, such as the level of the medical institution (e.g., Grade III Class A, Grade II Class B), the number of approved beds, and the number of key departments. This vector reflects the inherent characteristics of the medical institution and will not change significantly in the short term.

[0157] The mapping model refers to the regression model f(·) trained based on the set of static attribute vectors of all medical institutions in the region and the corresponding set of individual contribution coefficients. Its core function is to establish the statistical mapping relationship between the inherent attributes of medical institutions and individual contribution coefficients, and to realize the quantitative derivation from static attributes to structural contribution coefficients.

[0158] Structural contribution coefficient This refers to the static attribute vector of medical institutions. The individual contribution coefficient obtained by the mapping model f(·) is the theoretical contribution level determined by the inherent attributes of the medical institution. It reflects the contribution level of medical insurance reimbursement behavior that should match the regional medical structure, based on the institution's level, scale, number of key departments, and other characteristics.

[0159] Behavioral contribution coefficient The individual contribution coefficient 'a' is obtained from the actual solution. i Subtract structural contribution coefficient The portion obtained is the actual contribution deviation determined by the actual operation and management behavior of medical institutions. It reflects the difference between the institution's actual medical insurance reimbursement behavior and its theoretical behavior based on its inherent attributes, and is the core indicator for measuring the rationality of the institution's behavior.

[0160] The risk of medical practice distortion refers to the risk of abnormal medical insurance reimbursement behavior caused by abnormal actual operation and management behavior of medical institutions. It is a behavioral abnormality that is unrelated to the inherent attributes of medical institutions and is caused by the institution's proactive behavior (such as over-treatment, fictitious diagnosis and treatment, and splitting charges).

[0161] The risk of mismatch between scale and positioning refers to the abnormal risk caused by the mismatch between the actual medical insurance reimbursement behavior of medical institutions and their inherent attributes (scale, positioning, service capacity). It is a structural abnormality, reflecting that the actual operation of the institution exceeds the reasonable range corresponding to its inherent attributes. For example, small medical institutions carry out diagnosis and treatment projects beyond their service capacity, resulting in excessive consumption of medical insurance funds.

[0162] Behavioral contribution coefficient threshold τ b It refers to the preset structural contribution coefficient threshold, which is one of the core bases for judging the risk of alienation of diagnosis and treatment behavior. It is used to screen out medical institutions with certain inherent service capabilities and avoid over-judging small and low-level institutions.

[0163] The contribution ratio coefficient λ is a preset proportional constant used to measure the relative size relationship between the behavioral contribution coefficient and the structural contribution coefficient. It is the core basis for distinguishing the risk of alienation of diagnosis and treatment behavior and the risk of mismatch between scale and positioning.

[0164] Structural contribution coefficient threshold τ s It refers to the preset structural contribution coefficient threshold, which is one of the core bases for judging the risk of mismatch between scale and positioning, and is used to screen out medical institutions with a certain inherent scale and positioning.

[0165] The scale positioning index refers to the static attribute vector of medical institutions. The calculated quantitative indicators reflect the inherent scale and service positioning of medical institutions, and their values ​​are highly positively correlated with the level of medical institutions, the number of approved beds, and the number of key departments.

[0166] In the specific implementation, the static attribute vector of the target medical institution i is first obtained. The selection of medical institution level, approved number of beds, and number of key departments as core indicators is because these indicators are the core manifestation of the inherent attributes of medical institutions, directly determining their service capacity, scale positioning, and role in regional medical services. They also show a significant statistical correlation with the contribution level of medical institutions to medical insurance reimbursement.

[0167] For example, tertiary-level Class A medical institutions have high standards, a large number of approved beds, and numerous key departments, resulting in strong service capabilities. Consequently, their contribution to regional total expenditure through medical insurance reimbursement should be at a reasonably high level. Conversely, smaller medical institutions such as community health service centers should contribute at a relatively low and reasonable level. (Static attribute vector) The acquisition of this data ensures that the foundational data for subsequent mapping model training is highly compatible with the actual service capabilities and positioning of medical institutions.

[0168] Then, based on the static attribute vector set {Z1, Z2, ..., Z...} of all N medical institutions within the area to be regulated, N} and its corresponding set of individual contribution coefficients {a1, a2, ..., a N Train a regression model f(·), whose input is a vector of static attributes. The output is the structural contribution coefficient. The core logic of the training is to uncover the statistical regularity between the inherent attributes of medical institutions and individual contribution coefficients. The specific operation is as follows: through learning from a large number of samples, a quantitative mapping relationship is established from "inherent attributes" to "theoretical contribution levels," enabling the calculated... This accurately reflects the institution's reasonable contribution level within the regional healthcare structure based on its inherent attributes. After training, the Z-value of the target healthcare institution will be... i Substituting into the model f(·), we can obtain its structural contribution coefficient. .

[0169] The individual contribution coefficient a obtained from the actual solution i Decomposed into structural contribution coefficients and behavioral contribution coefficient The core logic of this decomposition is to break down the individual contribution coefficient into a "theoretical part determined by inherent attributes" and a "biased part determined by actual behavior," where... It is the "theoretical contribution value" of medical institutions, representing their reasonable contribution level based on their scale and positioning; This is the "actual deviation value" of medical institutions, representing the difference between their actual medical insurance reimbursement behavior and the theoretically reasonable level. A significant deviation from 0 indicates a discrepancy between the actual behavior of the medical institution and its reasonable behavior based on its inherent attributes, suggesting an abnormal risk. This decomposition allows for a refined breakdown of individual contribution coefficients, enabling the root cause of abnormal risks to be traced back to the structural or behavioral level.

[0170] In the multidimensional anomaly determination in step S5, the method is no longer based solely on robust risk scores and individual contribution coefficient sequence characteristics, but rather combines... and A refined assessment of abnormal risk types is conducted, with a core distinction between the risk of distorted medical practices (behavioral abnormalities) and the risk of mismatch between scale and positioning (structural abnormalities). The assessment logic for these two types of risks has different emphases, as detailed below:

[0171] When a medical institution's robust risk score exceeds a preset risk threshold (indicating an overall anomaly), and simultaneously meets the following conditions... >τ b and >λ· At that time, it was determined that there was a risk of distortion of medical treatment behavior. >τ b This indicates that the medical institution has certain inherent service capabilities, its theoretical contribution level has reached a certain standard, and it has the basis for the occurrence of behavioral abnormalities. >λ· This indicates that the actual behavioral deviation of the medical institution is far greater than its theoretical contribution level. Its abnormal risks are not caused by inherent attributes, but by abnormal actual operation and management behaviors, such as over-treatment, fictitious diagnosis and treatment, and splitting charges. These behaviors have led to a huge deviation between its actual medical insurance reimbursement behavior and the theoretically reasonable level.

[0172] When a medical institution's robust risk score exceeds a preset risk threshold (indicating an overall abnormality), and simultaneously meets the following conditions... >τ s and >λ· If the change in the scale positioning index is within the preset error range over M consecutive accounting periods, it is determined that there is a risk of scale and positioning mismatch. >τ s This indicates that the medical institution has a certain inherent scale and positioning; >λ· This indicates that the theoretical contribution level of the medical institution is far greater than the actual behavioral deviation, and its abnormal risk is not caused by behavioral abnormalities. The continuous stability of the scale positioning index indicates that the inherent attributes of the institution have not changed significantly in the short term, excluding changes in contribution level caused by changes in inherent attributes such as institutional upgrades and expansions. At this time, the abnormal risk stems from the fact that the actual operation of the institution exceeds the reasonable range corresponding to its inherent attributes. For example, a small medical institution carries out large-scale diagnosis and treatment projects beyond its service capacity, resulting in a serious mismatch between medical insurance reimbursement behavior and its own scale and positioning.

[0173] This embodiment achieves an upgrade from qualitative identification to quantitative diagnosis in medical insurance anomaly identification through "structure-behavior" coefficient decomposition and differentiated risk assessment. Compared with traditional single anomaly identification methods, it has the following significant advantages:

[0174] Traditional methods for identifying medical insurance anomalies can only determine whether an institution has an anomaly, but cannot pinpoint whether the root cause lies in the institution's inherent attributes or actual behavior, leading to a lack of targeted regulatory measures. This method, by decomposing individual contribution coefficients into structural and behavioral components, precisely distinguishes between the risk of distorted treatment behavior (behavioral anomalies) and the risk of mismatch between scale and positioning (structural anomalies), enabling regulatory authorities to identify the core reasons for institutional anomalies. For behavioral anomalies, regulatory authorities can take measures such as special investigations, penalties for violations, and behavioral rectification; for structural anomalies, regulatory authorities can take measures such as service capacity assessments, positioning adjustments, and resource allocation optimization, achieving differentiated and precise regulation and improving the effectiveness and relevance of regulatory measures.

[0175] Different medical institutions naturally differ in their level, size, and number of key departments, and their contribution levels to medical insurance reimbursement should also exhibit reasonable structural differences. Traditional methods fail to consider these inherent differences, easily misjudging the reasonablely low contribution levels of low-level, small medical institutions as abnormal, or the reasonablely high contribution levels of high-level medical institutions as abnormal, leading to regulatory injustice. This method uses a structural contribution coefficient... Quantify the reasonable theoretical contribution level of medical institutions, and then use behavioral contribution coefficients. By measuring deviations in actual behavior, the system effectively shields the structural differences inherent in medical institutions, identifying only significant deviations between actual behavior and theoretical levels as abnormalities. This avoids misjudging structural differences as behavioral abnormalities, making the determination of medical insurance abnormalities more fair and accurate.

[0176] Furthermore, this method, through statistical analysis of the static attribute vectors, structural contribution coefficients, and behavioral contribution coefficients of all medical institutions within a region, can accurately characterize the features of the regional medical structure. For example, it can identify which medical institutions within the region exhibit a mismatch between scale and positioning, reflecting irrational allocation of regional medical resources; it can also statistically analyze the distribution of institutions with a risk of distorted medical behavior within the region, reflecting key areas for regional medical insurance fund supervision. These data analysis results not only provide a basis for the abnormal supervision of medical insurance funds but also offer scientific quantitative data support for optimizing the regional medical structure, rationally allocating medical resources, and improving the service capacity of medical institutions, thus achieving synergistic advancement of medical insurance supervision and regional medical development.

[0177] In some embodiments, constructing the background mask in step S2 includes:

[0178] Acquire patient transfer data among medical institutions within the regulatory region over a consecutive T accounting periods;

[0179] For any medical institution i to be verified and the accounting period t, a directed weighted network is constructed based on the flow data; wherein, the nodes of the directed weighted network are medical institutions, and if a patient flows from medical institution u to medical institution v within the period t, there exists an edge from u to v, with the weight being the number of patients transferred.

[0180] Starting with medical institution i, in the directed weighted network, all medical institutions that have direct or indirect patient transfer relationships with medical institution i are selected to form the dynamic background institution set B of medical institution i in period t. i,t Among them, the path length between selected medical institutions is less than the preset length;

[0181] Calculate the total amount of reimbursement under dynamic background The calculation formula is as follows:

[0182] ;

[0183] Where j represents the dynamic background mechanism set B i,t Index of Chinese medical institutions This represents the independent reimbursement amount for medical institution j in the dynamic background institution set during the t-th accounting period;

[0184] In the restricted regression equation, the total amount of reimbursement under the dynamic background is used. Replace the original background reimbursement total amount x i,t Perform the calculation.

[0185] In this embodiment, patient cross-institutional transfer data refers to the total amount of data generated by insured patients traveling from one medical institution to another for diagnosis, follow-up examination, referral, etc. within the region to be regulated over T consecutive accounting cycles. It includes core fields such as the initiating institution, receiving institution, number of transferred patients, and transfer time, and is the core data source for characterizing the business relationships between medical institutions.

[0186] A directed weighted network is a network model constructed with medical institutions within a region to be regulated as nodes, the patient flow relationship between institutions as directed edges, and the number of patients flowing as the edge weights. The directionality reflects the direction of patient flow (such as from medical institution A to medical institution B), and the weighting reflects the closeness of the flow relationship between institutions. It is the core carrier for quantifying the business connections between medical institutions.

[0187] Dynamic background organization set B i,t It refers to the set of all medical institutions that have a direct or indirect patient flow relationship with medical institution i and whose path length is less than the preset length, selected from the directed weighted network starting from the medical institution i to be verified. This set is a personalized background that changes dynamically with the institution and the period, and is more in line with the actual business relationship scenarios of medical institutions.

[0188] Path length refers to the shortest edge between two medical institution nodes in a directed weighted network. It is used to measure the degree of indirect connection between institutions in patient transfer. For example, the path length of direct transfer is 1, and the path length of transfer through an institution is 2. The core function of preset path length is to limit the scope of association in dynamic background and avoid including institutions with no actual business relationship.

[0189] Dynamic background reimbursement total amount Refers to the dynamic background mechanism set B i,t The sum of independent reimbursement amounts for all medical institutions in the t-th accounting period reflects the medical insurance reimbursement scale of the actual business-related background of the medical institution i to be verified.

[0190] This embodiment constructs a dynamic, personalized background based on the business-related characteristics of patient cross-institutional transfers, making the background mask more closely match the actual operational ecosystem of medical institutions and achieving precise matching between individuals and backgrounds. The specific steps and principles are as follows:

[0191] First, we acquire patient inter-institutional transfer data for the monitored region over T consecutive accounting periods. This data forms the basis for constructing the dynamic context because patient transfers between medical institutions directly reflect their business interrelationships. Medical institutions with patient transfers have a high degree of correlation in terms of diagnosis and treatment services, patient resources, and medical needs, and their medical insurance reimbursement behaviors also influence each other. Conversely, medical institutions without patient transfers typically have no actual business relationship, and including them in the context would distort the match between individual behaviors and the context. Therefore, acquiring transfer data lays the data foundation for subsequently constructing a realistic dynamic context.

[0192] Then, for any medical institution i to be verified and an accounting period t, a directed weighted network is constructed based on the flow data. The nodes of the network are all medical institutions within the region. If a patient flows from medical institution u to medical institution v within period t, a directed edge is constructed from u to v, and the weight of the edge is the number of patients flowing from u to v within that period. The core logic of this construction method is to quantify the business relationships between institutions using a network model, specifically as follows: the directed edges accurately depict the direction of patient flow, reflecting the business relationships such as referrals and referrals between institutions; the weight reflects the tightness of the relationship. The more patients flow, the tighter the business relationship between the two institutions, and the greater the mutual influence of their medical insurance reimbursement behavior. Through this network, the implicit business relationships between medical institutions are transformed into quantifiable and analyzable explicit network relationships.

[0193] Then, starting with the medical institution i to be verified, all institutions with path lengths less than a preset length are selected in the directed weighted network to form the dynamic background institution set B. i,t The preset path length is set by considering factors such as regional medical structure characteristics and referral rules (e.g., a preset length of 2 typically includes institutions with direct referrals and single-transfer referrals). Its core logic is to limit the scope of background associations and exclude institutions with no actual business impact. This is because a path length that is too short will result in an overly narrow background that fails to reflect the actual association environment of the institution; a path length that is too long will lead to background generalization, reverting to a static, global background. By filtering through path length, B... i,t It only includes institutions that have direct or close indirect business relationships with medical institution i, thus achieving personalized and precise background information.

[0194] Then, the dynamic background reimbursement total is calculated. This indicator reflects the overall medical insurance reimbursement scale of background institutions that have actual business relationships with medical institution i, and the calculated dynamic background reimbursement total is used as the basis for calculation. Replace the original background reimbursement total amount x i,tA restricted regression equation is introduced. The core logic of the replacement is to make the "individual-background" matching of the regression model more closely reflect actual business scenarios. This is because a static global background includes institutions with no business connection, which amplifies background noise and makes it impossible to accurately extract the true behavioral characteristics of medical institution i; while a dynamic background only includes business-related institutions, whose medical insurance reimbursement behavior is influenced by the same regional medical needs, patient groups, and other factors as medical institution i, making it more reliable. Based on this, the restricted regression equation can more accurately characterize the individual contribution coefficient 'a' of medical institution i. i This improves the accuracy of parameter estimation.

[0195] This embodiment solves the problem of generalization matching of static global backgrounds by constructing a dynamic background mask based on the patient flow network. This makes the association between individuals and backgrounds more consistent with the actual operational ecosystem of medical institutions, further improving the accuracy and anti-interference ability of the model. The specific beneficial effects are as follows:

[0196] Static global backgrounds include all institutions except those under investigation, inevitably encompassing a large number of institutions with no patient transfers or actual business connections. The medical insurance reimbursement behavior of these institutions has no significant correlation with the institution under investigation, and their inclusion in the background generates substantial invalid noise, leading to distorted estimations of individual contribution coefficients. Dynamic backgrounds, on the other hand, only include institutions with direct or closely related indirect patient transfers to the institution under investigation. The background and individual have a strong business connection, and the influencing factors on medical insurance reimbursement behavior are highly similar. This effectively avoids assessment biases caused by unrelated institutions, making the "individual-background" deconstruction more meaningful for practical business applications and improving the rationality and accuracy of model parameter estimation.

[0197] Based on the dynamic context of patient transfer networks, this approach not only enables the identification of anomalies in individual institutions but also indirectly characterizes the business relationships between institutions, thereby identifying correlational anomalies that are difficult to detect using traditional methods. For example, a medical institution attracts patients through abnormal referral methods, leading to simultaneous anomalies in its and related institutions' medical insurance reimbursement behavior; or multiple related institutions in a region form a "community of interest," colluding to misuse medical insurance funds. Such correlational anomalies are masked by global noise in a static context, but in a dynamic context, due to the close relationship between the background and individuals, these anomalies are directly reflected in changes in individual contribution coefficients, achieving a dimensional expansion from "identifying anomalies in single institutions" to "identifying anomalies in related institutions."

[0198] In some embodiments, performing multidimensional anomaly assessment on the target medical institution and triggering corresponding risk warnings includes the following steps:

[0199] Calculate the behavioral correlation w between any two medical institutions p and q based on the time series of historical individual contribution coefficients or the business similarity between medical institutions. pqThis forms a matrix of relationships among medical institutions;

[0200] For the current accounting period t, calculate the change in the individual contribution coefficient of each medical institution pair (p, q). and And calculate the correlation change score of medical institutions. The calculation formula is as follows: ;in, This represents the individual contribution coefficient of medical institution p within the current accounting period t. This represents the individual contribution coefficient of medical institution p in the previous accounting period t-1. This represents the individual contribution coefficient of medical institution q within the current accounting period t. This represents the individual contribution coefficient of medical institution q in the previous accounting period t-1;

[0201] If a medical institution pair (p, q) that meets the third condition is identified, an association anomaly alert is triggered. The third condition includes the behavioral association degree w between medical institutions p and q. pq The score for the association change of (p,q) is higher than the preset association threshold. The absolute value is higher than the preset correlation change threshold, and its sign is negative.

[0202] In this embodiment, the behavioral correlation degree w pq It refers to an indicator used to quantify the behavioral similarity between any two medical institutions p and q within a region to be regulated. It is calculated based on the time series of historical individual contribution coefficients or the business similarity between medical institutions. The larger the value, the closer the correlation between the medical insurance reimbursement behaviors of the two institutions and the more similar the trend of change. It is a core quantitative indicator for characterizing the behavioral correlation between institutions.

[0203] Let N medical institutions within the area to be regulated be the rows and columns, and let w be the behavioral correlation between any two institutions. pq The N-order square matrix constructed for the matrix elements is a holistic quantification of the behavioral correlations among all medical institutions in the region, which can intuitively and systematically reflect the behavioral correlation network characteristics of medical institutions.

[0204] Change in individual contribution coefficient / These refer to the difference between the individual contribution coefficients of a medical institution p / q in the t-th accounting period and the individual contribution coefficients in the (t-1)-th accounting period. They are used to characterize the changing trend and magnitude of a single institution's medical insurance reimbursement behavior in adjacent accounting periods. A positive difference indicates an increase in the degree of contribution, while a negative difference indicates a decrease in the degree of contribution.

[0205] Related Change Score It refers to an indicator calculated based on the behavioral correlation and individual contribution coefficient changes of two institutions. It is used to quantify the degree and direction of change of the behavioral correlation pattern of institutions to (p,q) in the t-th accounting period and is a core indicator for identifying inter-institutional collaboration anomalies.

[0206] The association threshold refers to a preset threshold value for the degree of behavioral association based on the behavioral association characteristics of regional medical institutions. It is used to screen out pairs of institutions with close behavioral associations. Only pairs of institutions with a degree of behavioral association exceeding the threshold are included in the scope of collaborative anomaly identification.

[0207] The correlation change threshold refers to a preset absolute value threshold for the correlation change score, which is used to measure whether the change in the behavioral correlation pattern of an institution reaches a significantly abnormal level. Exceeding this threshold indicates that the behavioral change pattern of the institution deviates significantly from the normal correlation characteristics.

[0208] The associated anomaly alert is a special warning triggered for institutions with coordinated anomalies. Unlike single-institution anomaly alerts, this alert focuses on the coordination and correlation anomalies between institutions, providing clues for regulatory authorities to investigate collective violations such as collusion to cheat and abnormal competition.

[0209] This embodiment constructs a collaborative anomaly identification mechanism based on the behavioral correlations between medical institutions, enabling accurate identification of correlational anomalies such as collusion and cheating among institutions and abnormal competition. The specific steps and principles are as follows:

[0210] First, the behavioral correlation w between any two institutions p and q is calculated based on the time series of historical individual contribution coefficients or the business similarity between medical institutions. pq Furthermore, a correlation matrix of medical institutions was constructed. The core logic of calculating the historical individual contribution coefficient time series is the quantification of behavioral similarity. If the trends of individual contribution coefficient changes of two medical institutions within a historical period are highly consistent (e.g., rising or falling simultaneously, with similar fluctuation amplitudes), it indicates that their medical insurance reimbursement behaviors are influenced by the same factors, and the behavioral correlation w is high. pq The higher the value, the higher the business similarity calculation. The core logic of business similarity calculation is the quantification of business relevance—if two institutions have highly similar medical specialties, service populations, regional locations, etc., or have close referral or cooperative relationships, then their business relevance is high, and their behavioral relevance w is high. pq The larger the value, the better. The construction of the correlation matrix systematizes and matrixes the behavioral correlation between pairs of institutions, providing overall data support for subsequent collaborative anomaly identification across the entire domain.

[0211] For the current accounting period t, calculate the change in the individual contribution coefficient of all medical institutions to (p,q), and calculate the correlation change score. When the behavioral correlation between two medical institutions is w... pqWhen the value is high, under normal circumstances, the individual contribution coefficients of the two should maintain the same trend, that is... and If both are positive or both are negative, their product is positive, and the correlation change score is calculated. It is also positive; if the trends of the two changes are opposite, that is... and One positive and one negative, the product is negative, related change score A negative value indicates an abnormal deviation in the behavioral correlation pattern between the two. Meanwhile, w pq As a weight, the product of the changes in the behavior of the pair of organizations with higher correlation has a greater impact on the correlation change score, thus realizing the judgment logic that "the higher the degree of correlation, the stricter the requirement for coordination".

[0212] When a medical institution determines that there is a coordination anomaly and triggers an association anomaly alert when (p,q) meets three core conditions: first, the behavioral association degree w pq The criteria for inclusion in the assessment are: 1) exceeding the correlation threshold, ensuring that only institutional pairs with closely related behaviors are included, and excluding institutional pairs with no actual correlation; and 2) correlation change score. The absolute value is higher than the correlation change threshold to ensure that the changes in the behavioral correlation pattern of the institution reach a significant abnormal level, excluding small deviations caused by random fluctuations; thirdly A negative sign indicates that two closely related institutions have exhibited opposite behavioral trends, which is the core characteristic of collaborative anomalies. The business essence of this judgment logic is: institutions with closely related behaviors should have their medical insurance reimbursement behaviors influenced by the same regional medical needs, policy environment, and other factors, showing synchronized trends; if opposite trends appear, and the magnitude of the changes is significant, then abnormal behavior is highly likely. For example, two institutions engaging in abnormal competition for patient resources, or two institutions conspiring to manipulate medical insurance funds (such as one institution inflating reimbursements while the other deliberately reduces reimbursements to evade supervision), such collaborative anomalies are difficult to detect under a single-institution judgment model, but can be accurately identified through this method.

[0213] This embodiment overcomes the limitations of traditional single-institution anomaly identification by constructing a collaborative anomaly detection mechanism based on behavioral correlation analysis. It achieves accurate identification of correlation and collaborative anomalies among medical institutions, further improving the multi-dimensional anomaly judgment system. The specific beneficial effects are as follows:

[0214] Traditional methods for identifying medical insurance anomalies typically focus on individual institutions, examining deviations in their behavior but failing to identify coordinated anomalies such as collusion, fraud, or abnormal competition among institutions. For example, multiple related institutions within a region might collude to defraud medical insurance funds by splitting treatment items and cross-inflating reimbursements. While the reimbursement amounts and behavioral characteristics of a single institution might not reach anomaly thresholds, the coordinated behavioral changes among the institutions demonstrate clear anomalies. Similarly, two closely related institutions might employ abnormal charging and treatment strategies to compete for patient resources, resulting in opposing behavioral trends. These types of coordinated anomalies are a key challenge in medical insurance fund supervision. This method, through behavioral correlation analysis, achieves accurate identification of such anomalies, filling a gap in traditional supervisory methods.

[0215] The method described in this embodiment focuses on the relative changes among medical institutions, examining whether the behavioral change patterns of related institutions are abnormal. Even if the individual contribution coefficient of a single institution changes only slightly, as long as there are opposite trends among related institutions that reach a significant level, it can be considered abnormal. This method significantly improves the sensitivity of identifying hidden collaborative anomalies, allowing medical insurance supervision to detect potential collective violations earlier.

[0216] By calculating behavioral correlation and constructing a correlation matrix, medical institutions within the regulated area are structured as a behavioral correlation network. Regulatory authorities can then analyze the behavioral characteristics of these institutions from a network perspective. This includes identifying core nodes within the network, discovering abnormal behavioral subnetworks, and tracing the propagation paths of collaborative anomalies. Compared to traditional point-to-point, single-institution regulation, network-based regulation provides a more systematic and comprehensive understanding of the risk characteristics of the regional medical insurance fund. This facilitates the development of comprehensive regulatory strategies by regulatory authorities, enabling an upgrade from "partial regulation" to "comprehensive regulation."

[0217] For alerts triggered by abnormal collaboration, the system directly identifies pairs of institutions exhibiting linked abnormal behavior. Regulatory authorities can then conduct targeted audits based on these alerts, eliminating the need for a comprehensive review of all institutions within the region. For instance, when an alert is triggered indicating an abnormal association between medical institutions p and q, regulatory authorities can focus on examining their medical records, patient transfer records, and reimbursement data to investigate potential collusion, fraud, or unfair competition. This targeted auditing approach significantly reduces ineffective work by regulatory authorities, improves the efficiency and accuracy of manual audits, and allows limited regulatory resources to be focused on high-risk pairs of related institutions.

[0218] In some embodiments, in step S4, constructing a statistical baseline based on the set of individual contribution coefficients and calculating the robust risk score for each medical institution includes:

[0219] Based on preset association rules, the direct association strength ρ between any two medical institutions p and q within the regulatory area is calculated. pq , where 0≤ρ pq ≤1, ρ pp = 1, and construct an N-row, N-column correlation strength matrix R;

[0220] For any medical institution i, calculate its individual abnormality intensity. The calculation formula is as follows:

[0221] ;

[0222] in, and These are sets of individual contribution coefficients based on historical cycles, {a i The calculated benchmark mean and benchmark standard deviation;

[0223] Calculate the native robust risk score of healthcare institution i The native robust risk score By analyzing the individual abnormal intensity of medical institution i The result is obtained after smoothing and standardization.

[0224] Based on the aforementioned correlation strength matrix R, the network infection risk bonus of medical institution i is calculated. The calculation formula is:

[0225] ;

[0226] in, and The preset attenuation coefficient, >0, <1; This represents the additional attenuation factor for secondary transmission, 0 < <1, Let be the shortest path length from medical institution j to medical institution k;

[0227] Calculate the comprehensive robust risk score R of medical institution i i The calculation formula is as follows:

[0228] .

[0229] In this embodiment, the direct correlation strength ρ pq This refers to an indicator used to quantify the degree of direct business relationship between any two medical institutions p and q within a regulated area, with a value range of 0 ≤ ρ. pq ≤1, ρ pp= 1 (the correlation strength between medical institutions is 1). The larger the value, the closer the direct business relationship between the two medical institutions. It is the core element for constructing the correlation strength matrix.

[0230] The correlation strength matrix R refers to a matrix with N medical institutions in the region to be regulated as rows and columns, and the direct correlation strength ρ between any two institutions. pq The N-order square matrix constructed for the matrix elements is the core carrier for characterizing the direct business relationship network between medical institutions in the region, and provides a foundation for subsequent network infection risk calculation.

[0231] Individual abnormal intensity This refers to the individual contribution coefficient a based on medical institution i. i Compared with historical benchmark average Benchmark standard deviation The calculated indicators are used to quantify the inherent anomaly degree of a single institution in the t-th accounting period, which facilitates the subsequent calculation of the native robust risk score.

[0232] Native robust risk score This refers to the individual abnormality intensity of medical institution i. The indicators obtained after smoothing and standardization are used to quantify the basic anomaly risk of a single institution.

[0233] Increased risk of online transmission It refers to an indicator calculated based on the correlation strength matrix and the native robust risk scores of other institutions, used to quantify the degree of risk of abnormal risks from other institutions in the region spreading to medical institution i through the business correlation network.

[0234] Attenuation coefficient and These are preset constant coefficients, with values ​​ranging from 0 < α to β < 1. β is the primary transmission attenuation coefficient, used to control the risk transmission intensity of directly related institutions; β is the secondary transmission attenuation coefficient, used to control the risk transmission intensity of indirectly related institutions. The core function of the attenuation coefficient is to gradually reduce the risk transmission intensity as the level of association increases, which conforms to the actual risk transmission law.

[0235] Secondary transmission additional attenuation factor This refers to a preset constant coefficient, with a value range of 0 < γ < 1, used to further control the risk contagion intensity between indirectly related institutions. The contagion intensity increases with the length of the shortest path between institutions. The exponential decline in risk contagion due to the increase in risk makes the quantification of risk contagion more realistic.

[0236] Overall robust risk score R i This refers to the native robust risk score of medical institution i. Increased risk of online transmission The resulting index is a comprehensive risk quantification indicator that takes into account both the institution's own basic risks and the risks of network contagion, and more comprehensively reflects the institution's actual abnormal medical insurance risks.

[0237] This embodiment is a deep optimization of step S4, "Constructing a statistical baseline and calculating a robust risk score." It introduces the risk contagion theory, upgrading the independent risk score of a single institution to a comprehensive risk score of "individual risk + network contagion risk," realizing the transformation from "independent risk assessment" to "networked risk assessment," which is more in line with the actual propagation characteristics of medical insurance fund risks. The specific steps and principles are as follows:

[0238] First, based on preset association rules (such as patient inter-institutional transfer data, business cooperation data, regional location data, etc.), the direct association strength ρ between any two institutions p and q is calculated. pq And construct an N-order association strength matrix R. The core of the preset association rules is to fit the actual medical business. For example, the proportion of patients transferred can be used as the association rule. The higher the proportion of transfers, the stronger the direct association ρ. pq The larger the ρ, the higher the ρ; the more collaborative projects, the higher the ρ. pq The larger. ρ pq The value range of ρ is limited to 0-1, which standardizes the correlation strength and facilitates cross-institutional comparisons; pp =1 is used to ensure the integrity of the matrix. The construction of the association strength matrix R systematizes the direct business relationships between medical institutions within the region, providing a network foundation for subsequent quantitative calculations of risk contagion.

[0239] For any medical institution i, the individual abnormality intensity is calculated. The degree to which the individual contribution coefficient of medical institution i deviates from its historical normal level is considered; a larger deviation indicates a higher degree of inherent individual abnormality. Subsequently, the individual abnormality intensity is smoothed and standardized to obtain the native robust risk score. The purpose of smoothing is to eliminate abrupt changes in abnormal intensity caused by random fluctuations, allowing the score to better reflect the long-term stable abnormal characteristics of the institution; the purpose of standardization is to convert individual abnormal intensity into dimensionless scoring indicators, facilitating comparisons across institutions and time periods, ultimately resulting in... It accurately quantifies the underlying abnormal risks of medical institutions themselves.

[0240] Then, the network infection risk bonus of medical institution i is calculated based on the correlation strength matrix R. By dividing risk contagion into primary and secondary contagion, the risk contagion effect of directly and indirectly related institutions is comprehensively quantified, specifically including:

[0241] Level 1 Infection Section: This describes the risk of transmission from a directly related medical institution to institution i, meaning that medical institution j is a directly related institution to medical institution i. The strength of the direct correlation between the two. Let be the native robust risk score of medical institution j, and the product of the two scores represent the amount of abnormal risk of medical institution j that is directly transmitted to i. is the first-order transmission attenuation coefficient, used to control the intensity of direct transmission. Summing it up yields the total first-order transmission risk of i from all directly related institutions.

[0242] Secondary transmission: This describes the risk of transmission from an indirectly related medical institution to medical institution i, that is, medical institution k is an indirectly related institution of medical institution i (transferring through medical institution j). This indicates the strength of the indirect association between medical institution i and medical institution k. The native robust risk score for medical institution K. This is an additional attenuation factor for secondary transmission, varying with the shortest path length from medical institution j to k. The increase shows an exponential decline, reflecting the principle that "the further the indirect link, the weaker the risk contagion." This is the secondary transmission attenuation coefficient, used to control the overall intensity of indirect transmission. The total secondary transmission risk from all indirectly related institutions to healthcare institution i is obtained after double summation. The sum of the two yields... It accurately quantifies the total risk level of abnormal risks from other institutions within the region spreading to i through the business-related network, and realizes a systematic and quantitative calculation of the risk contagion effect.

[0243] The comprehensive robust risk score for medical institution i is then calculated, specifically by integrating its own fundamental risks with network contagion risks to comprehensively reflect the actual medical insurance anomaly risks of medical institution i. In actual medical insurance fund supervision scenarios, the anomaly risks of medical institutions do not exist in isolation but spread between institutions through business networks. For example, if a core referral institution engages in medical insurance violations, its anomaly risk will be transmitted to downstream receiving institutions through patient transfers, causing anomalies in the medical insurance reimbursement behavior of downstream institutions as well. The comprehensive robust risk score incorporates this risk contagion effect into its assessment scope, making the risk score more closely reflect actual risk characteristics and providing a more comprehensive and accurate basis for subsequent anomaly identification.

[0244] This embodiment, by constructing a networked risk assessment system that considers the contagion effect of risks, breaks through the limitations of traditional independent risk assessment, and achieves a comprehensive and dynamic assessment of abnormal risks in medical institutions. It is more in line with the actual propagation patterns of risks in the medical insurance fund, and its specific beneficial effects are as follows:

[0245] Traditional risk assessment focuses solely on the abnormal characteristics of individual medical institutions, neglecting the risk contagion effect between institutions. This leads to the underreporting of potentially high-risk institutions that, while not exhibiting obvious abnormalities themselves, are susceptible to contagion from abnormal risks of related institutions. This method incorporates medical institutions into a business network for risk assessment. By calculating the network contagion risk additive, it accurately quantifies the degree of contagion of abnormal risks from other institutions to the target institution. It can identify potentially high-risk institutions with low inherent risk but facing higher contagion risks due to abnormalities in related institutions. This represents a shift from "independent risk assessment" to "networked risk assessment," making risk assessment more comprehensive and accurate.

[0246] Systemic risks to the medical insurance fund often stem from abnormal risks in a few core institutions that spread rapidly within a region through business networks, ultimately triggering systemic fund risks across the entire region. Traditional independent risk assessments cannot capture this risk propagation trend and are insufficient for effective early warning of systemic risks. This method, by constructing a networked risk assessment system, can track the propagation path and intensity of risks among institutions in real time. For example, by identifying risk sources, core transmission nodes, and high-risk transmission paths within a region, regulatory authorities can use this information to proactively control core transmission nodes and high-risk transmission paths, preventing further spread of risks, effectively preventing systemic and systemic risks to the medical insurance fund, and enhancing the overall security of the medical insurance fund.

[0247] Comprehensive and robust risk assessment i This comprehensive approach reflects an institution's inherent risks and network contagion risks, enabling regulatory authorities to more accurately classify institutions' risk levels. For example, institutions with high inherent risks and high network contagion risks are classified as extremely high-risk institutions, requiring immediate special audits; institutions with low inherent risks but high network contagion risks are classified as potentially high-risk institutions, requiring key monitoring; and institutions with low inherent risks and low network contagion risks are classified as low-risk institutions, allowing for reduced regulatory frequency. Based on this more refined risk level classification, regulatory authorities can prioritize the allocation of limited regulatory resources to extremely high-risk institutions and core contagion nodes, achieving optimal allocation of regulatory resources and significantly improving the efficiency of medical insurance supervision.

[0248] This embodiment divides risk transmission into primary and secondary transmission, and introduces an attenuation coefficient. and additional attenuation factor This method allows the intensity of risk contagion to gradually decrease with increasing levels of association, and the intensity of indirect contagion to decrease exponentially with path length, perfectly aligning with actual risk propagation patterns. In healthcare business networks, the risk contagion intensity of directly associated institutions is far higher than that of indirectly associated institutions, and the longer the indirect association path, the weaker the risk contagion intensity. This realistic quantitative calculation method makes the calculation results of network contagion risk additives more reasonable and accurate, avoiding overestimation or underestimation of the risk contagion effect and improving the reliability of the comprehensive robust risk score.

[0249] In some embodiments, the method further includes:

[0250] Obtain a list of known medical insurance policy shock events that occur within T consecutive accounting periods, and generate a policy shock identifier vector D for each accounting period t. t , where D t It is a multi-dimensional vector, where each dimension corresponds to a specific policy shock event. If the policy shock event continues to be effective in period t or later, the corresponding dimension value is 1; otherwise, it is 0.

[0251] The policy impact identification vector D t Introducing the restricted regression equation as a control variable yields the extended restricted regression equation:

[0252] ;

[0253] Where G is the policy shock identifier vector D t The coefficient vector corresponding to the dimension is used to capture the impact of various policy shocks on the total medical insurance reimbursement expenditure C in the regulated area. t The average systemic impact, It is the transpose of G;

[0254] In step S3, based on the time series samples and the extended restricted regression equation, a robust parameter estimation mechanism is used to simultaneously estimate the individual contribution coefficient 'a'. i The policy shock coefficient vector G is then used to obtain a stable solution.

[0255] In this embodiment, a medical insurance policy shock event refers to a medical insurance policy adjustment event issued by the national or local medical insurance department within a consecutive T accounting cycles that has a systematic and overall impact on the reimbursement behavior and expenditure scale of the regional medical insurance fund. This includes, but is not limited to, adjustments to the medical insurance reimbursement ratio, updates to the catalog of medical treatment items, reforms to medical insurance payment methods, adjustments to critical illness insurance policies, adjustments to the deductible / cap of the medical insurance fund, and adjustments to the medical insurance price limits for drugs and consumables. These are important systemic factors that cause fluctuations in the total expenditure of regional medical insurance.

[0256] Policy shock identification vector D tIt refers to a multidimensional binary vector generated for each accounting period t. Its dimension is consistent with the number of medical insurance policy shock events, and each dimension corresponds to a specific medical insurance policy shock event. If a policy shock event continues to be effective in accounting period t or thereafter, the value of that dimension is 1; otherwise, the value is 0. It is the core carrier for transforming qualitative policy shock events into quantitative model control variables.

[0257] The extended restricted regression equation refers to the basic restricted regression equation that incorporates a policy shock identifier vector D. t A new regression equation was obtained using these as control variables; this equation can effectively isolate the systematic impact of medical insurance policy shocks on total regional medical insurance expenditures, allowing the individual contribution coefficient 'a' to be determined. i The estimates are more likely to reflect the true characteristics of medical institutions' medical insurance reimbursement behavior.

[0258] The policy shock coefficient vector G refers to the vector corresponding to the policy shock identifier D. t The coefficient vector corresponding to each dimension, where the coefficient value of each dimension is used to quantify the total medical insurance reimbursement expenditure C in the regulated region for the corresponding medical insurance policy shock event. t The average systemic impact; a positive coefficient value indicates that the policy shock event leads to an increase in the total regional medical insurance expenditure, while a negative coefficient value indicates a decrease in the total expenditure, and the absolute value of the coefficient value indicates the intensity of the policy shock.

[0259] This embodiment addresses the problem of distorted individual contribution coefficient estimation caused by policy shocks in traditional models by introducing the concept of control variables from econometrics. Medical insurance policy shocks systematically affect the reimbursement behavior of all medical institutions within a region, leading to significant fluctuations in total regional medical insurance expenditures. Traditional models cannot isolate this systematic impact, misinterpreting normal policy-driven behavioral changes as abnormal institutional behavior. The specific steps and principles are as follows:

[0260] First, a comprehensive review of all medical insurance policy impact events issued by national, provincial, and municipal medical insurance departments over T consecutive accounting periods is conducted to create a list of policy impact events. This review must clearly identify core information for each policy impact event, including its release date, implementation date, content of policy adjustments, and scope of impact. Simultaneously, the policy impact events are standardized by removing policy adjustments not significantly related to regional medical insurance fund expenditures and merging policy impact events with similar content and implementation dates. This ensures that the list of policy impact events accurately reflects systemic policy changes affecting regional medical insurance fund expenditures.

[0261] Then, based on the compiled list of policy shock events, a policy shock identifier vector D is generated for each accounting period t. tDuring construction, the vector's dimensions are first determined to be the number of policy shock events, with each dimension corresponding to a specific policy shock event. Then, values ​​are assigned to each dimension based on the policy's official implementation time and its continued effectiveness. Specifically, if a policy is officially implemented or has been implemented and remains effective in accounting period t, the dimension is assigned a value of 1; if a policy has not yet been implemented or has been repealed in accounting period t, the dimension is assigned a value of 0. Simultaneously, considering the lag in policy implementation, some medical insurance policies do not immediately have a significant impact in the implementation period but have a lag of 1-2 accounting periods. In this case, the corresponding dimension value needs to be assigned a value of 1 starting from the accounting period in which the policy actually has an impact, based on the policy's actual effect, so that the policy shock indicator vector can more accurately reflect the actual impact of the policy on medical insurance fund expenditures.

[0262] The policy impact identification vector D t Introducing these variables into the basic restricted regression equation yields the extended restricted regression equation: Where G is the policy shock identifier vector D. t The coefficient vector corresponding to the dimension is used to capture the impact of various policy shocks on the total medical insurance reimbursement expenditure C in the regulated area. t The average systemic impact, It is the transpose of G; Let C be the inner product of the transpose of the policy shock identifier vector and the policy shock coefficient vector, used to quantify the impact of all effective medical insurance policy shock events on the total regional medical insurance reimbursement expenditure C within the t-th accounting period. t The overall systemic impact. The core logic of constructing this extended equation is to isolate the systemic impact of policy shocks: in the basic regression equation, the fluctuations in regional total medical insurance expenditure caused by medical insurance policy shocks are incorporated into the environmental constant offset c and the regression residuals. In this context, the individual contribution coefficient a cannot be effectively separated from the behavioral characteristics of medical institutions. i The estimates include interference from policy shocks and cannot truly reflect the actual behavior of institutions; while D t After being included as a control variable, the policy shock affects C. t The systemic impact was partially and precisely quantified, effectively separating the policy shock effect from the behavioral effect of medical institutions, allowing the basic regression part to more accurately depict the true relationship between the behavior of medical institutions and the total regional medical insurance expenditure.

[0263] The coefficients of each dimension of the policy shock coefficient vector G can accurately quantify the average systematic impact of the corresponding policy shock event on the total regional medical insurance expenditure; while the individual contribution coefficient a obtained after removing the policy shock effect... iThis eliminates systemic interference caused by policy changes and more accurately reflects the actual medical insurance reimbursement behavior of medical institutions in the absence of policy impacts, thus avoiding misjudging normal policy-driven behavioral changes as abnormal institutional behavior.

[0264] After obtaining the policy impact coefficient vector G, it can be interpreted in an operational manner: a positive coefficient value for a policy indicates that the implementation of the policy leads to an increase in the total regional medical insurance expenditure; a negative coefficient value indicates that the implementation of the policy effectively controls the growth of the total medical insurance expenditure; the larger the absolute value of the coefficient value, the stronger the impact of the policy on the total medical insurance expenditure. These quantitative results can provide empirical data support for medical insurance policy-making departments to evaluate the implementation effect of existing policies and optimize subsequent medical insurance policy adjustment strategies.

[0265] Adjustments to medical insurance policies have a global and systemic impact on the reimbursement behavior of all medical institutions within a region, leading to significant fluctuations in total regional medical insurance expenditures. For example, if the medical insurance department increases the reimbursement ratio for oncology treatment, the reimbursement amount for all medical institutions offering oncology treatment within the region will increase, altering individual contribution coefficients. Traditional models cannot isolate this systemic impact, misjudging such policy-driven normal behavioral changes as abnormal institutional behavior, and also resulting in a lack of comparability in assessing institutional behavior across different policy cycles.

[0266] The method described in this embodiment incorporates policy shocks as control variables into the model, accurately quantifies and isolates the impact of policy shocks on total regional medical insurance expenditures, and calculates the individual contribution coefficient 'a'. i It fully reflects the behavioral characteristics of medical institutions themselves, rather than the results driven by policies. This not only avoids misjudging normal policy changes as institutional abnormalities, but also provides a unified benchmark for assessing institutional behavior across different policy cycles, significantly improving the accuracy and cross-cycle comparability of assessment results.

[0267] The coefficients of each dimension of the policy shock coefficient vector G precisely quantify the average systematic impact and direction of the corresponding medical insurance policy shock event on the total regional medical insurance reimbursement expenditure, providing valuable empirical data support for medical insurance policy formulation and evaluation departments. The policy shock coefficient can be used to assess the effectiveness of existing policies, such as determining whether medical insurance payment reform has achieved the goal of controlling medical insurance fund expenditure, or whether adjustments to medical insurance price limits for drugs and consumables have effectively reduced the total regional medical insurance expenditure.

[0268] This invention addresses the parameter insolvability problem caused by the large number of institutions and short observation periods in regional medical insurance supervision, proposing a robust modeling path based on an "individual-background" binary deconstruction. The specific implementation process is as follows:

[0269] Step 1: Multi-source data preprocessing and feature alignment. First, acquire summary data for T consecutive accounting periods (e.g., calendar months) within the regulated region, including all outpatient and inpatient medical invoices generated by all medical institutions within the regulated region during each accounting period. The medical invoice data must include at least the fields for issuing unit, invoice type, settlement time, and detailed cost breakdown. For each medical invoice, identify and extract the amount paid by the medical insurance fund from its cost breakdown, specifically including:

[0270] Global parameter extraction: Obtaining the total expenditure reimbursed by regional medical insurance .

[0271] Individual parameter extraction: Taking medical institutions as the statistical subject, and according to a preset accounting period, the medical insurance fund payment amounts corresponding to all medical invoices under the same medical institution are aggregated to obtain the total amount of medical insurance fund payments within the region. Independent reimbursement amount for each medical institution .

[0272] Global parameter modeling: Since the total expenditure of the medical insurance fund in the regulatory region is obtained by summing up the medical insurance fund payment amount of all medical bills in the regulatory region from bottom to top, under the premise of consistent statistical caliber, it theoretically satisfies the following formula:

[0273] ;

[0274] Alignment and calibration processing: The time series data constructed above are subjected to consistency verification and standardization processing to remove abnormal records caused by factors such as missing invoices, settlement delays, cross-period accounting or statistical rule adjustments, so as to ensure that the data between different medical institutions and between different accounting cycles are comparable, and provide a stable data foundation for subsequent modeling and analysis.

[0275] Step 2: Deconstructing and balancing the "target-background" space. The system does not directly perform full-scale analysis. Instead of a return to the original state, this applies to each medical institution that needs to be verified. Perform the following operations:

[0276] Constructing a background mask: This involves masking areas other than medical institutions. All medical institutions other than those mentioned above are considered as a whole for calculating the total background reimbursement amount: Establish a restricted regression equation: Construct a mathematical model that can describe the proportion of an individual's contribution to the overall regulatory area.

[0277] ;

[0278] in, This represents the individual contribution coefficient of the target institution. Under normal compliance circumstances, changes in reimbursement for a single medical institution should be consistent with changes in the overall region, i.e., when the total medical insurance reimbursement expenditure in the regulated region... When the amount increases, right The contribution relationship is stable. Approximately the group median level; when insurance fraud, excessive medical treatment, or fraudulent hospitalization occur, the independent reimbursement amount for this medical institution in the t-th accounting period is... More of the reimbursements stemmed from internal operations than from genuine regional healthcare needs. After deconstructing the "individual-background" relationship, their behavior and the total amount reimbursed based on their background... The synergistic relationship weakens, making it easier to fit. The model can only suppress The significantly smaller value reflects a decrease in the consistency between the medical institution's reimbursement behavior and the overall medical treatment structure in the regulated region. This phenomenon statistically corresponds to an increase in the proportion of abnormal endogenous drivers, thus indicating a higher probability of fund compliance risk.

[0279] The background contribution coefficient (reflects the stability of the macro-level medical treatment structure and generally approaches 1). This is to account for environmental constant offsets (absorbing global noise such as changes in medical insurance policies and seasonal adjustments). The regression residuals for each accounting period.

[0280] Step 3: Considering that medical insurance fund data may be affected by non-structural factors such as seasonal disease outbreaks and centralized settlement during actual operation, leading to short-term abnormal fluctuations in total regional medical insurance fund expenditures within certain time windows, directly using the ordinary least squares method for parameter estimation could easily cause model parameters to be dominated by a few extreme samples, thereby weakening the stable expression of individual behavioral characteristics. Therefore, this invention introduces a robust parameter estimation mechanism without changing the "target medical institution—background medical institution" modeling structure, to estimate the individual contribution coefficients in the regression model. A stable solution is obtained.

[0281] The robust parameter estimation mechanism adaptively adjusts the weights of time series samples, ensuring that the estimated model parameters primarily reflect the long-term stable behavioral patterns of the target medical institution across multiple time windows, rather than being affected by short-term fluctuations in individual abnormal months. For time window samples that deviate significantly from the overall fitting trend, their influence weight in the parameter estimation process is automatically suppressed, thereby avoiding interference from extreme cases with the individual consistency coefficient estimation results.

[0282] It should be noted that the robust estimation method introduced in this step is not used to directly identify abnormal time windows or abnormal sample points, but only as a numerical solution tool to obtain statistically stable and comparable individual contribution coefficients under conditions of limited sample size and significant background noise. .

[0283] Through the above method, even when the time sample size T is much smaller than the number of medical institutions N, the model still only involves the individual contribution coefficients of the target medical institution. Background contribution coefficient The three parameters to be estimated, including the constant term c, satisfy the parameter identifiability condition, ensuring that subsequent consistency analysis based on coefficient distribution has a reliable statistical basis.

[0284] Step 4: Anomaly feature extraction based on population distribution. When Individual coefficient of family institution After all calculations are completed, the system enters the risk assessment phase, which includes the following steps:

[0285] Constructing a statistical baseline: Calculating the median of the set of coefficients. and absolute median deviation .

[0286] The Robust Risk Score is calculated using the following formula:

[0287] ;

[0288] in, The absolute median deviation of the set of individual contribution coefficients is used to characterize the robustness of the coefficient distribution. A coefficient of 1.4826 is a scaling factor used to convert the absolute median deviation into a robust scaling estimate with dimensions consistent with the standard deviation under approximately normal distribution conditions. This ensures the comparability of the deviations in individual contribution coefficients across different medical institutions, thereby enabling the... Mapped to the risk range.

[0289] Behavioral consistency analysis: If a certain medical institution corresponds to A preset risk score threshold indicates that the medical institution's behavioral logic is significantly different from the overall "background pattern" of the regulated area.

[0290] Step 5: Multidimensional Anomaly Detection and Early Warning Trigger Mechanism. After obtaining the set of individual contribution coefficients of all medical institutions within the region and completing robust distribution modeling, the system further combines horizontal distribution deviation characteristics and vertical time stability characteristics to perform multidimensional anomaly detection on the medical insurance reimbursement behavior of the target medical institutions and trigger corresponding risk early warning mechanisms, specifically including:

[0291] (1) Static anomaly warning based on lateral distribution deviation. The system uses the statistical distribution of individual contribution coefficients of regional medical institutions within the same accounting period as a reference background to determine the individual contribution coefficients of target medical institutions. A horizontal comparative analysis was conducted. This involved analyzing the individual contribution coefficients of the target medical institution within a specific accounting period. A significant deviation from the central interval of the distribution (e.g., significantly lower or higher than the robust reference interval) indicates a significant decrease in the consistency between the medical institution's medical insurance reimbursement behavior and the overall regional medical service structure during that period. This static deviation does not directly confirm the existence of abnormal behavior, but rather serves as a risk signal based on statistical consistency, indicating that the medical institution may exhibit atypical reimbursement behavior characteristics during the current period, thereby triggering a static anomaly warning marker.

[0292] (2) Dynamic trend early warning based on time series evolution. Building upon static deviation analysis, the system further introduces a longitudinal time dimension to analyze the changing trends of individual contribution coefficients of the target medical institution across multiple consecutive accounting periods. Specifically, the system analyzes the individual contribution coefficient sequence {a1, a2, ..., a...} of the target medical institution in the t-th verification period. N}

[0293] Then, a trend and stability assessment is conducted. A dynamic anomaly is identified when one or more of the following conditions are met:

[0294] The individual contribution coefficient shows a unilateral and continuous downward or upward trend over multiple consecutive accounting periods, and the degree of deviation gradually increases;

[0295] The individual contribution coefficient fluctuated frequently and significantly between adjacent accounting periods, which was significantly higher than the normal fluctuation level of the regional medical institutions as a whole.

[0296] The individual contribution coefficient has long been in an abnormal range of regional distribution and has not regressed with changes in the overall regional medical structure.

[0297] The aforementioned dynamic characteristics reflect the structural instability of the target medical institution's medical insurance reimbursement behavior over time, suggesting that it may have a continuous impact on the structure of medical insurance fund consumption through methods such as splitting charging items, fabricating or exaggerating medical treatment behavior, and abnormal centralized settlement.

[0298] (3) Multidimensional Risk Fusion and Early Warning Triggering. The system integrates static distribution deviation results and dynamic trend analysis results to generate a comprehensive abnormal risk assessment conclusion for the target medical institution, and triggers corresponding level risk warning signals according to preset rules. Among them, single-cycle static deviation is mainly used to indicate short-term abnormal behavior risks, while dynamic trend warning is used to identify potential systemic and persistent compliance risks. Through the multidimensional judgment mechanism, the probability of false alarms caused by occasional peak medical visits or sudden events can be effectively reduced, and the reliability and regulatory interpretability of abnormal identification results can be improved. Finally, the system outputs the abnormal risk warning results to the medical insurance supervision platform or manual audit system as a reference for subsequent special verification, key monitoring or joint supervision.

[0299] Unlike the traditional approach of simultaneously incorporating all medical institutions within a region into regression modeling, this invention constructs a restricted regression relationship model between any target medical institution and the overall behavior of the other medical institutions, transforming the originally unsolvable multivariate regression problem into a low-dimensional, identifiable parameter estimation problem.

[0300] By abstracting the "other medical institutions" as a whole into a background reference system that changes dynamically over time, the technical bottleneck of parameters being unpredictable and results being unstable under the condition that "the number of medical institutions is much greater than the number of time samples" is effectively avoided, and a structural basis is provided for the stable quantification of the behavioral characteristics of a single medical institution.

[0301] In the aforementioned dimensionality reduction modeling structure, this invention introduces a background contribution coefficient. This allows for a unified characterization of the overall reimbursement behavior of other medical institutions within the regulatory area, ensuring that the parameter is no longer simply a mathematical fit but is endowed with clear business and regulatory semantics.

[0302] The background contribution coefficient is used to characterize the public impact of changes in the overall regional medical treatment structure, disease spectrum distribution, and macro-level medical demand on the consumption of medical insurance funds, thereby forming a stable "background anchor" at the model level. Through this anchor design, the influence of global fluctuations such as seasonal epidemics on the model can be absorbed to a certain extent, allowing the individual behavioral characteristics of the target medical institution to be effectively separated from complex background noise.

[0303] This invention, in its anomaly identification logic, shifts from the traditional method of judgment "based on reimbursement amount or regression residual size" to a paradigm "based on the degree of consistency deviation of individual contribution coefficients in the group distribution." This invention does not directly focus on the absolute level of a medical institution's reimbursement amount, but rather determines whether it maintains statistical consistency with other medical institutions in the same region by analyzing the relative contribution of that medical institution's medical insurance reimbursement behavior to the total expenditure of the regional medical insurance fund.

[0304] The above methods can effectively identify hidden abnormal behaviors where the reimbursement amount appears to be close to the regional average, but its internal contribution structure has undergone systematic shifts. For example, it can circumvent traditional threshold rules by fine-tuning the charging structure and dispersing the reimbursement amount, thereby significantly improving the sensitivity and accuracy of anomaly identification.

[0305] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical insurance anomaly identification method based on background structure as described in the first aspect of the present invention.

[0306] The computer-readable storage medium may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0307] The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device.

[0308] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of the present invention are intended to include these and any other suitable types of memory.

[0309] like Figure 2 As shown, in a third aspect, the present invention provides an electronic device 10, including a processor 101 and a storage medium 102, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the medical insurance anomaly identification method based on background structure as described in the first aspect of the present invention.

[0310] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor, thereby enabling the processor to execute some or all of the steps or any combination thereof in the medical insurance anomaly identification method based on background structure described in the various embodiments of the present invention.

[0311] Finally, it should be noted that although the above embodiments have been described in the description and drawings of this invention, this should not limit the scope of patent protection of this invention. Any technical solutions that are based on the essential concept of this invention, utilize the content described in the description and drawings of this invention to make equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this invention.

Claims

1. A method for identifying medical insurance anomalies based on background structure, characterized in that, The method includes: S1: Obtain medical invoice data for T consecutive accounting periods within the regulatory region, extract the medical insurance fund payment amount from the medical invoice data, and aggregate them to obtain the total medical insurance reimbursement expenditure C within the regulatory region. t The individual reimbursement amount for each medical institution within the regulated region x i,t Where T is the number of accounting cycles, i is the medical institution index, and t is the accounting cycle index; S2: For any medical institution i to be verified within the regulatory area, construct a background mask, treating all medical institutions other than medical institution i as a whole, and calculate the total background reimbursement amount X corresponding to medical institution i. -i,t The calculation formula is as follows: X -i,t =C t -x i,t And establish a restricted regression equation: ； in, For the total medical insurance reimbursement expenditure in the regulatory region during the t-th accounting period, X represents the independent reimbursement amount for medical institution i in the t-th accounting period. -i,t For the total amount of background reimbursement, Let be the individual contribution coefficient of medical institution i, used to characterize the degree to which the medical insurance reimbursement behavior of this medical institution is driven by the demand for medical treatment in the external region. Here, c represents the background contribution coefficient, and c represents the environmental constant offset. Let be the regression residual for the t-th accounting period; S3: Time series samples based on T accounting periods {(x i,1 ,X -i,1 ,C1),(x i,2 ,X -i,2 ,C2),...,(x i,T ,X -i,T C T A robust parameter estimation mechanism is used to estimate the individual contribution coefficients in the restricted regression equation. To achieve a stable solution, the robust parameter estimation mechanism includes: S31: Perform iterative weighted least squares estimation on the time series samples, and assign adaptive weights w to each accounting period t. t ; S32: Based on the regression residuals corresponding to each accounting period t Size dynamically adjusts weight w t For outlier sample points with large residuals, their weights should be reduced. S33: Through multiple rounds of iteration, the individual contribution coefficient is increased. The estimated value remains stable even with reduced weights for outlier sample points with large residuals, thus obtaining the individual contribution coefficient that reflects the stable behavioral characteristics of the target medical institution across multiple accounting periods. ; S4: Calculate the set of individual contribution coefficients {a1, a2, ..., a...} for all N medical institutions within the regulatory area. N Based on the set of individual contribution coefficients, a statistical baseline is constructed, and a robust risk score for each medical institution is calculated. S5: Based on the time series variation characteristics of the robust risk score and / or individual contribution coefficient, perform multidimensional anomaly determination on the target medical institution and trigger corresponding risk warnings. The time series variation characteristics include the trend variation characteristics and stability variation characteristics of the individual contribution coefficient in multiple consecutive accounting periods.

2. The medical insurance anomaly identification method based on background structure as described in claim 1, characterized in that, In step S4, constructing the statistical baseline specifically includes: Calculate the median Med(a) and absolute median deviation MAD(a) of the set of individual contribution coefficients; Calculate robust sizing estimates The calculation formula is as follows: Where k is a preset scale calibration factor used to adjust the absolute median deviation. Convert it into a robust measure of dispersion with dimensions consistent with the standard deviation under a normal distribution; Calculate robust risk score The calculation formula is as follows: 。 3. The medical insurance anomaly identification method based on background structure as described in claim 1 or 2, characterized in that, Step S5 includes static anomaly determination and dynamic anomaly determination; The static anomaly determination includes: based on the robust risk score, when the robust risk score of the target medical institution in a certain accounting period exceeds the preset static threshold, it is determined that there is a static anomaly and an early warning is triggered; The dynamic anomaly determination includes: evaluating the individual contribution coefficient sequence of the target medical institution over multiple consecutive accounting periods based on the time series change characteristics of the individual contribution coefficient; when the change characteristics of the individual contribution coefficient sequence meet a preset dynamic anomaly pattern, a dynamic anomaly is determined to exist and an early warning is triggered; the dynamic anomaly pattern includes at least one of the following: The individual contribution coefficient sequence shows a statistically significant monotonic trend, and its deviation from the center of the population distribution increases as the trend develops; The fluctuation range of the individual contribution coefficient sequence between consecutive accounting periods exceeds the dynamic fluctuation threshold determined by the concurrent fluctuation level of the medical institutions in the regulatory region. The individual contribution coefficient remains in the abnormal state range determined by the group distribution for more than the preset number of durations.

4. The medical insurance anomaly identification method based on background structure as described in claim 1, characterized in that, In step S3, the individual contribution coefficient is obtained by solving... Then, the following steps are also included: Obtain the static attribute vector Z of the target medical institution i. i The static attribute vector includes one or more of the following: medical institution level, approved number of beds, and number of key departments; Constructing a mapping model between medical institution attributes and contribution coefficients, specifically including: based on the static attribute vector set {Z1, Z2, ..., Z...} of all N medical institutions within the regulatory area. N } and its corresponding set of individual contribution coefficients {a1, a2, ..., a N Train a regression model f(·) to obtain the structural contribution determined by the inherent attributes of the medical institution. The calculation formula is as follows: ; The individual contribution coefficient a i Decomposed into structural contribution coefficients and behavioral contribution coefficient ,in: ； In step S5, a multidimensional anomaly assessment is performed on the target medical institution, and corresponding risk warnings are triggered, including: Combination and Perform multidimensional anomaly detection: When the first condition is met, it is determined to be a risk of distorted medical practice. The first condition includes: the robustness risk score exceeds a preset risk threshold, and simultaneously meets the following conditions: >τ b and >λ· , where τ b λ is the threshold for the behavioral contribution coefficient, and λ is the contribution ratio coefficient. When the second condition is met, it is determined to be a risk of mismatch between scale and positioning. The second condition includes: the robustness risk score exceeds a preset risk threshold, and simultaneously meets the following conditions: >τ s and >λ· And based on the static attribute vector Z i The calculated change in the medical institution size positioning index over M consecutive accounting periods is within the preset error range, where τ s The threshold for the structural contribution coefficient.

5. The medical insurance anomaly identification method based on background structure as described in claim 1, characterized in that, In step S2, constructing the background mask includes: Acquire patient transfer data among medical institutions within the regulatory region over a consecutive T accounting periods; For any medical institution i to be verified and the accounting period t, a directed weighted network is constructed based on the flow data; wherein, the nodes of the directed weighted network are medical institutions, and if a patient flows from medical institution u to medical institution v within the period t, there exists an edge from u to v, with the weight being the number of patients transferred. Starting with medical institution i, in the directed weighted network, all medical institutions that have direct or indirect patient transfer relationships with medical institution i are selected to form the dynamic background institution set B of medical institution i in period t. i,t Among them, the path length between selected medical institutions is less than the preset length; Calculate the total amount of reimbursement under dynamic background The calculation formula is as follows: ； Where j represents the dynamic background mechanism set B i,t Index of Chinese medical institutions This represents the independent reimbursement amount for medical institution j in the dynamic background institution set during the t-th accounting period; In the restricted regression equation, the total amount of reimbursement under the dynamic background is used. Replace the original background reimbursement total amount x i,t Perform the calculation.

6. The medical insurance anomaly identification method based on background structure as described in claim 1, characterized in that, The process of identifying multidimensional anomalies in target medical institutions and triggering corresponding risk warnings includes the following steps: Calculate the behavioral correlation w between any two medical institutions p and q based on the time series of historical individual contribution coefficients or the business similarity between medical institutions. pq This forms a matrix of relationships among medical institutions; For the current accounting period t, calculate the change in the individual contribution coefficient of each medical institution pair (p, q). and And calculate the correlation change score of medical institutions. The calculation formula is as follows: ;in, This represents the individual contribution coefficient of medical institution p within the current accounting period t. This represents the individual contribution coefficient of medical institution p in the previous accounting period t-1. This represents the individual contribution coefficient of medical institution q within the current accounting period t. This represents the individual contribution coefficient of medical institution q in the previous accounting period t-1; If a medical institution pair (p, q) that meets the third condition is identified, an association anomaly alert is triggered. The third condition includes the behavioral association degree w between medical institutions p and q. pq The score for the association change of (p,q) is higher than the preset association threshold. The absolute value is higher than the preset correlation change threshold, and its sign is negative.

7. The medical insurance anomaly identification method based on background structure as described in claim 1, characterized in that, In step S4, a statistical baseline is constructed based on the set of individual contribution coefficients, and the robust risk score for each medical institution is calculated, including: Based on preset association rules, the direct association strength ρ between any two medical institutions p and q within the regulatory area is calculated. pq , where 0≤ρ pq ≤1, ρ pp = 1, and construct an N-row, N-column correlation strength matrix R; For any medical institution i, calculate its individual abnormality intensity. The calculation formula is as follows: ； in, and These are sets of individual contribution coefficients based on historical cycles, {a i The calculated benchmark mean and benchmark standard deviation; Calculate the native robust risk score of healthcare institution i The native robust risk score By analyzing the individual abnormal intensity of medical institution i The result is obtained after smoothing and standardization. Based on the aforementioned correlation strength matrix R, the network infection risk bonus of medical institution i is calculated. The calculation formula is: ； in, and The preset attenuation coefficient, >0, <1; This represents the additional attenuation factor for secondary transmission, 0 < <1, Let be the shortest path length from medical institution j to medical institution k; Calculate the comprehensive robust risk score R of medical institution i i The calculation formula is as follows: 。 8. The medical insurance anomaly identification method based on background structure as described in claim 1, characterized in that, The method further includes: Obtain a list of known medical insurance policy shock events that occur within T consecutive accounting periods, and generate a policy shock identifier vector D for each accounting period t. t , where D t It is a multi-dimensional vector, where each dimension corresponds to a specific policy shock event. If the policy shock event continues to be effective in period t or later, the corresponding dimension value is 1; otherwise, it is 0. The policy impact identification vector D t Introducing the restricted regression equation as a control variable yields the extended restricted regression equation: ； Where G is the policy shock identifier vector D t The coefficient vector corresponding to the dimension is used to capture the impact of various policy shocks on the total medical insurance reimbursement expenditure C in the regulated area. t The average systemic impact, It is the transpose of G; In step S3, based on the time series samples and the extended restricted regression equation, a robust parameter estimation mechanism is used to simultaneously estimate the individual contribution coefficient 'a'. i The policy shock coefficient vector G is then used to obtain a stable solution.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the medical insurance anomaly identification method based on background structure as described in any one of claims 1 to 8.

10. An electronic device having a computer program stored thereon, characterized in that, It includes a processor and a storage medium, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the medical insurance anomaly identification method based on background structure as described in any one of claims 1 to 8.

Images