Precise monitoring and intervention method for medical insurance cost of medical institution

Abstract

The invention discloses a medical institution medical insurance cost accurate monitoring and intervention method, and relates to the technical field of medical informationization, and the method comprises the following steps: collecting multi-source medical insurance data in real time, carrying out the standardized cleaning, and generating a standardized data flow; the standardized data flow is converted into a multi-dimensional feature vector, and time dimension deviation is eliminated through a dynamic baseline; based on the generated multi-dimensional feature vector, calculating a dynamic association weight of each feature and medical insurance cost change in real time; in combination with the feature vectors and the dynamic weights, analyzing a time sequence association relationship between the features, and constructing a multi-dimensional association network to capture a combined abnormal behavior pattern; integrating the feature vector, the dynamic weight and the association network result to generate a real-time anomaly score; triggering a grading response mechanism according to the real-time abnormal score, and executing an automatic intervention measure from observation recording to forced interception; and on the basis of intervention records and historical data, periodically updating a feature baseline and a weight calculation rule to form self-optimization closed-loop control.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical informatization, and particularly relates to a medical institution medical insurance expense accurate monitoring and intervention method. BACKGROUND

[0002] The monitoring and intervention method for medical insurance expenses of medical institutions in the prior art generally has problems of insufficient data integration capability, lagging real-time response, low abnormal behavior recognition accuracy, and the like, and traditional methods mostly rely on static rules or manual auditing, and are difficult to cope with complex and changeable diagnosis and treatment behaviors and dynamic expense fluctuations, resulting in high abnormal detection false positive rate and great missed detection risk, some systems attempt to introduce data analysis technology, but the feature processing dimension is single, time series bias cannot be effectively eliminated, and weight distribution is fixed, so the systems cannot adapt to the dynamic correlation of medical insurance expense changes, and the existing schemes lack multi-dimensional correlation network analysis, and are difficult to capture combined abnormal behavior patterns, and intervention measures are mostly post-processing, so real-time blocking and closed-loop optimization cannot be realized; these defects result in low medical insurance fund supervision efficiency and risk of unreasonable expense expenditure, and therefore the application provides a medical institution medical insurance expense accurate monitoring and intervention method. SUMMARY

[0003] To solve the above technical problems, the application is implemented through the following technical scheme:

[0004] The application is a medical institution medical insurance expense accurate monitoring and intervention method, comprising the following steps:

[0005] Step S1: real-time collection of multi-source medical insurance data and standardization cleaning, including diagnosis and treatment projects, drug use, patient information and diagnosis codes, to generate standardized data flow;

[0006] Step S2: conversion of the standardized data flow into a multi-dimensional feature vector, including patient attributes, diagnosis and treatment behaviors and time series change characteristics, elimination of time dimension bias through dynamic baseline;

[0007] Step S3: real-time calculation of the dynamic correlation weight of each feature and medical insurance expense change based on the generated multi-dimensional feature vector, and weight initialization through historical data correlation;

[0008] Step S4: analysis of the time series correlation between features in combination with the feature vector and the dynamic weight, construction of a multi-dimensional correlation network to capture combined abnormal behavior patterns;

[0009] Step S5: generation of a real-time abnormal score by comprehensively combining the feature vector of step two, the dynamic weight of step three and the correlation network result of step four, to quantify the risk level of diagnosis and treatment behaviors;

[0010] Step S6: Trigger a hierarchical response mechanism according to the real-time anomaly score, and perform automated intervention measures from observation recording to forced interception according to the risk level;

[0011] Step S7: Based on the intervention record of step six and the historical data of step one, periodically update the feature baseline and weight calculation rule to form a self-optimizing closed-loop control.

[0012] Further, the step S1 includes the following steps:

[0013] Step S11, real-time acquisition of medical insurance settlement data stream through medical institution HIS interface, including diagnosis and treatment item code, drug code, quantity, unit price, patient basic information and diagnosis code;

[0014] Step S12, standardization processing of unstructured diagnosis description text, using ICD-10 code converter to standardize the diagnosis name;

[0015] Step S13, perform data cleaning operation, delete duplicate records, complete missing fields, correct abnormal values, and records with unit price exceeding the preset threshold trigger manual review process.

[0016] Further, the step S2 includes the following steps:

[0017] Step S21, define the features of individual dimension, diagnosis and treatment dimension and time sequence dimension:

[0018] Individual dimension: patient age segmentation, past medical history;

[0019] Diagnosis and treatment dimension: single visit item book, high-priced drug proportion, inspection item clustering degree;

[0020] Time sequence dimension: same department visit frequency week-on-week change rate, same kind of drug dosage 3-day moving average, same diagnosis item price fluctuation standard deviation;

[0021] Step S22, eliminate dimensional difference and enhance feature comparability, dynamic baseline standardization of original data, formula as follows:

[0022]

[0023] In the formula, is the value of the normalized feature i at time t, is the untreated value of feature i in the original data at time t, is the mean value of feature i in the past t hours, used for dynamic baseline calibration, is the standard deviation of feature i in the past 7 days, used for standardization scaling, 3 times standard deviation is used to exclude the influence of extreme values.

[0024] Further, the step S3 comprises the following steps:

[0025] Step S31, assign initial weights to features according to their correlation with cost in historical data, important features are given priority in calculation, formula as follows:

[0026]

[0027] In the formula, is the initial weight of feature i, based on the correlation between historical data and cost, is the Pearson correlation coefficient between feature i and total cost cost, the calculation range is historical data;

[0028] Step S32, update the dynamic weight using the latest cost data collected in step S1, capture the real-time correlation between features and cost fluctuations, adapt to changes in network environment, dynamic weight update formula as follows:

[0029]

[0030] In the formula, is the dynamic weight of feature i at time t, reflecting real-time correlation, λ is the forgetting factor, controlling the retention proportion of historical weight, is the covariance between feature i and cost instantaneous volatility rate Δc (t) , is the variance of feature i at time t;

[0031] Δc( t) is the cost instantaneous volatility rate, defined as follows:

[0032]

[0033] In the formula, cost (t) is the actual cost value at time t, is the average value of cost in the previous time period.

[0034] Further, the step S4 comprises the following steps:

[0035] Step S41, capture hidden association patterns between features, construct a time series association graph based on the feature matrix of step S2, the edge weight calculation formula as follows:

[0036]

[0037] In the formula, is the Pearson correlation coefficient between feature i and k within the time window T, used to quantify the time series correlation strength between the two, is the mean value of feature i within the window;

[0038] Step S42, quantifying the influence of associated features on target features, calculating feature propagation effect based on the association graph of step S41, the propagation coefficient formula is as follows:

[0039]

[0040] In the formula, γ k is the normalized influence coefficient of feature k, reflecting its global importance in the association network, PageRank(G) υk is the PageRank value of feature k in graph G, measuring its node importance, G = V, E is the association graph, where node V is the feature, and the weight of edge E is defined by .

[0041] Further, the step S5 comprises the following steps:

[0042] Step S51, integrating the real-time abnormal risk, integrating the standardized features of step S2, the dynamic weight of step S3 and the association effect of step S4, generating an abnormal score, the formula is as follows:

[0043]

[0044] In the formula, S t is the real-time abnormal score at time t, the larger the value, the higher the abnormal risk, is the Z-score standardized deviation value, measuring the instantaneous abnormality of feature i, is the time decay factor, θ = 0.1 is the control decay speed, Δt i is the time interval of the last triggered abnormality of feature i, is the weighted contribution of associated features, m is the number of associated features, β is the association effect weight, controlling the proportion of associated items in the score;

[0045] Step S52, dynamically adapting to the change of feature distribution, updating the reference value according to the real-time data of step S5, the reference update formula is as follows:

[0046]

[0047] In the formula, is the dynamic mean of feature i, updated by exponential smoothing, is the dynamic standard deviation of feature i, updated by exponential smoothing.

[0048] Further, the step S6 comprises the following steps:

[0049] Step S61, based on the abnormal score S t of step S5, triggering a hierarchical response mechanism, the threshold is set as follows:

[0050] Yellow alert (S t ∈[1.5,2.5]):Generate observation records and store in the database of step S1 for subsequent analysis;

[0051] Orange alert (S t ∈[2.5,4.0]):Freeze the current prescription for 30 minutes through the HIS interface of step S1, and block the abnormal settlement process;

[0052] Red alert (S t ≥4.0):Suspend medical insurance settlement and generate a review work order, and associate the feature data of step S2 for manual review;

[0053] Step S62, for quick intervention, freeze settlement through API, generate an electronic review work order, automatically associate patient historical visit records, similar case comparison data, and push the warning details to the responsible physician through the message middleware, and require a situation explanation to be submitted within 2 hours.

[0054] Further, the step S7 comprises the following steps:

[0055] Step S71, based on the warning records of step S6 and the historical data of step S1, correct the baseline parameters daily If If the dynamic mean deviates from the baseline by more than 2 times the standard deviation, adjust the forgetting factor in step S3 to λ=0.7 to accelerate the weight update;

[0056] In the formula, is the baseline mean, calculated based on the historical 7-day data;

[0057] Step S72, based on the weight update records of step S3 and the correlation analysis results of step S4, generate a feature importance report every week for the reviewer to adjust the feature whitelist of step S2.

[0058] The present application has the following beneficial effects:

[0059] 1、The present application collects and standardizes multiple sources of medical insurance data in real time, eliminates time dimension bias using dynamic baseline, combines dynamic weight calculation and inter-feature time sequence correlation analysis, improves the real-time and accuracy of medical insurance cost anomaly detection, updates the weight and baseline parameters in real time, dynamically captures the correlation between cost fluctuations and medical behavior, ensures that the abnormal score can quickly reflect the latest risk changes, and thus realizes automated intervention from observation records to forced interception in yellow, orange and red alert grading responses, effectively curbing unreasonable cost expenditure.

[0060] 2、The application adopts multi-dimensional correlation network analysis technology, breaks through the limitation of single feature recognition, quantifies the hidden relationship between features through the construction of time sequence correlation graph, combines the PageRank algorithm to evaluate the global influence of features, and can accurately capture the combined abnormal behavior mode, the application reduces the missed detection rate and improves the accuracy of abnormal behavior recognition through multi-dimensional feature propagation effect calculation.

[0061] 3、The application periodically corrects the feature baseline and weight calculation rule through the closed loop self-optimizing mechanism, dynamically adjusts the forgetting factor and feature white list combined with the early warning record and historical data, ensures that the system continuously adapts to the change of medical insurance cost mode, updates the dynamic mean and standard deviation through exponential smoothing, and generates a feature importance report every week, the system can automatically optimize the parameter configuration, reduces the cost of manual intervention, and the self-adaptive ability makes the monitoring model maintain high sensitivity and stability in long-term operation.

[0062] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed for the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0064] Figure 1 The flowchart of the medical institution medical insurance cost precise monitoring and intervention method of the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0066] Please refer to Figure 1 The present application is a medical institution medical insurance cost precise monitoring and intervention method, comprising the following steps:

[0067] Step S1: Real-time acquisition of multi-source medical insurance data and standardization cleaning, including diagnosis and treatment projects, drug use, patient information and diagnosis code, generating standardized data flow;

[0068] Step S2: Convert the normalized data stream into a multi-dimensional feature vector, including patient attributes, diagnosis and treatment behavior, and time series change characteristics, and eliminate time dimension bias by dynamic baseline;

[0069] Step S3: Based on the generated multi-dimensional feature vector, real-time calculation of the dynamic correlation weight of each feature and medical insurance cost change, and weight initialization through historical data correlation;

[0070] Step S4: Combine the feature vector and dynamic weight to analyze the time series correlation between features, and build a multi-dimensional correlation network to capture combined abnormal behavior patterns;

[0071] Step S5: Integrate the feature vector of step two, the dynamic weight of step three, and the correlation network result of step four to generate a real-time anomaly score, quantifying the risk level of diagnosis and treatment behavior;

[0072] Step S6: Trigger a hierarchical response mechanism according to the real-time anomaly score, and execute automatic intervention measures from observation recording to forced interception according to the risk level;

[0073] Step S7: Based on the intervention records of step six and the historical data of step one, periodically update the feature baseline and weight calculation rules to form a self-optimizing closed-loop control.

[0074] Step S1 includes the following steps:

[0075] Step S11, real-time acquisition of medical insurance settlement data stream through medical institution HIS interface, including diagnosis and treatment project code, drug code, quantity, unit price, patient basic information and diagnosis code;

[0076] Step S12, standardization processing of unstructured diagnosis description text, using ICD-10 code converter to standardize diagnosis name;

[0077] Step S13, perform data cleaning operation, delete duplicate records, complete missing fields, correct abnormal values, records with unit price exceeding preset threshold trigger manual review process.

[0078] Step S2 includes the following steps:

[0079] Step S21, define individual dimension, diagnosis and treatment dimension and time series dimension features:

[0080] Individual dimension: patient age segmentation, past medical history;

[0081] Diagnosis and treatment dimension: single visit project book, high-priced drug proportion, inspection item clustering degree;

[0082] Time dimension: same department visit frequency week-on-week change rate, same type drug dosage 3-day moving average, same diagnosis item price fluctuation standard deviation;

[0083] Step S22, eliminate dimensional differences and enhance feature comparability, dynamic baseline standardization of original data, formula as follows:

[0084]

[0085] In the formula, is the value of the normalized feature i at time t, is the unprocessed value of feature i at time t in the original data, is the mean value of feature i in the past t hours, used for dynamic baseline calibration, is the standard deviation of feature i in the past 7 days, used for standardization scaling, 3 times the standard deviation is used to exclude the influence of extreme values.

[0086] Step S3 includes the following steps:

[0087] Step S31, assign initial weights according to the correlation between features and costs in historical data, so that important features are given priority in calculation, formula as follows:

[0088]

[0089] In the formula, is the initial weight of feature i, based on the correlation between historical data and cost, is the Pearson correlation coefficient between feature i and total cost cost, calculated in the range of historical data;

[0090] Step S32, update the weight dynamically using the latest cost data collected in step S1, capture the real-time correlation between features and costs, adapt to changes in network environment, dynamic weight update formula as follows:

[0091]

[0092] In the formula, is the dynamic weight of feature i at time t, reflecting real-time correlation, λ is the forgetting factor, controlling the retention proportion of historical weight, is the covariance between feature i and the instantaneous fluctuation rate Δc (t) of cost, is the variance of feature i at time t;

[0093] Δc (t) is the instantaneous fluctuation rate of cost, defined as follows:

[0094]

[0095] In the formula, cost (t) is the actual cost value at time t, The average value of the cost in the previous time period.

[0096] The step S4 includes the following steps:

[0097] Step S41, capture the hidden association pattern between features, construct a time sequence association graph based on the feature matrix of step S2, and the edge weight calculation formula is as follows:

[0098]

[0099] In the formula, is the Pearson correlation coefficient of features i and k in the time window T, which is used to quantify the time sequence association strength of the two, T is the length of the sliding time window, which is used to limit the time range of the association analysis, is the mean value of feature i in the window;

[0100] Step S42, quantify the influence of the associated features on the target feature, calculate the feature propagation effect based on the association graph of step S41, and the propagation coefficient formula is as follows:

[0101]

[0102] In the formula, γ k is the normalized influence coefficient of feature k, reflecting its global importance in the association network, PageRank (G) υk is the PageRank value of feature k in graph G, which measures the node importance of feature k, G = V, E is the association graph, wherein the node V is the feature, and the weight of the edge E is defined by

[0103] The step S5 includes the following steps:

[0104] Step S51, comprehensively quantify the real-time abnormal risk, integrate the standardized features of step S2, the dynamic weight of step S3, and the association effect of step S4, generate an abnormal score, and the formula is as follows:

[0105]

[0106] In the formula, S t is the real-time abnormal score at time t, the larger the value, the higher the abnormal risk, is the Z-score standardized deviation value, which measures the instantaneous abnormal degree of feature i, is the time decay factor, θ = 0.1 is the control decay speed, Δt i is the time interval of the last triggered abnormality of feature i, is the weighted contribution of the associated features, m is the number of associated features, β is the association effect weight, which controls the proportion of the associated items in the score;

[0107] ​Step S52, dynamically adapt to feature distribution changes, update the baseline value according to the real-time data of step S5, and the baseline update formula is as follows:

[0108]

[0109] In the formula, is the dynamic mean value of feature i, which is updated by exponential smoothing, is the dynamic standard deviation of feature i, which is updated by exponential smoothing.

[0110] Step S6 includes the following steps:

[0111] Step S61, based on the anomaly score S of step S5 t , trigger the hierarchical response mechanism, and the threshold is set as follows:

[0112] Yellow warning (S t ∈[1.5,2.5]): generate observation records and store in the database of step S1 for subsequent analysis;

[0113] Orange warning (S t ∈[2.5,4.0]): freeze the current prescription for 30 minutes through the HIS interface of step S1, and block the abnormal settlement process;

[0114] Red alert (S t ≥4.0): suspend medical insurance settlement and generate a review work order, and associate feature data of step S2 for manual review;

[0115] Step S62, for quick intervention, freeze settlement through API, generate electronic audit work order, automatically associate patient historical visit records, similar case comparison data and push warning details to the responsible physician through message middleware, and require to submit situation explanation within 2 hours.

[0116] Step S7 includes the following steps:

[0117] Step S71, based on the warning record of step S6 and the historical data of step S1, correct the baseline parameters daily If If the dynamic mean value deviates from the baseline by more than 2 times the standard deviation, adjust the forgetting factor in step S3 to λ=0.7 to accelerate weight update;

[0118] In the formula, is the baseline mean value, which is calculated based on the historical 7-day data;

[0119] Step S72, based on the weight update record of step S3 and the correlation analysis result of step S4, generate a feature importance report every week for the reviewer to adjust the feature whitelist of step S2.

[0120] One specific application of this embodiment is:

[0121] Step one,

[0122] Implementation:

[0123] Real-time access to medical insurance settlement data streams through medical institution HIS (hospital information system) interface, including diagnosis and treatment item code (such as "CT examination: W0201"), drug code (such as "Atorvastatin: Y0035"), quantity, unit price, patient basic information (age, gender, medical insurance card number) and diagnosis code; for example, a patient is diagnosed with "hypertension (ICD-10 code I10)", the system automatically converts the unstructured text "hypertension" into standardized ICD-10 code;

[0124] Data cleaning operation:

[0125] Delete duplicate records: such as duplicate settlement records of the same patient at the same time point;

[0126] Complete missing fields: if the drug unit price is missing, automatically complete it from the drug directory library;

[0127] Correct abnormal values: when the unit price of a certain drug exceeds the preset threshold (such as unit price > 500 yuan), trigger the manual review process; for example, the unit price of a certain antibiotic is recorded as 800 yuan, the system freezes the record and notifies the administrator to check;

[0128] Step two,

[0129] Feature definition:

[0130] Individual dimension: patient age segmentation (such as "<18 years old", "18-60 years old", ">60 years old"), medical history (such as diabetes, cardiovascular disease);

[0131] Diagnosis and treatment dimension: number of single visit items (such as one visit containing 5 items), high-priced drug proportion (such as high-priced drug cost accounting for 30% of total drug cost), examination item clustering degree (such as 3 similar examinations issued by the same department);

[0132] Time sequence dimension: same department visit frequency week-on-week change rate (such as this week's visit frequency increasing by 50% compared with last week), same type of drug use amount 3-day moving average (such as the average daily use amount of a certain antihypertensive drug in the past 3 days is 10 boxes), same diagnosis item price fluctuation standard deviation (such as CT examination price standard deviation is 50 yuan);

[0133] Dynamic baseline standardization:

[0134] The formula is as follows:

[0135]

[0136] For example, if a patient has 8 examination items on the day, the average number of examination items in the same department in the past 24 hours is 5, and the standard deviation in the past 7 days is 2, then the standardized eigenvalue is:

[0137]

[0138] Step 3:

[0139] Initial weight distribution:

[0140] Weights are assigned based on the Pearson correlation coefficient between features and costs in historical data. For example, if the correlation coefficient between "proportion of high-priced drugs" and total costs is 0.8, and the correlation coefficient between "number of inspection items" is 0.5, the initial weights are:

[0141]

[0142] Dynamic weight update:

[0143] Assume that the characteristic “proportion of high-priced drugs” at the current time t is related to the cost volatility Δc( t) The covariance of is 0.6, the variance is 0.4, and the forgetting factor λ is 0.8, then the dynamic weight update is:

[0144]

[0145] Step 4:

[0146] Timing association diagram construction:

[0147] Calculate the Pearson correlation coefficient between features. For example, the correlation coefficient between "proportion of high-priced drugs" and "number of inspection items" in the past 24 hours It means that the two are highly positively correlated;

[0148] Calculation of characteristic propagation effect:

[0149] The PageRank algorithm is used to evaluate the global influence of features. If the PageRank value of “proportion of high-priced drugs” is 0.9 (the highest is 1), then its normalized influence coefficient γ k =0.9;

[0150] Step 5:

[0151] Scoring formula application:

[0152] Assume that the standardized characteristic deviation value of a visit Weighted contribution of association effect ∑γ k R i,k =1.2, time decay factor e -0.1×2 =0.82, then the anomaly score is:

[0153] St = 0.53 x (2.5 x 0.82 + 0.3 x 1.2) = 1.45;

[0154] Dynamic baseline update:

[0155] If the current feature mean New observation The updated mean:

[0156]

[0157] Assume the dynamic mean at the previous time Dynamic standard deviation New observation Updated dynamic mean Substitute into the formula:

[0158]

[0159] Step six,

[0160] Yellow alert (S t = 1.5): The system automatically generates observation records and stores them in the database for subsequent analysis.

[0161] Orange alert (S t = 3): Freeze the current prescription for 30 minutes through the HIS interface and block the settlement process; for example, a prescription is frozen due to high proportion of high-priced drugs, which needs to be confirmed by the physician again.

[0162] Red alert (S t = 4.5): Stop medical insurance settlement and generate a review work order, associate historical patient data (such as the last 3 months of medical records) and push to the reviewer, requiring processing within 2 hours.

[0163] Step seven,

[0164] Parameter correction:

[0165] If the dynamic mean Deviation from the baseline mean More than 2 times the standard deviation Adjust the forgetting factor λ = 0.7 to accelerate weight update.

[0166] Feature importance report:

[0167] Generate a report every week, for example, "high-priced drug proportion" weight is increased to 0.7, prompting the reviewer to add it to the feature whitelist for priority monitoring.

[0168] In the description of the specification, reference to "one embodiment", "an example", "a specific example" or the like means that a particular feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. The appearances of the phrases "in one embodiment", "an example", "a specific example" or the like in various places in the specification are not necessarily referring to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0169] The preferred embodiments of the application disclosed above are only to help explain the application. The preferred embodiments do not describe all the details of the application and limit the application to the specific embodiments described. Obviously, many modifications and variations can be made in light of the contents of the specification. The specification selects and specifically describes these embodiments in order to better explain the principles and practical application of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution, characterized by: The following steps are involved: Step S1: Real-time collection of multi-source medical insurance data and standardization cleaning, including diagnosis and treatment items, drug use, patient information and diagnosis codes, to generate standardized data streams; Step S2: Convert the normalized data stream into a multi-dimensional feature vector, which includes patient attributes, diagnosis and treatment behavior, and temporal change characteristics, and eliminate the time dimension deviation through dynamic baseline; Step S3: Based on the generated multi-dimensional feature vector, the dynamic correlation weights between each feature and the changes in medical insurance costs are calculated in real time, and the weights are initialized through the correlation of historical data; Step S4: Combine the feature vectors and dynamic weights to analyze the temporal correlation between features and construct a multi-dimensional correlation network to capture combined abnormal behavior patterns; Step S5: Combining the feature vectors of step 2, the dynamic weights of step 3, and the association network results of step 4, a real-time abnormality score is generated to quantify the risk level of the diagnosis and treatment behavior; Step S6: triggering a graded response mechanism based on the real-time anomaly score, and executing automated intervention measures from observation and recording to mandatory interception according to the risk level; Step S7: Based on the intervention record of step 6 and the historical data of step 1, the feature baseline and weight calculation rules are periodically updated to form a self-optimizing closed-loop control.

2. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S1 includes the following steps: Step S11: Obtain medical insurance settlement data stream in real time through the medical institution HIS interface, including diagnosis and treatment item code, drug code, quantity, unit price, patient basic information and diagnosis code; Step S12: standardize the unstructured diagnosis description text and use an ICD-10 code converter to normalize the diagnosis name; Step S13: Perform data cleaning operations to delete duplicate records, fill in missing fields, and correct abnormal values. Records with unit prices exceeding the preset threshold will trigger a manual review process.

3. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S2 includes the following steps: Step S21: Define the characteristics of individual dimension, diagnosis and treatment dimension, and time series dimension: Individual dimension: patient age group and medical history; Diagnosis and treatment dimensions: single visit item list, proportion of high-priced drugs, and clustering degree of examination and testing items; Time series dimension: Week-on-week change rate of frequency of visits to the same department, 3-day moving average of the dosage of similar drugs, standard deviation of price fluctuations of the same diagnostic items; Step S22: Eliminate dimensional differences and enhance feature comparability by performing dynamic baseline standardization on the original data. The formula is as follows: Where, is the value of the standardized feature i at time t, is the unprocessed value of feature i at time t in the original data, is the mean value of feature i in the past t hours, used for dynamic baseline calibration, is the standard deviation of feature i in the past 7 days, used for standardization and scaling, and 3 times the standard deviation is used to eliminate the impact of extreme values.

4. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S3 includes the following steps: Step S31: Assign initial weights based on the correlation between features and costs in historical data, giving priority to important features in the calculation. The formula is as follows: Where, is the initial weight of feature i, assigned based on the correlation between historical data and costs, is the Pearson correlation coefficient between feature i and total cost, calculated based on historical data; Step S32: Use the latest cost data collected in step S1 to update the weight dynamics, capture the real-time correlation between features and cost fluctuations, and adapt to changes in the network environment. The dynamic weight update formula is as follows: Where, is the dynamic weight of feature i at time t, reflecting the real-time relevance, λ is the forgetting factor, controlling the retention ratio of historical weights, is the characteristic i and the instantaneous volatility of the fee Δc (t) The covariance of is the variance of feature i at time t; Δc (t) is the instantaneous volatility of fees, defined as follows: Where cost (t) is the actual cost value at time t, The average cost in the previous time period.

5. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S4 includes the following steps: Step S41: Capture the hidden correlation pattern between features and construct a temporal correlation graph based on the feature matrix of step S2. The edge weight calculation formula is as follows: Where, is the Pearson correlation coefficient between features i and k in the time window T, which is used to quantify the temporal correlation strength between the two. T is the length of the sliding time window, which is used to limit the time range of the correlation analysis. is the mean of feature i in the window; Step S42: quantify the influence of the associated features on the target features, and calculate the feature propagation effect based on the association graph of step S41. The propagation coefficient formula is as follows: Where, γ k is the normalized influence coefficient of feature k, reflecting its global importance in the association network, PageRank(G) υk is the PageRank value of feature k in graph G, which measures its node importance. G = V, E is the association graph, where node V is the feature and the weight of edge E is given by definition.

6. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S5 includes the following steps: Step S51: Comprehensively quantify the real-time abnormal risk by integrating the standardized features of step S2, the dynamic weights of step S3, and the correlation effects of step S4 to generate an abnormality score. The formula is as follows: Where S t is the real-time anomaly score at time t. A larger value indicates a higher anomaly risk. is the Z-score standardized deviation value, which measures the instantaneous abnormality of feature i, e - θΔt i is the time attenuation factor, θ=0.1 is the control attenuation speed, Δt i is the time interval since feature i last triggered an exception. is the weighted contribution of the associated features, m is the number of associated features, and β is the association effect weight, which controls the proportion of the associated items in the score; Step S52: Dynamically adapt to changes in feature distribution and update the benchmark value based on the real-time data from step S5. The benchmark update formula is as follows: Where, is the dynamic mean of feature i, updated by exponential smoothing, is the dynamic standard deviation of feature i, updated by exponential smoothing.

7. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S6 includes the following steps: Step S61: Abnormal score S based on step S5 t , triggering the hierarchical response mechanism, the thresholds are set as follows: Yellow Warning (S t ∈[1.5,2.5]): Generate observation records and store them in the database of step S1 for subsequent analysis; Orange Alert (S t ∈[2.5,4.0]): Freeze the current prescription for 30 minutes through the HIS interface in step S1 to block the abnormal settlement process; Red Alert (S t ≥4.0): suspend medical insurance settlement and generate a review work order, and associate the feature data in step S2 for manual review; Step S62: To quickly implement intervention, freeze settlement through the API, generate an electronic review work order, automatically associate the patient's historical medical records, comparative data of similar cases, and push warning details to the responsible physician through the message middleware, requiring submission of a situation report within 2 hours.

8. A method for accurately monitoring and intervening in medical insurance expenses of a medical institution according to claim 1, characterized in that: The step S7 includes the following steps: Step S71: Based on the warning record of step S6 and the historical data of step S1, the benchmark parameters are calibrated daily. like If the dynamic mean deviates from the benchmark by more than 2 standard deviations, the forgetting factor in step S3 is adjusted to λ = 0.7 to accelerate the weight update; Where, It is the benchmark average, calculated based on the historical 7-day data; Step S72: Based on the weight update record of step S3 and the association analysis result of step S4, a feature importance report is generated every week for the reviewer to adjust the feature whitelist of step S2.

Images