Medical insurance anomaly detection method based on multi-dimensional anomaly indexes

Abstract

The invention discloses a medical insurance anomaly detection method based on a multi-dimensional anomaly index, and the method comprises the following steps: S1, obtaining desensitized medical insurance structured transaction data; S2, carrying out the processing of the desensitized medical insurance structured transaction data, obtaining a doctor-seeing frequency abnormal index, a doctor-seeing cost abnormal index and a doctor-seeing behavior abnormal index, and constructing an abnormal index vector R (ui) based on the abnormal indexes; and S3, constructing an abnormal index vector R (ui) for the abnormal index, and carrying out calculation processing to obtain a medical insurance abnormal detection result. According to the method, the multi-dimensional anomaly index system is constructed, various behavior characteristics are quantitatively scored, and the anomaly degree of the patient in each dimension is systematically measured, so that the detection result has higher discrimination ability, the anomaly source and risk degree can be visually presented through quantitative scoring, and the interpretability of the model output result is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of big data, and particularly relates to a medical insurance abnormality detection method based on multi-dimensional abnormality indexes. BACKGROUND

[0002] As an important part of legal supervision, medical insurance fund supervision plays a key role in ensuring fund safety, regulating fund use, maintaining medical insurance order and protecting citizens' legal rights and interests. The problem of medical insurance drug reselling is a typical problem in medical insurance fund supervision. Traditional methods often rely on static rule matching or manual verification to detect such behavior. With the rapid accumulation of medical behavior data and the continuous evolution of fraud methods in recent years, medical insurance drug reselling behavior is becoming increasingly complex and hidden. Therefore, it is urgent to carry out systematic and comprehensive modeling and joint analysis of patient abnormal characteristics. In this context, legal supervision models constructed with advanced computing technologies such as data mining and machine learning have gradually become an important means to improve the efficiency of medical insurance fund supervision. These models can mine potential abnormal behavior patterns through comprehensive analysis of medical treatment data, drug purchase data and other information, and assist in the detection and identification of medical insurance drug reselling behavior. Relying on big data analysis and artificial intelligence algorithms, legal supervision models have shown good application prospects in the practice of procuratorial organs and medical insurance management departments.

[0003] Early methods mainly rely on expert experience to define association rules and threshold rules. Viveros et al. obtained association rules through the Apriori algorithm and used them for medical insurance supervision clue mining. Aral et al. designed a feature association matrix (such as drug and diagnosis, drug and patient age) and a corresponding risk score calculation function, and performed supervision clue mining through threshold filtering. Zhou Jiehui et al. found abnormal patient medical behavior by associating patient medical behavior characteristics. The Chinese patent with publication number CN109615547A identifies abnormal drug purchasing behavior by judging the correlation between medical records and purchased drugs. The Chinese patent with publication number CN112289403A first uses an isolation forest model to screen patients based on annual medical visit frequency, annual medical insurance reimbursement amount, and annual number of drug types, and then further analyzes the abnormality of different high-frequency purchased drugs of the obtained abnormal patients based on monthly drug frequency and monthly drug type. The Chinese patent with publication number CN110781222A uses an isolation forest to detect abnormal drug use behavior and a box plot to detect abnormal spending behavior, and combines the two to realize drug reselling behavior clue mining. However, fixed thresholds and manual weighting are difficult to adapt to the dynamic changes of fraud patterns, with high false positive and false negative rates, and single abnormal rules or multiple simple rule combinations are difficult to cope with complex fraud behavior scenarios in reality.

[0004] In recent years, with the development of deep learning, it has shown strong pattern mining ability in medical insurance supervision clue mining. Yi Dongyi et al. extracted patient-doctor graph structure information based on graph convolution network, and effectively learned the fraud behavior distribution from the scarce fraud behavior samples through variational autoencoder (VAE); Lu et al. constructed a patient-hospital-drug relationship graph, and obtained patient features by weighting and aggregating neighbor features through hierarchical attention mechanism for supervision clue mining; Hong et al. constructed a heterogeneous graph based on medical insurance data and used graph structure learning to extract topological structure, features and semantic information, and detected fraud behavior through multi-channel feature fusion; The Chinese patent with publication number CN113657548A normalizes numerical data such as medical insurance data costs and hospitalization days, and clusters and labels type data such as drug use records and laboratory records to extract feature vectors and input them into a neural network for classification. However, the deep learning-based medical insurance supervision clue mining method generally has the problem of insufficient model explainability, and its black box characteristics limit its practicality in medical insurance fund supervision and other application scenarios that highly depend on decision transparency and traceability.

[0005] In general, previous methods faced two challenges: First, the existing reselling of medical insurance drugs is becoming increasingly covert, and abnormal characteristics are often concealed by circumventing rule thresholds, controlling the frequency of medical treatment and the distribution of institutions, making it difficult for single-dimensional rules to be effective. This is because they lack a multi-dimensional, systematic, scenario-specific and explainable anomaly analysis mechanism for the reselling of medical insurance drugs, making it difficult to comprehensively characterize the potential risk characteristics of abnormal medical insurance cards in multiple behavioral dimensions; second, the existing indicator fusion methods mostly use simple weighting or rule logic combination, which fails to fully consider the importance and discrimination of different abnormal characteristics, resulting in the fused indicator scores being difficult to accurately measure the degree of abnormality of medical insurance cards.

[0006] In response to the problems existing in previous methods, this technical solution designs a multi-dimensional comprehensive anomaly calculation system, and performs multi-level and multi-angle abnormal behavior indicator score calculations under each abnormal dimension, so as to comprehensively characterize the multi-dimensional abnormal performance of the medical insurance card and realize the quantitative definition of the abnormal degree of the medical insurance card; at the same time, the entropy weight method and TOPSIS algorithm are used to fuse the multi-dimensional abnormal feature scores, realize the reasonable and effective weighted aggregation of multiple abnormal behavior indicator scores, and then realize the identification and discovery of suspected drug reselling behavior. At the same time, the fused behavior indicator scores can present the abnormal degree of the medical insurance card in an intuitive and reliable manner, and better assist prosecutors in their prosecution work.

[0007] Therefore, technicians in this field are committed to developing a medical insurance anomaly detection method based on multi-dimensional anomaly indicators. Summary of the Invention

[0008] In view of the above-mentioned defects of the prior art, the technical problems to be solved by the present invention are as follows:

[0009] 1) To address the increasingly complex and insidious patterns of drug resale involving medical insurance, there is a lack of a multi-dimensional, systematic anomaly analysis mechanism. Specifically, most studies focus on detecting single fraud scenarios or indicators, lacking a unified mechanism for integrating multiple types of anomaly signals. Criminals circumvent these single-behavior rule-based detection methods by switching medical institutions, circumventing rule thresholds, and manipulating visit distribution.

[0010] 2) Existing methods ignore the differences and importance of various indicators in anomaly detection when integrating abnormal indicators, making it difficult to accurately reflect the overall abnormality of the patient, which can easily lead to false positives and missed reports. In particular, the recognition accuracy is insufficient when facing fraud patterns with complex behaviors and strong concealment.

[0011] To achieve the above objectives, the present invention provides a medical insurance anomaly detection method based on multi-dimensional anomaly indicators, comprising the following steps:

[0012] S1: Obtain desensitized medical insurance structured transaction data

[0013]

[0014] wherein, is a certain medical record of a patient u i , t j is the timestamp corresponding to the jth medical record, h j is the corresponding desensitized medical institution code, c j is the medical insurance settlement fee amount of the transaction, med j is the code of the transaction item, JYLX j is the type of visit, which includes outpatient and emergency, FYLB j is the fee category, which includes drug purchase fee, examination fee and laboratory fee;

[0015] S2: Process the desensitized medical insurance structured transaction data to obtain visit frequency abnormality indicators visit fee abnormality indicators and visit behavior abnormality indicators wherein k is the corresponding kth visit frequency abnormality indicator, m is the corresponding mth visit fee abnormality indicator, and n is the corresponding nth visit behavior abnormality indicator,

[0016] construct an abnormality indicator vector R(u i ) based on the abnormality indicators,

[0017] R(u i ) = [R freq (u i ), R cost (u i ), R beh (u i )]

[0018] wherein is the visit frequency abnormality indicator vector, is the fee abnormality indicator vector, is the behavior abnormality indicator vector;

[0019] S3: Calculate and process the abnormality indicator vector R(u i ) to obtain a medical insurance abnormality detection result.

[0020] Further, the desensitized medical insurance structured transaction data is processed to obtain visit frequency abnormality indicators wherein k = 1-5, and the processing judgment condition is:

[0021] 1: The monthly outpatient visits total more than 15 times;

[0022] 2: The number of visits to the emergency department per month is more than 20;

[0023] 3: The outpatient clinic occurs more than 4 times a day and accumulates for more than 3 days;

[0024] 4. Monthly outpatient visits to more than 3 hospitals on a single day and for more than 3 consecutive days;

[0025] 5. The number of outpatient visits per year is more than 100;

[0026] Statistics on whether the patient meets the set conditions under each judgment condition, recorded as Where k is the corresponding k-th abnormal indicator of medical frequency.

[0027] Furthermore, the desensitized medical insurance structured transaction data Processing to obtain abnormal indicators of medical expenses Where m=1-3, the judgment condition for processing is:

[0028] 1: Monthly outpatient expenses total more than RMB 5,000;

[0029] 2: The cumulative outpatient expenses within the year are more than 20,000 yuan;

[0030] 3: The annual outpatient and emergency expenses are more than 25,000 yuan;

[0031] Count whether the patient meets the set conditions under each judgment condition and record it as Where m is the corresponding mth abnormal medical expense indicator.

[0032] Furthermore, the desensitized medical insurance structured transaction data Processing to obtain abnormal behavior indicators Where n=1-4, the judgment conditions for processing are:

[0033] 1: Settle the same drug at three or more medical institutions within one week;

[0034] 2: Visit more than 3 medical institutions within 1 week and settle more than 10 types of drugs;

[0035] 3. The total annual medical insurance settlement amount is greater than 30,000 yuan, and the drug settlement amount accounts for more than 80% of the total medical insurance settlement amount, and the sum of the examination and laboratory test fees accounts for less than 10% of the total medical insurance settlement amount;

[0036] 4: In July of the past two years, the medical insurance settlement amount of a 60-year-old male or a 55-year-old female is greater than 1680 yuan, and the medical insurance settlement amount of other personnel is greater than 700 yuan;

[0037] Statistical patients under each condition whether to achieve set conditions, recorded as abnormal indicators Wherein n is the corresponding nth visit behavior abnormal index.

[0038] Further, the algorithm of the calculation processing in step S3 is a clustering algorithm.

[0039] Further, the step S3 comprises:

[0040] S301: Normalizing the abnormal indicator vector R(u i );

[0041] S302: Distributing the objective weight using the entropy weight method;

[0042] S303: Using the TOPSIS algorithm to process to obtain the relative abnormal degree of the patient;

[0043] S304: Using a clustering method to calculate the threshold of the relative abnormal degree of the patient, and the patient with a relative abnormal degree higher than the threshold is recorded as a medical insurance abnormal patient.

[0044] Further, the step S301 comprises the following steps:

[0045] S3011: Constructing an overall abnormal indicator matrix with the abnormal indicator vector R(u i ) = [R freq (u i ), R cost (u i ), R beh (u i )]. Wherein n represents n patients,

[0046] S3012: Normalizing the matrix X using a range normalization form, and normalizing the jth abnormal indicator column to: Obtain the normalized abnormal indicator matrix X' = [x' ij ] ∈ [0, 1] n×12 .

[0047] Further, the step S302 comprises the following steps:

[0048] S3021: Calculating the information entropy of the jth abnormal indicator dimension

[0049]

[0050] Wherein,

[0051] S3022: Calculate the weight of the jth index dimension

[0052]

[0053] S3023: Finally, the weight vector w = [w1, w2, …, w 12 ] T , satisfying

[0054] Further, the step S303 specifically includes:

[0055] The positive ideal solution is defined as:

[0056]

[0057] The negative ideal solution is defined as:

[0058]

[0059] The Euclidean distance of the final abnormal score of the patient to the positive and negative ideal solutions is calculated in combination with the weight vector w:

[0060]

[0061] The relative abnormality degree of the patient

[0062] Further, the clustering algorithm in the step S304 is the Kmeans algorithm, and the cluster center of the Kmeans algorithm is 2.

[0063] In a second aspect, the present application also provides a computer device, comprising:

[0064] a memory;

[0065] one or more processors coupled to the memory;

[0066] one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the one or more processors, and the one or more application programs are configured to execute the medical insurance abnormality detection method described above.

[0067] The present application can achieve the following technical effects by providing the above method and computer device:

[0068] 1) The present invention combines the characteristics of typical cases in the actual case handling process with the knowledge of experts in the field, and based on the systematic statistical analysis and behavioral pattern mining of real medical insurance treatment data, it constructs an abnormal indicator system covering multiple dimensions such as frequency of medical treatment, medical expenses and medical behavior. The system systematically depicts the overall picture of the patient's medical behavior from the aspects of the frequency distribution of medical behavior, cost level, drug purchasing behavior pattern and medical institution selection pattern. Multi-dimensional joint modeling can effectively break through the interference caused by patients who resell drugs by circumventing a single rule threshold or controlling the distribution of medical treatment, making it difficult to conceal their abnormal behavior in the overall behavioral portrait, providing a more comprehensive and robust quantitative evaluation basis for abnormal card detection. This solution constructs a systematic abnormal indicator system based on the three core dimensions of medical frequency, medical expenses and medical behavior. For each core dimension, we further designed multi-angle abnormality indicators, and refined the quantitative modeling from different analytical perspectives, thereby achieving a structured characterization of patient behavior and the joint identification of abnormal signals: the abnormal indicator of medical frequency reveals abnormally intensive or high-frequency medical behavior patterns from different time scales; the abnormal indicator of medical expenses identifies unreasonable medical consumption from the cost dimension; and the abnormal indicator of medical behavior examines atypical medical behavior patterns from dimensions such as the distribution of medical institutions and the type of medicines purchased. Through the systematic design and quantitative definition of multi-dimensional abnormal indicators, the present invention enhances the coverage and precision of abnormality identification, and overcomes the limitations of the traditional method of "single indicator, single event" analysis. By comparing the method based on association rules, the present invention verifies the effectiveness and pertinence of the method in capturing abnormal behavior patterns. By constructing a multi-dimensional abnormality indicator system and quantitatively scoring various behavioral characteristics, the degree of abnormality of patients in each dimension is systematically measured, so that the detection results not only have stronger discrimination capabilities, but also can intuitively present the source of abnormality and risk level through quantitative scoring, thereby enhancing the interpretability of the model output results.

[0069] 2) The present invention objectively weights each abnormal indicator based on the entropy weight method, and combines the TOPSIS algorithm to avoid the interference of extreme values, fully considering the information entropy of different abnormal signals, and helping to improve the accuracy of fraud identification. The importance of different abnormal indicators needs to be measured objectively and discriminatively, especially for the multi-dimensional abnormal indicators proposed in the present invention. The entropy weight method introduced in this solution can dynamically adjust the weights according to the differences of each indicator in the real data, and more objectively reflect the importance of different abnormal indicators to the overall judgment results. At the same time, the TOPSIS algorithm avoids the interference of extreme values ​​on the scoring results on the basis of retaining multi-indicator information, making the final abnormal ranking result more stable and more accurate. By comparing the effects of different indicator fusion methods, the effectiveness of the verification method in indicator fusion is verified, and abnormal behavior and normal behavior can be effectively distinguished.

[0070] The concept, specific structure and generated technical effects of the present application will be further described below in combination with the drawings, so as to fully understand the purposes, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 is a medical insurance abnormality detection method flowchart based on multi-dimensional abnormality indicators of a preferred embodiment of the present application;

[0072] Figure 2 is a medical insurance abnormality detection method flowchart based on multi-dimensional abnormality indicators of another preferred embodiment of the present application. DETAILED DESCRIPTION

[0073] The following describes a plurality of preferred embodiments of the present application with reference to the drawings of the specification, so as to make the technical content of the present application clearer and easier to understand. The present application can be embodied in many different forms of embodiments, and the protection scope of the present application is not limited to the embodiments mentioned in the text.

[0074] In the drawings, the same components have the same reference numerals, and components with similar structures or functions have similar reference numerals.

[0075] To solve the problem in the detection of medical insurance medicine reselling and insurance fraud behavior, two targets are summarized and refined: multi-dimensional abnormality indicator design and objective score weighting method design. Correspondingly, the present scheme innovatively designs for the two targets, and finally ideal effects are obtained. The first embodiment is a multi-dimensional abnormality indicator system, through which the patient is calculated for multi-dimensional abnormality behavior, and the extracted abnormality indicator vector can be clustered to realize the abnormality prediction of the patient. The comprehensiveness and effectiveness of the indicator system are verified through experiments. In order to more intuitively show the abnormality degree of the patient and assist the prosecutor in the prosecution work, the second embodiment, i.e. the objective score weighting method design, is designed. The fused score is used as the abnormality degree of the patient for display. Specifically, the score fusion method combining entropy weight method and TOPSIS is established based on the designed abnormality indicator system.

[0076] Embodiment 1:

[0077] As Figure 1As shown, the present embodiment combines the typical case characteristics and field expert knowledge in the actual case handling process, and based on the systematic statistical analysis and behavior pattern mining of real medical insurance medical data, a medical insurance abnormality detection method based on multi-dimensional abnormality index is proposed, which is used to identify the suspected drug reselling behavior of patients. The specific abnormal index subset is designed from three dimensions of abnormality of medical frequency, medical cost and medical behavior, and finally the abnormal behavior quantitative index is formed. In terms of medical frequency, the number of medical visits of medical insurance card holders within a certain period of time is analyzed and a threshold is set. Frequent medical visits beyond the normal range may indicate abnormal drug purchase behavior; in terms of medical cost, the cost threshold is set by evaluating the single medical cost and the cumulative medical cost. Abnormally high cost may indicate abnormal drug purchase behavior; in terms of medical behavior, the rationality of medical behavior is further examined, including frequent medical visits across regions, multiple medical visits in different hospitals within a short period of time, and high proportion of drug settlement amount.

[0078] The present embodiment is based on the analysis of desensitized medical insurance structured transaction data, including desensitized patient card number, medical date, medical type, desensitized medical institution code, item code, medical insurance settlement cost, and cost category information. The patient u i has a medical record :

[0079]

[0080] Wherein, t j is the timestamp corresponding to the jth medical record, h j is the corresponding desensitized medical institution code, med j is the code of the transaction item, c j is the medical insurance settlement cost of the transaction, JYLX j is the medical type (outpatient or emergency), FYLB j is the cost category (drug purchase, examination, laboratory test, etc.).

[0081] Medical frequency abnormality:

[0082] This dimension analyzes whether the patient has excessive frequent medical behavior, excessive scattered medical institutions and other atypical characteristics from the frequency point of view.

[0083] Indicator 1 Cumulative monthly outpatient visits 15 times or more Indicator 2 Cumulative monthly outpatient and emergency visits 20 times or more Indicator 3 Monthly outpatient visits with more than 4 visits on a single day and cumulative for more than 3 days Indicator 4 Monthly outpatient visits with more than 3 visits at a single hospital on a single day and cumulative for more than 3 days Indicator 5 Cumulative annual outpatient visits 100 times or more

[0084] The number of months / years of patients reaching the set conditions under each index is counted and recorded as Where k is the corresponding kth abnormal index, k = 1-5.

[0085] Medical cost abnormality

[0086] This dimension analyzes whether the insured person's cost expenditure in a certain time window is far beyond the norm from the cost perspective.

[0087] Indicator 1 Cumulative monthly outpatient expenses 5000 yuan or more Indicator 2 Cumulative annual outpatient expenses 20000 yuan or more Indicator 3 Cumulative annual outpatient and emergency expenses 25000 yuan or more

[0088] The number of months / years of patients reaching the set conditions under each indicator is counted and recorded as Where m is the corresponding mth abnormal indicator, m = 1-3.

[0089] Abnormal clinic behavior

[0090] This dimension analyzes whether the patient's clinic behavior has atypical characteristics from the behavior perspective, including high drug purchase proportion, too dispersed drug purchase behavior, etc.

[0091]

[0092]

[0093] The number of months / years of patients reaching the set conditions under each indicator is counted and recorded as Where n is the corresponding nth abnormal indicator, n = 1-4.

[0094] Based on the above indicators, each patient u i The following abnormal indicator vector is constructed:

[0095] R(u i ) = [R freq (u i ), R cost (u i ), R beh (u i )]

[0096] Where is the clinic frequency abnormal indicator vector, is the clinic cost abnormal indicator vector, is the clinic behavior abnormal indicator vector.

[0097] Finally, the abnormal indicator vector R(u i ) is calculated and processed to obtain the medical insurance abnormality detection result. In this embodiment, the algorithm for calculation and processing is a clustering algorithm.

[0098] Performance comparison

[0099] Table 1 Comparison of different abnormal medical insurance card detection methods on real medical insurance data set under different training: test division

[0100]

[0101] In the experimental comparison part, the commonly used abnormal medical card detection methods are selected for comparison, and the performance evaluation is carried out according to different training set:test set division methods (8:2, 7:3, 6:4), and the results are shown in table 1. Relative to the method based on Apriori algorithm for mining association rules, the index system proposed in the scheme can efficiently and accurately detect abnormal medical cards, which shows that the method proposed in the scheme can show strong scene adaptability and pertinence in dealing with the illegal behavior of medical insurance drug reselling which has concealment and regularity.

[0102] Embodiment 2:

[0103] As Figure 2 shown, this embodiment is based on the first embodiment, introduces entropy weight method for abnormal score fusion, combines TOPSIS algorithm to avoid the interference of extreme value, and is used for constructing the overall abnormal behavior score index of patients, and further confirms whether it is a suspect of reselling and cheating insurance. The difference between the first embodiment and this embodiment is the processing of the output results of the index system. In this embodiment, the different index results are fused in the form of fusion score to more directly indicate the final abnormal degree of the patient.

[0104] Data preprocessing:

[0105] Set the total sample as n patients, and the abnormal index vector extracted from each patient is

[0106] R(u i )=[R freq (u i ),R cost (u i ),R beh (u i )]

[0107] Then the overall abnormal index matrix is defined as:

[0108]

[0109] In order to avoid the influence of the dimensionless of each index on the score, the matrix needs to be normalized. The range normalization form is adopted, and the j-dimensional abnormal index column is normalized as:

[0110]

[0111] After normalization, the abnormal index matrix X'=[x' ij ]∈[0,1] n×12 .

[0112] Abnormal score fusion based on entropy weight method and TOPSIS:

[0113] In the weighted aggregation of each abnormal behavior indicator score, to avoid the influence of subjective preference, the entropy weight method is used to allocate the objective weight. The entropy weight method is an objective weighting method used to determine the weight of multiple indicators in comprehensive evaluation. Its basic principle is based on the concept of information entropy. The weight of each indicator is determined by calculating the entropy value. The smaller the entropy value, the stronger the discriminant of the indicator in distinguishing samples, the higher the reference value of the comprehensive evaluation result, and the greater the weight should be given. Conversely, the greater the entropy value, the lower the reference value of the indicator, and the smaller the weight should be given. First, the information entropy Hj of the jth abnormal indicator dimension is calculated j :

[0114]

[0115] The weight wj of the jth indicator dimension is defined as: f

[0116]

[0117] The final weight vector w = [w1, w2, …, w 12 ] T satisfies

[0118] To further improve the sorting discrimination of abnormal identification, the TOPSIS algorithm is used to measure the relative abnormality degree of the patient. By constructing the "positive ideal solution" and "negative ideal solution", and calculating the relative closeness of each score to the positive and negative ideal solutions, the relative abnormality degree is finally obtained. The positive ideal solution is defined as:

[0119]

[0120] The negative ideal solution is defined as:

[0121]

[0122] Then, the Euclidean distance between the final abnormal score of the patient and the positive and negative ideal solutions is calculated:

[0123]

[0124] The relative abnormality degree of the patient is defined as:

[0125]

[0126] At this time, the higher the relative abnormality degree, the more likely it is a patient who sells drugs, and the lower the more likely it is a normal use pattern.

[0127] ​Finally, the relative abnormality degree threshold is calculated by clustering methods such as Kmeans, that is, the cluster centers of different clusters are calculated by clustering methods, and the number of clusters set in the implementation process of the embodiment is 2, which can achieve good clustering effect. The average value of the cluster center is taken as the threshold of the relative abnormality degree, and the patient who is higher than the threshold is recorded as a suspected reseller of medicines.

[0128] Performance comparison

[0129] Table 2 Comparison of each index fusion method on real medical insurance data set under different training-test division

[0130]

[0131] In the experimental comparison part, the commonly used index fusion methods are selected for comparison:

[0132] 1. Simple weighted sum: the scores of all indexes are simply summed up;

[0133] 2. CRITIC: the weight is determined according to the standard deviation of each index (reflecting the contrast) and the correlation between indexes (reflecting the conflict), and the weight of each index is calculated as:

[0134]

[0135] Where r ji is the correlation coefficient of index j and index i.

[0136] 3. Coefficient of Variation Method (COV): the coefficient of variation (CV) is the ratio of the standard deviation to the mean, which is used to measure the dispersion degree of the distribution of index data. For the index j with mean μ j and standard deviation σ j , the weight of the index is calculated as:

[0137]

[0138] 4. PCA weighting method: the eigenvalue decomposition of the covariance matrix of the normalized data matrix is calculated, and the load (eigenvector) e1=(e 11 ,e 12 ,…,e 1m ) of the first principal component reflects the contribution of each index in the “maximum variance direction”, and the weight of each index is calculated as:

[0139]

[0140] According to different training set: test set division mode (8:2, 7:3, 6:4), the performance evaluation results are shown in Table 2, the weighted method proposed in the scheme can not only well consider the importance of each index, but also well distinguish the abnormal drug purchase behavior and the normal drug purchase behavior.

[0141] The preferred embodiments of the present application are described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations without departing from the concept of the present application. Therefore, any technical solutions obtained by logical analysis, reasoning or limited experiments based on the concept of the present application and the prior art within the scope of the present application should be within the protection scope defined by the claims.

Claims

1. A medical insurance anomaly detection method based on multi-dimensional anomaly indicators, characterized by: The following steps are involved: S1: Obtaining desensitized medical insurance structured transaction data in, For patients i A medical record of j is the timestamp corresponding to the j-th medical record, h j is the corresponding desensitized medical institution code, c j The amount of medical insurance settlement for this transaction, med j The code of the transaction item, JYLX j The type of medical consultation includes outpatient and emergency, FYLB j is the expense category, which includes medicine purchase expenses, examination expenses and laboratory test expenses; S2: Desensitized medical insurance structured transaction data Processing to obtain abnormal index of frequency of medical consultation Abnormal indicators of medical expenses and abnormal behavior indicators Among them, k is the corresponding k-th abnormal index of medical frequency, m is the corresponding m-th abnormal index of medical expense, and n is the corresponding n-th abnormal index of medical behavior. Based on the abnormal index, an abnormal index vector R(u i ), R(u i )=[R freq (u i ),R cost (u i ),R beh (u i )] in is the abnormal index vector of medical frequency, is the cost anomaly indicator vector, is the behavioral abnormality indicator vector; S3: For the abnormal indicator vector R(u i ) Calculate and process to obtain the medical insurance anomaly detection results.

2. The medical insurance anomaly detection method according to claim 1, wherein: The desensitized medical insurance structured transaction data Processing to obtain abnormal index of frequency of medical consultation Where k = 1-5, the judgment conditions for processing are: 1: The monthly outpatient visits total more than 15 times; 2: The number of visits to the emergency department per month is more than 20; 3: The outpatient clinic occurs more than 4 times a day and accumulates for more than 3 days; 4. Monthly outpatient visits to more than 3 hospitals on a single day and for more than 3 consecutive days; 5. The number of outpatient visits per year is more than 100; Statistics on whether the patient meets the set conditions under each judgment condition, recorded as Where k is the corresponding k-th abnormal indicator of medical frequency.

3. The medical insurance anomaly detection method according to claim 1, wherein: The desensitized medical insurance structured transaction data Processing to obtain abnormal indicators of medical expenses Where m=1-3, the judgment condition for processing is: 1: Monthly outpatient expenses total more than RMB 5,000; 2: The cumulative outpatient expenses within the year are more than 20,000 yuan; 3: The annual outpatient and emergency expenses are more than 25,000 yuan; Count whether the patient meets the set conditions under each judgment condition and record it as Where m is the corresponding mth abnormal medical expense indicator.

4. The medical insurance anomaly detection method according to claim 1, wherein: The desensitized medical insurance structured transaction data Processing to obtain abnormal behavior indicators Where n=1-4, the judgment conditions for processing are: 1: Settle the same drug at three or more medical institutions within one week; 2: Visit more than 3 medical institutions within 1 week and settle more than 10 types of drugs; 3. The total annual medical insurance settlement amount is greater than 30,000 yuan, and the drug settlement amount accounts for more than 80% of the total medical insurance settlement amount, and the sum of the examination and laboratory test fees accounts for less than 10% of the total medical insurance settlement amount; 4. In July of two consecutive years, the medical insurance settlement amount for a 60-year-old male or a 55-year-old female is greater than RMB 1,680, and the medical insurance settlement amount for other individuals is greater than RMB 700; Statistics are collected to determine whether the patient meets the set conditions under each judgment condition and recorded as abnormal indicators. Where n is the corresponding nth abnormal indicator of medical treatment behavior.

5. The medical insurance anomaly detection method according to claim 1, wherein: The algorithm used for the calculation process in step S3 is a clustering algorithm.

6. The medical insurance anomaly detection method according to claim 1, wherein: The step S3 comprises: S301: The abnormality indicator vector R(u i ) for normalization; S302: Allocating objective weights using entropy weight method; S303: Using TOPSIS algorithm to obtain the relative abnormality degree of the patient; S304: A clustering method is used to calculate a threshold value of the relative abnormality level of the patient, and patients whose relative abnormality level is higher than the threshold value are recorded as abnormal medical insurance patients.

7. The medical insurance anomaly detection method according to claim 6, wherein: The step S301 includes the following steps: S3011: Use the abnormal indicator vector R(u i )=[R freq (u i ),R cost (u i ),R beh (u i )]Build the overall abnormal indicator matrix Where n represents n patients, S3012: normalize the matrix X using the range normalization method, and normalize the j-th dimension abnormality index column to: Get the normalized abnormal indicator matrix X′=[x′ ij ]∈[0,1] n×12 .

8. The medical insurance anomaly detection method according to claim 7, wherein: The step S302 includes the following steps: S3021: Calculate the information entropy of the j-th abnormal indicator dimension in, S3022: Calculate the weight of the j-th indicator dimension S3023: Finally, the weight vector w=[w1,w2,…,w 12 ] T ,satisfy 9. The medical insurance anomaly detection method according to claim 8, wherein: The clustering algorithm in step S304 is the Kmeans algorithm, and the number of cluster centers of the Kmeans algorithm is 2.

10. A computer device, characterized in that: include: Memory; one or more processors coupled to the memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by one or more processors, and the one or more applications are configured to execute the medical insurance anomaly detection method according to any one of claims 1 to 9.

Images