Medical insurance fraud detection method and system based on block chain and artificial intelligence

Abstract

The invention discloses a medical insurance fraud detection method and system based on a block chain and artificial intelligence, and belongs to the technical field of information detection, and the method comprises the steps: carrying out the embedded coding of original data, obtaining a structure embedded vector, and evaluating a structure risk score; extracting a medical insurance object list behavior sequence, calculating a behavior template coding vector, and calculating a behavior deviation score through a deviation score function; constructing a payment path diagram, outputting a payment jump index, and calculating to obtain a consistency scoring index; performing structure alignment detection on the payment path sequence and a behavior template coding vector, and outputting a structure offset score; outputting a final risk score; and carrying out risk grade division according to the interval to which the final risk score belongs, and executing a corresponding preset strategy. According to the method, a potential risk list is identified from a field structure, and abnormal combination and frequency characteristics are identified based on historical behavior modeling and rule offset analysis, so that early discovery of complex and evolutionary fraudulent behaviors is realized.

Description

Technical Field

[0001] The present invention belongs to the field of information detection technology, and in particular to a medical insurance fraud detection method and system based on blockchain and artificial intelligence. Background Art

[0002] Currently, there are two main approaches to fraud detection in the industry: audit systems based on manually configured rules, which statically match field anomalies or catalog discrepancies; and machine learning-based model training methods, which build predictive models from labeled data. These two approaches are effective in addressing regular, single-point fraud, but they face significant limitations when dealing with complex, evolving fraud patterns. Rule-based systems are costly to maintain and lack flexibility, making them inadequate for addressing the continuous evolution of fraudulent behavior. AI models, on the other hand, face challenges such as scarce training samples, uninterpretable results, and a lack of verifiable evidence for behavior. Furthermore, data is fragmented across various modules of the current medical insurance system, particularly payment information flows and invoice structure data, which often reside in different systems. This results in a broken chain of conduct, making it difficult to fully reconstruct fraudulent activity. While some regions have explored the use of blockchain in medical insurance systems for data storage, these applications remain largely static, such as tamper-proofing and on-chain archiving. They have yet to be integrated with intelligent identification systems and lack dynamic analysis and risk assessment mechanisms for on-chain behavior.

[0003] To this end, we propose a medical insurance fraud detection method and system based on blockchain and artificial intelligence to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of how to identify abnormal behavior chains in the existing technology, and to propose a medical insurance fraud detection method and system based on blockchain and artificial intelligence.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] Medical insurance fraud detection methods based on blockchain and artificial intelligence include:

[0007] S1: Input the original data collected by the medical insurance list system;

[0008] Embed the original data to obtain a structural embedding vector, and evaluate the structural embedding vector to obtain a structural risk score;

[0009] S2: extracting a list behavior sequence within a preset time window of the medical insurance subject, extracting a latent state vector at each time point from the list behavior sequence, and calculating a behavior template encoding vector based on the latent state vector;

[0010] Calculate the behavior deviation score by using a deviation scoring function; the deviation scoring function represents the degree of deviation of the current behavior sequence by calculating the cosine similarity between the list behavior sequence and the historical record and the distance between the two vectors;

[0011] S3: Input the on-chain payment operation record, which is consistent with the time window of the structure embedding vector behavior sequence;

[0012] After mapping the on-chain payment operation records into a payment status sequence, a payment path graph is constructed, where nodes of the payment path graph represent payment statuses and edges represent time.

[0013] Output a payment jump index, which is used to measure the degree of state jump in the payment path graph. The payment jump index is obtained by calculating the ratio of abnormal jump times to total jump opportunities. The abnormal jump is obtained by calculating the distance between two payment path graph nodes.

[0014] Calculating a consistency score index, the consistency score index is obtained by fusing a payment jump index and a behavior-payment deviation, the behavior-payment deviation is obtained by calculating the offset between the projection of the payment state sequence and the list behavior sequence in the payment space;

[0015] S4: Performing structural alignment detection on the payment path sequence and the behavior template encoding vector, outputting a structural deviation score; and outputting a final risk score based on the structural deviation score;

[0016] S5: Risk levels are divided according to the interval to which the final risk score belongs, and corresponding preset strategies are executed.

[0017] Preferably, the original data includes diagnosis code, drug code, doctor ID, fee amount, department code, and catalog category.

[0018] Preferably, the structural risk score is calculated by combining the complexity information entropy with the business conflict item; the business conflict item is 1 when there is a conflict combination between the fields, and 0 otherwise.

[0019] Preferably, the behavior template encoding vector is a structural behavior template embedding of the entire list behavior sequence; the attention weight of the behavior template encoding vector is calculated by structural risk scoring.

[0020] Preferably, the on-chain payment operation record includes a timestamp, operation type, source of funds, reimbursement catalog, hospital grade, and settlement code.

[0021] Preferably, a risk adjustment factor is introduced into the structural deviation score to more easily trigger a high score of alignment inconsistency in the case of path anomalies, thereby improving the detection sensitivity.

[0022] Preferably, the risk level classification is based on the following rules:

[0023] Setting a first threshold and a second threshold;

[0024] When the final risk score is lower than the first threshold, it is considered as a low-risk list and automatically released;

[0025] When the final risk score is higher than or equal to the first threshold and lower than the second threshold, it is identified as a medium-risk list and submitted to the manual review system;

[0026] When the final risk score is higher than or equal to the second threshold, it is identified as a high-risk list, and the list is immediately frozen and pushed to the audit system.

[0027] Preferably, after being identified as a high-risk list, the high-risk list is marked as suspected fraud, and the final risk score and behavior-path deviation are transmitted to the medical insurance audit clue system, and a risk record containing the following fields is generated: list number, score value, freezing reason, and freezing time.

[0028] The blockchain and AI-based medical insurance fraud detection system includes:

[0029] A feature extraction module is used to input the original data collected by the medical insurance list system; embed the original data to obtain a structural embedding vector, and evaluate the structural embedding vector to perform a structural risk score;

[0030] A behavior detection module is used to extract a list of behavior sequences within a preset time window of a medical insurance subject, extract a latent state vector at each time point from the list of behavior sequences, and calculate a behavior template encoding vector based on the latent state vector; a behavior deviation score is calculated using a deviation scoring function; the deviation scoring function represents the degree of deviation of the current behavior sequence by calculating the cosine similarity between the list of behavior sequences and the historical records, as well as the distance between the two vectors;

[0031] A payment path extraction module is configured to input on-chain payment operation records, which are consistent with the time window of the structural embedding vector behavior sequence; after mapping the on-chain payment operation records into a payment state sequence, construct a payment path graph, wherein the nodes of the payment path graph are payment states and the edges are time; output a payment jump index, which is used to measure the degree of state jumps in the payment path graph, and is obtained by calculating the ratio of abnormal jumps to total jump opportunities, and the abnormal jumps are obtained by calculating the distance between two payment path graph nodes; and calculate a consistency score index, which is obtained by fusing the payment jump index with the behavior-payment deviation, and the behavior-payment deviation is obtained by calculating the offset between the projection of the payment state sequence and the list behavior sequence in the payment space;

[0032] an alignment verification module, the alignment verification module being used to perform structural alignment detection on the payment path sequence and the behavior template encoding vector, outputting a structural deviation score; and outputting a final risk score based on the structural deviation score;

[0033] A strategy generation module is used to divide the risk levels according to the interval to which the final risk score belongs, and to execute corresponding preset strategies.

[0034] In summary, the technical effects and advantages of the present invention are as follows: the present invention identifies a list of potential risks starting from the field structure, and identifies abnormal combinations and frequency characteristics based on historical behavior modeling and rule deviation analysis, thereby achieving early detection of complex and evolving fraudulent behaviors. At the same time, the present invention further utilizes the path data in the medical insurance payment process as behavioral evidence to establish a correspondence between the list items and the actual payment behavior, dynamically restore the evolution trajectory of the payment path, and based on this, determine whether there are fraudulent deviations such as abnormal payment methods and path jumps, thereby solving the core bottleneck of the inability to verify identification results in existing systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method steps in the present invention;

[0036] Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0038] like Figure 1 As shown in the figure, the medical insurance fraud detection method based on blockchain and artificial intelligence includes:

[0039] S1: Input the original data collected by the medical insurance list system;

[0040] Embed the original data to obtain a structural embedding vector, and evaluate the structural embedding vector to obtain a structural risk score;

[0041] S2: extracting a list behavior sequence within a preset time window of the medical insurance subject, extracting a latent state vector at each time point from the list behavior sequence, and calculating a behavior template encoding vector based on the latent state vector;

[0042] Calculate the behavior deviation score by using a deviation scoring function; the deviation scoring function represents the degree of deviation of the current behavior sequence by calculating the cosine similarity between the list behavior sequence and the historical record and the distance between the two vectors;

[0043] S3: Input the on-chain payment operation record, which is consistent with the time window of the structure embedding vector behavior sequence;

[0044] After mapping the on-chain payment operation records into a payment status sequence, a payment path graph is constructed, where nodes of the payment path graph represent payment statuses and edges represent time.

[0045] Output a payment jump index, which is used to measure the degree of state jump in the payment path graph. The payment jump index is obtained by calculating the ratio of abnormal jump times to total jump opportunities. The abnormal jump is obtained by calculating the distance between two payment path graph nodes.

[0046] Calculating a consistency score index, the consistency score index is obtained by fusing a payment jump index and a behavior-payment deviation, the behavior-payment deviation is obtained by calculating the offset between the projection of the payment state sequence and the list behavior sequence in the payment space;

[0047] S4: Performing structural alignment detection on the payment path sequence and the behavior template encoding vector, outputting a structural deviation score; and outputting a final risk score based on the structural deviation score;

[0048] S5: Risk levels are divided according to the interval to which the final risk score belongs, and corresponding preset strategies are executed.

[0049] The specific steps are as follows:

[0050] Step 1: Structural embedding generation and anomaly feature extraction

[0051] This step aims to identify structural anomalies that may exist in medical insurance settlement invoices, focusing on detecting atypical, unreasonable, or rule-breaking combinations between fields. Examples include "prescribing anticancer drugs in non-oncology departments," "prescribing emergency drugs for chronic disease diagnoses," and "frequently prescribing low-priced and high-priced combinations of drugs." Automatic detection of structural anomalies is achieved by modeling and entropy analysis of invoice field combinations, combined with a conflict knowledge base in the medical insurance field. The output of this step not only provides a structural risk score but also a structural embedding representation for subsequent behavioral sequence analysis.

[0052] Enter the structured field set collected for the medical insurance list system:

[0053] X={x1,x2,...,x n}

[0054] Among them, each x i Represents a standard field in a list, such as diagnosis code, drug code, doctor ID, fee amount, department code, catalog category, etc. Data is output from the hospital HIS system or medical insurance platform using a structured interface and enters the model after ETL processing. All fields are uniformly coded and do not include text or image fields.

[0055] Each medical insurance settlement statement contains multiple semantically related fields, which combine to form a structural relationship. For example, in a statement with the diagnosis code "diabetes," the medication should be a hypoglycemic drug and the service department should be endocrinology-related. If antidepressants or obstetrics and gynecology are included, there may be a structural anomaly.

[0056] In order to capture this type of structural anomaly, we first embed all fields to form a structural representation E S , which is a low-dimensional dense vector obtained by concatenating or mapping all manifest fields after encoding. This embedding does not require pre-training and can be obtained through end-to-end training using a frequency statistics table or random initialization.

[0057] Next, we evaluate how common this structural combination is in historical inventory data. We use structural entropy to measure the complexity of this structural combination:

[0058]

[0059] in:

[0060] H s Indicates the complexity information entropy of the current structural combination;

[0061] P(x i ) is the field x i marginal probability of appearing in the historical compliance list;

[0062] n is the number of structure fields.

[0063] The higher the value, the rarer the structure, representing a greater potential for abnormality.

[0064] However, the structural anomalies of medical insurance fraud are often not just "uncommon", but also conflict with medical insurance rules. For example, a certain type of drug can only be prescribed for a specific diagnosis or department. We introduce the business conflict penalty function C(x i ,x j ), provided by the medical insurance knowledge base, is used to mark field combinations that violate industry standards. We define the structural risk score R s :

[0065]

[0066] in:

[0067] R s is the structural risk score;

[0068] α and β are adjustable coefficients used to balance the importance of entropy terms and rule conflict terms;

[0069] C(x i ,x j )∈{0,1}, if the field x i with x j If it is a conflicting combination, it is 1, otherwise it is 0;

[0070] Sum all possible field pairs (x i ,x j );

[0071] If there is a known conflicting combination between diagnosis and drug, such as "pharyngitis + paclitaxel", the score will be significantly improved.

[0072] For example:

[0073] The list structure is: Diagnosis = Pharyngitis, Drug = Paclitaxel, Cost = 5000 yuan, Department = Otolaryngology;

[0074] This combination has never been seen in history. s Higher;

[0075] At the same time, diagnosis + drug is a conflict combination defined in the medical insurance rules, that is, C(x i ,x j )=1;Finally R s If it is higher, the system will judge it as a structural abnormality list.

[0076] Output:

[0077] R s : Structural risk score, the value range is [0,1], the higher the value, the more abnormal the structure;

[0078] E s : Structural embedding vector, which is the vectorized representation of the current list in the structural space and is used for behavioral sequence analysis.

[0079] Step 2: Structural sequence modeling and behavioral template detection

[0080] Medical insurance fraud often does not occur in isolation, but is carried out in batches within a short period of time through strategies such as "template copying," "structural drift," or "intermittent disguise." For example, the same doctor may repeatedly prescribe medications using lists with highly similar structures but slightly modified to circumvent structural rule verification; or the same patient may be frequently prescribed high-risk structural combination lists within a short period of time, resulting in false reimbursement. To this end, the structural risk score R output in the first step is used in this step. s With the structure embedding vector E s Based on this, the behavior sequence of the same object (such as a doctor or a patient) within a time window is modeled, and the risk trajectory at the behavioral level is captured through behavioral dynamic modeling and template deviation measurement, and the behavior template code E is output. b and behavioral risk score R t .

[0081] enter:

[0082] The input of this step is the output of step 1, that is, the following for each medical insurance settlement list:

[0083] E s : Structural embedding vector, which represents the low-dimensional dense representation of the combination of manifest structure fields;

[0084] R s : Structural risk score, which indicates the rarity and conflict of the structural field combination in history.

[0085] At the same time, from the list management system, we extract the continuous lists within the recent time window (such as all lists within 30 days) by doctor ID or patient ID to form a behavior sequence:

[0086]

[0087] Where T is the length of the behavior sequence, and are the structural embedding and risk score at time point t, respectively.

[0088] To model dynamic structural usage from this behavioral sequence, we designed a risk-guided attention sequence model, whose core is a bidirectional GRU (BiGRU) combined with a structural risk-weighted attention mechanism. The model aims to identify potential behavioral anomalies such as "fake template usage," "high-risk repetition," and "structural drift" based on the evolution of the inventory structure.

[0089] We first process the structure embedding sequence through a bidirectional GRU Get the hidden state vector h at each time point t , represents the local dynamic meaning of the list in the behavioral trajectory. Subsequently, we introduce a structural risk-weighted attention mechanism to improve the model's sensitivity to high-risk structural fragments.

[0090] Behavior template encoding vector E b The calculation of is as follows:

[0091] in

[0092] in:

[0093] h t It is the list behavior representation of GRU output;

[0094] Score the inventory structure risk at that point in time;

[0095] δ is the temperature coefficient, which is used to regulate the impact of risk score on attention weight;

[0096] γ t is the attention weight guided by structural risk;

[0097] E b Embed the structural behavior template for the entire list behavior sequence.

[0098] The innovation of this design is that the attention weights are no longer learned from data, but are directly derived from the structural risk score. Calculation improves the ability to focus on high-risk structural segments, which is very critical in medical insurance fraud scenarios because most fraudulent behaviors are concentrated in behavioral segments with repeated use of risk structures.

[0099] Next, to identify the degree of deviation between this behavior template and historical normal behavior, we designed an innovative deviation scoring function. We extracted multiple doctor / patient behavior template sets ε from the compliance behavior database. norm , used to represent the normal behavior trajectory distribution. We define the behavior deviation score R t :

[0100]

[0101] in:

[0102] cos(E b ,e) represents the cosine similarity between the current behavior template and the historical template;

[0103] ||E b -e||1 is the L1 distance between the two vectors, which also indicates the degree of sparse offset of the template;

[0104] λ is the sparsity regularization weight, which is used to amplify the sparse deviation;

[0105] R t Indicates the degree of deviation between the current behavior and the closest normal template. The larger the value, the more likely it is abnormal behavior.

[0106] This design, unlike simply using cosine similarity, is particularly well-suited for the "low-amplitude, high-frequency" structural drift fraud typically seen in healthcare scenarios. This involves behaviors that appear similar but deviate from compliance guidelines. By introducing L1 distance and a regularization term, the model's ability to detect behavioral boundary deviations is enhanced.

[0107] Output:

[0108] E b : The behavior template of the current list sequence is embedded for subsequent payment path alignment and abnormal jump detection;

[0109] R t : Structural behavior deviation score, which indicates the degree of deviation between the current behavior sequence and the compliance behavior template, ranging from [0,1]. The higher the score, the more abnormal it is.

[0110] Step 3: On-chain reconstruction of behavioral paths and extraction of payment traces

[0111] A prominent feature of medical insurance fraud is the disconnect between behavior and payment paths: that is, the behavior appears to comply with medical insurance review rules, but the actual settlement path shows extreme jumps, such as frequent changes in payment methods, switching catalogs, circumventing limit items, and cross-hospital and cross-regional reporting. These abnormal paths are often not directly reflected in the list data, but are hidden in the interaction logs between the medical insurance payment system and the settlement chain. This step uses the behavior template code E output in step 2. b and behavioral risk score R t As a "behavioral fingerprint", real payment data is called in the medical insurance chain platform or settlement system to extract the payment path sequence P associated with the list behavior. b , and reconstruct the payment path graph G p , used for subsequent path deviation detection.

[0112] This step is the key bridge for the transition from "intelligent behavior recognition" to "credible evidence verification" in the patented system. It not only integrates the results of artificial intelligence behavior modeling, but also directly connects to the payment link infrastructure, reflecting the coupling innovation of AI and blockchain in actual anti-fraud applications.

[0113] The input for this step is:

[0114] E b : The behavior template encoding output by step 2 is a dynamic aggregation expression of the list behavior structure in the time dimension;

[0115] R t : Behavioral risk score, which indicates the degree of deviation between the behavior sequence and the normal behavior template. The higher the value, the more significant the abnormality.

[0116] In addition, it is necessary to obtain on-chain payment operation records consistent with the time window of the behavior sequence from the medical insurance blockchain payment platform, reconciliation system, HIS payment log interface, etc. Each record contains fields such as timestamp, operation type, source of funds, reimbursement catalog, hospital level, settlement code, etc.

[0117] These operation records undergo field standardization and structure mapping to form an ordered payment track sequence:

[0118] P b ={p1,p2,…,p L}

[0119] where p i =(time i ,paytype i ,catalog i ,org i ) represents a state node in the payment path, which includes key fields such as payment time, method, medical insurance catalog, and hospital type, and L is the trajectory length.

[0120] Details of the steps:

[0121] Path data mapping and standardization:

[0122] The structure of the payment link log is highly dependent on the platform deployment, so we designed a set of platform-independent payment status mapping rules. Unify the mapping of fields in different log formats to standard status fields. For example:

[0123] Operation code 1A → "Directory 1";

[0124] Hospital level C → "tertiary hospital";

[0125] Payment method code P5 → "Self-payment + comprehensive payment".

[0126] This process is completed by a preset rule library, without the need for model learning, ensuring cross-platform portability.

[0127] Path diagram structure reconstruction and behavior alignment:

[0128] According to the standardized payment status sequence P b , we construct the payment path graph G with nodes as states and time as edges p =(V,E), where:

[0129] V={p1,p2,...,p L} is the payment status node;

[0130] E={(p i ,p i+1 )} is the time series connection edge, preserving the causal order of the behavior process.

[0131] On this basis, we propose an innovative payment jump indicator C p , which is used to measure the degree of state transition in the payment path graph:

[0132]

[0133] in:

[0134] d(p i ,p i+1 ) is the field distance between two payment status nodes;

[0135] The distance function is defined as the normalized weighted L1 distance of the state vector difference;

[0136] θ is the slight change tolerance threshold set by the system;

[0137] 1[·] is the indicator function, which determines whether it is a “substantial jump”;

[0138] C p The value range is [0,1], indicating the abnormal jump ratio in the path.

[0139] Payment behavior stability regularization term design:

[0140] To enhance the model's ability to detect abnormal payment behavior, we added an innovative regularization term to the path complexity score to penalize sections in the payment path where the structure and behavior are inconsistent. This regularization term is designed as follows:

[0141]

[0142] in:

[0143] f(p i ) is the payment state pi A function that maps to a behavior space representation (via a layer of linear projection or rule-based encoding);

[0144] g(E b ) is the projection of the behavior template in the payment space;

[0145] ||·||1 is the L1 distance, which measures the deviation between the behavior prediction and the actual payment;

[0146] The higher the value, the greater the difference between the payment path and the predicted behavior template, and there may be a situation where "behavior masks payment anomalies."

[0147] Finally, the structured output of the payment path consists of two parts: the actual path sequence P b and scoring index S p :

[0148] S p =αC p +βΩ misalign

[0149] Among them, α and β are weight coefficients, which are set based on experience or fitted through historical data.

[0150] For example:

[0151] If a behavior template E b It is displayed as "stable chronic disease management type", but the payment trajectory frequently switches from pooling → self-payment → special disease catalog, and C p =0.7 (far exceeding the mean), while the behavior prediction and payment status deviate from Ω misalign =0.5, then the comprehensive S p It indicates that this behavior is strongly inconsistent with payment and has a high risk of fraud.

[0152] Output:

[0153] P b : Payment path state sequence, used for subsequent offset analysis;

[0154] S p : Payment behavior consistency score, which integrates the payment trajectory jump rate and behavior-payment deviation, is used to determine whether there are path fraud signals.

[0155] Step 4: Verify the alignment between the payment path and the behavior template

[0156] The goal of this step is to extract the payment path sequence P from step 3. b The behavior template encoding E obtained by modeling in step 2 bConduct structural alignment analysis. We aim to verify whether the medical insurance payment pathway aligns with behavioral patterns through alignment testing, thereby identifying potential fraudulent behaviors known as "behavior-path mismatches." For example, while behaviors may indicate chronic disease management, the payment pathway frequently involves cross-institutional settlements and changes in payment methods.

[0157] This is a structural matching modeling step, not a final judgment or visualization step. Its purpose is to construct a scoring mechanism that quantifies the structural similarity between the behavior template and the payment path, and outputs a structured "alignment offset" result to support the subsequent system integration scoring and automated linkage decision-making. Therefore, all modeling structures must focus on the calculation of the structural alignment between behavior and path, and cannot extend beyond the policy output layer.

[0158] enter:

[0159] P b : Payment path state sequence, output by step 3, represents the state combination sequence of the actual payment behavior sorted by time;

[0160] E b : Behavior template encoding, output by step 2, is an embedding vector reflecting the behavior pattern of the list structure sequence;

[0161] S p : Payment consistency score (output of step 3), which represents the degree of abnormality of the payment path itself. This step uses it as a dynamic adjustment factor in the deviation modeling.

[0162] Specific steps:

[0163] Payment path mapping to structure space:

[0164] We first need to make the payment path P b Converted into behavior template E b A comparable structural representation is to encode the path sequence into a vector E p This step is not to build a new model, but to perform structural normalization and vector embedding on the existing payment status sequence.

[0165] Each payment status node p i ∈P b Consists of the following fields:

[0166] p i =(dir i ,pay i ,org i )

[0167] where dir i Indicates the reimbursement catalog code (such as "Catalog 1" or "Special Disease Catalog"), pay iIndicates payment method (such as "pooling" or "self-payment"), org i Indicates the grade of the hospital (such as "community" or "top three").

[0168] We use dictionary encoding to convert the above fields into dense vectors and construct the path vector by simple average pooling:

[0169]

[0170] Among them, Embed(p i ) is a structural vector encoding implemented by embedding lookup table, and L is the payment path length. p is the expression of the payment trajectory in the structural space, which is consistent with E b Maintain the same dimensions to ensure comparability.

[0171] For example, a payment path is:

[0172] p1 = ("Directory 1", "Coordination", "Community Hospital");

[0173] p2 = ("Catalog 1", "Self-payment", "Thirteen-level hospital");

[0174] p3 = ("Special Disease Catalog", "Planning", "Thirteen-Level Hospital")

[0175] After encoding through the field lookup table, for each p i Do vector splicing and average pooling to get E p This is a bridging operation from the medical insurance payment trajectory to the structural behavior space.

[0176] Behavior-path alignment offset score calculation:

[0177] To compare the matching degree between behavior and path, we design a joint offset scoring function R align , which combines the vector direction deviation (cosine angle) and the value deviation (L1 difference):

[0178] R align =α(1-cos(E b ,E p ))+β||E b -E p ||1

[0179] in:

[0180] cos(E b ,E p ) is the cosine similarity between the behavior template and the path encoding, reflecting the direction consistency;

[0181] ||E b -E p ||1 is the absolute difference between the two vectors, indicating the average structural difference between the overall behavior and the payment state;

[0182] α, β are balance weights, and the recommended settings are α=0.5, β=0.5;

[0183] R align is the behavior-path structure deviation score, with a larger range indicating greater mismatch.

[0184] We further enhance the fraud sensitivity of the score: when the payment consistency score S p If the risk is high (i.e. the path itself is extremely complex), we should be more stringent in demanding its behavior matching. Therefore, we designed the risk adjustment factor:

[0185] R adj =R align ·(1+γ·S p )

[0186] Here, γ is the risk adjustment coefficient (e.g., 0.5 to 1.0), which makes it easier to trigger high scores for alignment inconsistencies in the case of path anomalies, thereby improving detection sensitivity.

[0187] Examples of reproducible real-life scenarios:

[0188] Assume that a checklist behavior template is E b =(0.2, 0.3, 0.1, 0.1), indicating the high-frequency "catalog 1 + overall planning + community" model; and its payment path is p1 = (catalog 1, overall planning, community), p2 = (catalog 2, self-payment, three-level), path code E p =(0.4,0.1,0.2,0.3).

[0189] The calculated cosine difference is about 0.4, and the L1 difference is 0.5. Assuming S p =0.8, then we get:

[0190] R align =0.5(1-0.6)+0.5(0.5)=0.45;

[0191] R adj =0.45·(1+0.8·0.8)≈0.77;

[0192] If the threshold is higher than 0.6, it can be judged that the behavior is highly inconsistent with the path and there is a potential risk of fraud.

[0193] Output:

[0194] R align : Structural alignment score between behavior and path, higher values ​​represent greater misalignment;

[0195] R adj : The final alignment risk score after path complexity adjustment for subsequent strategy fusion.

[0196] Step 5: Risk integration decision-making and system linkage execution

[0197] The core goal of this step is to classify the risk levels of the behavioral path alignment score results calculated in the previous step (Step 4), and drive the medical insurance business system to complete the linkage execution operation based on this.

[0198] According to R adj For the scoring value, we define the following three-stage judgment rule:

[0199] If R adj <0.3: Identified as a “low risk” list and automatically released;

[0200] If 0.3≤R adj <0.6: Identified as a "medium risk" list and needs to be submitted to the manual review system;

[0201] If R adj ≥0.6: Identified as a "high-risk" list, the list is immediately frozen and pushed to the audit system.

[0202] The system execution logic is as follows:

[0203] Release processing (low risk)

[0204] The system directly updates the status of the list to "automatically reviewed and passed", and the medical insurance settlement process continues to process it without any intervention.

[0205] Submit for manual review (medium risk)

[0206] The system adds the list to the manual review task queue, marks it as "There is a deviation between the behavior and the payment path, please review", and generates a report containing R align and R adj The rating record field.

[0207] Automatic freeze and audit (high risk)

[0208] The system sets the status of the list to "Frozen" and calls the following three sub-actions:

[0209] Suspend the listing in the listing management system and mark it as "suspected fraud";

[0210] Send the scores and reasons for behavior-path deviation to the medical insurance audit clue system;

[0211] Automatically generate risk records containing the following fields: list number, score value, freeze reason (such as "path jump + behavior deviation"), freeze time, etc., and write them into the audit log table.

[0212] Examples of reproducible real-life scenarios:

[0213] Assume a list is scored as follows:

[0214] R align =0.50;

[0215] R adj =0.72;

[0216] According to the policy rules, this list belongs to the "high-risk range" and will trigger the following linkage actions:

[0217] The status is marked as "frozen" by the system;

[0218] The risk label is “inconsistent behavioral path”;

[0219] The list data is transmitted to the audit platform;

[0220] The audit record field is updated to: "Source score = Behavior path deviation", value = 0.72."

[0221] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: The present invention identifies a list of potential risks based on the field structure, and identifies abnormal combinations and frequency characteristics based on historical behavior modeling and rule deviation analysis, thereby achieving early detection of complex and evolving fraudulent behaviors. At the same time, the present invention further utilizes the path data in the medical insurance payment process as behavioral evidence to establish a correspondence between the list items and the actual payment behavior, dynamically restore the evolution trajectory of the payment path, and based on this, determine whether there are fraudulent deviations such as abnormal payment methods and path jumps, thereby solving the core bottleneck of the existing system that cannot verify the identification results.

[0222] The present application also provides a medical insurance fraud detection system based on blockchain and artificial intelligence, such as Figure 2 Shown, including:

[0223] A feature extraction module is used to input the original data collected by the medical insurance list system; embed the original data to obtain a structural embedding vector, and evaluate the structural embedding vector to perform a structural risk score;

[0224] A behavior detection module is used to extract a list of behavior sequences within a preset time window of a medical insurance subject, extract a latent state vector at each time point from the list of behavior sequences, and calculate a behavior template encoding vector based on the latent state vector; a behavior deviation score is calculated using a deviation scoring function; the deviation scoring function represents the degree of deviation of the current behavior sequence by calculating the cosine similarity between the list of behavior sequences and the historical records, as well as the distance between the two vectors;

[0225] A payment path extraction module is configured to input on-chain payment operation records, which are consistent with the time window of the structural embedding vector behavior sequence; after mapping the on-chain payment operation records into a payment state sequence, construct a payment path graph, wherein the nodes of the payment path graph are payment states and the edges are time; output a payment jump index, which is used to measure the degree of state jumps in the payment path graph, and is obtained by calculating the ratio of abnormal jumps to total jump opportunities, and the abnormal jumps are obtained by calculating the distance between two payment path graph nodes; and calculate a consistency score index, which is obtained by fusing the payment jump index with the behavior-payment deviation, and the behavior-payment deviation is obtained by calculating the offset between the projection of the payment state sequence and the list behavior sequence in the payment space;

[0226] an alignment verification module, the alignment verification module being used to perform structural alignment detection on the payment path sequence and the behavior template encoding vector, outputting a structural deviation score; and outputting a final risk score based on the structural deviation score;

[0227] A strategy generation module is used to divide the risk levels according to the interval to which the final risk score belongs, and to execute corresponding preset strategies.

[0228] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A medical insurance fraud detection method based on blockchain and artificial intelligence, characterized by: include: S1: Input the original data collected by the medical insurance list system; Embed the original data to obtain a structural embedding vector, and evaluate the structural embedding vector to obtain a structural risk score; S2: extracting a list behavior sequence within a preset time window of the medical insurance subject, extracting a latent state vector at each time point from the list behavior sequence, and calculating a behavior template encoding vector based on the latent state vector; Calculate the behavior deviation score by using a deviation scoring function; the deviation scoring function represents the degree of deviation of the current behavior sequence by calculating the cosine similarity between the list behavior sequence and the historical record and the distance between the two vectors; S3: Input the on-chain payment operation record, which is consistent with the time window of the structure embedding vector behavior sequence; After mapping the on-chain payment operation records into a payment status sequence, a payment path graph is constructed, where nodes of the payment path graph represent payment statuses and edges represent time. Output a payment jump index, which is used to measure the degree of state jump in the payment path graph. The payment jump index is obtained by calculating the ratio of abnormal jump times to total jump opportunities. The abnormal jump is obtained by calculating the distance between two payment path graph nodes. Calculating a consistency score index, the consistency score index is obtained by fusing a payment jump index and a behavior-payment deviation, the behavior-payment deviation is obtained by calculating the offset between the projection of the payment state sequence and the list behavior sequence in the payment space; S4: Performing structural alignment detection on the payment path sequence and the behavior template encoding vector, and outputting a structural deviation score; outputting a final risk score based on the structural deviation score; S5: Risk levels are divided according to the interval to which the final risk score belongs, and corresponding preset strategies are executed.

2. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 1 is characterized in that: The original data includes diagnosis code, drug code, doctor ID, fee amount, department code, and catalog category.

3. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 1 is characterized in that: The structural risk score is calculated by combining the complexity information entropy with the business conflict item; the business conflict item is 1 when there is a conflicting combination between fields, and 0 otherwise.

4. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 1 is characterized in that: The behavior template encoding vector is a structural behavior template embedding of the entire list behavior sequence; the attention weight of the behavior template encoding vector is calculated through the structural risk score.

5. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 1 is characterized in that: The on-chain payment operation record includes timestamp, operation type, source of funds, reimbursement catalog, hospital grade, and settlement code.

6. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 1 is characterized in that: A risk adjustment factor is introduced into the structural deviation score to make it easier to trigger a high score for alignment inconsistency in the case of path anomalies, thereby improving detection sensitivity.

7. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 1 is characterized in that: The risk level classification is based on the following rules: Setting a first threshold and a second threshold; When the final risk score is lower than the first threshold, it is considered as a low-risk list and automatically released; When the final risk score is higher than or equal to the first threshold and lower than the second threshold, it is identified as a medium-risk list and submitted to the manual review system; When the final risk score is higher than or equal to the second threshold, it is identified as a high-risk list, and the list is immediately frozen and pushed to the audit system.

8. The medical insurance fraud detection method based on blockchain and artificial intelligence according to claim 7 is characterized in that: After being identified as a high-risk list, the high-risk list will be marked as suspected fraud, and the final risk score and behavior-path deviation will be transmitted to the medical insurance audit clue system, and a risk record containing the following fields will be generated: list number, score value, freezing reason, and freezing time.

9. A medical insurance fraud detection system based on blockchain and artificial intelligence, characterized by: include: A feature extraction module, which is used to input raw data collected by the medical insurance list system; Embed the original data to obtain a structural embedding vector, and evaluate the structural embedding vector to obtain a structural risk score; A behavior detection module is used to extract a list of behavior sequences within a preset time window of a medical insurance subject, extract a latent state vector at each time point from the list of behavior sequences, and calculate a behavior template encoding vector based on the latent state vector; a behavior deviation score is calculated using a deviation scoring function; the deviation scoring function represents the degree of deviation of the current behavior sequence by calculating the cosine similarity between the list of behavior sequences and the historical records, as well as the distance between the two vectors; A payment path extraction module, which is used to input on-chain payment operation records that are consistent with the time window of the structural embedding vector behavior sequence; after mapping the on-chain payment operation records into a payment state sequence, construct a payment path graph, wherein the nodes of the payment path graph are payment states and the edges are time; Output a payment transition index, which is used to measure the degree of state transitions in the payment path graph. The payment transition index is obtained by calculating the ratio of abnormal transitions to total transition opportunities. The abnormal transition is obtained by calculating the distance between two payment path graph nodes. Calculate a consistency score index, which is obtained by fusing the payment transition index with the behavior-payment deviation. The behavior-payment deviation is obtained by calculating the offset between the projection of the payment state sequence and the list behavior sequence in the payment space. an alignment verification module, the alignment verification module being used to perform structural alignment detection on the payment path sequence and the behavior template encoding vector, and output a structural deviation score; outputting a final risk score based on the structural deviation score; A strategy generation module is used to divide the risk levels according to the interval to which the final risk score belongs, and to execute corresponding preset strategies.

Images