Dip and drg medical insurance risk prediction method, device and equipment and medium

Abstract

The application discloses a DIP and DRG medical insurance risk prediction method, device, equipment and medium, the method comprises the following steps: preprocessing the collected multi-source heterogeneous data to obtain an initial time series data set, extracting dynamic features from the initial time series data set to obtain a key time series feature combination of each insured user, training a pre-constructed original model based on the key time series feature combination to obtain a fraud risk time series prediction model with a double-branch coding structure, inputting the data after target area preprocessing and dynamic feature extraction into the fraud risk time series prediction model, outputting a dynamic fraud risk prediction result, accurately mining and time-series representing fraud features, predicting fraud risk from the time and space dimensions, improving the accurate prediction of complex fraud behavior, dynamically and real-timely predicting abnormal behavior in medical insurance transactions, controlling risk transactions in advance, and improving the accuracy and efficiency of medical insurance transaction supervision.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a method, apparatus, device, and medium for predicting DIP and DRG medical insurance risks. Background Technology

[0002] The medical insurance regulatory system includes two payment methods: Diagnosis Related Groups (DRG) payment and Diagnosis-Intervention Packet (DIP) payment, which are two core disease-based bundled payment methods. The current medical insurance regulatory system is facing unprecedented challenges under the DIP / DRG payment method.

[0003] Traditional manual review methods rely heavily on the experience and judgment of regulators. This not only results in low efficiency in identifying professional indicators such as disease grouping and diagnosis-treatment correlation, but is also susceptible to subjective factors, leading to a high rate of missed detections of covert fraudulent activities. Furthermore, it requires significant human resources, resulting in high regulatory costs. While rule-based systems achieve some automated screening, their static rule nature makes them unable to dynamically adapt to the rapid evolution of fraud methods. They exhibit significant lag in prevention and control when facing adjustments to DIP / DRG grouping rules and new fraud patterns, making it difficult to accurately identify new violations such as high-scoring fraud and substitution of treatment items. In addition, current data integration efforts have not broken down data silos. Multi-source heterogeneous data from medical insurance settlement, medical institution operations, and patient visits are only simply collected, lacking deep integration and dynamic feature extraction mechanisms. Moreover, the overall regulatory model is mostly reactive, failing to accurately predict and proactively control fraudulent activities before they occur. Therefore, how to break down data barriers, achieve dynamic fraud feature extraction, and establish an intelligent medical insurance risk prediction system with pre-emptive prediction capabilities to adapt to the complex fraud prevention and control needs under the DIP / DRG payment model and ensure the security of the medical insurance fund has become an urgent problem to be solved. Summary of the Invention

[0004] The main objective of this invention is to provide a method, device, equipment, and medium for predicting medical insurance risks using DIP and DRG, aiming to solve the technical problems of existing technologies being unable to effectively mine dynamic features from multi-source heterogeneous data, unable to complete accurate prediction and early control before fraudulent activities occur, resulting in low accuracy and significant lag in fraud prediction.

[0005] To achieve the above objectives, the present invention provides a method for predicting DIP and DRG medical insurance risks, the method comprising the following steps:

[0006] The collected multi-source heterogeneous data is preprocessed to obtain an initial time-series dataset, which includes medical insurance settlement data, medical institution operation data and patient medical record data.

[0007] Dynamic feature extraction is performed on the initial time-series dataset to obtain key time-series feature combinations for each insured user. The key time-series feature combinations include fraud types and multiple key fraud time-series features associated with the fraud types.

[0008] The pre-constructed original model is trained based on the combination of key temporal features to obtain a fraud risk temporal prediction model. The original model includes a dual-branch feature encoding module, a feature fusion module, and a prediction output module. The first branch of the dual-branch feature encoding module is a graph neural network, and the second branch is a bidirectional long short-term memory network. The first branch is configured for graph structure data learning, and the second branch is configured for temporal feature learning. The fraud risk temporal prediction model includes a first-branch feature encoding structure and a second-branch feature encoding structure. The first-branch feature encoding structure is constructed based on a graph neural network, and the second-branch feature encoding structure is constructed based on a bidirectional long short-term memory network.

[0009] Data monitoring is performed on the target area, and the collected monitoring data is preprocessed and dynamic features are extracted. The preprocessed and dynamically extracted data is input into the fraud risk time series prediction model, and the dynamic fraud risk prediction result is output. The dynamic fraud risk prediction result includes the risk level of each transaction behavior in the target area.

[0010] Optionally, the step of dynamically extracting features from the initial time-series dataset to obtain key time-series feature combinations for each insured user includes:

[0011] The initial time-series dataset is input into a semantic encoding model for semantic encoding to obtain semantic vectors. The semantic encoding model is pre-built based on a dictionary dataset in the medical insurance field.

[0012] Analyze the correlation strength of semantic vectors of each insured user at different time steps to obtain time-series trend features;

[0013] Based on the semantic vector analysis, the semantic vector association strength of each entity is obtained to obtain entity association features. The entities include insured users, medical institutions and medical services.

[0014] Based on the time-series trend features and the semantic vector correlation strength, a feature subset is generated for each insured user, and the feature subset contains multiple feature matrices;

[0015] Based on historical samples, the fraud feature weights of each feature matrix are determined, and based on the fraud feature weights, each feature matrix is ​​associated with the fraud type to generate multiple fraud feature combinations. The historical samples include feature matrix samples of fraud labels and feature matrix samples of normal labels.

[0016] Feature filtering is performed on the fraud feature combination to obtain key time-series feature combinations.

[0017] Optionally, the step of determining the fraud feature weights of each feature matrix based on historical samples, and then associating and combining each feature matrix with the fraud type based on the fraud feature weights to generate multiple fraud feature combinations, includes:

[0018] A weighted analysis model is constructed based on historical samples. The weighted analysis model includes an input layer, a multi-head causal attention layer, and a fusion layer. The input layer is configured to map the input feature matrix into a causal feature vector based on an intervention function. The multi-head causal attention layer includes multiple causal attention heads, including a fraud type attention head and a causal hierarchy attention head. The fraud type attention head is configured to calculate the attention score of the causal feature vector based on the fraud type dimension. The causal hierarchy attention head is configured to calculate the attention score of the causal feature vector based on the causal hierarchy dimension. The fusion layer is configured to fuse the outputs of the multiple causal attention heads and output the initial feature weights.

[0019] The feature subsets of each insured user are input into the weight analysis model to obtain the initial feature weights;

[0020] Fact analysis is performed on the historical samples to obtain fact analysis results, including the change in fraud probability caused by feature loss;

[0021] Based on the results of the fact analysis, the initial feature weights are adjusted to obtain the fraud feature weights of each feature matrix;

[0022] Based on the fraud feature weights, priority analysis is performed on the feature matrix to determine the core feature matrix and auxiliary feature matrix in the feature matrix;

[0023] The core feature matrix and the auxiliary feature matrix are associated and combined with the fraud type to obtain the fraud feature combination.

[0024] Optionally, the step of performing feature filtering on the fraud feature combination to obtain key time-series feature combinations includes:

[0025] Based on medical insurance rule information, reasonable business constraints are generated, and the reasonableness of the fraud feature combination is screened based on the reasonable business constraints to obtain the first feature combination;

[0026] Based on the time step analysis of each feature in the first feature combination, the time fluctuation information of each feature is analyzed, and the time stability of the first feature combination is screened based on the time fluctuation information to obtain the second feature combination.

[0027] Multicollinearity among the features in the second feature combination is analyzed to determine the mutual information value among the features. Based on the mutual information value, feature redundancy screening is performed on the second feature combination to obtain key time-series feature combinations.

[0028] Optionally, training the pre-constructed original model based on the combination of key temporal features to obtain a fraud risk temporal prediction model includes:

[0029] Based on the combination of key temporal features, a logical analysis of fraud behavior is performed to generate a fraud feature chain. The fraud feature chain is a logical chain of fraud behavior composed of multiple key fraud temporal features connected in series or in parallel.

[0030] Based on the medical insurance treatment process, the fraud feature links are analyzed in a time-series manner, and the fraud feature links are adjusted based on the results of the time-series behavior analysis to obtain the target feature links;

[0031] Based on the target feature links, causal analysis is performed to construct a fraud causal time sequence graph structure. The nodes of the fraud causal time sequence graph structure include time sequence feature nodes, fraud behavior nodes, fraud type nodes, and intervention nodes. The edges of the fraud causal time sequence graph structure include causal driving edges, behavior classification edges, time sequence driving edges, and intervention adjustment edges.

[0032] The pre-constructed original model is trained based on the fraud causal time series graph structure and the combination of key time series features to obtain a fraud risk time series prediction model.

[0033] Optionally, the fraud risk time series prediction model further includes a betweenness center module, a feature fusion module, and a prediction output module;

[0034] The betweenness center module is used to calculate the betweenness center value of each time-series feature node based on the node characteristics of each node in the fraud causal time-series graph structure, and to perform betweenness optimization on the fraud causal time-series graph structure and the key time-series feature combination based on the betweenness center value, and output the betweenness-optimized fraud causal time-series graph structure and key time-series feature combination.

[0035] The first branch feature encoding structure is used to encode the features of the betweenness-optimized fraud causal time series graph structure and output the graph structure embedding vector.

[0036] The second branch feature encoding structure is used to encode the key temporal feature combination after betweenness optimization and output a temporal-aware embedding vector.

[0037] The feature fusion module is used to perform weighted fusion of the graph structure embedding vector and the time-aware embedding vector to obtain an initial fused feature, calculate the importance weight of each feature dimension in the initial fused feature, determine the feature contribution of each feature dimension based on the importance weight, perform attention calibration on the initial fused feature based on the feature contribution, and output the target fused feature.

[0038] The prediction output module is used to perform feature transformation and regularization on the target fusion features to obtain high-level abstract features, calculate the probability distribution information of each risk level based on the high-level abstract features, determine the risk level of each transaction behavior in the target area based on the probability distribution information, and output the risk level of each transaction behavior in the target area as a dynamic fraud risk prediction result.

[0039] Optionally, the preprocessing of the collected multi-source heterogeneous data to obtain an initial time-series dataset includes:

[0040] The first dataset is obtained by preprocessing the multi-source heterogeneous data collected locally. The preprocessing includes data cleaning, standardization, and time-series alignment.

[0041] Send data acquisition requests to multiple data sharing nodes so that each data sharing node can preprocess the multi-source heterogeneous data collected locally, encrypt the second dataset obtained from the preprocessing, and respond to the data acquisition request based on the encrypted data;

[0042] The system acquires response information from each data sharing node, and performs community aggregation on each data sharing node based on the response information to generate key communities and edge communities. The key community aggregation has multiple first nodes, and the edge community aggregation has multiple second nodes. The first nodes are data sharing nodes whose response speed, historical data quality, and / or historical credibility meet preset conditions, and the second nodes are data sharing nodes whose response speed, historical data quality, and historical credibility do not meet preset conditions. The response information includes encrypted data, identity information, and response speed of the data sharing node response.

[0043] The encrypted data of the responses from each first node in the key community is decrypted to obtain the second dataset;

[0044] The node data weights are determined based on the response speed, historical data quality, and historical reliability of each first node.

[0045] The first dataset and the second dataset are aggregated based on the node data to obtain an initial time-series dataset.

[0046] Furthermore, to achieve the above objectives, the present invention also proposes a DIP and DRG medical insurance risk prediction device, the DIP and DRG medical insurance risk prediction device comprising:

[0047] The data processing module is used to preprocess the collected multi-source heterogeneous data to obtain an initial time-series dataset. The multi-source heterogeneous data includes medical insurance settlement data, medical institution operation data, and patient medical record data.

[0048] The dynamic feature extraction module is used to perform dynamic feature extraction on the initial time series dataset to obtain key time series feature combinations for each insured user. The key time series feature combinations include fraud types and multiple key fraud time series features associated with the fraud types.

[0049] The model building module is used to train the pre-built original model based on the key temporal feature combination to obtain a fraud risk temporal prediction model. The original model includes a dual-branch feature encoding module, a feature fusion module, and a prediction output module. The first branch of the dual-branch feature encoding module is a graph neural network, and the second branch is a bidirectional long short-term memory network. The first branch is configured for graph structure data learning, and the second branch is configured for temporal feature learning. The fraud risk temporal prediction model includes a first-branch feature encoding structure and a second-branch feature encoding structure. The first-branch feature encoding structure is built based on a graph neural network, and the second-branch feature encoding structure is built based on a bidirectional long short-term memory network.

[0050] The dynamic risk prediction module is used to monitor data in the target area, preprocess the collected monitoring data and extract dynamic features, input the preprocessed and dynamically extracted data into the fraud risk time series prediction model, and output dynamic fraud risk prediction results. The dynamic fraud risk prediction results include the risk level of each transaction behavior in the target area.

[0051] In addition, to achieve the above objectives, this application also proposes a DIP and DRG medical insurance risk prediction device, the device comprising: a memory, a processor, and a DIP and DRG medical insurance risk prediction program stored on the memory, the processor being used to run the DIP and DRG medical insurance risk prediction program, the computer program being configured to implement the steps of the DIP and DRG medical insurance risk prediction method as described above.

[0052] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the DIP and DRG medical insurance risk prediction methods described above.

[0053] This invention constructs a dual-branch coding-based temporal prediction model for fraud risk by combining temporal feature mining and graph structure association learning. This breaks through the technical constraints of traditional medical insurance supervision, which relies on manual experience and static rules, leading to lagging risk supervision. It achieves precise targeted screening of fraud features through dynamic feature extraction, and simultaneously captures the spatial correlation logic and temporal evolution patterns of fraudulent behavior through dual-branch coding. This model is effectively adapted to the risk prediction scenarios of DIP / DRG payment models, enabling dynamic real-time prediction of abnormal behaviors in medical insurance transactions, and achieving early control of risky transactions, thereby improving the accuracy and efficiency of medical insurance transaction supervision. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0055] Figure 1 This is a flowchart illustrating the first embodiment of the DIP and DRG medical insurance risk prediction method of the present invention.

[0056] Figure 2 This is a schematic diagram of the weighting analysis model in one embodiment of the DIP and DRG medical insurance risk prediction method of the present invention;

[0057] Figure 3 This is a flowchart illustrating the second embodiment of the DIP and DRG medical insurance risk prediction method of the present invention.

[0058] Figure 4 This is a schematic diagram of the structure of a fraud risk time series prediction model in one embodiment of the DIP and DRG medical insurance risk prediction method of the present invention;

[0059] Figure 5 This is a structural block diagram of the first embodiment of the DIP and DRG medical insurance risk prediction device of the present invention.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0062] This invention provides a method for predicting DIP and DRG medical insurance risks, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the DIP and DRG medical insurance risk prediction method of the present invention.

[0063] In this embodiment, the DIP and DRG medical insurance risk prediction method includes the following steps:

[0064] Step S10: Preprocess the collected multi-source heterogeneous data to obtain the initial time series dataset.

[0065] It should be understood that the executing entity of this embodiment can be a terminal electronic device (such as a computer) with data processing, network communication, and program execution functions. The following uses the DIP and DRG medical insurance risk prediction device (hereinafter referred to as the prediction device) as an example to describe this embodiment and the following embodiments.

[0066] It should be noted that multi-source heterogeneous data refers to a set of medical insurance-related data with different sources, data formats and / or structures, which can cover three core types of data: medical insurance settlement, medical institution operation, and patient visits.

[0067] Among them, medical insurance settlement data refers to various data generated when medical insurance agencies, medical institutions, and insured persons conduct cost accounting after insured persons complete medical services, including settlement amount, reimbursement ratio, settlement status, social security card usage information, etc.

[0068] Medical institution operation data refers to data generated by medical institutions during the course of daily diagnosis and treatment services, which reflects the operational status of the institution and covers operational indicators such as patient volume, departmental scheduling plans, frequency of use of medical equipment, and inventory of medicines and consumables.

[0069] Patient medical record data refers to the complete medical documents and service records generated when a patient receives medical services at a medical institution. It includes chief symptoms, clinical diagnosis results, implemented medical procedures, doctor's orders, and other relevant medical information.

[0070] It should be noted that the initial time series dataset refers to a standardized dataset that has been preprocessed from multi-source heterogeneous data, and then standardized and integrated according to the time dimension, possessing a unified format and time series attributes.

[0071] It is understandable that multi-source heterogeneous data collection can be carried out through multiple channels such as medical institution system interfaces, medical insurance management databases, and mobile storage devices. It can also collect medical insurance settlement data, medical institution operation data, and patient medical record data simultaneously. During the collection process, data traceability tags will be established to ensure that the data source is traceable.

[0072] In practice, the prediction device first de-identifies the collected data, removing sensitive information and retaining only the anonymized identifier; secondly, it performs data cleaning to remove invalid and duplicate data; then, it transforms unstructured data into structured data; finally, it performs time-series alignment, associating the three types of data with the same time dimension according to a preset time granularity (e.g., day / week), forming an initial time-series dataset with time-series attributes. This breaks down data silos, achieves effective integration of multi-dimensional data, eliminates interference from invalid data, and provides a unified data dimension for subsequent fraud detection.

[0073] Step S20: Perform dynamic feature extraction on the initial time series dataset to obtain the key time series feature combination of each insured user.

[0074] It should be noted that fraud type refers to the specific classification of various fraudulent and illegal behaviors in the medical insurance field. Key fraud temporal characteristics refer to core features that are highly correlated with specific fraud types and have time evolution attributes, which can reflect the behavioral trajectory and correlation patterns of fraudulent and illegal behaviors in the time dimension.

[0075] In some embodiments, the prediction device first organizes and encodes all data features (e.g., diagnosis codes, treatment items, consumable usage, hospitalization cycles, settlement weights, etc.) of the same insured user in the initial time series dataset to form a unique feature subset; then, based on the weights of the feature subset labeled by historical fraud cases in DRG / DIP, it extracts basic features strongly correlated with typical DRG / DIP fraud behaviors (e.g., the trajectory of changes in primary and secondary diagnosis codes, matching records of high-value consumable charges and usage, time series information of multiple hospitalizations for the same disease, etc.). The extracted basic features are then matched and associated with typical DRG / DIP fraud types (e.g., low-code high-encoding, split hospitalization, fictitious treatment / consumable records, low-indication hospitalization, etc.) to form feature combinations classified according to DRG / DIP fraud types; subsequently, multi-level screening is carried out, first based on DRG / DIP payment rules and clinical diagnosis and treatment. The system generates reasonable constraints for business operations, eliminates features that do not conform to the logic of diagnosis and treatment and settlement (such as linking mild cases in children to DRG group codes for severe cases in adults, or including general outpatient treatment items in the scope of inpatient DRG settlement), analyzes the temporal fluctuation information of features (such as the frequency of diagnosis code adjustment within a period of time, and the temporal changes in the number of hospitalizations) to complete the temporal stability screening, and finally eliminates redundant features (such as highly correlated features like "adjustment of primary diagnosis code" and "change of DRG group") by calculating the mutual information value between features. Finally, it obtains the key temporal feature combination in the DRG / DIP scenario. Through dynamic feature engineering, it achieves accurate targeted screening of DRG / DIP fraud-related features, effectively preserves the temporal evolution law of features, eliminates invalid, temporally unstable and redundant features, and reduces the interference of invalid information on subsequent model training.

[0076] Step S30: Train the pre-constructed original model based on the combination of key time-series features to obtain a fraud risk time-series prediction model.

[0077] Understandably, the prediction device pre-builds a basic model architecture including a feature encoding module. The first branch feature encoding structure is built using a Graph Neural Network (GNN) to learn the relationship structure between features; the second branch feature encoding structure is built based on a Bidirectional Long Short-Term Memory (Bi-LSTM) network to capture the bidirectional temporal dependencies of features. Key temporal features are combined and input into the original model, and gradient descent is used to optimize the model parameters, while a regularization mechanism is introduced to avoid overfitting. During training, the first branch feature encoding structure learns the relationship graph structure between insured individuals, medical institutions, and treatment projects, while the second branch feature encoding structure mines the temporal progression patterns of the features. Once the model converges, a fraud risk temporal prediction model is obtained.

[0078] Step S40: Monitor the target area, preprocess the collected monitoring data and extract dynamic features, input the preprocessed and dynamically extracted data into the fraud risk time series prediction model, and output the dynamic fraud risk prediction result.

[0079] It should be noted that the target area refers to a specific geographical jurisdiction or medical insurance management jurisdiction where special monitoring of medical insurance transaction fraud risks needs to be carried out. The monitoring data refers to relevant data on medical insurance settlement, medical institution operation, and patient visits collected in real time within the target area, which serves as real-time input data for conducting dynamic fraud risk prediction.

[0080] It should be noted that the dynamic fraud risk prediction result refers to the risk assessment conclusion output after the monitoring data is analyzed by the model, which includes the risk level of each medical insurance transaction behavior within the target area. The risk level refers to the hierarchical classification of the degree of fraud risk of medical insurance transactions based on preset thresholds (e.g., high, medium, and low levels), which can intuitively reflect the risk level of different transactions and provide a basis for decision-making in subsequent medical insurance supervision.

[0081] In some embodiments, for a designated medical insurance regulatory jurisdiction (target area), medical insurance transaction monitoring data within the jurisdiction is acquired through a real-time data acquisition channel. This includes settlement data, operational data, and patient visit data of medical institutions within the area. The preprocessing and dynamic feature extraction processes described above are then repeated on the collected monitoring data to ensure that the format and dimensions of the input data are consistent with the model training data. The processed monitoring data is then input into a fraud risk time-series prediction model, which outputs the risk level of each transaction within the target area. Regulatory personnel can trigger a freezing mechanism for high-risk transactions based on the risk level, automatically triggering a freezing mechanism for high-risk transactions, and sending medium- and low-risk transactions for manual review, forming a collaborative prevention and control closed loop of "machine review + manual intervention".

[0082] This embodiment is applied to predict and control fraudulent activities such as insurance fraud before they occur, using a model to prevent and control such activities through accurate prediction. Traditional regulatory models often lag behind fraudulent activities, relying on post-event manual review and static screening by rule engines, which is insufficient to cope with the rapid evolution and concealment of fraud methods. This embodiment, through deep fusion of multi-source heterogeneous data, breaks down information silos across dimensions such as medical insurance settlement data, medical institution operational data, and patient medical records. By employing temporal feature mining and graph structure association learning, it constructs a dual-branch coding-based temporal prediction model for fraud risk. This model overcomes the technical constraints of traditional medical insurance supervision, which relies on manual experience and static rules leading to lagging risk supervision. Dynamic feature extraction enables precise targeted screening of fraud features, while dual-branch coding simultaneously captures the spatial correlation logic and temporal evolution patterns of fraudulent behavior. This effectively adapts to risk prediction scenarios in DIP / DRG payment models, enabling dynamic real-time prediction of abnormal behavior in medical insurance transactions. This allows for proactive control of risky transactions, improving the accuracy and efficiency of medical insurance transaction supervision, reducing manual review costs and the risk of misjudgment, and nipping violations in the bud through pre-emptive intervention. This effectively safeguards the medical insurance fund, promotes standardized medical practices in medical institutions, and fosters a healthy medical ecosystem.

[0083] Furthermore, in order to capture the characteristic patterns of various covert fraudulent insurance claims, in one embodiment, step S20 further includes:

[0084] Step S201: Input the initial time series dataset into the semantic coding model for semantic coding to obtain semantic vectors.

[0085] It should be noted that the semantic encoding model can refer to an encoding model based on a general pre-trained language model (such as BERT), adjusted by combining it with a medical insurance domain dictionary dataset, and possessing the ability to understand the semantics of medical insurance professional terms. This model is configured to transform different types of medical insurance data into semantic vectors of a unified dimension. The medical insurance domain dictionary dataset can be formed by the prediction device first integrating standard medical insurance industry data, including disease diagnosis codes, medical insurance treatment item catalogs, DIP / DRG disease grouping rules, medical insurance settlement terminology tables, etc., to create a dedicated medical insurance domain dictionary dataset, while also annotating the relationships between terms (e.g., "pneumonia" corresponds to the "chest CT" treatment item).

[0086] In some embodiments, the prediction device can adjust the model parameters based on the aforementioned medical insurance domain dictionary dataset, adjusting the parameters of the word embedding layer and attention layer of the model to adapt to the semantic understanding of medical insurance professional terms, and complete the pre-construction of the semantic encoding model.

[0087] In some embodiments, the prediction device inputs the initial time-series dataset into a pre-built semantic coding model according to the sample dimension (single insured user, single time step data). The model performs unified semantic representation on textual and structured information such as diagnosis and treatment descriptions, settlement types, and disease names in the data, and outputs a fixed-dimensional semantic vector corresponding to each sample. This achieves vector unification between unstructured and structured data, thereby solving the semantic representation problem of multi-source heterogeneous medical insurance data (such as medical record text and structured settlement data). Through a domain-adaptive semantic coding model, vector unification of different types of data is achieved.

[0088] Step S202: Analyze the correlation strength of semantic vectors of each insured user at different time steps to obtain time-series trend features.

[0089] It should be noted that time-series trend features refer to the characteristics used to mine and reflect the evolution of medical insurance data over time by analyzing the correlation strength of semantic vectors of insured users at different time steps. They can reflect the time-series anomalies in the medical treatment behavior of insured users.

[0090] A time step refers to a time unit obtained by segmenting the initial time series dataset according to its time dimension; it is the basic time granularity for conducting time series feature analysis. The aforementioned correlation strength refers to an indicator used to quantify the semantic similarity or correlation between two or more semantic vectors; it serves as a quantitative basis for mining time series trends.

[0091] In its implementation, the prediction device segments the initial time-series dataset according to a preset time granularity (e.g., day, week, month), determining the semantic vector sequence for each insured user at different time steps. It calculates the semantic vector association strength of the same insured user at adjacent and cross-time steps (e.g., using algorithms such as cosine similarity and Pearson correlation coefficient), quantifying the semantic association degree of data at different time steps. Based on the association strength data, it analyzes the evolution of semantic vectors over time, extracting temporal trend features, such as a continuous increase in semantic vector association strength (representing abnormal convergence of medical behavior), a sudden drop (representing a sudden change in medical behavior), and periodic fluctuations (representing abnormally high frequency of medical visits). This allows for the mining of the evolution patterns of insured users' medical insurance data from a time dimension, transforming discrete time-step data into continuous temporal trend features, and capturing temporal anomalies (such as sudden changes in medical behavior in a short period) that traditional static features cannot reflect.

[0092] Step S203: Analyze the semantic vector association strength of each entity based on the semantic vector to obtain entity association features.

[0093] It should be noted that entity association features refer to features obtained by analyzing the semantic vector association strength of three types of entities: insured users, medical institutions, and treatment items. These features can reflect abnormal relationships between entities and can indicate abnormal entity collaboration in medical insurance transactions. The aforementioned entities can be the core participants and service content in the medical insurance transaction scenario, specifically including insured users, medical institutions, and treatment items, which are the core objects for constructing entity associations.

[0094] In its implementation, the prediction device extracts three types of entities from the initial time-series dataset: insured users, medical institutions, and treatment items. It aggregates all semantic vectors corresponding to each entity (e.g., mean aggregation, attention-weighted aggregation) to obtain the global semantic vector of each entity. It calculates the association strength of the global semantic vectors between different entity pairs (insured user-medical institution, insured user-treatment item, medical institution-treatment item) to quantify the degree of semantic association between entities. Based on the association strength results, it extracts entity association features, such as the association strength between an insured user and a certain medical institution being far above the normal level (representing high-frequency visits to designated medical institutions) and the strong association between the semantic vector of an insured user and an unsuitable treatment item (representing that the treatment item does not match the condition). This enables the spatial association patterns of medical insurance data to be mined from the dimension of entity association, breaking through the feature limitations of a single user dimension and being able to identify abnormal collaboration patterns between entities (e.g., abnormal binding of an insured user with multiple medical institutions).

[0095] Step S204: Generate a feature subset for each insured user based on the time-series trend features and the semantic vector association strength.

[0096] It should be noted that a feature subset refers to a set of features formed by integrating the time-series trend features and entity association features of a single insured user. This set contains feature matrices with multiple dimensions. A feature matrix is ​​a matrix formed by structuring features of a certain dimension according to a preset row and column dimension, which can intuitively present the distribution and association relationships of features.

[0097] In its implementation, the prediction device can aggregate the extracted time-series trend features and entity association features according to the insured user dimension, ensuring that the feature set of each insured user contains information in both the time dimension and the entity association dimension. The aggregated features of each insured user are matrixed, for example, a time-series feature matrix is ​​constructed with time step as the row dimension and various features as the column dimension; an entity association feature matrix is ​​constructed with entity type as the row dimension and association strength as the column dimension, forming a user-specific feature subset containing multiple feature matrices. This achieves user-level aggregation and structuring of multi-dimensional features, and constructs a complete feature profile for each insured user.

[0098] Step S205: Determine the fraud feature weights of each feature matrix based on historical samples, and associate and combine each feature matrix with the fraud type based on the fraud feature weights to generate multiple fraud feature combinations. The historical samples include feature matrix samples of fraud labels and feature matrix samples of normal labels.

[0099] It should be noted that historical samples refer to a set of samples that includes feature matrices corresponding to past medical insurance transaction data and has been labeled as fraudulent or normal transactions, which are used to support the calculation of fraud feature weights.

[0100] Fraud feature weights can be an indicator that quantifies the contribution of each feature matrix to fraud detection. The higher the weight value, the stronger the ability of the feature matrix to distinguish fraud types.

[0101] Fraud feature combinations can be feature sets formed by associating and matching feature matrices of different priorities according to fraud types. Each combination corresponds to a specific type of fraudulent insurance behavior.

[0102] In some embodiments, the prediction device can clean and label historical case samples related to DRG / DIP settlement, ensuring that the samples simultaneously contain feature matrices with typical DRG / DIP fraud labels (such as low code high encoding, split hospitalization, fictitious consumables, etc.) and normal settlement labels, and divide the samples into training and validation sets proportionally; build a weight analysis model (e.g., a multi-head causal attention model), input the training set into the model, and the model calculates the attention score of each feature matrix from two dimensions: DRG / DIP fraud type (low code high encoding, high code low encoding, split hospitalization, fictitious treatment / consumables, low indication hospitalization, etc.) and causal level. Combined with the validation results of the validation set, the fraud feature weight of each feature matrix is ​​obtained. The higher the weight, the greater the contribution of the feature matrix to DRG / DIP fraud identification; prioritize the feature matrices based on the fraud feature weights, and prioritize the core feature matrices with high weights (such as main... Records of minor diagnostic code adjustments, hospital stay breakdown records, and matching degree of consumable usage and charges are associated with medium-weighted auxiliary feature matrices (such as new complication code records and admission criteria compliance status) according to different DRG / DIP fraud types. For example, the feature matrix of "frequent adjustments of primary and secondary diagnostic codes + no clinical evidence for complication codes" is associated with the "low code, high code" fraud type, and the feature matrix of "multiple hospitalizations for the same disease in a short period of time + repetitive treatment items" is associated with the "decomposed hospitalization" fraud type. This generates a fraud feature combination library covering various DRG / DIP fraud types. The feature contribution is accurately quantified through historical case samples and weight models, ensuring a strong correlation between feature combinations and DRG / DIP fraud types. The fraud feature combinations generated by weight have a clear DRG / DIP fraud orientation, which greatly improves the feature's ability to distinguish DRG / DIP fraud behavior.

[0103] Among them, "high-weighted classification" (high-weighted disease group) refers to classifying low-weighted cases into high-weighted disease groups. For example, "chronic obstructive pulmonary disease with exacerbation" is classified as "chronic obstructive pulmonary disease with acute lower respiratory tract infection." "Low-standard admission" (i.e., low-standard admission) refers to admitting patients who could be treated on the outpatient basis to the hospital. For example, admitting patients with simple conditions such as chronic gastritis or pelvic inflammatory disease, whose treatment is mainly oral medication, to the hospital.

[0104] Step S206: Perform feature filtering on the fraud feature combination to obtain key time-series feature combinations.

[0105] It should be noted that the key time-series feature combination can be a feature set that has been filtered through multiple dimensions and has strong fraud detection, business compliance and time-series stability. It is the core feature that supports subsequent model training.

[0106] Understandably, feature selection can be carried out by filtering fraud feature combinations layer by layer according to multiple constraints such as business rationality, temporal stability, and low redundancy, with the aim of retaining high-value fraud identification features.

[0107] In some embodiments, the prediction device can generate constraints based on business specifications such as medical insurance treatment processes and DIP / DRG grouping rules, and eliminate fraudulent feature combinations that do not conform to business logic (such as a combination where the patient's primary diagnosis is hypertension and the associated complication is anemia) to obtain a first feature combination; analyze the temporal fluctuation of each feature in the first feature combination, eliminate features with irregular temporal fluctuations and poor stability, and retain feature combinations with clear temporal patterns to obtain a second feature combination; calculate the mutual information value between each feature in the second feature combination, eliminate features with excessively high mutual information values ​​(representing highly redundant features), and finally obtain a key temporal feature combination that takes into account business rationality, temporal stability and low redundancy. Through a multi-level screening mechanism, invalid, temporally unstable and redundant features are eliminated, which greatly reduces the interference of invalid features on subsequent model training.

[0108] This embodiment extracts features from three dimensions: semantics, temporal sequence, and entity association. This breaks through the limitations of traditional single-dimensional feature mining and constructs a comprehensive feature system covering "semantic expression, temporal evolution, and entity collaboration." It can capture the characteristic patterns of various covert fraudulent insurance behaviors. The resulting key temporal feature combination not only retains the core semantic and temporal attributes of medical insurance data but also has a clear ability to distinguish fraudulent behaviors. This provides solid feature support for the accurate training of the fraud risk temporal prediction model and ensures the fraud identification effect of the overall solution.

[0109] Furthermore, in order to accurately detect potential fraudulent activities, in one embodiment, step S205 further includes:

[0110] Step S2051: Construct a weighted analysis model based on historical samples.

[0111] It should be noted that, referring to Figure 2 , Figure 2 This is a schematic diagram of the weighted analysis model. The weighted analysis model includes an input layer, a multi-head causal attention layer, and a fusion layer. The input layer is configured to map the input feature matrix into a causal feature vector based on an intervention function. The multi-head causal attention layer includes multiple causal attention heads, including a fraud type attention head and a causal hierarchy attention head. The fraud type attention head is configured to calculate the attention score of the causal feature vector based on the fraud type dimension. The causal hierarchy attention head is configured to calculate the attention score of the causal feature vector based on the causal hierarchy dimension. The fusion layer is configured to fuse the outputs of the multiple causal attention heads and output the initial feature weights.

[0112] It should be noted that the intervention function is a function in the input layer used to eliminate confounding variables and extract the direct causal relationship between features and fraudulent behavior, enabling the transformation of features from statistical correlation to causal correlation. The causal attention head is the basic computational unit of the multi-head causal attention layer. In this embodiment, the fraud type attention head and the causal hierarchy attention head are used to quantify the fraud identification contribution of features from different dimensions.

[0113] Among them, the fraud type attention head is used to calculate the attention score of the causal feature vector from the fraud type dimension, which measures the feature's ability to distinguish a specific fraud type.

[0114] The causal hierarchy attention head is used to calculate attention scores from the causal hierarchy dimension of "feature-intermediate behavior-fraudulent result", measuring the contribution of a feature in the causal chain.

[0115] In its implementation, the prediction device first standardizes historical samples, unifying the dimensions and units of the feature matrix, and then divides them into training, validation, and test sets according to a preset ratio. Next, it classifies and labels fraud tags in the samples, clarifying the correlation between feature matrices for each fraud type, providing a fully labeled dataset for model training. The training set is then input into the original model, and backpropagation optimizes the model parameters. The validation set is used to verify the model's weight prediction accuracy, and the test set is used to verify the model's generalization ability, ensuring that the model stably outputs reliable initial feature weights, thus obtaining the weight analysis model.

[0116] In some embodiments, the input layer has a built-in intervention function (such as a backdoor adjustment function in causal intervention) that maps the input feature matrix to a causal feature vector. Specifically, the intervention function eliminates the influence of confounding factors in the feature matrix, separates the direct causal relationship between the feature and the fraudulent behavior, and realizes the transformation from "statistical correlation features" to "causal correlation features".

[0117] In some embodiments, a multi-head causal attention layer is constructed, which includes a fraud type attention head and a causal hierarchy attention head. The fraud type attention head presets feature anchors for various fraud types and obtains the attention score for the fraud type dimension by calculating the similarity between the causal feature vector and the anchor. The causal hierarchy attention head calculates the contribution ratio of the causal feature vector in different levels according to the causal hierarchy of "feature-intermediate behavior-fraud result" and obtains the attention score for the causal hierarchy dimension.

[0118] In some embodiments, the fusion layer uses a weighted average or attention-weighted approach to fuse the attention scores output by the two causal attention heads, eliminating the bias of a single-dimensional score, and finally outputting the initial feature weights corresponding to each feature matrix.

[0119] It is understandable that this embodiment, by constructing a weight analysis model, achieves the evaluation of feature weights from a causal dimension, breaking through the limitation of traditional statistical weights that only measure surface correlations; the multi-head attention mechanism quantifies feature contribution from two dimensions, improving the comprehensiveness of weight evaluation; and the introduction of the intervention function ensures the causal reliability of feature weights.

[0120] Step S2052: Input the feature subsets of each insured user into the weight analysis model to obtain the initial feature weights.

[0121] In practice, the prediction device can perform format verification and dimension alignment on the feature subsets of each insured user to ensure that the dimension of the feature matrix within the feature subset is consistent with the input requirements of the weight analysis model, thus eliminating the interference of format errors on weight calculation. The preprocessed feature subsets are then input into the trained weight analysis model in batches according to the insured user dimension. The model will sequentially perform operations such as causal vector mapping, two-dimensional attention score calculation, and score fusion on each feature matrix, and finally output the initial feature weights corresponding to each feature matrix.

[0122] Step S2053: Perform fact analysis on the historical samples to obtain the fact analysis results.

[0123] It should be noted that the fact analysis results include the change in fraud probability caused by feature loss, reflecting the actual fraud detection value of the features.

[0124] In some embodiments, the prediction device can conduct feature ablation experiments on historical samples, that is, sequentially remove a certain feature matrix from the sample, and then input the sample after removal into the benchmark fraud detection model to calculate the fraud probability before and after feature removal; statistically analyze the magnitude and direction of the change in fraud probability after removal of each feature matrix to obtain the fraud probability change value corresponding to that feature. If the fraud detection probability decreases significantly after removal, it indicates that the feature has a high actual contribution to fraud detection; if the change is small, it indicates a low contribution. The fraud probability change values ​​of all feature matrices are classified and summarized according to feature type to form factual analysis results, while also annotating the differentiated impact of feature loss on the identification of different fraud types.

[0125] Step S2054: Adjust the initial feature weights based on the fact analysis results to obtain the fraud feature weights of each feature matrix.

[0126] In practical implementation, the prediction device can formulate weight adjustment rules based on the results of fact analysis. If the fraud probability of a certain feature changes significantly (i.e., the missing feature has a significant impact on recognition), its initial feature weight is increased by a preset ratio; if the change is small, the weight is appropriately decreased; if the change is negative (the missing feature increases the recognition probability, indicating that it is a interference feature), the weight is directly reduced to the low contribution range. The initial feature weights are batch corrected according to the adjustment rules, and the adjusted weights are normalized to ensure that the sum of the weights of each feature matrix meets the model input requirements.

[0127] In some embodiments, the prediction device may extract a portion of the samples and substitute the adjusted weights into the baseline model to verify the change in fraud detection performance after weight adjustment, ensuring that the adjusted weights have higher practical application value.

[0128] Step S2055: Perform priority analysis on the feature matrix based on the fraud feature weights to determine the core feature matrix and auxiliary feature matrix in the feature matrix.

[0129] It should be noted that the core feature matrix can be a feature matrix with high fraud feature weights that plays a decisive role in fraud identification and is the core indicator set for identifying fraudulent insurance behavior; the auxiliary feature matrix can be a feature matrix with moderate fraud feature weights that complements the core feature matrix and can enhance the ability to identify covert fraudulent behavior, thereby uncovering potential fraudulent behavior.

[0130] In practical implementation, the prediction device can set weight threshold ranges based on business needs and weight distribution. For example, feature matrices with weights higher than a preset high threshold are classified as core feature matrices, those with weights in the middle threshold range are classified as auxiliary feature matrices, and those with weights lower than the low threshold are judged as low-value feature matrices and removed.

[0131] In some embodiments, the prediction device can sort all feature matrices in descending order according to fraud feature weights, complete the priority division by combining the weight threshold range, and mark the adaptability of each feature matrix to different fraud types, clarifying the main fraud identification direction of the core feature matrix; cross-validate the priority division results with the experience judgment of medical insurance supervision experts, correct unreasonable division results, and ensure that the division of core and auxiliary feature matrices conforms to actual supervision needs.

[0132] Step S2056: Associate and combine the core feature matrix and the auxiliary feature matrix with the fraud type to obtain a fraud feature combination.

[0133] In some embodiments, the prediction device can first sort out common fraud types in the DRG / DIP medical insurance field (such as under-reporting, under-standard admission, splitting hospitalization, falsely invoicing medical treatment items, and inflating DIP / DRG scores, etc.), and establish exclusive feature requirement labels for each fraud type: the core feature matrix and the auxiliary feature matrix are matched according to the feature requirement labels of the fraud type. For example, the core feature matrix of "falsely invoicing medical treatment items" is associated with the fraud type of "inflating and over-reporting" and matched with the auxiliary feature matrix of "rational drug use" to form a feature combination for the fraud type, that is, the fraud feature combination.

[0134] In some embodiments, the prediction device can establish a feature combination index for each fraud type, classify and store the matched fraud feature combinations according to fraud type, and label the weight ratio and role of each feature matrix in the combination, forming a clearly structured and quickly searchable fraud feature combination library.

[0135] This embodiment uses a weighted analysis model to perform causal attention analysis and adjusts the initial feature weights output by the model in conjunction with the results of fact analysis. This achieves a dual weight determination mechanism, eliminating the one-sidedness of a single model evaluation. The feature weights are supported by causal logic and are aligned with actual fraud detection scenarios, significantly improving the accuracy of the weights. By dividing the feature matrix into a core feature matrix and an auxiliary feature matrix, the priority division clarifies the functional positioning of the core and auxiliary feature matrices. This highlights the leading role of high-value features while leveraging the supplementary value of auxiliary features, avoiding the waste of feature resources and enabling the discovery of potential fraudulent behavior.

[0136] Furthermore, to ensure business rationality and improve timing stability, in one embodiment, step S206 further includes:

[0137] Step S2061: Generate reasonable business constraints based on medical insurance rule information, and perform reasonableness screening on the fraud feature combination based on the reasonable business constraints to obtain the first feature combination.

[0138] It should be noted that the medical insurance rules information can be a collection of regulations issued by medical insurance regulatory agencies, such as the requirements of the "three medical insurance catalogs" (drugs, treatment items, and medical service facilities), covering core content such as DIP / DRG grouping rules and treatment item adaptation specifications, and is used to determine the rationality of feature combination business.

[0139] The aforementioned reasonable business constraints refer to quantifiable and verifiable screening conditions based on the transformation of medical insurance rule information, used to eliminate invalid fraud feature combinations at the business logic level.

[0140] The aforementioned first feature combination refers to the set of fraud feature combinations that meet the compliance requirements of medical insurance business after being screened under reasonable business constraints, eliminating invalid feature combinations at the business level.

[0141] In some embodiments, the predictive device can integrate various business rules in the medical insurance field, including DIP / DRG disease grouping rules, medical insurance treatment item adaptation specifications, medical treatment process logic (such as the basic time sequence process of "registration-treatment-settlement"), medical insurance reimbursement scope definition standards, etc., to form a structured medical insurance rule information database; then the rule information is transformed into quantifiable and verifiable reasonable business constraints.

[0142] In some embodiments, the prediction device can traverse all fraud feature combinations to be screened, compare the feature dimensions of each combination with reasonable business constraints one by one, eliminate combinations containing features that violate the medical insurance business logic (such as the feature combination of "insured persons who seek medical treatment in other places without filing but enjoy the reimbursement ratio for hospitalization in other places"), and collect feature combinations that meet all business constraints to obtain the first feature combination. This achieves the initial filtering of feature combinations from the perspective of medical insurance business logic, effectively eliminating invalid combinations that are technically related but have no practical significance at the business level, avoiding interference from business violation features in subsequent model training, and ensuring that the feature combination is compatible with the actual business scenario of medical insurance supervision.

[0143] Step S2062: Analyze the time fluctuation information of each feature based on the time step of each feature in the first feature combination, and perform time series stability screening on the first feature combination based on the time fluctuation information to obtain the second feature combination.

[0144] It should be noted that time fluctuation information refers to the changes in the value of a feature at different time steps. It can be quantitatively characterized by indicators such as variance and coefficient of variation, and is the core data for determining the time series stability of a feature.

[0145] Temporal stability screening refers to the screening process based on the temporal fluctuation information of features, eliminating features with chaotic temporal patterns and retaining features with stable temporal patterns.

[0146] The second feature combination refers to the set of fraud feature combinations that, after being screened for temporal stability, possess clear and stable temporal evolution patterns.

[0147] In practical implementation, the prediction device can define the time step granularity of time series analysis (such as preset time units such as day, week, month, etc.), and then extract the value data of each feature in the first feature combination at different time steps; then select appropriate time fluctuation quantification indicators, including variance, coefficient of variation, and time series trend fit, to measure the degree of fluctuation of features in the time dimension. For example, the coefficient of variation can reflect the relative dispersion of feature values, and the trend fit can determine whether the feature has a stable time series evolution pattern.

[0148] Then, based on the temporal characteristics of medical insurance fraud, a temporal stability threshold is set (such as a preset coefficient of variation critical value or trend fit critical value). The temporal fluctuation information of each feature in the first feature combination is analyzed, and features with time fluctuations exceeding the threshold (i.e., chaotic temporal patterns and poor stability) are eliminated. For example, if a feature has a large fluctuation range in value at consecutive time steps and no clear trend, it is determined to be a temporally unstable feature. The retained temporally stable features are then recombined to obtain the second feature combination.

[0149] Step S2063: Analyze the multicollinearity among the features in the second feature combination, determine the mutual information value among the features, and perform feature redundancy screening on the second feature combination based on the mutual information value to obtain the key time series feature combination.

[0150] It should be noted that multicollinearity refers to a strong linear correlation between multiple features in a feature combination, which leads to feature information duplication and reduces the model's feature learning efficiency. Mutual information value is a quantitative indicator used to measure the degree of information overlap between two features. It can accurately characterize the redundancy relationship between features; a higher value indicates a higher degree of feature redundancy.

[0151] It should be noted that the key temporal feature combination refers to the core feature combination that has business compliance, temporal stability and high information density after being screened through three layers of criteria: business rationality, temporal stability and low redundancy. It is the key feature resource that supports subsequent model training.

[0152] Understandably, the prediction device can perform multicollinearity analysis on all features in the second feature combination. If the variance inflation factor of a certain feature exceeds the preset critical value, it is determined that it has a strong linear correlation with other features, and its redundancy needs to be further verified.

[0153] For the suspected redundant features and all feature pairs initially diagnosed, the mutual information value between each pair of features is calculated. The mutual information value is used to quantify the degree of information overlap between two features. The higher the value, the higher the information redundancy between features and the smaller the incremental contribution to fraud detection.

[0154] In practical implementation, the prediction device can set a critical threshold for mutual information value. For feature pairs whose mutual information value exceeds the threshold, features with higher fraud feature weights and greater contribution to fraud identification are retained first, while redundant features are eliminated. After completing the redundancy screening of all features, the remaining high independence and high information density features are recombined to finally obtain the key time series feature combination, thereby effectively eliminating information redundancy between features and reducing the computational complexity of subsequent model training.

[0155] This embodiment achieves feature filtering from three core dimensions—business rationality, temporal stability, and low redundancy—by using a three-tiered progressive screening approach. This significantly improves the overall quality of the final feature combination. Business rationality screening ensures that the feature combination aligns with the actual regulatory scenarios of medical insurance, temporal stability screening adapts to the technical requirements of temporal risk prediction, and redundancy screening balances the efficiency and accuracy of model training, thus realizing a deep synergy between business requirements and technical implementation.

[0156] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the DIP and DRG medical insurance risk prediction method of the present invention.

[0157] Based on the above embodiments, in this embodiment, step S30 further includes:

[0158] Step S301: Perform fraud behavior logic analysis based on the combination of key time-series features to generate fraud feature chains.

[0159] It should be noted that a fraud feature chain is a logical link of fraudulent behavior composed of multiple key fraud temporal features connected in series or parallel. The fraud feature chain is used to represent fraudulent behavior as a complete logical feature sequence; it is a concrete mapping of fraudulent behavior at the feature level.

[0160] In this context, a sequential link refers to a link structure where features are connected in the chronological order of fraudulent behavior, reflecting the sequential dependency between features and reconstructing the gradual occurrence of fraudulent behavior. A parallel link refers to a link structure where multiple features do not have a clear order but must exist simultaneously to determine fraud, reflecting the collaborative coexistence relationship between features and covering fraud patterns involving multiple linked features.

[0161] It should be noted that fraudulent behavior is a single, concrete violation in the medical insurance scenario (a specific "event"), such as "assigning low-weight cases to high-weight disease groups", "falsely ordering CT examination items for insured persons who have not sought medical treatment", and "submitting reimbursement materials for the same hospitalization repeatedly".

[0162] Fraud types are abstract classifications of multiple similar fraudulent behaviors (a macro-level "category"). For example, behaviors such as "inflating the absence of complications" and "falsely reporting the degree of disease verification" are collectively classified as "high-level fraud," while behaviors such as "falsely ordering examination items" and "prescribing excessive medications" are classified as "fraudulent medical treatment."

[0163] It is understandable that this embodiment transforms the originally discrete key time-series features into logically related feature chains, breaking the isolation of a single feature and realizing the upgrade from feature points to feature chains; at the same time, through serial and parallel topology design, it accurately restores the feature association logic of different types of fraud behavior.

[0164] In practical implementation, the prediction device can classify key temporal feature combinations according to the corresponding fraud types, and clarify the core key fraud temporal features corresponding to different fraud types (such as low-standard admission, split hospitalization, and high DIP / DRG scores). Then, for each type of fraud, the fraud behavior logic is decomposed, and the sequential dependence or parallel coexistence relationship between features is analyzed.

[0165] For features with sequential dependencies, a serial approach is used to construct links, such as a progressive fraud feature link of "using a social security card other than one's own → seeking medical treatment in another location without registration → high-frequency small-amount settlement," to reconstruct the temporal evolution of fraudulent behavior. For features without a clear sequence but that need to appear simultaneously, a parallel approach is used to construct links, such as a feature link where "off-label treatment projects" and "over-prescription of drugs and consumables" coexist, covering the feature association patterns of collaborative fraud. Finally, the serial / parallel features corresponding to all fraud types are integrated to form a multi-type fraud feature link set.

[0166] Step S302: Perform time-series behavior analysis on the fraud feature links based on the medical insurance treatment process, and adjust the fraud feature links based on the time-series behavior analysis results to obtain the target feature links.

[0167] It should be noted that the medical insurance treatment process can be the standard business process for insured persons within the medical insurance system to receive medical services. It can include the diagnosis and treatment and settlement steps in different scenarios such as outpatient and inpatient services, and is used as a business benchmark to determine the rationality of the timing of characteristic links.

[0168] For example, the medical insurance process may include the basic outpatient process of "registration → consultation → examination / treatment → prescription of medicine / consumables → settlement", as well as the inpatient process of "admission registration → diagnosis → development of treatment plan → implementation of treatment → discharge settlement".

[0169] It should be noted that temporal behavior analysis refers to the process of comparing fraud-related feature chains with the standard medical insurance treatment process, analyzing the compliance of the feature temporal arrangement, and identifying temporal anomalies, in order to correct the business adaptability of the chain. The target feature chain refers to the feature chain that, after being adapted and adjusted to the medical insurance treatment process, possesses both temporal business compliance and fraud anomaly representation capabilities.

[0170] It is understandable that this embodiment eliminates temporal logic errors in the feature chain by adapting and verifying the medical insurance treatment process, ensuring that the chain conforms to the actual medical insurance business scenario and avoiding the model learning invalid feature associations that are detached from the business; at the same time, it retains the temporal abnormality attributes of fraudulent behavior, so that the feature chain has both business rationality and accurate fraud identification.

[0171] In some embodiments, the predictive device first aggregates standard medical treatment process specifications in the medical insurance field, and simultaneously incorporates the diagnosis and treatment behavior specifications corresponding to DIP / DRG groupings to form a standardized medical insurance treatment process benchmark; it then compares the initially generated fraud feature links with the standard medical insurance treatment process to analyze whether the temporal arrangement of features in the links conforms to the actual medical treatment logic (e.g., identifying feature links that violate the medical treatment process, such as "settling before registration" or "prescribing high-priced drugs before treatment"); for links with disordered temporal logic, it adjusts the order of features according to the standard process; for abnormal feature nodes that exceed the normal process (e.g., completing 3 types of large-scale examinations within 1 day of hospitalization), it retains their abnormal attributes and marks them with temporal abnormality tags; it then aggregates the feature links that have been adjusted to comply with the temporal logic and retain the fraud abnormality attributes to form target feature links, ensuring that the links not only conform to the medical insurance business process but also accurately represent the temporal abnormality of fraudulent behavior.

[0172] Step S303: Perform causal analysis based on the target feature links to construct a fraud causal time sequence graph structure.

[0173] It should be noted that the nodes of the fraud causal time sequence graph structure include time sequence feature nodes, fraud behavior nodes, fraud type nodes, and intervention nodes, and the edges of the fraud causal time sequence graph structure include causal driving edges, behavior classification edges, time sequence driving edges, and intervention adjustment edges.

[0174] In this graph, the temporal feature nodes are the foundational nodes representing key temporal features of fraud. Fraud behavior nodes are intermediate nodes representing specific fraudulent insurance claims, serving as the causal drivers for the temporal feature nodes. Fraud type nodes are top-level nodes representing categories of fraudulent behavior, used for classifying and grouping fraudulent activities. Intervention nodes are moderating nodes representing external policies, regulations, and other intervention factors, used to adjust the strength of causal relationships.

[0175] Causal-driven edges connect temporal feature nodes to fraud behavior nodes, representing the direct causal triggering relationship between features and fraud behavior. Behavior classification edges connect fraud behavior nodes to fraud type nodes, implementing a category mapping from fraud behavior to fraud type. Temporal-driven edges connect temporal feature nodes, representing the temporal dependency relationship between features. Intervention-modulation edges connect intervention nodes to causal-driven edges, representing the moderating effect of external intervention on the strength of causal associations.

[0176] In its implementation, the predictive device performs causal graph analysis based on the target feature links. First, it identifies four types of core nodes in the graph structure: key fraud time-series features are defined as time-series feature nodes; specific fraudulent behaviors (such as fabricating medical treatment or tampering with test results) are defined as fraud behavior nodes; the categories of fraudulent behaviors (such as splitting hospitalizations or overstating scores) are defined as fraud type nodes; and external influencing factors such as adjustments to medical insurance policies and upgrades to regulatory measures are defined as intervention nodes. Then, causal inference algorithms (such as propensity score matching and backdoor adjustment) are used to mine the causal relationships between nodes.

[0177] Then, four types of edges are defined for the relationships between nodes: edges pointing from temporal feature nodes to fraudulent behavior nodes are causal-driven edges, representing the triggering effect of features on fraudulent behavior; edges pointing from fraudulent behavior nodes to fraud type nodes are behavior classification edges, realizing the classification of fraudulent behavior into fraud type; edges between temporal feature nodes arranged according to the medical treatment process are temporal-driven edges, reflecting the temporal dependence of features; edges pointing from intervention nodes to causal-driven edges are intervention-regulating edges, representing the regulating effect of external intervention on the causal relationship between "feature-fraudulent behavior"; finally, all nodes and edges are integrated according to the association logic to construct a complete fraud causal temporal graph structure.

[0178] Step S304: Train the pre-built original model based on the fraud causal time series graph structure and the combination of key time series features to obtain a fraud risk time series prediction model.

[0179] It is understood that the pre-built original model in this embodiment includes a dual-branch feature encoding module (the first branch is a graph neural network GNN and the second branch is a bidirectional long short-term memory network Bi-LSTM), a feature fusion module, and a prediction output module. The GNN branch is adapted to the learning of graph structure data, and the Bi-LSTM branch is adapted to the learning of temporal features.

[0180] In its implementation, the prediction device can preprocess the adjacency matrix of the fraud causal time series graph structure with the time series sequence of node features and key time series features. The preprocessed data is then divided into training set, validation set and test set according to a preset ratio to provide data support for model training.

[0181] In some embodiments, the prediction device can input training set data into the original model, with the GNN branch learning the structural associations and causal driving relationships of the fraud causal time series graph, and the Bi-LSTM branch learning the bidirectional temporal dependencies of key temporal features; setting a composite loss function (such as cross-entropy loss combined with causal consistency loss), and optimizing the model parameters through backpropagation; using the validation set to monitor the model training process, adjusting hyperparameters such as the learning rate and regularization coefficient to avoid model overfitting; finally, verifying the model's generalization ability through the test set, and outputting the fraud risk time series prediction model after the model performance meets the standards.

[0182] This embodiment transforms discrete key temporal features into a graph structure with business logic, temporal dependencies, and causal relationships through a progressive process of feature link construction, process adaptation, and causal graph construction. This significantly improves the representation ability and information density of the features, thereby enhancing the model's multi-dimensional feature learning ability, strengthening the model's business adaptability and interpretability. By integrating medical insurance treatment process norms and causal relationship logic, it not only makes the prediction results fit the business scenario of medical insurance supervision, but also allows the core feature links of risk to be traced through the causal time series graph.

[0183] Furthermore, to reduce noise interference, improve the accuracy of fraud risk prediction, and achieve deep integration of multi-dimensional features, referencing Figure 4 , Figure 4 This is a schematic diagram of the structure of a fraud risk time series prediction model in one embodiment. The fraud risk time series prediction model also includes a betweenness center module, a feature fusion module, and a prediction output module.

[0184] It should be noted that the betweenness center module is a preprocessing module in the fraud risk time series prediction model used to refine the fraud causal time series graph structure and key time series feature combinations. It is used to calculate the betweenness center value and complete the screening and optimization of features and graph nodes.

[0185] Betweenness centrality is a quantitative indicator that measures the pivotal role of temporal feature nodes in information transmission within a fraudulent causal time-series graph. It comprehensively reflects a node's control over information flow within the graph and the coreity of features in fraud identification. The betweenness-optimized fraudulent causal time-series graph structure refers to a graph structure that retains core causal relationships after removing low-betweenness nodes and their corresponding edges, possessing a more streamlined topology and stronger fraudulent information representation capabilities. The betweenness-optimized key temporal feature combination refers to a feature set that retains core fraudulent temporal features after removing low-betweenness redundant features, reducing interference from invalid features.

[0186] In its implementation, the betweenness center module determines the node characteristics (such as the fraud contribution and temporal correlation strength) of the temporal feature nodes in the fraud causal temporal graph structure and calculates the betweenness center value of each temporal feature node using the Brandes algorithm. Specifically, the Brandes algorithm quantifies the betweenness center value by counting the number of times a node is a necessary point on the shortest path between any two nodes, combined with the node feature weights, thereby measuring the node's role as an information transmission hub in the graph structure. Simultaneously, it assigns betweenness center values ​​consistent with the corresponding graph nodes to features in key temporal feature combinations.

[0187] In some embodiments, the prediction device can set a betweenness optimization threshold range based on the business needs of medical insurance fraud identification and the distribution pattern of betweenness centrality values, and determine nodes / features with betweenness centrality values ​​higher than a preset high threshold as core nodes / core features; otherwise, they are determined as low-value nodes / redundant features.

[0188] Understandably, the betweenness center module removes low betweenness temporal feature nodes and their corresponding associated edges from the fraud causal time series graph structure, while retaining the complete association logic of the core nodes, thus forming a betweenness-optimized fraud causal time series graph structure. For key time series feature combinations, it simultaneously removes redundant features with low betweenness, retains the temporal association relationship of the core features, and outputs the betweenness-optimized key time series feature combinations, achieving simultaneous purification of graph structure and feature combinations.

[0189] In the specific implementation, the adjacency matrix and node feature matrix of the fraud causal time sequence graph structure optimized by betweenness are input into the first branch feature encoding structure. The first layer network performs initial embedding of nodes. The middle layer aggregates the features of the neighboring nodes of each node through a multi-head attention mechanism to strengthen the information weight of the core nodes. The last layer outputs the high-dimensional embedding vector of each node, and then integrates the embedding information of all nodes through a global pooling operation to generate a graph structure embedding vector that represents the association logic of the entire graph structure.

[0190] The betweenness center module organizes the key temporal features after betweenness optimization into a temporal feature sequence along the time step dimension, and inputs it into the second branch feature encoding structure. In the second branch feature encoding structure, the forward LSTM learns the temporal progression of features from early to late stages (such as the positive fraud chain of "non-personal social security card use → high frequency of cross-institutional medical treatment"), and the backward LSTM learns the reverse verification logic of features from late to early stages (such as "high DIP score" to reverse trace the rationality of "false invoicing of medical items"). Finally, the hidden state of the last moment of the bidirectional LSTM is concatenated to generate a temporal-aware embedding vector representing the temporal dependency pattern.

[0191] In its implementation, the feature fusion module can learn the fusion weights of graph structure embedding vectors and time-aware embedding vectors based on validation set data (e.g., assigning preset weights to graph structure embedding vectors and corresponding weights to time-aware embedding vectors based on their contributions to the fraud detection task). The two types of vectors are then concatenated according to their weights to generate initial fusion features. A feature importance evaluation algorithm (e.g., gradient-based feature importance analysis) is introduced to calculate the importance weight of each dimension in the initial fusion features for the fraud detection task. The higher the weight value, the greater the contribution of that dimension to fraud detection. An attention calibration layer is then built, and the initial fusion features are input into this layer. Based on the calculated importance weights, differentiated attention coefficients are assigned to different feature dimensions. Features with high contribution dimensions are weighted and enhanced, while features with low contribution dimensions are weakened. Finally, the target fusion features after attention calibration are output.

[0192] In some embodiments, the prediction output module is constructed based on a multilayer perceptron (MLP) network, which includes an input layer, multiple hidden layers and an output layer. The dimension of the input layer is consistent with the dimension of the target fused features. The hidden layers use the ReLU activation function to achieve non-linear transformation of features. At the same time, Dropout and L2 regularization mechanisms are introduced in the hidden layers to suppress overfitting.

[0193] In some embodiments, the feature fusion module inputs the target fused features into the input layer of a prediction network constructed based on an MLP network. After passing through the feature transformation and regularization processes of each hidden layer, the shallow fused features are transformed into high-level abstract features with deep fraud association logic. The high-level abstract features are then input into the MLP output layer, and the probability distribution information of low-risk, medium-risk, and high-risk levels corresponding to each transaction behavior in the target area is calculated through the Softmax activation function. Based on a preset risk probability threshold (e.g., if the high-risk probability is higher than the preset threshold, it is judged as high-risk), the risk level of each transaction behavior is determined in conjunction with the medical insurance supervision business rules. Finally, the risk level and corresponding probability are integrated into a dynamic fraud risk prediction result, which is output to the regulatory end through an interface or visualization platform.

[0194] Understandably, this embodiment refines core features through betweenness optimization, learns fraud patterns in structural and temporal dimensions through dual-branch encoding, and completes dimensional complementarity and core feature enhancement through a feature fusion module. Ultimately, it forms a comprehensive feature system that combines spatial correlation, temporal trend, and feature contribution analysis, breaking through the limitations of traditional models' single-dimensional feature learning. Betweenness optimization reduces noise interference, dual-branch encoding ensures the comprehensiveness of feature representation, attention calibration enhances the value of core features, and the deep transformation of the prediction output module achieves accurate risk classification. This ensures that the fraud risk temporal prediction model can effectively identify covert and collaborative DIP / DRG medical insurance fraud, reducing the false negative rate and false positive rate.

[0195] Furthermore, to improve data quality and provide high-quality data for subsequent dynamic feature extraction and time-series risk prediction, thereby improving the accuracy of fraud risk prediction, in one embodiment, step S10 further includes:

[0196] Step S101: Preprocess the locally collected multi-source heterogeneous data to obtain the first dataset. The preprocessing includes data cleaning, standardization, and time-series alignment.

[0197] In practice, the predictive device can collect local multi-source heterogeneous data such as medical insurance settlement data, medical institution operation data, and patient medical record data through local database interfaces, medical service system terminals, etc. At the same time, it adds a collection timestamp and data source tag to each data, completes the initial screening of data, and removes invalid data with completely incorrect format or unknown source.

[0198] In its implementation, the predictive equipment can perform cleaning operations on the data after initial screening. This includes removing duplicate data (such as duplicate settlement records for the same insured person at the same time) using preset rules, filling in missing data using interpolation (such as the historical average of similar treatments for the same insured person) or business rules (e.g., missing treatment item codes in some medical records can be supplemented through medical record text mapping), and filtering out abnormal data (such as invalid documents with settlement amounts far exceeding the normal range). The cleaned data is then standardized in format and content, converting unstructured data (such as free text medical records and medical order image reports) into structured data (e.g., mapped to diagnostic codes and standardized medical insurance treatment item codes). Numerical features (such as settlement amount and frequency of visits) are normalized / standardized to eliminate differences in units of measurement. Categorical features (such as medical institution level and insured person type) are uniquely encoded or labeled to ensure a unified data format. Finally, according to preset time granularity (such as day, week, month), all local data is associated with a unified time dimension; for example, all medical records and settlement data for an insured person on a single day are aggregated to the same "day". At each time step, the first dataset with temporal attributes is formed.

[0199] Step S102: Send data acquisition requests to multiple data sharing nodes so that each data sharing node preprocesses the locally collected multi-source heterogeneous data, encrypts the preprocessed second dataset, and responds to the data acquisition request based on the encrypted data.

[0200] It should be noted that data sharing nodes can be external data holders with relevant medical insurance data sharing qualifications and capable of providing multi-source heterogeneous data, such as medical insurance agencies in other regions or cooperative medical institutions. The second dataset refers to the standardized dataset generated by each data sharing node after local preprocessing; it is the core data source for cross-institutional data aggregation. Encrypted data refers to the ciphertext data formed after the second dataset has been processed by a preset encryption mechanism, ensuring the security of cross-institutional data transmission.

[0201] In some embodiments, the predictive device can select external nodes with data sharing qualifications (such as off-site medical insurance agencies, cooperative medical institutions, regional medical data centers, etc.) based on business needs, and send standardized data acquisition requests to these nodes, specifying the type, time range, format requirements, and data protection specifications of the required data in the request. After receiving the request, each data sharing node performs a cleaning, standardization, and time-series alignment process on its own multi-source heterogeneous data locally, consistent with local preprocessing, to generate a second dataset that meets the request requirements, ensuring data format compatibility with the first dataset.

[0202] In some embodiments, the shared node uses a preset encryption mechanism (such as isomorphic encryption or AES symmetric encryption) to encrypt the second dataset. During encryption, a node identity identifier and an encryption timestamp are added to the data. The encrypted data, node identity information, response time, and other content are then integrated into response information and fed back to the data requester to achieve secure data transmission. This ensures the security of data transmission, avoids leakage of original data, and ensures format compatibility between cross-organizational data and local data.

[0203] Step S103: Obtain the response information of each data sharing node, and perform community aggregation on each data sharing node based on the response information to generate key communities and edge communities.

[0204] It should be noted that a critical community refers to a community formed by the aggregation of multiple first nodes, whose nodes provide data with high reliability and high quality. A peripheral community refers to a community formed by the aggregation of multiple second nodes, whose nodes provide data whose overall quality and credibility cannot meet the core data requirements. Specifically, the first nodes are data-sharing nodes whose response speed, historical data quality, and / or historical credibility meet preset conditions, while the second nodes are data-sharing nodes whose response speed, historical data quality, and historical credibility do not meet preset conditions. The response information refers to a comprehensive information package fed back by the data-sharing nodes, containing encrypted data, node identity information, response speed, etc., and may include the encrypted data, identity information, and response speed of the data-sharing node's response.

[0205] It is understandable that this embodiment achieves hierarchical filtering of data sharing nodes through community aggregation. The identification of key communities accurately locks in high-quality data sources, while the differentiation of peripheral communities avoids the interference of low-quality data on the overall dataset. This ensures the overall quality of subsequent cross-institutional data aggregation from the data source level, while improving the automation and accuracy of data filtering.

[0206] In some embodiments, the predictive device can determine the evaluation dimensions and preset conditions of the shared node, including response speed (such as whether the response time is lower than a preset threshold), historical data quality (such as whether the completeness and accuracy of historically provided data meet the standards), and historical credibility (such as compliance indicators such as whether historical data has abnormal behavior and whether the data matches the actual business), and set quantitative scoring standards for each dimension.

[0207] In some embodiments, the predictive device quantifies and scores each data-sharing node based on response information, identifying nodes with fast response speeds, high historical data quality, and acceptable historical reliability as first nodes; and identifying nodes with slow response speeds, unacceptable historical data quality, and questionable historical reliability as second nodes. A community discovery algorithm (such as the Louvain algorithm) is used to aggregate all shared nodes into communities, grouping multiple first nodes into key communities where node data possesses high quality and high reliability; and grouping multiple second nodes into peripheral communities where the overall quality and reliability of node data are insufficient, thereby achieving hierarchical management of nodes.

[0208] Step S104: Decrypt the encrypted data of the responses from each first node in the key community to obtain the second dataset.

[0209] In some embodiments, the prediction device may establish a key management mechanism with the data contributing nodes in advance, and only grant decryption permissions to operating terminals with legitimate decryption permissions. Before decryption, the device verifies the identity information of the first node in the key community and the integrity of the encrypted data to confirm that the data has not been tampered with and that the node's identity is legitimate.

[0210] In the specific implementation, the prediction device uses a decryption algorithm (such as AES decryption or homomorphic decryption) that matches the encryption mechanism to decrypt the encrypted data fed back by each first node in the key community and restore it to readable standardized data. The decrypted data is then subjected to secondary verification to check whether the data format, time range, and feature dimensions meet the preset requirements. Invalid data with format errors or missing information after decryption is removed, and finally integrated to form the second dataset.

[0211] Step S105: Determine the node data weights based on the response speed, historical data quality, and historical reliability of each first node.

[0212] It should be noted that the node data weight refers to the comprehensive weight calculated based on the response speed, historical data quality, and historical reliability of the first node, which is used to quantify the proportion of reference value of different node data during aggregation.

[0213] In practical implementation, the predictive device can pre-construct a three-dimensional weighted evaluation system that includes response speed, historical data quality, and historical reliability. Basic weights are assigned to each dimension (e.g., historical data quality has the highest weight, followed by response speed). Based on this three-dimensional weighted evaluation system, the predictive device quantifies and scores each first node's response speed (e.g., shorter response time results in a higher score), historical data quality (e.g., historical data completeness and accuracy compliance rate), and historical reliability (e.g., historical data compliance and relevance to actual business), obtaining the dimension score for each node. Based on the dimension's basic weight and the node's dimension score, a weighted summation algorithm is used to calculate the comprehensive score for each first node. The comprehensive score is then normalized to obtain the node data weight corresponding to each first node; a higher weight indicates greater reference value for the node's data.

[0214] Step S106: Aggregate the first dataset and the second dataset based on the node data to obtain an initial time series dataset.

[0215] In practical implementation, the prediction device can first align the feature dimensions and time granularity of the first dataset and the second dataset to ensure that the feature names, encoding rules and time step divisions of the two datasets are completely consistent. For example, the encoding rule of "medical institution level" in the second dataset is unified to the standard of the first dataset, and the time step is unified to the "day" level granularity.

[0216] Then, the second dataset is weighted and integrated according to the node data weights of each first node. For example, if the data weight of a certain node is 0.3, its contribution to the aggregation is 30%. The weighted and integrated second dataset is then merged with the first dataset according to time steps to achieve the organic integration of local data and cross-institutional data. The aggregated data is then subjected to final time-series normalization to verify the data integrity and logical consistency at each time step, and to remove a small amount of redundant data generated after fusion. Finally, an initial time-series dataset covering high-quality local and cross-institutional data with unified time-series attributes is formed, thus ensuring the core position of high-value data. Time-series normalization ensures the integrity of the time-series attributes of the dataset, and the final generated initial time-series dataset has both data breadth and data quality.

[0217] This embodiment achieves secure data sharing across institutions through mechanisms such as cross-node data encryption transmission and access control decryption. It breaks down data silos, complies with relevant medical insurance data security protection standards, and avoids the risk of data leakage. At the data source level, it selects high-quality nodes through community aggregation and then achieves differentiated aggregation through node data weights. Combined with multi-stage data governance operations, it significantly improves the quality of the initial time-series dataset, eliminates invalid and low-value data, strengthens the core value density of the data, and provides high-quality data for subsequent dynamic feature extraction and time-series risk prediction, thereby improving the accuracy of fraud risk prediction.

[0218] In addition, to achieve the above objectives, this application also proposes a DIP and DRG medical insurance risk prediction device, the device comprising: a memory, a processor, and a DIP and DRG medical insurance risk prediction program stored on the memory, the processor being used to run the DIP and DRG medical insurance risk prediction program, the computer program being configured to implement the steps of the DIP and DRG medical insurance risk prediction method as described above.

[0219] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a DIP and DRG medical insurance risk prediction program, wherein when the DIP and DRG medical insurance risk prediction program is executed by a processor, it implements the steps of the DIP and DRG medical insurance risk prediction method described above.

[0220] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof.

[0221] The aforementioned computer-readable storage medium may be included in the DIP and DRG medical insurance risk prediction device; or it may exist independently and not be assembled into the DIP and DRG medical insurance risk prediction device.

[0222] Reference Figure 5 , Figure 5 This is a structural block diagram of the first embodiment of the DIP and DRG medical insurance risk prediction device of the present invention. Figure 5 As shown, the DIP and DRG medical insurance risk prediction device proposed in this embodiment of the invention includes:

[0223] Data processing module 10 is used to preprocess the collected multi-source heterogeneous data to obtain an initial time-series dataset. The multi-source heterogeneous data includes medical insurance settlement data, medical institution operation data, and patient medical record data.

[0224] The dynamic feature extraction module 20 is used to perform dynamic feature extraction on the initial time series dataset to obtain key time series feature combinations for each insured user. The key time series feature combinations include fraud types and multiple key fraud time series features associated with the fraud types.

[0225] The model building module 30 is used to train the pre-built original model based on the combination of key temporal features to obtain a fraud risk temporal prediction model. The original model includes a dual-branch feature encoding module, a feature fusion module, and a prediction output module. The first branch of the dual-branch feature encoding module is a graph neural network, and the second branch is a bidirectional long short-term memory network. The first branch is configured for graph structure data learning, and the second branch is configured for temporal feature learning. The fraud risk temporal prediction model includes a first branch feature encoding structure and a second branch feature encoding structure. The first branch feature encoding structure is built based on a graph neural network, and the second branch feature encoding structure is built based on a bidirectional long short-term memory network.

[0226] The dynamic risk prediction module 40 is used to monitor the target area, preprocess the collected monitoring data and extract dynamic features, input the preprocessed and dynamically extracted data into the fraud risk time series prediction model, and output the dynamic fraud risk prediction result, which includes the risk level of each transaction behavior in the target area.

[0227] This embodiment constructs a dual-branch coding-based temporal prediction model for fraud risk by combining temporal feature mining and graph structure association learning. This breaks through the technical constraints of traditional medical insurance supervision, which relies on manual experience and static rules, leading to lagging risk supervision. It achieves precise targeted screening of fraud features through dynamic feature extraction, and simultaneously captures the spatial correlation logic and temporal evolution patterns of fraudulent behavior through dual-branch coding. This model is effectively adapted to the risk prediction scenarios of DIP / DRG payment models, enabling dynamic real-time prediction of abnormal behaviors in medical insurance transactions, and achieving early control of risky transactions, thereby improving the accuracy and efficiency of medical insurance transaction supervision.

[0228] The DIP and DRG medical insurance risk prediction device provided in this application, employing the DIP and DRG medical insurance risk prediction method in the above embodiments, can solve the technical problems of DIP and DRG medical insurance risk prediction. Compared with the prior art, the beneficial effects of the DIP and DRG medical insurance risk prediction device provided in this application are the same as the beneficial effects of the DIP and DRG medical insurance risk prediction method provided in the above embodiments, and other technical features in the DIP and DRG medical insurance risk prediction device are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0229] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0230] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0231] In addition, for technical details not described in detail in this embodiment, please refer to the DIP and DRG medical insurance risk prediction methods provided in any embodiment of the present invention, which will not be repeated here.

[0232] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

Claims

1. A DIP and DRG medical insurance risk prediction method, characterized in that, The method comprises: preprocessing the collected multi-source heterogeneous data to obtain an initial time series dataset, wherein the multi-source heterogeneous data comprises medical insurance settlement data, medical institution operation data and patient medical record data; extracting dynamic features from the initial time series dataset to obtain key time series feature combinations of each insured user, wherein the key time series feature combinations comprise a fraud type and a plurality of key fraud time series features associated with the fraud type; training a pre-constructed original model based on the key time series feature combinations to obtain a fraud risk time series prediction model, wherein the original model comprises a double-branch feature encoding module, a feature fusion module and a prediction output module, a first branch of the double-branch feature encoding module is a graph neural network and a second branch is a bidirectional long short-term memory network, wherein the first branch is configured for graph structure data learning and the second branch is configured for time series feature learning, the fraud risk time series prediction model comprises a first branch feature encoding structure and a second branch feature encoding structure, the first branch feature encoding structure is constructed based on a graph neural network, and the second branch feature encoding structure is constructed based on a bidirectional long short-term memory network; monitoring data in a target area and preprocessing and extracting dynamic features from the collected monitoring data, inputting the preprocessed and dynamically extracted data into the fraud risk time series prediction model, and outputting a dynamic fraud risk prediction result, wherein the dynamic fraud risk prediction result comprises risk levels of each transaction behavior in the target area; The method comprises: inputting the initial time series dataset into a semantic encoding model for semantic encoding to obtain a semantic vector, wherein the semantic encoding model is pre-constructed based on a medical insurance field dictionary dataset; analyzing the correlation strength of the semantic vectors of each insured user at different time steps to obtain time series trend features; analyzing the semantic vector correlation strength of each entity based on the semantic vector to obtain entity correlation features, wherein the entities include insured users, medical institutions and diagnosis and treatment projects; generating a feature subset of each insured user based on the time series trend features and the semantic vector correlation strength, wherein the feature subset comprises a plurality of feature matrices; determining fraud feature weights of each feature matrix based on historical samples, and associating and combining each feature matrix with a fraud type based on the fraud feature weights to generate a plurality of fraud feature combinations, wherein the historical samples comprise feature matrix samples with fraud labels and feature matrix samples with normal labels; performing feature screening on the fraud feature combinations to obtain key time series feature combinations; The method comprises: determining fraud feature weights of each feature matrix based on historical samples, and associating and combining each feature matrix with a fraud type based on the fraud feature weights to generate a plurality of fraud feature combinations, wherein the historical samples comprise feature matrix samples with fraud labels and feature matrix samples with normal labels. The weight analysis model is constructed based on historical samples, and the weight analysis model includes an input layer, a multi-head causal attention layer, and a fusion layer. The input layer is configured to map an input feature matrix to a causal feature vector based on an intervention function. The multi-head causal attention layer includes multiple causal attention heads. The causal attention heads include a fraud type attention head and a causal level attention head. The fraud type attention head is configured to calculate an attention score of the causal feature vector based on a fraud type dimension. The causal level attention head is configured to calculate an attention score of the causal feature vector based on a causal level dimension. The fusion layer is configured to fuse output results of the multiple causal attention heads to output initial feature weights. The feature subsets of each insured user are input into the weight analysis model to obtain initial feature weights. Fact analysis is performed on the historical samples to obtain fact analysis results. The fact analysis results include a fraud probability change value caused by feature missing. The initial feature weights are adjusted based on the fact analysis results to obtain fraud feature weights of each feature matrix. Priority analysis is performed on the feature matrix based on the fraud feature weights to determine a high-weight core feature matrix and a medium-weight auxiliary feature matrix in the feature matrix. The core feature matrix and the auxiliary feature matrix are associated and combined with fraud types to obtain a fraud feature combination. The fraud feature combination is subjected to feature screening to obtain a key time sequence feature combination, including: Business reasonable constraint conditions are generated based on medical insurance rule information, and the fraud feature combination is subjected to reasonable screening based on the business reasonable constraint conditions to obtain a first feature combination. The time fluctuation information of each feature in the first feature combination is analyzed based on the time step of each feature, and the first feature combination is subjected to time sequence stability screening based on the time fluctuation information to obtain a second feature combination. The multi-collinearity between features in the second feature combination is analyzed, the mutual information value between features is determined, and the second feature combination is subjected to feature redundancy screening based on the mutual information value to obtain a key time sequence feature combination.

2. The DIP and DRG Medicare risk prediction method of claim 1, wherein, The pre-constructed original model is trained based on the key time sequence feature combination to obtain a fraud risk time sequence prediction model, including: Fraud feature links are generated based on fraud behavior logic analysis of the key time sequence feature combination. The fraud feature links are logical links of fraud behaviors composed of multiple key fraud time sequence features in series or in parallel. The fraud feature links are subjected to time sequence behavior analysis based on medical insurance medical process, and the fraud feature links are adjusted based on the time sequence behavior analysis results to obtain target feature links. Causal analysis is performed based on the target feature links to construct a fraud causal time sequence graph structure. The nodes of the fraud causal time sequence graph structure include time sequence feature nodes, fraud behavior nodes, fraud type nodes, and intervention nodes. The edges of the fraud causal time sequence graph structure include causal driving edges, behavior classification edges, time sequence driving edges, and intervention adjustment edges. Train a pre-constructed original model based on the fraud causal time series graph structure and the key time series feature combination to obtain a fraud risk time series prediction model.

3. The DIP and DRG Medicare risk prediction method of claim 2, wherein, The fraud risk time series prediction model further comprises a betweenness center module, a feature fusion module and a prediction output module. The betweenness center module is configured to calculate betweenness center values of each time series feature node based on node features of nodes in the fraud causal time series graph structure, and perform betweenness optimization on the fraud causal time series graph structure and the key time series feature combination based on the betweenness center values, and output the fraud causal time series graph structure and the key time series feature combination after betweenness optimization. The first branch feature encoding structure is configured to encode the fraud causal time series graph structure after betweenness optimization to output a graph structure embedding vector. The second branch feature encoding structure is configured to encode the key time series feature combination after betweenness optimization to output a time series perception embedding vector. The feature fusion module is configured to weight and fuse the graph structure embedding vector and the time series perception embedding vector to obtain initial fusion features, calculate importance weights of each feature dimension in the initial fusion features, determine feature contribution degrees of each feature dimension based on the importance weights, and perform attention calibration on the initial fusion features based on the feature contribution degrees to output target fusion features. The prediction output module is configured to perform feature transformation and regularization processing on the target fusion features to obtain high-level abstract features, calculate probability distribution information of each risk level based on the high-level abstract features, determine risk levels of each transaction behavior in a target area based on the probability distribution information, and output the risk levels of each transaction behavior in the target area as dynamic fraud risk prediction results.

4. The DIP and DRG Medicare risk prediction method as claimed in claim 1, wherein, The preprocessing of the collected multi-source heterogeneous data to obtain an initial time series dataset comprises: Preprocessing the locally collected multi-source heterogeneous data to obtain a first dataset, wherein the preprocessing comprises data cleaning, standardization processing and time series alignment processing; Sending a data acquisition request to a plurality of data sharing nodes, so that each data sharing node preprocesses the locally collected multi-source heterogeneous data, encrypts a second dataset obtained by preprocessing, and responds to the data acquisition request based on the encrypted data; Obtaining response information of each data sharing node, aggregating each data sharing node based on the response information to generate a key community and an edge community, wherein the key community aggregates a plurality of first nodes, and the edge community aggregates a plurality of second nodes, the first nodes are data sharing nodes whose response speed, historical data quality and / or historical credibility meet preset conditions, the second nodes are data sharing nodes whose response speed, historical data quality and historical credibility do not meet preset conditions, and the response information comprises encrypted data, identity information and response speed of the data sharing node. Decrypting the encrypted data responded by each first node in the key community to obtain a second dataset; Determining node data weights based on the response speed, historical data quality and historical credibility of each first node; Based on the node data, the first data set is aggregated with the second data set to obtain an initial time series data set.

5. A DIP and DRG medical insurance risk prediction device using the DIP and DRG medical insurance risk prediction method according to any one of claims 1 to 4, characterized by The DIP and DRG medical insurance risk prediction device comprises: A data processing module is configured to preprocess the collected multi-source heterogeneous data to obtain an initial time series data set, wherein the multi-source heterogeneous data comprises medical insurance settlement data, medical institution operation data and patient medical record data. A dynamic feature extraction module is configured to extract dynamic features from the initial time series data set to obtain a key time series feature combination of each insured user, wherein the key time series feature combination comprises a fraud type and a plurality of key fraud time series features associated with the fraud type. A model construction module is configured to train a pre-constructed original model based on the key time series feature combination to obtain a fraud risk time series prediction model, wherein the original model comprises a double-branch feature encoding module, a feature fusion module and a prediction output module, the first branch of the double-branch feature encoding module is a graph neural network, and the second branch is a bidirectional long short-term memory network, wherein the first branch is configured for graph structure data learning, and the second branch is configured for time series feature learning, the fraud risk time series prediction model comprises a first branch feature encoding structure and a second branch feature encoding structure, the first branch feature encoding structure is constructed based on a graph neural network, and the second branch feature encoding structure is constructed based on a bidirectional long short-term memory network. A dynamic risk prediction module is configured to monitor data in a target area, preprocess and extract dynamic features from the collected monitoring data, input the preprocessed and dynamically extracted data into the fraud risk time series prediction model, and output a dynamic fraud risk prediction result, wherein the dynamic fraud risk prediction result comprises a risk level of each transaction behavior in the target area.

6. A DIP and DRG Medicare risk prediction device characterized by, The DIP and DRG medical insurance risk prediction device comprises a memory, a processor and a DIP and DRG medical insurance risk prediction program stored in the memory, the processor is configured to run the DIP and DRG medical insurance risk prediction program, and the DIP and DRG medical insurance risk prediction program is configured to implement the DIP and DRG medical insurance risk prediction method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a DIP and DRG medical insurance risk prediction program, and the DIP and DRG medical insurance risk prediction program is executed by a processor to implement the DIP and DRG medical insurance risk prediction method according to any one of claims 1 to 4.

Images