Compliance review methods, media, and equipment based on chronological causal tracing of medical invoices
By constructing a multidimensional time series matrix and a standard diagnosis and treatment causal template diagram, the causal relationships between diagnosis and treatment items are identified, which solves the problem of insufficient analysis of causal transmission chains in medical insurance compliance review, generates reliable law enforcement evidence, reduces false alarm rate and improves the stability of review conclusions.
Patent Information
- Application Number
- CN202611132063.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-07-29
AI Technical Summary
Existing medical insurance compliance review methods lack the ability to analyze the causal transmission chain between treatment items, cannot distinguish between normal clinical complex medication use and malicious causal bundling arbitrage behavior, and the review conclusions cannot be used as law enforcement evidence. They also lack a multi-subset replication verification mechanism, resulting in a high false alarm rate and unstable conclusions.
By constructing a multidimensional time series matrix and a standard diagnosis and treatment causal template diagram, the causal relationships between diagnosis and treatment items are identified, the propagation entropy value is calculated, abnormal propagation edges are identified and a white-box causal evidence chain is generated, logical consistency verification and cross-time window reproducibility stability assessment are performed, and traceable and reproducible compliance review evidence is formed.
It enables reliable traceability and review of medical insurance funds, identifies covert causal arbitrage activities, generates reliable law enforcement evidence, reduces false alarm rates, and improves the stability of review conclusions.
Smart Images

Figure CN122675587B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent supervision technology for medical big data, specifically to a compliance review method, medium, and equipment based on the chronological causal tracing of medical invoices. Background Technology
[0002] Currently, the industry's automated compliance review of medical invoice details mainly relies on two technical approaches: "hard-coded rule matching" and "statistical anomaly detection." The former verifies the intersection of the underlying invoice detail fields with expert-preset medical insurance payment rules and is widely used when dealing with single-point violations with clear boundaries. The latter introduces isolated forest, local anomaly factor, or deep learning models to identify anomalous individuals that significantly deviate from the statistical distribution by performing group modeling of the cost distribution of massive historical invoices.
[0003] However, existing solutions have the following drawbacks: First, the anomaly detection logic relies on static cost baselines or single-point outlier detection, lacking the ability to analyze the micro-causal transmission chain between treatment items, and cannot distinguish between normal complex clinical medication and malicious causal bundling arbitrage behavior, resulting in a high false alarm rate; Second, the anomaly probability or classification label output by the algorithm is disconnected from the underlying physical invoices, lacking row-level data traceability capabilities, and the review conclusions cannot serve as direct evidence for on-site audits; Third, there is a lack of multi-subset reproduction and verification mechanisms, and the anomaly conclusions output after a single static processing are highly sensitive to data disturbances, making it difficult to stably reproduce under different time windows. Summary of the Invention
[0004] In view of the above problems, the present invention provides a technical solution for compliance review based on the temporal causal tracing of medical invoices, in order to solve the technical problems of existing medical insurance compliance review methods, which lack the ability to analyze the causal transmission chain between medical treatment items, lack traceable data support for abnormal conclusions, and lack statistical reproduction and verification mechanisms for conclusions, resulting in difficulty in identifying hidden causal arbitrage behavior, and review conclusions that cannot be used as law enforcement evidence and have insufficient reliability.
[0005] To achieve the above objectives, in a first aspect, the present invention provides a compliance review method based on chronological causal tracing of medical invoices, the method comprising:
[0006] S1: Obtain the original medical settlement invoice details of the medical institution under review in multiple consecutive accounting periods, extract the project code and billing time of each diagnosis and treatment item from the original medical settlement invoice details, and sort the diagnosis and treatment items in a standardized manner according to the billing time to construct a multi-dimensional time series matrix. The multi-dimensional time series matrix contains the diagnosis and treatment item code sequence and corresponding cost characteristics.
[0007] S2: Obtain basic medical knowledge data, extract the temporal constraint relationship between diagnosis and treatment items from the basic medical knowledge data, construct a standard diagnosis and treatment causal template graph and set a temporal hard constraint operator. The standard diagnosis and treatment causal template graph includes diagnosis and treatment item nodes and directed edges. Each directed edge is used to indicate that there is a temporal sequence relationship between the two connected diagnosis and treatment items. The temporal hard constraint operator is used to constrain the causal inference results that do not conform to the temporal sequence relationship indicated by the directed edge.
[0008] S3: Combine the diagnosis and treatment items in the multidimensional time series matrix into pairs to form diagnosis and treatment item groups. Decompose each diagnosis and treatment item group into atomic causal propositions to characterize the candidate causal transmission relationship and transmission direction between the two diagnosis and treatment items. Using the standard diagnosis and treatment causal template graph as a prior constraint, filter the atomic causal propositions. Select the atomic propositions that are retained after filtering as effective causal associations. Perform topology optimization on the effective causal associations. Use the transfer entropy algorithm to calculate the transfer entropy value of the optimized effective causal associations. Output the ticket instance causal graph. The ticket instance causal graph includes the causal topology structure and the transfer entropy value.
[0009] S4: Perform a topological comparison between the invoice instance causal graph and the standard treatment causal template graph to identify redundant edges in the invoice instance causal graph that are different from those in the standard treatment causal template graph; for the identified redundant edges, combine the transmission entropy value corresponding to the redundant edge in the invoice instance causal graph, and the first cost data of the treatment items at both ends of the redundant edge when including the previous treatment item and the second cost data when not including the previous treatment item, calculate the cost change risk value based on the difference between the first cost data and the second cost data, and confirm the redundant edges whose cost change risk value exceeds the preset baseline as abnormal transmission edges, aggregate all abnormal transmission edges, and extract the illegal arbitrage subgraph;
[0010] S5: Perform topological analysis on the illegal arbitrage subgraph, extract abnormal causal paths, and decompose the abnormal causal paths into atomic causal jump sequences; for each atomic causal jump, combine the transmission entropy value corresponding to the atomic causal jump in the causal graph of the invoice instance and the billing data of the treatment items involved to quantify the amount of medical insurance fund loss caused by the atomic causal jump; map the atomic causal jump sequence to the physical data rows of the original database through an inverted index, bind and generate data anchors, and assemble the abnormal causal path, the amount of medical insurance fund loss, and the data anchors into a white-box causal evidence chain;
[0011] S6: Perform logical consistency checks on the white-box causal evidence chains sequentially. The logical consistency checks include causal direction checks and data anchor validity checks. Perform a cross-time window reproducibility stability assessment based on bootstrapping sampling on the white-box causal evidence chains that pass the logical consistency checks. Score the white-box causal evidence chains according to the results of the logical consistency checks and the cross-time window reproducibility stability assessment. Archive the white-box causal evidence chains that reach the preset score threshold to the confirmation database, and archive the white-box causal evidence chains that do not reach the threshold to the clue database.
[0012] Furthermore, in step S1, the standardized ranking of treatment items based on billing time includes:
[0013] For each medical record in the details of each original medical settlement invoice, extract the billing timestamp corresponding to each treatment item, sort the treatment items in ascending order according to the order of the billing timestamps, and generate the treatment item code sequence.
[0014] For medical records spanning multiple days, they are sorted in ascending order of billing timestamp values.
[0015] Furthermore, in step S2, extracting the temporal constraints between treatment items from the medical knowledge base data and constructing a standard treatment causal template diagram specifically includes:
[0016] By analyzing unstructured text in drug instructions, medical insurance catalog payment scope texts, and clinical operation guidelines using a medical entity recognition mechanism, the system identifies the names of treatment items and descriptive words describing the sequential execution order or mutual exclusion relationship between treatment items. The recognition results are then converted into structured temporal triples, which include a first treatment item identifier, a second treatment item identifier, and a temporal relationship type between the first and second treatment items. The temporal relationship type includes any one of the following: the first treatment item occurs before the second treatment item, the second treatment item occurs before the first treatment item, or the first and second treatment items occur mutually exclusively.
[0017] Aggregate all the aforementioned time-series triples, using the diagnosis and treatment item identifiers in the time-series triples as nodes and the time-series relationship type as the direction basis for directed edges, and construct a standard diagnosis and treatment causal template graph stored in the form of a directed acyclic graph.
[0018] Furthermore, in step S3, the atomic causal propositions are filtered, and the retained atomic propositions are considered as valid causal relationships, specifically including:
[0019] The atomic causal proposition is matched with the standard diagnostic causal template diagram to determine whether there is a directed path in the standard diagnostic causal template diagram that is consistent with the causal transmission direction indicated by the atomic causal proposition.
[0020] If a consistent directed path exists, the conditional independence of the atomic causal proposition is tested using the first significance threshold.
[0021] If there is no consistent directed path in the standard diagnosis and treatment causal template graph, but there is an indirect path connecting the two diagnosis and treatment items, then the conditional independence test is performed using the second significance threshold. The significance level value corresponding to the second significance threshold is less than the significance level value corresponding to the first significance threshold.
[0022] If there is no path connecting the two treatment items in the standard diagnosis and treatment causal template graph, then a third significance threshold is used to perform a conditional independence test, and the significance level value corresponding to the third significance threshold is less than the significance level value corresponding to the second significance threshold.
[0023] Atomic propositions that pass the conditional independence test are retained as valid causal relationships.
[0024] Furthermore, in step S2, after setting the time-series hard constraint operator, the method further includes dynamically updating the time-series hard constraint operator, specifically including the following steps:
[0025] The number of billing records corresponding to each treatment item in the multidimensional time series matrix within the current accounting period is counted as the frequency of occurrence of that treatment item.
[0026] Treatment items that occur less frequently than a preset low-frequency threshold are marked as low-frequency treatment items. For directed edges in the standard treatment causal template graph where both ends are low-frequency treatment items, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a first coefficient value. For directed edges where only one end is a low-frequency treatment item, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a second coefficient value. For directed edges where neither end is a low-frequency treatment item, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a third coefficient value. The first coefficient value is less than the second coefficient value, and the second coefficient value is less than the third coefficient value.
[0027] Substituting the adjusted constraint coefficients into step S3, after the atomic causal propositions have passed the conditional independence test, for atomic causal propositions that pass the conditional independence test but whose causal transmission direction is inconsistent with the direction of the directed edge corresponding to the standard diagnosis and treatment causal template diagram, the product of the constraint coefficients and the significance threshold used in the conditional independence test is used as the adjusted significance threshold, and atomic causal propositions that fail the adjusted significance threshold test are removed from the effective causal associations.
[0028] Furthermore, in step S4, calculating the cost variation risk value based on the difference between the first cost data and the second cost data includes:
[0029] For each redundant edge, obtain the transit entropy value of the previous treatment item to the next treatment item in the causal graph of the ticket instance in the redundant edge;
[0030] The total billing amount of the next treatment item in all treatment item coding sequences that include the previous treatment item in the multidimensional time series matrix is obtained as the first cost data, and the total billing amount of the next treatment item in all treatment item coding sequences that do not include the previous treatment item in the multidimensional time series matrix is obtained as the second cost data.
[0031] The difference between the first cost data and the second cost data is multiplied by the transmission entropy value to obtain the cost change risk value.
[0032] Furthermore, in step S4, redundant edges whose cost variation risk values exceed a preset baseline are identified as abnormal propagation edges. All abnormal propagation edges are aggregated to extract the illegal arbitrage subgraph, specifically including:
[0033] Obtain the first-level category label and the second-level category label corresponding to the diagnosis and treatment items at both ends of each redundant edge. The first-level category label is the department category code to which the diagnosis and treatment item belongs, and the second-level category label is the item category code to which the diagnosis and treatment item belongs.
[0034] Redundant edges with the same first-level category label for the treatment items at both ends are grouped into a first-level redundant edge set, and the arithmetic mean of the cost variation risk values of all redundant edges in the first-level redundant edge set is calculated to obtain the first-level risk mean.
[0035] For each first-level redundant edge set, redundant edges with the same second-level category label for both ends of the diagnosis and treatment items in the first-level redundant edge set are merged into a second-level redundant edge set, and the arithmetic mean of the cost variation risk values of all redundant edges in the second-level redundant edge set is calculated to obtain the second-level risk mean.
[0036] Each redundant edge in the first-level redundant edge set whose first-level risk average exceeds the preset first-level baseline, and each redundant edge in the second-level redundant edge set whose second-level risk average exceeds the preset second-level baseline, are identified as abnormal propagation edges.
[0037] All redundant edges identified as anomalous propagation edges are aggregated to extract the illegal arbitrage subgraph. Multiple anomalous propagation edges with shared diagnosis and treatment nodes belong to the same illegal arbitrage subgraph.
[0038] Furthermore, in step S5, for each atomic causal jump, combining the transit entropy value corresponding to that atomic causal jump in the causal graph of the invoice instance and the billing data related to the medical treatment items, the amount of medical insurance fund loss caused by that atomic causal jump is quantified, specifically including:
[0039] For each atomic causal jump, obtain the transmission entropy value of the previous treatment item to the next treatment item in the causal graph of the invoice instance in the atomic causal jump, and obtain the total billing amount of the previous treatment item in the original medical settlement invoice details data of all abnormal causal paths that trigger the atomic causal jump in the current accounting period, and obtain the medical insurance fund payment ratio coefficient corresponding to the previous treatment item.
[0040] Multiplying the transfer entropy value by the total billing amount and the medical insurance fund payment ratio coefficient yields the amount of medical insurance fund loss caused by the atomic causal jump.
[0041] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the compliance review method based on time-series causal tracing of medical invoices as described in the first aspect of the present invention.
[0042] In a third aspect, the present invention provides an electronic device having a computer program stored thereon, including a processor and a storage medium, wherein the computer program is stored on the storage medium, and when executed by the processor, the computer program implements the compliance review method based on medical invoice chronological causal tracing as described in the first aspect of the present invention.
[0043] Unlike existing technologies, the above technical solution involves a compliance review method, medium, and equipment based on time-series causal tracing of medical invoices. The method includes: extracting the item codes and billing times of treatment items from original medical settlement invoices, constructing a multi-dimensional time-series matrix containing the sequence of treatment item codes and cost characteristics; extracting time-series constraint relationships from basic medical knowledge data, constructing a standard treatment causal template graph, and setting time-series hard constraint operators; performing atomic causal proposition decomposition on treatment items, using the standard template graph as a priori constraint for screening and topology optimization, quantifying causal strength through a transfer entropy algorithm, and outputting an invoice instance causal graph; performing a topological comparison between the invoice instance causal graph and the standard template graph, identifying redundant edges, calculating cost variation risk values based on transfer entropy values and cost data, and extracting an illegal arbitrage subgraph; performing topological analysis on the illegal arbitrage subgraph, quantifying the amount of medical insurance fund losses, and generating a white-box causal evidence chain with data anchors through inverted indexing; performing logical consistency verification and cross-time window reproducibility stability assessment based on bootstrapping sampling on the evidence chain, and hierarchical archiving. This invention enables full-chain medical insurance compliance review, from invoice data to traceability, reproducibility, and enforceability.
[0044] The above description of the invention is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical solution of the present invention and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of the present invention easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of the present invention. Attached Figure Description
[0045] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on the present invention.
[0046] In the accompanying drawings of the instruction manual:
[0047] Figure 1 This is a flowchart of the compliance review method based on time-series causal tracing of medical invoices as described in the first exemplary embodiment of the present invention;
[0048] Figure 2 This is a schematic diagram of an electronic device according to an exemplary embodiment of the present invention;
[0049] The reference numerals used in the above figures are explained as follows:
[0050] 10. Electronic equipment; 101. Processor; 102. Storage medium. Detailed Implementation
[0051] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that can be achieved by this invention, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this invention, and are therefore only examples, and should not be used to limit the scope of protection of this invention.
[0052] In the first aspect, such as Figure 1 As shown, this application provides a compliance review method based on the chronological causal tracing of medical invoices, the method comprising:
[0053] S1: Obtain the original medical settlement invoice details of the medical institution under review in multiple consecutive accounting periods, extract the project code and billing time of each diagnosis and treatment item from the original medical settlement invoice details, and sort the diagnosis and treatment items in a standardized manner according to the billing time to construct a multi-dimensional time series matrix. The multi-dimensional time series matrix contains the diagnosis and treatment item code sequence and corresponding cost characteristics.
[0054] S2: Obtain basic medical knowledge data, extract the temporal constraint relationship between diagnosis and treatment items from the basic medical knowledge data, construct a standard diagnosis and treatment causal template graph and set a temporal hard constraint operator. The standard diagnosis and treatment causal template graph includes diagnosis and treatment item nodes and directed edges. Each directed edge is used to indicate that there is a temporal sequence relationship between the two connected diagnosis and treatment items. The temporal hard constraint operator is used to constrain the causal inference results that do not conform to the temporal sequence relationship indicated by the directed edge.
[0055] S3: Combine the diagnosis and treatment items in the multidimensional time series matrix into pairs to form diagnosis and treatment item groups. Decompose each diagnosis and treatment item group into atomic causal propositions to characterize the candidate causal transmission relationship and transmission direction between the two diagnosis and treatment items. Using the standard diagnosis and treatment causal template graph as a prior constraint, filter the atomic causal propositions. Select the atomic propositions that are retained after filtering as effective causal associations. Perform topology optimization on the effective causal associations. Use the transfer entropy algorithm to calculate the transfer entropy value of the optimized effective causal associations. Output the ticket instance causal graph. The ticket instance causal graph includes the causal topology structure and the transfer entropy value.
[0056] S4: Perform a topological comparison between the invoice instance causal graph and the standard treatment causal template graph to identify redundant edges in the invoice instance causal graph that are different from those in the standard treatment causal template graph; for the identified redundant edges, combine the transmission entropy value corresponding to the redundant edge in the invoice instance causal graph, and the first cost data of the treatment items at both ends of the redundant edge when including the previous treatment item and the second cost data when not including the previous treatment item, calculate the cost change risk value based on the difference between the first cost data and the second cost data, and confirm the redundant edges whose cost change risk value exceeds the preset baseline as abnormal transmission edges, aggregate all abnormal transmission edges, and extract the illegal arbitrage subgraph;
[0057] S5: Perform topological analysis on the illegal arbitrage subgraph, extract abnormal causal paths, and decompose the abnormal causal paths into atomic causal jump sequences; for each atomic causal jump, combine the transmission entropy value corresponding to the atomic causal jump in the causal graph of the invoice instance and the billing data of the treatment items involved to quantify the amount of medical insurance fund loss caused by the atomic causal jump; map the atomic causal jump sequence to the physical data rows of the original database through an inverted index, bind and generate data anchors, and assemble the abnormal causal path, the amount of medical insurance fund loss, and the data anchors into a white-box causal evidence chain;
[0058] S6: Perform logical consistency checks on the white-box causal evidence chains sequentially. The logical consistency checks include causal direction checks and data anchor validity checks. Perform a cross-time window reproducibility stability assessment based on bootstrapping sampling on the white-box causal evidence chains that pass the logical consistency checks. Score the white-box causal evidence chains according to the results of the logical consistency checks and the cross-time window reproducibility stability assessment. Archive the white-box causal evidence chains that reach the preset score threshold to the confirmation database, and archive the white-box causal evidence chains that do not reach the threshold to the clue database.
[0059] In this embodiment, the original medical settlement invoice details data refers to the bottom details of all invoices generated by medical institutions for outpatient and inpatient settlements, including multi-source heterogeneous fields such as codes for drugs, examinations, and consumables, billing amounts, billing timestamps, length of hospital stay, medical insurance payment ratio, visit serial number, detail line number, and accounting cycle identifier.
[0060] A multidimensional time series matrix is a structured data matrix that has been standardized and aligned with time series data. The row dimension is a single outpatient visit log, and the column dimension includes standardized treatment item coding sequences, relative hospitalization progress time series intervals, billing amounts for each item, medical insurance co-payment ratios, and other cost characteristics, thus eliminating time series comparison bias caused by differences in the total number of hospitalization days among different patients.
[0061] Basic medical knowledge data refers to publicly available medical standard datasets consisting of the National Medical Insurance Drug Catalog, drug instructions, clinical operation guidelines, and texts on contraindications for diagnosis and treatment, which are used to extract compliant clinical time-series logic.
[0062] The standard clinical causal template graph refers to a compliant clinical topology model stored in a directed acyclic graph (DAG). The nodes are standardized clinical treatment item codes, and the directed edge S→M represents the requirement that S must occur before M according to clinical standards. The temporal hard constraint operator is the underlying logical judgment rule. Once the algorithm infers a causal relationship that violates the edge direction, it will directly restrict the causal edge from entering the valid association set.
[0063] Atomic causal propositions refer to triplets (X,Y,D) formed by decomposing any two medical items X and Y, where D represents three candidate causal relationships: X→Y, Y→X, and no direct causal edge. All atomic causal propositions can be computed in parallel, solving the computational explosion problem caused by high-dimensional item combinations.
[0064] Transmission entropy (TE) is a quantitative indicator based on time-series information theory algorithms. It is used to characterize the cascading transmission strength of the prescribing behavior of pre-treatment item X to the cost scale of subsequent treatment item Y. The higher the value, the stronger the bundled triggering effect.
[0065] The invoice instance causal graph refers to the full causal topology network obtained by autonomous mining based on real invoice data. It contains all directed edges of valid causal relationships and corresponding transit entropy values, and is used for differential comparison with standard templates.
[0066] Redundant edges refer to directed causal edges where the causal graph of real invoices exists but the standard diagnosis and treatment template graph is completely missing. These correspond to bundled prescriptions without clinical indications and arbitrage through excessive examinations. The cost change risk value integrates the transmission strength and cost difference to quantify the degree of medical insurance fund loss. Edges exceeding the baseline are identified as abnormal transmission edges.
[0067] The illegal arbitrage subgraph refers to an independent topological network formed by aggregating abnormal propagation edges of multiple shared nodes using Connected Component Analysis (CCA), representing systematic and large-scale illegal practices in medical institutions.
[0068] The data anchor triple is represented as [medical visit serial number, detailed table row number, accounting cycle identifier]. The underlying database bidirectional retrieval engine enables precise binding between the abstract causal path and the original physical detailed line of the invoice. The white-box causal evidence chain integrates abnormal paths, fund losses, and physical anchors, which can be directly used for auditing and evidence collection.
[0069] The working principle of this embodiment is as follows:
[0070] In step S1, a multi-dimensional time series matrix is constructed, specifically as follows: First, full settlement details for T consecutive accounting cycles are collected in batches. For each medical visit, the codes, billing timestamps, amounts, and medical insurance ratios of all medical items are extracted. For differences in hospitalization days, a relative hospitalization progress percentage mapping algorithm is used to convert the absolute hospitalization days into a unified time series interval (initial admission, stable treatment period, discharge stage). For example, if a patient hospitalized for 3 days has a medical item progress of 66% on day 2 and a patient hospitalized for 30 days has a medical item progress of 60% on day 18, the two are grouped into the same time series interval. The medical item code sequence is generated by arranging the standardized time series in ascending order. Finally, the cost vector is concatenated to construct a multi-dimensional time series matrix, which unifies the time series comparison benchmark for the entire institution and all cases.
[0071] In step S2, a standard diagnosis and treatment causal template graph and temporal hard constraints are constructed. Specifically, this includes: parsing unstructured medical text through medical entity recognition, extracting temporal triples (diagnosis and treatment item A, diagnosis and treatment item B, and the temporal relationship between diagnosis and treatment items A and B), and aggregating all triples to generate a standard diagnosis and treatment causal template graph; at the same time, a temporal hard constraint operator t(Y)-t(X)>0 is defined. If X→Y exists in the standard causal template graph, then the billing time of diagnosis and treatment item Y in the real invoice must be later than that of diagnosis and treatment item X. When reverse causality occurs during the algorithm mining stage, the circuit breaker is directly triggered to correct the causal direction.
[0072] In step S3, all diagnosis and treatment items in the multidimensional time series matrix are paired to generate atomic causal propositions, and the search space is narrowed down to the polynomial level based on the standard template. A three-level differential significance threshold is used to perform conditional independence testing: the template-compliant path is relaxed α=0.1, the indirect path is standard α=0.01, and the unrelated path is strict α=0.001. The GES greedy topology is used to optimize and embed a structural deviation penalty term to penalize causal edges that violate the clinical time series. The propagation entropy is calculated for each valid directed edge, and the ticket instance causal graph carrying the propagation strength value is output.
[0073] In step S4, redundant edges are extracted by topological difference between the invoice instance causal graph and the standard diagnosis and treatment causal template graph. The risk value of cost variation is calculated by combining the propagation entropy and the total cost of subsequent projects when there are no preceding projects. Edges exceeding the preset baseline are marked as abnormal propagation edges. Connected component analysis is used to aggregate the abnormal edges of shared nodes into independent illegal arbitrage subgraphs to lock the large-scale bundled arbitrage network of medical institutions.
[0074] In step S5, a topological traversal is performed on the violation subgraph to distinguish between source nodes and cost terminal nodes, and the atomic causal jump sequence is split. The amount of medical insurance fund loss caused by a single jump is calculated based on the transfer entropy, the total cost of the preceding projects, and the medical insurance pooled payment ratio. The database is retrieved in reverse using the inverted index, and physical row-level data anchors are bound to assemble the topological path, fund loss, and invoice anchors into a complete and verifiable data chain.
[0075] In step S6, a three-stage audit verification and archiving process is performed, which includes: the first step is logical verification, which verifies the causal sequence, the authenticity and completeness of anchor data, and directly isolates logical conflicts; the second step is bootstrap sampling; the third step is comprehensive weighted scoring, with high-scoring evidence stored in the confirmation database for on-site medical insurance administrative law enforcement penalties, and low-scoring clues archived for manual review; verified high-confidence causal chains are automatically used as prior constraints for the next round of modeling, continuously compressing the algorithm search space to achieve system self-evolution.
[0076] Specifically, the complete execution logic of cross-time window bootstrapping sampling is as follows: The system uses the accounting month as the unit of division, randomly selects 100 independent time window data subsets, and each subset independently and completely executes the full-process time-series causal tracing operation from S1 to S5; it counts the number of times the white-box causal evidence chain recurs in 100 repeated operations, calculates the recurrence rate, and sets a hard judgment threshold: if the recurrence rate is ≥50%, it is judged as a systematic and organized arbitrage behavior of medical institutions and archived in the confirmation database; if the recurrence rate is <50%, it is judged as occasional data noise such as medication change, data entry error, and special conditions in single visit records, and is only stored in the clue database for manual review.
[0077] The highly reliable white-box causal evidence chain, which is archived to the confirmation library after three stages of verification, will automatically serve as a priori constraint for the next round of atomic causal proposition screening, continuously compressing the causal edge search range. As the total amount of original medical settlement invoice details processed increases, the computational complexity of the algorithm continues to decrease, and the efficiency of the review operation is continuously optimized.
[0078] The aforementioned solution first establishes prior clinical constraints using a standardized time-series matrix and a medical knowledge-based DAG template. Combined with a three-level significance threshold, it distinguishes between reasonable treatment and bundled arbitrage, significantly reducing the high false alarm rate caused by single-point cost baselines. Secondly, it quantifies causal transmission strength using transfer entropy and generates row-level physical anchors through inverted indexes, completely breaking the limitations of black-box output in machine learning. Each abnormal conclusion can retrieve the original invoice via SQL retrieval, forming white-box evidence with administrative enforcement validity. Finally, it introduces bootstrap sampling multi-window stability verification to filter out high noise such as medication returns, changes, and cross-period accounting, ensuring the statistical credibility of the review conclusions. Furthermore, the solution relies only on basic invoice details, requiring no medical records or DRG data, resulting in low implementation costs. The algorithm efficiency is continuously optimized as business data accumulates, balancing recognition accuracy, evidence-gathering applicability, and long-term performance.
[0079] In some embodiments, in step S1, standardizing the ranking of medical services based on billing time includes:
[0080] For each medical record in the details of each original medical settlement invoice, extract the billing timestamp corresponding to each treatment item, sort the treatment items in ascending order according to the order of the billing timestamps, and generate the treatment item code sequence.
[0081] For medical records spanning multiple days, they are sorted in ascending order of billing timestamp values.
[0082] In this embodiment, the billing timestamp refers to the precise settlement time identifier that comes with the original invoice details. Outpatient records are accurate to the hour and minute, while inpatient records are marked with the specific number of days of hospitalization.
[0083] Medical records refer to the detailed data of a single original medical settlement invoice or the complete set of medical treatment details corresponding to a single inpatient transaction record; medical records spanning multiple days: inpatient transaction records, with medical treatment items distributed across multiple natural days of admission, hospitalization, and discharge.
[0084] The working principle of this embodiment is as follows:
[0085] Before constructing the multidimensional time-series matrix, the billing timestamps of all treatment items in a single medical record are extracted and strictly arranged in ascending order according to the timestamp values to generate an ordered sequence of treatment item codes. For scenarios where hospitalization spans multiple days, no time dimension truncation or merging is performed; the original timestamp sorting logic is directly retained, and then the relative hospitalization progress percentage conversion logic described above is superimposed to double ensure the accuracy of time-series alignment.
[0086] For example, the system obtains detailed data of original medical settlement invoices corresponding to two different medical records. The first medical record corresponds to patient Zhang San, with a total hospitalization duration of 3 days. The billing time sequence of the treatment items in the details is: blood routine code EX-01 on the morning of the first day of admission, chest CT code EX-02 on the afternoon of the same day, and high-value traditional Chinese medicine DR-99 on the second day of admission. The second medical record corresponds to patient Li Si, with a total hospitalization duration of 30 days. The billing records for the treatment items are generated on the 2nd, 6th, and 18th days, respectively. If only the absolute number of hospitalization days is used for horizontal time sequence comparison, the time sequence scale of the treatment behavior of the two types of medical records cannot be aligned. The method in this embodiment, after sorting the billing timestamps of the treatment items in ascending order, executes an adaptive partitioning algorithm based on the relative hospitalization progress days. The standardized progress calculation formula for a single treatment item is as follows:
[0087] The relative progress of treatment items is calculated as: (Number of days the treatment item occurred) ÷ (Total number of days of hospitalization for that medical record). After conversion, Zhang San's progress for each treatment item is 33%, 33%, and 66%, respectively, while Li Si's progress for each treatment item is 6.6%, 20%, and 60%, respectively. The system uses a unified standard for defining time series intervals: progress ≤ 33% is classified as the initial admission interval, and progress ≥ 60% is classified as the stable treatment interval. EX-01 and EX-02 in both medical records are classified as initial admission, while DR-99 is classified as the stable treatment interval. This completes the time series scale normalization, constructs a unified benchmark multidimensional time series matrix, provides standardized input for subsequent calculations of propagation entropy causality strength, and eliminates time series comparison bias caused by differences in hospitalization duration among different medical records.
[0088] The above scheme uses the original timestamp ascending order as the basic sorting rule to fully preserve the hospital's true settlement timeline and prevent the loss of original business sequence information during the relative progress mapping process. At the same time, it distinguishes between outpatient single-day documents and inpatient cross-day documents and unifies the processing logic, resulting in stronger consistency in timeline sorting. The timeline benchmark of the input data for subsequent causal discovery algorithms is more reliable, reducing the generation of false causal associations caused by the simplification of sorting logic and further reducing the probability of false alarms.
[0089] In some embodiments, in step S2, extracting the temporal constraint relationships between diagnosis and treatment items from the medical knowledge base data and constructing a standard diagnosis and treatment causal template diagram specifically includes:
[0090] By analyzing unstructured text in drug instructions, medical insurance catalog payment scope texts, and clinical operation guidelines using a medical entity recognition mechanism, the system identifies the names of treatment items and descriptive words describing the sequential execution order or mutual exclusion relationship between treatment items. The recognition results are then converted into structured temporal triples, which include a first treatment item identifier, a second treatment item identifier, and a temporal relationship type between the first and second treatment items. The temporal relationship type includes any one of the following: the first treatment item occurs before the second treatment item, the second treatment item occurs before the first treatment item, or the first and second treatment items occur mutually exclusively.
[0091] Aggregate all the aforementioned time-series triples, using the diagnosis and treatment item identifiers in the time-series triples as nodes and the time-series relationship type as the direction basis for directed edges, and construct a standard diagnosis and treatment causal template graph stored in the form of a directed acyclic graph.
[0092] In this embodiment, the medical entity recognition mechanism refers to a named entity extraction model adapted to medical terminology, which can identify the standardized coded names of drugs, examinations, and consumables.
[0093] Unstructured medical texts refer to natural language texts of medical insurance policies, drug instructions, and clinical operation guidelines that do not have a fixed table format.
[0094] A temporal triplet refers to a basic data unit that is structured to store clinical sequence or mutual exclusion relationships.
[0095] The working principle of this embodiment is as follows:
[0096] The first step involves using a medical entity recognition tool to parse the original medical text, automatically distinguishing between treatment item entities and temporal relationship descriptors, identifying three types of temporal logic: A precedes B, B precedes A, and A and B are mutually exclusive, and encapsulating them into standardized temporal triples. The second step involves batch aggregating all temporal triples, using the treatment item identifiers within the temporal triples as graph nodes, drawing directed edges in the corresponding directions according to the temporal relationship type, and finally storing them in the system's underlying layer in DAG format.
[0097] The standard clinical causal template graph constructed in this embodiment, along with its accompanying time-series hard constraint operator, possesses automatic circuit breaker verification capabilities. The underlying judgment rule of the operator is t(Y)-t(X)>0; if the standard clinical causal template graph contains a standard directed edge X→Y, then the billing time of the corresponding original medical settlement invoice item Y must be later than that of the treatment item X. Taking the standardized clinical time-series link EX-08 (skin test) → DR-05 (cephalosporin drugs) → NS-02 (infusion service fee) as an example, if unsupervised causal mining is interfered with by data noise, and a reverse candidate causal proposition pointing from DR-05 to EX-08 is generated, the time-series hard constraint operator directly triggers the circuit breaker mechanism, automatically eliminating the atomic causal proposition that violates the clinical time sequence, without the need for manual maintenance of the rule base. Simultaneously, the standard clinical causal template graph supports batch entry of fixed sequence constraints such as surgical treatment codes and postoperative care codes, and automatic verification.
[0098] The above solution automatically completes the structured conversion of massive medical texts through entity recognition, eliminating the need for manual entry of clinical rules line by line, and significantly reducing the cost of building and maintaining the template library. The DAG storage mode naturally avoids the contradiction of cyclic timing, and with the mutual exclusion triple constraint, it provides a rigorous medical knowledge framework for subsequent causal graph learning, reduces false causal edges that violate clinical common sense generated by unsupervised mining, and improves the overall accuracy of anomaly identification.
[0099] In some embodiments, in step S3, the atomic causal propositions are screened, and the retained atomic propositions after screening are taken as valid causal associations, specifically including:
[0100] The atomic causal proposition is matched with the standard diagnostic causal template diagram to determine whether there is a directed path in the standard diagnostic causal template diagram that is consistent with the causal transmission direction indicated by the atomic causal proposition.
[0101] If a consistent directed path exists, the conditional independence of the atomic causal proposition is tested using the first significance threshold.
[0102] If there is no consistent directed path in the standard diagnosis and treatment causal template graph, but there is an indirect path connecting the two diagnosis and treatment items, then the conditional independence test is performed using the second significance threshold. The significance level value corresponding to the second significance threshold is less than the significance level value corresponding to the first significance threshold.
[0103] If there is no path connecting the two treatment items in the standard diagnosis and treatment causal template graph, then a third significance threshold is used to perform a conditional independence test, and the significance level value corresponding to the third significance threshold is less than the significance level value corresponding to the second significance threshold.
[0104] Atomic propositions that pass the conditional independence test are retained as valid causal relationships.
[0105] In this embodiment, atomic causal proposition path matching refers to retrieving the connection link between two diagnostic items in a standard DAG template.
[0106] Conditional independence testing is a basic algorithm in statistics used to determine whether there is a true causal relationship between two items.
[0107] Significance threshold It refers to the statistical test criteria; the smaller the value, the more stringent the test criteria.
[0108] A direct directed edge is a directed link that connects two nodes in one step; an indirect path is a link that connects two nodes through at least one intermediate node; a disconnected path is a link in the standard causal diagnosis template graph where there is absolutely no link between the two nodes.
[0109] The working principle of this embodiment is as follows:
[0110] For each group of diagnostic and treatment items, a three-level path matching hierarchical verification is performed on the (X,Y) atomic proposition: At the first level, if a direct directed path exists in the standard causal diagnostic and treatment template graph that is completely consistent with the causal direction of the atomic proposition, it means that the association fully complies with clinical norms, and a relaxed significance threshold is used. To prevent the erroneous exclusion of genuine and compliant causal relationships from sparse subsets of original medical settlement invoice details, conditional independence tests are conducted. At the second level, if the standard causal diagnosis template has no direct edges but contains indirect connected paths of length ≥2, it indicates an indirect clinical relationship between the two, and a standard significance threshold is used. Standard validation; Level 3: If there is no connection between diagnostic items X and Y in the standard causal diagnostic template, it means that there is no clinical correlation between the two, which is suspected of being a bundled arbitrage, and a strict significance threshold is applied. Strict verification is required. Only atomic propositions that pass all significance tests at the corresponding level will be retained as valid causal relationships and proceed to the subsequent topology optimization process.
[0111] This embodiment decomposes all diagnosis and treatment items within the multidimensional time-series matrix into pairwise (X,Y,D) atomic causal proposition triples that can be computed in parallel. Relying on a standard diagnosis and treatment causal template to limit the candidate search range, it compresses the original exponential causal edge search space to a polynomial scale, solving the problems of dimensionality curse and computational overload in high-dimensional operations under massive amounts of original medical settlement invoice details. After screening all valid causal associations through a three-level significance test, an initial candidate causal graph is constructed. A Greedy Equivalence Search (GES) algorithm is used for iterative topology optimization. The optimization evaluation formula is: Evaluation Score = Data Fit Score - Structure Deviation Penalty Term. If the candidate directed edges generated in the iteration conflict with the time-series direction of the standard diagnosis and treatment causal template, the penalty term significantly reduces the comprehensive score, forcing the algorithm to converge to a causal graph of invoice instances that fully conforms to clinical common sense.
[0112] For each line in the cause-effect graph of the invoice instance For directed edges, the temporal transit entropy is used to quantify the causal transmission strength between treatment items. The formula for calculating the transit entropy is as follows:
[0113] ;
[0114] in, and These represent the probability distributions of the historical billing status of treatment items X and Y within the standardized time interval, respectively. The current status of subsequent treatment projects; transfer entropy. The higher the value, the stronger the cascading amplification effect of the pre-treatment item X on the cost expansion of the subsequent treatment item Y. It is a core input indicator for quantifying the risk of subsequent cost changes and the loss of medical insurance payment funds.
[0115] The above scheme adopts a three-level differentiated significance threshold to achieve a hierarchical verification logic of "relaxed release of compliant associations and strict control of suspicious associations". It will not lose the causal relationship of normal clinical medication due to sparse samples, and can accurately intercept the bundled prescription associations without medical basis. It reduces the missed detection rate of excessive medical behavior and the false alarm rate of normal diagnosis and treatment from the source of causal mining, and significantly improves the authenticity and compliance of the causal graph topology of invoices.
[0116] In some embodiments, after setting the time-series hard constraint operator in step S2, the method further includes dynamically updating the time-series hard constraint operator, specifically including the following steps:
[0117] The number of billing records corresponding to each treatment item in the multidimensional time series matrix within the current accounting period is counted as the frequency of occurrence of that treatment item.
[0118] Treatment items that occur less frequently than a preset low-frequency threshold are marked as low-frequency treatment items. For directed edges in the standard treatment causal template graph where both ends are low-frequency treatment items, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a first coefficient value. For directed edges where only one end is a low-frequency treatment item, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a second coefficient value. For directed edges where neither end is a low-frequency treatment item, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a third coefficient value. The first coefficient value is less than the second coefficient value, and the second coefficient value is less than the third coefficient value.
[0119] Substituting the adjusted constraint coefficients into step S3, after the atomic causal propositions have passed the conditional independence test, for atomic causal propositions that pass the conditional independence test but whose causal transmission direction is inconsistent with the direction of the directed edge corresponding to the standard diagnosis and treatment causal template diagram, the product of the constraint coefficients and the significance threshold used in the conditional independence test is used as the adjusted significance threshold, and atomic causal propositions that fail the adjusted significance threshold test are removed from the effective causal associations.
[0120] In this embodiment, low-frequency medical services items refer to drugs, examinations, or consumables items whose total number of records is lower than the system's preset low-frequency threshold after statistically analyzing all billing records of each medical service item within the multi-dimensional time-series matrix of a single accounting period.
[0121] The constraint coefficient refers to the weight adjustment parameter bound to the temporal hard constraint operator, which is used to dynamically adjust the strictness of the removal of causal edges that violate the standard temporal order.
[0122] The adjusted significance threshold is a secondary criterion obtained by multiplying the significance threshold of the atomic causal proposition by the corresponding directed edge constraint coefficient. The smaller the constraint coefficient, the more stringent the secondary criterion.
[0123] Dynamic updating of temporal hard constraint operators refers to an adaptive mechanism that adjusts the constraint strength according to the frequency of occurrence of diagnosis and treatment items, thereby achieving differentiated temporal control for high-frequency routine diagnosis and treatment and low-frequency unpopular items.
[0124] In this embodiment, the time-series hard constraint operator can dynamically and adaptively adjust based on the frequency of occurrence of medical treatment items, thus perfectly adapting to the differentiated data characteristics of sparse samples of infrequent medical treatment items and abundant data of frequent medical treatment items in a large number of invoices. The specific working principle is as follows:
[0125] The system first traverses the multi-dimensional time series matrix generated in the current accounting cycle, counts the number of full billing records for each type of diagnosis and treatment item, and uses this as the frequency of the item's occurrence. Items with a frequency of occurrence lower than the preset low-frequency threshold are marked as low-frequency diagnosis and treatment items.
[0126] Subsequently, all directed edges in the standard diagnosis and treatment causal template graph are assigned hierarchical values: if both ends of a directed edge are low-frequency diagnosis and treatment items, the constraint coefficient of the temporal hard constraint operator corresponding to the edge is set to the first coefficient value; when only one end node is a low-frequency item, the constraint coefficient is set to the second coefficient value with a larger value; when both ends are high-frequency routine diagnosis and treatment items, the constraint coefficient is taken as the largest third coefficient value, and the three satisfy the condition that the first coefficient value < the second coefficient value < the third coefficient value.
[0127] The adjusted hierarchical constraint coefficients are simultaneously transmitted to the causal graph topology optimization process in step S3. For atomic causal propositions that pass the initial screening of level three significance but whose causal transmission direction conflicts in reverse with the directed edge of the standard template, the system multiplies the constraint coefficient of the corresponding level of the atomic causal proposition with the previous level three differential significance threshold to generate the adjusted significance threshold, and then conducts a second test. Atomic propositions that fail this second test are directly removed from the set of valid causal associations. To address the pain point of low-frequency diagnosis and treatment projects having few natural samples and being prone to generating false reverse causality, the system proactively tightens the temporal constraint strength to avoid sparse data noise generating causal edges that violate clinical common sense. For high-frequency diagnosis and treatment projects with sufficient data, the constraints are appropriately relaxed to retain real clinical specific temporal associations.
[0128] The above scheme achieves dynamic hierarchical adjustment of temporal hard constraint operators, setting differentiated constraint strengths for high-frequency routine medical services and low-frequency infrequent medical services, thus addressing the limitation of uniform fixed constraint coefficients failing to adapt to differences in sample sparsity. For causal links where both ends are low-frequency truth items, constraints are significantly tightened to filter out spurious reverse causality caused by insufficient sample size at the source; constraints are looser for high-frequency medical services, without interfering with the retention of genuine special clinical temporal relationships. A double-layer verification is achieved through secondary adjustment of the significance threshold, forming a dual temporal constraint barrier in conjunction with structural deviation penalty terms. This further reduces the number of abnormal edges within the causal graph that violate medical common sense, improving the authenticity and effectiveness of the causal graph of the invoice instance, while also reducing the probability of misidentification of abnormal subgraphs in subsequent steps and decreasing the workload of manual review by auditors.
[0129] In some embodiments, calculating the cost variation risk value based on the difference between the first cost data and the second cost data in step S4 includes:
[0130] For each redundant edge, obtain the transit entropy value of the previous treatment item to the next treatment item in the causal graph of the ticket instance in the redundant edge;
[0131] The total billing amount of the next treatment item in all treatment item coding sequences that include the previous treatment item in the multidimensional time series matrix is obtained as the first cost data, and the total billing amount of the next treatment item in all treatment item coding sequences that do not include the previous treatment item in the multidimensional time series matrix is obtained as the second cost data.
[0132] The difference between the first cost data and the second cost data is multiplied by the transmission entropy value to obtain the cost change risk value.
[0133] In this embodiment, redundant edges refer to abnormal directed edges where the causal graph of the ticket instance exists but the standard diagnosis and treatment causal template graph has no corresponding connected path, representing bundled behaviors without clinical indications.
[0134] The first cost data refers to the total cost of subsequent treatment items Y in the subset of all original medical settlement invoices containing redundant pre-treatment items X. The second cost data refers to the total cost of subsequent treatment items Y in the subset of all original medical settlement invoices that do not contain pre-treatment items X.
[0135] Cost variation risk value is a quantitative indicator that integrates the strength of causal transmission and the cost difference, and intuitively represents the degree of abnormal consumption of medical insurance funds caused by a single bundled edge.
[0136] The transfer entropy value, calculated in step S3 based on temporal information theory, represents the cascading amplification effect of the prescribing behavior of treatment item X on the cost scale of treatment item Y.
[0137] This embodiment extracts and performs numerical calculations step by step for each redundant edge identified by topological comparison. The specific workflow is as follows:
[0138] The first step is to read the transit entropy TEX→Y of the redundant edge X→Y stored in the causal graph of the ticket instance. This value is obtained by calculating the transit entropy formula from the standardized time series matrix, reflecting the strength of the causal transmission of the treatment item X to Y.
[0139] The second step is to traverse the complete multidimensional time series matrix and perform group statistics. Filter all treatment sequences that contain pre-treatment items X, and accumulate the total billing amount of post-treatment items Y in all samples to obtain the first cost data. Then, filter all treatment sequences that do not contain treatment items X, and accumulate the total billing amount of treatment items Y in the corresponding samples to obtain the second cost data.
[0140] The third step calculates the difference between the two sets of cost data, multiplies this difference by the transitive entropy value, and finally outputs the cost change risk value corresponding to the redundant edge. This calculation method, combined with the dimension of actual fund consumption, quantifies the severity of violations, avoiding the generation of invalid clues that only contain topological anomalies but lack actual medical insurance losses.
[0141] This embodiment couples the causal transmission strength with the difference in grouping costs to calculate the risk value, achieving a dual-dimensional quantitative assessment of topological anomalies and fund losses, thus overcoming the limitations of relying solely on topological differences to identify redundant edges. It can accurately distinguish between harmless noise associations that are "only structurally abnormal with no actual cost inflation" and genuine arbitrage links that "significantly increase the total cost of drugs / examinations through bundling," greatly reducing the output of invalid anomaly clues and lowering the computational cost of subsequent subgraph extraction and evidence chain generation. The risk value can intuitively reflect the severity of violations of a single bundling edge, providing auditors with a basis for tiered handling and improving the accuracy and audit value of identifying medical insurance fund losses.
[0142] In some embodiments, in step S4, redundant edges whose cost variation risk values exceed a preset baseline are identified as abnormal propagation edges. All abnormal propagation edges are aggregated to extract an illegal arbitrage subgraph, specifically including:
[0143] Obtain the first-level category label and the second-level category label corresponding to the diagnosis and treatment items at both ends of each redundant edge. The first-level category label is the department category code to which the diagnosis and treatment item belongs, and the second-level category label is the item category code to which the diagnosis and treatment item belongs.
[0144] Redundant edges with the same first-level category label for the treatment items at both ends are grouped into a first-level redundant edge set, and the arithmetic mean of the cost variation risk values of all redundant edges in the first-level redundant edge set is calculated to obtain the first-level risk mean.
[0145] For each first-level redundant edge set, redundant edges with the same second-level category label for both ends of the diagnosis and treatment items in the first-level redundant edge set are merged into a second-level redundant edge set, and the arithmetic mean of the cost variation risk values of all redundant edges in the second-level redundant edge set is calculated to obtain the second-level risk mean.
[0146] Each redundant edge in the first-level redundant edge set whose first-level risk average exceeds the preset first-level baseline, and each redundant edge in the second-level redundant edge set whose second-level risk average exceeds the preset second-level baseline, are identified as abnormal propagation edges.
[0147] All redundant edges identified as anomalous propagation edges are aggregated to extract the illegal arbitrage subgraph. Multiple anomalous propagation edges with shared diagnosis and treatment nodes belong to the same illegal arbitrage subgraph.
[0148] In this embodiment, the first-level category label refers to the department code to which the diagnosis and treatment item belongs, such as internal medicine, laboratory medicine, surgery, etc.; the second-level category label refers to the major category code of the sub-item under the same department, such as blood routine, CT imaging, and biochemical testing under the laboratory medicine department.
[0149] The first / second level redundant edge set refers to the grouping of redundant edges after clustering by two dimensions: department and project category. The level risk mean refers to the arithmetic mean of the cost change risk values of all redundant edges in the group.
[0150] Connected Component Analysis (CCA) is a graph theory clustering algorithm that aggregates multiple abnormal propagation edges with shared nodes into independent illegal arbitrage subgraphs.
[0151] This implementation addresses the technical problem that independent judgment of a single redundant edge is easily affected by individual cases, and that misjudging sporadic behavior as systemic arbitrage occurs by refining the hierarchical clustering logic for judging abnormal propagation edges and extracting illegal arbitrage subgraphs. The specific working principle is as follows:
[0152] The system first assigns two-level classification labels to the treatment items at both ends of each redundant edge: department as the primary classification and project category as the secondary classification. The first step involves clustering by the primary department label, grouping redundant edges with items belonging to the same department into the same first-level redundant edge set. The arithmetic mean of the cost variation risk values of all redundant edges within the group is calculated to obtain the first-level risk mean. The second step involves further clustering within each first-level redundant edge set by the secondary project category label, generating a second-level redundant edge set and calculating the corresponding secondary risk mean. The system configures two independent baseline thresholds for the department dimension and the project category dimension. All redundant edges within a group whose primary risk mean exceeds the department baseline or whose secondary risk mean exceeds the project category baseline are uniformly marked as abnormal propagation edges.
[0153] After completing the hierarchical screening, the connected component analysis algorithm is called to aggregate all abnormal propagation edges with shared diagnosis and treatment nodes, and divide them into independent topological networks that are not connected to each other. Each independent network is a complete illegal arbitrage subgraph, corresponding to the fixed routine over-medical behavior of medical institutions.
[0154] Topological differential operations simultaneously identify two types of abnormal topologies: the first is redundant edges, where the causal graph of the invoice instance exists, but the standard treatment causal template lacks corresponding directed edges, representing arbitrage through causal bundling of treatment items without indication; the second is missing edges, where the required clinical transmission directed edges specified in the standard treatment causal template are largely missing in the institution's full invoice causal network. For example, if clinical guidelines require core treatment items to be accompanied by safety testing items, and the proportion of missing corresponding standard directed edges exceeds a preset threshold, it is marked as a clue for arbitrage involving reduced compliance services and uniformly included in the subsequent tiered risk assessment process. This method simultaneously identifies both types of non-compliant topologies, overcoming the limitation of existing technologies that can only identify bundling arbitrage but cannot identify service reduction arbitrage.
[0155] Topological variations alone are insufficient for compliance assessment. To ensure the review conclusions have practical audit value for medical insurance audits, the system needs to incorporate a cascading amplification effect check based on cost dimensions. For each identified pair of topologically redundant items X→Y, the system uses the transit entropy value output from step S3 to quantitatively calculate the corresponding cost variation risk value. The calculation formula is as follows:
[0156] ;
[0157] in, The transitive entropy of treatment item X to treatment item Y; The total billing amount for treatment item Y is the total amount of all original medical settlement invoices containing treatment item X. This refers to the total billing amount for treatment Y from all original medical settlement invoice details that do not include treatment item X. If the cost variation risk value calculated from a single redundant edge is significantly higher than the preset overall cost baseline threshold, it can be determined that the arbitrage behavior corresponding to this topological redundancy without any indication of causal linkage will cause irreversible losses to the medical insurance pooled payment funds.
[0158] The aforementioned solution uses a two-tiered clustering approach—by department and project category—to calculate the average risk value. This filters out isolated, high-risk redundant edges that occasionally arise from individual documents, accurately identifying large-scale and routine bundled arbitrage activities at the department and project category levels, thus reducing false alarms caused by the specificity of individual cases. Layered baselines match the differences in cost baselines across different departments and treatment categories, avoiding misjudgments of specialized treatment behaviors based on a uniform baseline. Connectivity component analysis integrates scattered abnormal edges into a complete network of violations, visually presenting the overall topology of institutional arbitrage, facilitating regulators' complete understanding of the chain of violations and simplifying evidence collection.
[0159] In some embodiments, in step S5, for each atomic causal jump, the amount of medical insurance fund loss caused by the atomic causal jump is quantified by combining the transit entropy value corresponding to the atomic causal jump in the causal graph of the invoice instance and the billing data related to the medical treatment items. Specifically, this includes:
[0160] For each atomic causal jump, obtain the transmission entropy value of the previous treatment item to the next treatment item in the causal graph of the invoice instance in the atomic causal jump, and obtain the total billing amount of the previous treatment item in the original medical settlement invoice details data of all abnormal causal paths that trigger the atomic causal jump in the current accounting period, and obtain the medical insurance fund payment ratio coefficient corresponding to the previous treatment item.
[0161] Multiplying the transfer entropy value by the total billing amount and the medical insurance fund payment ratio coefficient yields the amount of medical insurance fund loss caused by the atomic causal jump.
[0162] In this embodiment, atomic causal jump refers to the binary jump unit of adjacent diagnosis and treatment items after the abnormal causal path of the illegal arbitrage subgraph is split.
[0163] The reimbursement ratio coefficient of the medical insurance pooled fund refers to the fixed proportion of the cost borne by the medical insurance pooled account to the total cost of each treatment item at the time of settlement.
[0164] The working principle of this embodiment is as follows:
[0165] For each decomposed atomic causal jump X→Y, a three-step indicator extraction and calculation is performed: First, the transit entropy TEX→Y of the redundant edge corresponding to the atomic causal jump is read as the causal binding trigger strength weight; second, all original invoices containing the causal path of this abnormal jump within the current accounting period are summarized, and the total billing amount of the preceding medical treatment item X in all invoices is accumulated; finally, the medical insurance fund payment ratio coefficient μ corresponding to the medical treatment item X is retrieved, which is automatically generated by the hospital settlement system based on medical insurance policies. The transit entropy, the total billing amount of the medical treatment item, and the payment ratio coefficient are multiplied to obtain the medical insurance fund loss amount corresponding to the atomic causal jump. This calculation method transforms the abstract graph theory causal relationship into a real absolute value of fund loss with audit value, enabling the loss amount of each illegal jump to be quantified and statistically summarized separately.
[0166] Preferred amount of medical insurance fund loss The calculation formula is as follows:
[0167] ;
[0168] in, Let X be the actual transfer entropy of the cost transmission from treatment item X to treatment item Y. For reference to conduction strength, in the case of illegal bundling where this side is completely missing from the standard template, ; This refers to the total actual billing amount for treatment item X in all medical settlement invoices that triggered this pattern within the current accounting period. This represents the corresponding actual reimbursement ratio coefficient of the medical insurance pooled fund.
[0169] The aforementioned solution integrates three indicators—causal transmission strength, total project cost, and medical insurance reimbursement ratio—to quantify fund losses, outputting precise loss amounts that can be directly used for auditing and accounting. This addresses the pain point of existing technologies, which can only qualitatively determine violations but cannot quantitatively calculate fund losses. Each bundled link can be individually tracked for losses, allowing regulators to summarize overall violation losses from multiple dimensions, including projects, departments, and institutions. This provides precise financial data support for medical insurance refusal to pay and administrative penalties, significantly enhancing the evidentiary value and practicality of intelligent audit conclusions in administrative enforcement.
[0170] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the compliance review method based on time-series causal tracing of medical invoices as described in the first aspect of the present invention.
[0171] The computer-readable storage medium may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.
[0172] The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device.
[0173] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of the present invention are intended to include these and any other suitable types of memory.
[0174] like Figure 2 As shown, in a third aspect, the present invention provides an electronic device 10, including a processor 101 and a storage medium 102, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the compliance review method based on medical invoice chronological causal tracing as described in the first aspect of the present invention.
[0175] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may use at least one of the following: circuit, single or multiple application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), central processing units (CPUs), controllers, microcontrollers, and microprocessors, thereby enabling the processor to execute some or all of the steps, or any combination thereof, of the compliance review method based on time-series causal tracing of medical invoices as described in the various embodiments of the present invention.
[0176] Finally, it should be noted that although the above embodiments have been described in the description and drawings of this invention, this should not limit the scope of patent protection of this invention. Any technical solutions that are based on the essential concept of this invention, utilize the content described in the description and drawings of this invention to make equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this invention.
Claims
1. A compliance review method based on chronological causal tracing of medical invoices, characterized in that, The method includes: S1: Obtain the original medical settlement invoice details of the medical institution under review in multiple consecutive accounting periods, extract the project code and billing time of each diagnosis and treatment item from the original medical settlement invoice details, and sort the diagnosis and treatment items in a standardized manner according to the billing time to construct a multi-dimensional time series matrix. The multi-dimensional time series matrix contains the diagnosis and treatment item code sequence and corresponding cost characteristics. S2: Obtain basic medical knowledge data, extract the temporal constraint relationship between diagnosis and treatment items from the basic medical knowledge data, construct a standard diagnosis and treatment causal template graph and set a temporal hard constraint operator. The standard diagnosis and treatment causal template graph includes diagnosis and treatment item nodes and directed edges. Each directed edge is used to indicate that there is a temporal sequence relationship between the two connected diagnosis and treatment items. The temporal hard constraint operator is used to constrain the causal inference results that do not conform to the temporal sequence relationship indicated by the directed edge. S3: Combine the diagnosis and treatment items in the multidimensional time series matrix into pairs to form diagnosis and treatment item groups. Decompose each diagnosis and treatment item group into atomic causal propositions to characterize the candidate causal transmission relationship and transmission direction between the two diagnosis and treatment items. Using the standard diagnosis and treatment causal template graph as a prior constraint, filter the atomic causal propositions. Select the atomic propositions that are retained after filtering as effective causal associations. Perform topology optimization on the effective causal associations. Use the transfer entropy algorithm to calculate the transfer entropy value of the optimized effective causal associations. Output the ticket instance causal graph. The ticket instance causal graph includes the causal topology structure and the transfer entropy value. S4: Perform a topological comparison between the invoice instance causal graph and the standard treatment causal template graph to identify redundant edges in the invoice instance causal graph that are different from those in the standard treatment causal template graph; for the identified redundant edges, combine the transmission entropy value corresponding to the redundant edge in the invoice instance causal graph, and the first cost data of the treatment items at both ends of the redundant edge when including the previous treatment item and the second cost data when not including the previous treatment item, calculate the cost change risk value based on the difference between the first cost data and the second cost data, and confirm the redundant edges whose cost change risk value exceeds the preset baseline as abnormal transmission edges, aggregate all abnormal transmission edges, and extract the illegal arbitrage subgraph; S5: Perform topological analysis on the illegal arbitrage subgraph, extract abnormal causal paths, and decompose the abnormal causal paths into atomic causal jump sequences; for each atomic causal jump, combine the transmission entropy value corresponding to the atomic causal jump in the causal graph of the invoice instance and the billing data of the treatment items involved to quantify the amount of medical insurance fund loss caused by the atomic causal jump; map the atomic causal jump sequence to the physical data rows of the original database through an inverted index, bind and generate data anchors, and assemble the abnormal causal path, the amount of medical insurance fund loss, and the data anchors into a white-box causal evidence chain; S6: Perform logical consistency checks on the white-box causal evidence chains sequentially. The logical consistency checks include causal direction checks and data anchor validity checks. Perform a cross-time window reproducibility stability assessment based on bootstrapping sampling on the white-box causal evidence chains that pass the logical consistency checks. Score the white-box causal evidence chains according to the results of the logical consistency checks and the cross-time window reproducibility stability assessment. Archive the white-box causal evidence chains that reach the preset score threshold to the confirmation database, and archive the white-box causal evidence chains that do not reach the threshold to the clue database.
2. The compliance review method based on the chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S1, the standardized ranking of medical services based on billing time includes: For each medical record in the details of each original medical settlement invoice, extract the billing timestamp corresponding to each treatment item, sort the treatment items in ascending order according to the order of the billing timestamps, and generate the treatment item code sequence. For medical records spanning multiple days, they are sorted in ascending order of billing timestamp values.
3. The compliance review method based on chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S2, the temporal constraints between diagnostic and treatment items are extracted from the basic medical knowledge data, and a standard diagnostic and treatment causal template diagram is constructed, specifically including: By analyzing unstructured text in drug instructions, medical insurance catalog payment scope texts, and clinical operation guidelines using a medical entity recognition mechanism, the system identifies the names of treatment items and descriptive words describing the sequential execution order or mutual exclusion relationship between treatment items. The recognition results are then converted into structured temporal triples, which include a first treatment item identifier, a second treatment item identifier, and a temporal relationship type between the first and second treatment items. The temporal relationship type includes any one of the following: the first treatment item occurs before the second treatment item, the second treatment item occurs before the first treatment item, or the first and second treatment items occur mutually exclusively. Aggregate all the aforementioned time-series triples, using the diagnosis and treatment item identifiers in the time-series triples as nodes and the time-series relationship type as the direction basis for directed edges, and construct a standard diagnosis and treatment causal template graph stored in the form of a directed acyclic graph.
4. The compliance review method based on chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S3, the atomic causal propositions are filtered, and the retained atomic propositions are considered as valid causal relationships. Specifically, this includes: The atomic causal proposition is matched with the standard diagnostic causal template diagram to determine whether there is a directed path in the standard diagnostic causal template diagram that is consistent with the causal transmission direction indicated by the atomic causal proposition. If a consistent directed path exists, the conditional independence of the atomic causal proposition is tested using the first significance threshold. If there is no consistent directed path in the standard diagnosis and treatment causal template graph, but there is an indirect path connecting the two diagnosis and treatment items, then the conditional independence test is performed using the second significance threshold. The significance level value corresponding to the second significance threshold is less than the significance level value corresponding to the first significance threshold. If there is no path connecting the two treatment items in the standard diagnosis and treatment causal template graph, then a third significance threshold is used to perform a conditional independence test, and the significance level value corresponding to the third significance threshold is less than the significance level value corresponding to the second significance threshold. Atomic propositions that pass the conditional independence test are retained as valid causal relationships.
5. The compliance review method based on chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S2, after setting the time-series hard constraint operator, the method further includes dynamically updating the time-series hard constraint operator, specifically including the following steps: The number of billing records corresponding to each treatment item in the multidimensional time series matrix within the current accounting period is counted as the frequency of occurrence of that treatment item. Treatment items that occur less frequently than a preset low-frequency threshold are marked as low-frequency treatment items. For directed edges in the standard treatment causal template graph where both ends are low-frequency treatment items, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a first coefficient value. For directed edges where only one end is a low-frequency treatment item, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a second coefficient value. For directed edges where neither end is a low-frequency treatment item, the constraint coefficient of the temporal hard constraint operator corresponding to the directed edge is adjusted to a third coefficient value. The first coefficient value is less than the second coefficient value, and the second coefficient value is less than the third coefficient value. Substituting the adjusted constraint coefficients into step S3, after the atomic causal propositions have passed the conditional independence test, for atomic causal propositions that pass the conditional independence test but whose causal transmission direction is inconsistent with the direction of the directed edge corresponding to the standard diagnosis and treatment causal template diagram, the product of the constraint coefficients and the significance threshold used in the conditional independence test is used as the adjusted significance threshold, and atomic causal propositions that fail the adjusted significance threshold test are removed from the effective causal associations.
6. The compliance review method based on chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S4, the calculation of the cost variation risk value based on the difference between the first cost data and the second cost data includes: For each redundant edge, obtain the transit entropy value of the previous treatment item to the next treatment item in the causal graph of the ticket instance in the redundant edge; The total billing amount of the next treatment item in all treatment item coding sequences that include the previous treatment item in the multidimensional time series matrix is obtained as the first cost data, and the total billing amount of the next treatment item in all treatment item coding sequences that do not include the previous treatment item in the multidimensional time series matrix is obtained as the second cost data. The difference between the first cost data and the second cost data is multiplied by the transmission entropy value to obtain the cost change risk value.
7. The compliance review method based on chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S4, redundant edges whose cost variation risk values exceed a preset baseline are identified as abnormal propagation edges. All abnormal propagation edges are aggregated to extract the illegal arbitrage subgraph, specifically including: Obtain the first-level category label and the second-level category label corresponding to the diagnosis and treatment items at both ends of each redundant edge. The first-level category label is the department category code to which the diagnosis and treatment item belongs, and the second-level category label is the item category code to which the diagnosis and treatment item belongs. Redundant edges with the same first-level category label for the treatment items at both ends are grouped into a first-level redundant edge set, and the arithmetic mean of the cost variation risk values of all redundant edges in the first-level redundant edge set is calculated to obtain the first-level risk mean. For each first-level redundant edge set, redundant edges with the same second-level category label for both ends of the diagnosis and treatment items in the first-level redundant edge set are merged into a second-level redundant edge set, and the arithmetic mean of the cost variation risk values of all redundant edges in the second-level redundant edge set is calculated to obtain the second-level risk mean. Each redundant edge in the first-level redundant edge set whose first-level risk average exceeds the preset first-level baseline, and each redundant edge in the second-level redundant edge set whose second-level risk average exceeds the preset second-level baseline, are identified as abnormal propagation edges. All redundant edges identified as anomalous propagation edges are aggregated to extract the illegal arbitrage subgraph. Multiple anomalous propagation edges with shared diagnosis and treatment nodes belong to the same illegal arbitrage subgraph.
8. The compliance review method based on the chronological causal tracing of medical invoices as described in claim 1, characterized in that, In step S5, for each atomic causal jump, the amount of medical insurance fund loss caused by the atomic causal jump is quantified by combining the transit entropy value corresponding to the atomic causal jump in the causal graph of the invoice instance and the billing data of the relevant medical treatment items. Specifically, this includes: For each atomic causal jump, obtain the transmission entropy value of the previous treatment item to the next treatment item in the causal graph of the invoice instance in the atomic causal jump, and obtain the total billing amount of the previous treatment item in the original medical settlement invoice details data of all abnormal causal paths that trigger the atomic causal jump in the current accounting period, and obtain the medical insurance fund payment ratio coefficient corresponding to the previous treatment item. Multiplying the transfer entropy value by the total billing amount and the medical insurance fund payment ratio coefficient yields the amount of medical insurance fund loss caused by the atomic causal jump.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the compliance review method based on the temporal causal tracing of medical invoices as described in any one of claims 1 to 8.
10. An electronic device having a computer program stored thereon, characterized in that, It includes a processor and a storage medium, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the compliance review method based on the time-series causal tracing of medical invoices as described in any one of claims 1 to 8.
Citation Information
Patent Citations
Intelligent medical insurance compliance examination method and system based on graph retrieval enhanced generation
CN121998779A
DRG high sleeve identification method, medium and equipment based on bill backstepping and causal blotting
CN122286194A
