A medical collaborative claim auditing method and system based on evidence chain driving

Abstract

The present application relates to a kind of medical collaborative claim auditing method and system based on evidence chain driving, belong to medical data processing technical field.The method includes: obtaining diagnosis and treatment, expense and voucher material, collection space adjacent field, establish cross material entity association and solidify as cross material reference edge, generate medical auditing evidence sequence;Reasoning node is constructed, upstream adoptable reference edge set is used as downstream evidence addressing constraint, form evidence reasoning chain;Evidence sequence is input into reasoning chain, node is positioned along reference edge diagnosis and the association of charging entity, executes indication matching and payment range check;The verification path subset that needs to be cooperated and verified is pushed to multi-role collaborative auditing end, and the evidence addressing constraint is updated by binding the weight of reference edge with the rejection decision, the association of node and downstream node is positioned and the adaptability check is re-executed, and payment instruction is generated.The evidence domain is gradually converged and end-to-end traceable audit is realized, and the efficiency and accuracy of claim settlement audit are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical data processing technology, specifically relating to a medical collaborative claims review method and system based on evidence chain-driven approach. Background Technology

[0002] With the continuous advancement of medical informatization and the ongoing expansion of commercial health insurance coverage, the number of medical insurance claims is growing rapidly, placing higher demands on the efficiency of claims review. Currently, medical claims review generally faces problems such as low review efficiency, high labor costs, and unstable review quality. Existing claims review technical solutions mainly have the following shortcomings.

[0003] First, serial review schemes based on rule chains struggle to dynamically narrow the scope of judgment. Existing technologies apply the chain-of-responsibility model to claims data review. This involves setting multiple claims data review rules arranged according to preset rules on the chain of responsibility, and then sequentially inputting the claims data to be reviewed into each rule node. This type of scheme achieves a structured organization of the review process through the rule chain, with clear logic and low error rates for each rule node. However, this type of scheme is essentially a "rule execution chain"—data flows sequentially through each node, and the judgments of each node are independent. The review conclusion of the previous node does not narrow the scope of judgment for the next node, and there is a lack of substantial logical dependency between nodes, making it impossible to dynamically narrow the scope of review.

[0004] Second, the reasoning process of intelligent review solutions based on large language models is opaque. In recent years, with the development of large model technology, some solutions have attempted to use large language models or thought chain technology to review claims materials. These solutions automatically extract claims information and generate review conclusions through the semantic understanding capabilities of large models, which improves review efficiency to some extent. However, the reasoning process of these solutions is opaque to the reviewers, and the generation path of the review conclusions is difficult to trace. In highly regulated scenarios such as medical claims, the interpretability of the review results faces significant challenges.

[0005] Third, existing solutions lack a mechanism for tracing the entire review process and for incremental review. When human intervention is required, current claims review technologies typically transfer the entire case to a human reviewer from scratch, failing to achieve incremental review while retaining existing reasoning results. Furthermore, the reasoning path in the review process lacks systematic recording and tracing methods, making it difficult to meet the auditing needs of heavily regulated scenarios.

[0006] To address the aforementioned issues, this invention provides a medical claims review method and system based on evidence chain-driven approach, thereby enabling complete traceability and incremental correction throughout the review process, and improving the efficiency and accuracy of medical claims review. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a medical collaborative claims review method and system based on evidence chain drive. The objective of this invention can be achieved through the following technical solutions: Obtain medical treatment materials, expense materials and voucher materials for medical claims, collect the spatial adjacency fields within the same material, establish cross-material entity associations, solidify entity associations that meet preset conditions into cross-material reference edges, and generate a medical review evidence sequence; The reasoning nodes are constructed according to the preset review order, and the set of admissible reference edges output by the upstream nodes is used as the evidence addressing constraint for the downstream nodes, thus forming an evidence reasoning chain. The medical audit evidence sequence is input into the evidence reasoning chain. The reasoning nodes locate the association pairs between the diagnostic entity and the billing entity in the evidence sequence along the reference edge. The association pairs are input into the pre-set clinical diagnosis and treatment standard library and medical insurance catalog to perform indication matching and payment scope verification. The verification path of each node is output. The subset of verification paths requiring collaborative verification is pushed to the multi-role collaborative review end. The rejection judgment of the review end is bound to the credibility weight of the corresponding reference edge to update the evidence addressing constraints. Based on the updated reference edge, the location and adaptability verification of the association pair of the node and downstream node are re-executed to generate a compensation instruction.

[0008] Specifically, the process of forming the cross-material reference edge includes: For a text area within the same material, multiple fields describing the same medical fact are grouped into a set of spatially adjacent fields based on the spatial adjacency distance and layout relationship between fields. Extract entity attributes from different materials that point to the same medical fact, and establish cross-material entity associations when the similarity between entity attributes exceeds a preset entity matching threshold; Calculate the association confidence of the cross-material entity association. When the association confidence reaches a preset solidification threshold, solidify the cross-material entity association into a cross-material reference edge. The reference edge includes the source entity identifier, the target entity identifier, and the initial confidence weight.

[0009] Specifically, the cross-material reference edges are represented by directed edges, with the direction pointing from the charging entity to the diagnostic entity; multiple reference edges constitute a reference edge set, and the medical review evidence sequence includes an evidence graph structure composed of a set of entity nodes and a set of reference edges.

[0010] Specifically, the construction process of the inference node includes: A reasoning chain topology graph is constructed with each review node as a vertex and the evidence addressing dependency relationship between each review node as a directed edge. The evidence addressing propagation direction is determined based on the direction of the directed edge. Starting from the initial node in the propagation direction, each vertex is extracted sequentially along the direction of the directed edge and arranged into a sequence of reasoning nodes. A transmission channel for a set of credible reference edges is established between adjacent reasoning nodes to form an evidence reasoning chain.

[0011] Specifically, the evidence addressing constraints include: After the upstream node completes the verification, it outputs a set of acceptable reference edges consisting of the verified reference edges; When a downstream node initiates verification, it determines the set of credible reference edges as the addressing range of the entity association pair of the current node, and locates the corresponding billing entity and diagnosis entity from the medical review evidence sequence along the source entity identifier and target entity identifier of each reference edge in the set of credible reference edges, thus forming the entity association pair to be verified.

[0012] Specifically, the execution of the evidence addressing constraint includes: Based on the set of credible reference edges received by the current node, an admission determination is performed on the reference edges in the medical review evidence sequence, and the reference edges that pass the determination are loaded into the associated pair positioning buffer of the current node.

[0013] Specifically, the process of locating the association pair between the diagnostic entity and the charging entity includes: Traverse each reference edge in the set of credible reference edges for the current node, and extract the corresponding charging entity and diagnostic entity from the entity nodes of the evidence sequence according to the source entity identifier and target entity identifier of the reference edge. The extracted billing entity and the diagnostic entity are combined into an association pair, which includes drug-indication pair, surgery-indication pair or consumable-indication pair.

[0014] Specifically, the indication matching and reimbursement scope verification includes: The disease code of the diagnostic entity is compared with the indication mapping table in the clinical diagnosis and treatment guidelines library to determine whether the charging entity belongs to the compliant treatment item under the clinical diagnosis, and an indication matching result is generated. The medical insurance catalog code of the charging entity is compared with the medical insurance payment scope rules to determine whether the charging entity is in the current medical insurance payment catalog and whether it meets the payment restriction conditions, and a payment scope verification result is generated. Based on the combined indication matching results and payment scope verification results, the verification path of the current node is output. The verification path includes the associated pair identifier, matching result, verification result, and node confidence.

[0015] Specifically, the update process of the evidence addressing constraints includes: Analyze the rejection judgment returned by the collaborative review end and locate the target reference edge corresponding to the rejected association pair; Based on the type of rejection judgment, the credibility weight of the target reference edge is reduced, and an updated reference edge weight is generated; Based on the updated reference edge weights, the evidence addressing constraints of downstream nodes are reconstructed. Each reference edge in the set of admissible reference edges is traversed, and reference edges with credibility weights lower than a preset weight threshold are removed from the set of admissible reference edges.

[0016] Specifically, the association between the node and its downstream nodes for location and compatibility verification includes: Using the node that received the rejection decision as the replay starting node, the location and adaptability verification of the associated pairs of the node and each downstream node are re-executed according to the updated evidence addressing constraints. During the replay process, the verification paths of each node upstream of the starting node remain unchanged, the verification paths of the starting node and downstream nodes are updated, and an updated complete case reasoning path is generated. The adaptability verification includes the process of verifying the matching of the diagnosis and the indications of the chargeable items based on the medical knowledge base, and verifying the payment scope of the chargeable items based on the medical insurance policy database.

[0017] Specifically, the generation of the compensation instruction includes: Based on the set of credible reference edges output by the final node, a compensation constraint domain is constructed, which includes the association pairs verified by all nodes and the corresponding chargeable entity amount data; Within the compensation constraint domain, compensation rule deduction is performed, and related pairs with no compensation space are gradually eliminated based on medical insurance payment policies and insurance product terms, thus shrinking the compensation space to the final compensation domain; The compensation amount is calculated within the final compensation domain, and a compensation instruction is generated that includes the compensation amount, a list of acceptable reference edges, a rule matching identifier, and a compensation condition code.

[0018] The beneficial effects of this invention are as follows: (1) This invention, by constructing an evidence reasoning chain and establishing evidence addressing constraints between adjacent reasoning nodes, realizes the step-by-step transmission and dynamic contraction of the evidence addressing range, significantly improving the automation and interpretability of claims review. In existing claims review schemes based on the chain of responsibility, the judgments of each rule node are independent of each other, and the review conclusion of the previous node does not have any contraction effect on the scope of judgment of the next node. This invention uses the set of admissible reference edges output by the upstream node as the evidence addressing constraint of the downstream node, so that the addressing range of entity association pairs of the downstream node is dynamically limited by the verification conclusion of the upstream node, and the set of admissible reference edges narrows step-by-step as the reasoning chain progresses. Specifically, after the upstream node completes the indication matching and payment range verification, it only includes the verified reference edges into the set of admissible reference edges. The downstream node only locates the association pairs of the diagnostic entity and the charging entity along the reference edges in this set, performs verification within the addressing range limited by this set, until the compensation calculation node generates a compensation instruction within the finally converged set of admissible reference edges. This "reference-edge driven constraint transmission" mechanism ensures that every step of the review process is based on evidence and traceable, overcoming the technical shortcomings of opaque reasoning processes and difficulty in tracing review conclusions in existing technologies, and improving the interpretability and credibility of medical claims review.

[0019] (2) This invention updates the evidence addressing constraints by binding the rejection judgment of the review end with the credibility weight of the corresponding reference edge, and re-executes the location and adaptability verification of the association pairs of the node and downstream nodes based on the updated reference edge. This realizes the structured injection of human review experience into the inference chain, avoiding the efficiency loss caused by the overall review from scratch in the traditional scheme. When it is necessary to push a subset of the verification path to the multi-role collaborative review end, this invention binds the rejection judgment of the review end with the credibility weight of the corresponding reference edge, lowers the weight of the target reference edge and reconstructs the evidence addressing constraints of the downstream node, and re-executes the location and adaptability verification of the association pairs of the node and downstream nodes based on the updated reference edge, while keeping the verification paths of the upstream nodes of the node unchanged. This "local replay based on reference edge weight update" mechanism enables human review experience to be injected into the inference chain in the form of structured weight parameters, and the inference correction can be completed without overall re-review, which significantly reduces the workload of manual review, improves the efficiency of collaborative review, and enables the inference chain to continuously learn and optimize from human feedback. Attached Figure Description

[0020] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating a medical collaborative claims review method based on evidence chain drive according to the present invention. Figure 2This is a schematic diagram of the evidence reasoning chain architecture and constraint transmission in this invention; Figures 3-5 This is a schematic diagram of the audit operation panel in this invention. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0023] Please see Figures 1-5 A medical claims review method based on evidence chain drive includes: Example 1: Generation of Medical Review Evidence Sequence This embodiment provides a medical claims review method based on evidence chain drive, which is applied to a medical insurance claims review system. The system includes a material acquisition module, an evidence encoding module, an inference engine module, a collaborative review interface module, and a claims execution module.

[0024] Step 1: Obtain the medical treatment materials, expense materials, and voucher materials for medical claims, collect the spatial adjacency fields within the same materials, establish cross-material entity associations, solidify the entity associations that meet the preset conditions into cross-material reference edges, and generate a medical review evidence sequence.

[0025] In this embodiment, the medical insurance claims review system first receives claims materials from hospital information systems, pharmacy settlement systems, or uploaded by users. These claims materials include medical records (such as medical records, diagnostic certificates, prescriptions, and examination reports), expense materials (such as detailed expense lists and inpatient settlement invoices), and supporting documents (such as invoices, electronic drug supervision codes, and payment vouchers).

[0026] Taking a claim case for targeted therapy drugs for lung cancer as an example, the system retrieves the following materials: medical records include the outpatient medical record and prescription issued by the hospital, which record the patient's diagnosis of "malignant lung tumor" and the prescription for "osimertinib 80mg"; expense materials include the inpatient expense details list and medical insurance settlement statement, which record the chargeable item "osimertinib (80mg×30 tablets)," the unit price "4980 yuan," and the quantity "1"; and supporting documents include a VAT electronic general invoice with the product name "osimertinib."

[0027] After the system completes the material acquisition, it enters the evidence sequence generation stage. This stage consists of three sub-steps: collection of spatial adjacency fields within the same material, establishment of cross-material entity associations, and solidification of entity associations into cross-material reference edges.

[0028] I. Collection of Spatial Adjacency Fields within the Same Material For text regions within the same material, the system identifies multiple fields describing the same medical fact and groups them into a set of spatially adjacent fields based on the spatial adjacency distance and layout relationship between fields.

[0029] Specifically, the system performs OCR recognition and layout analysis on each document to obtain the bounding box coordinates (x and y coordinates of the top left corner, and width and height) of each text field. For the medical document "Prescription", the system identifies the fields "Diagnosis: Malignant Lung Tumor" and "Drug: Osimertinib 80mg". These two fields are vertically adjacent on the page and the distance between them is less than the preset adjacency threshold (in this embodiment, the adjacency threshold is set as the upper limit of the field spacing in the same or adjacent rows). The system determines that the two fields are spatially adjacent and groups them into a set of spatially adjacent fields describing the medical fact that "the patient has malignant lung tumor and is being treated with osimertinib".

[0030] Similarly, for the expense material "Inpatient Expense Details List", the system identifies fields such as "Item Name: Osimertinib (80mg×30 tablets)", "Unit Price: 4980 yuan", "Quantity: 1", and "Amount: 4980 yuan". These fields are arranged horizontally in the same table row with zero spatial adjacency. The system collects them into a set of spatially adjacent fields that describe the medical fact of "Osimertinib drug charge details".

[0031] II. Establishment of Cross-Material Entity Associations The system extracts entity attributes pointing to the same medical fact from different materials and determines whether there is a cross-material entity association by calculating the similarity between entity attributes.

[0032] Specifically, the system extracts entity attribute values ​​corresponding to the same medical fact from medical records, expense records, and voucher records. Taking the drug entity "Osimertinib" as an example: the entity attribute extracted from the medical records is "Osimertinib 80mg" (from the drug name field of the prescription), the entity attribute extracted from the expense records is "Osimertinib (80mg×30 tablets)" (from the item name field of the expense list), and the entity attribute extracted from the voucher records is "Osimertinib" (from the product name field of the invoice).

[0033] The system calculates the string similarity between the above entity attributes. The similarity calculation formula is: Sim string (s1,s2)=(1-edit distance (s1,s2)) / (max(len(s1),len(s2)), Among them, edit distance (s1,s2) represents the minimum number of single-character editing operations (including insertion, deletion, and replacement) required to convert string s1 to string s2. len(s1) and len(s2) represent the character lengths of the two strings, respectively; `max` means taking the larger of the two values.

[0034] Calculations showed that the similarity between the medical document "osimertinib 80mg" and the cost document "osimertinib (80mg x 30 tablets)" was 0.76; the similarity between the medical document and the voucher document "osimertinib" was 0.82; and the similarity between the cost document and the voucher document was 0.71. When the similarity between entity attributes exceeds the preset entity matching threshold θ... entity (θ in this embodiment) entity =0.65 (based on the distribution of name variations of the same drug on different materials in historical claims data). When the system determines that these entity attributes point to the same medical fact, it establishes cross-material entity association.

[0035] III. Entity associations are solidified into cross-material reference edges. The system calculates the association confidence of each cross-material entity association. When the association confidence reaches the preset solidification threshold, the entity association is solidified into a cross-material reference edge.

[0036] Specifically, the calculation of association confidence comprehensively considers the following three factors: (1) the string similarity between the attributes of the entities involved in the association; (2) the complementarity of the material types involved in the association (the combination of treatment + cost + voucher is given a higher complementary weight); and (3) the completeness of the number of fields in the spatial adjacency field set. The weighted fusion formula of the three factors is as follows: Confidence = α·Sim attr +β·W complement +γ·N adj / N expected , Among them, Sim attr For entity attribute similarity, W complement Complementary weights for material types (1.0 for complete treatment, fees, and documentation; 0.7 for missing one; 0.3 for missing two), N adj N represents the number of fields actually aggregated in the spatial adjacency field set. expected To describe the expected number of fields required to describe the medical fact (e.g., the expected number of fields for a drug fact, including name, specifications, quantity, and amount), α, β, and γ are weighting coefficients. In this embodiment, α = 0.5, β = 0.3, γ = 0.2, and α + β + γ = 1.

[0037] Taking the aforementioned "Osimertinib" entity association as an example: Sim attr Taking the maximum pairwise similarity score of the three materials as 0.82, W complement =1.0 (all three materials are complete), N adj=4 (Name, specifications, quantity, and amount have all been collected), N expected =4, then Confidence = 0.5 × 0.82 + 0.3 × 1.0 + 0.2 × 1.0 = 0.91.

[0038] When Confidence ≥ preset curing threshold θ cure (θ in this embodiment) cure When the confidence level is 0.75, the system solidifies the cross-material entity association into a cross-material reference edge. The reference edge includes: a source entity identifier (the charging entity, such as "Osimertinib fee item"), a target entity identifier (the diagnostic entity, such as "lung malignant tumor diagnosis"), and an initial confidence weight (taking an association confidence level of 0.91 as the initial value). The direction of the reference edge points from the charging entity to the diagnostic entity, indicating whether "this fee item has a corresponding compliant diagnostic basis".

[0039] Similarly, the system executes the above three sub-steps for other medical facts of the patient (such as the correlation between examination items and diagnosis, the correlation between surgical items and diagnosis, etc.), generating multiple cross-material reference edges. These multiple reference edges, together with their corresponding entity nodes, constitute the evidence graph structure. The entity node set includes treatment entity nodes (diagnosis, symptoms, examination results, etc.) and billing entity nodes (medications, surgery, consumables, examinations, etc.), while the reference edge set contains each cross-material reference edge. This evidence graph structure serves as the medical review evidence sequence, providing standardized data input for the subsequent construction of the evidence reasoning chain and constraint-based verification.

[0040] Through the above steps, the system transforms multi-source, heterogeneous, unstructured claims materials into a structured evidence graph, achieving a standardized conversion from raw materials to computable evidence objects. Each reference edge carries a source entity identifier, a target entity identifier, and an initial credibility weight, providing a complete topological foundation for subsequent inference nodes to address entity association pairs along the reference edges.

[0041] Example 2: Construction of the Evidence Reasoning Chain Step 2: Construct reasoning nodes according to the preset review order, and use the set of admissible reference edges output by the upstream nodes as evidence addressing constraints for the downstream nodes, thus forming an evidence reasoning chain.

[0042] This embodiment builds upon the medical review evidence sequence generated in Embodiment 1 (i.e., an evidence graph structure composed of a set of entity nodes and a set of reference edges), and constructs an inference chain for this evidence graph structure. The construction of the inference chain requires determining three core elements: the order of the inference nodes, the constraint propagation relationships between nodes, and the propagation channels of the set of admissible reference edges between adjacent nodes.

[0043] I. Constructing the inference chain topology graph The system first constructs a reasoning chain topology graph. The reasoning chain topology graph is a directed graph C=(H,D), where H is the set of audit nodes and D is the set of directed edges. The directed edges represent the evidence addressing dependency relationship between upstream and downstream nodes, and the direction is from the upstream node to the downstream node.

[0044] The directed edges are determined based on the calling patterns between review nodes during the historical review process. The system extracts the calling order and dependency frequency between nodes from the historical review logs. When the dependency frequency exceeds a preset frequency threshold θ, the system will determine the origin of the edge. freq (θ) freq When setting the statistical significance of the call relationships between nodes based on historical audit logs, confirm the existence of an evidence addressing dependency between two nodes and establish a directed edge. like Figure 2 As shown, in this embodiment, the set of review nodes H contains five inference nodes: h1 (material integrity verification node), h2 (reasonableness verification node), h3 (indication matching node), h4 (payment scope verification node), and h5 (compensation calculation node). The evidence addressing dependency relationship between each node is as follows: h1 is the starting node of the inference chain, with an in-degree of 0, and does not depend on the output of other nodes; the determination of h2 depends on the evidence integrity information output by h1, with a directed edge h1→h2; the determination of h3 depends on the reasonableness determination conclusion output by h2, with a directed edge h2→h3; ​​the determination of h4 depends on the indication matching review conclusion output by h3, with a directed edge h3→h4; the determination of h5 depends on the set of reference edges that pass the payment scope verification output by h4, with a directed edge h4→h5. Therefore, the direction of the directed edges is: h1→h2→h3→h4→h5.

[0045] II. Determine the direction of evidence addressing and propagate the sequence of reasoning nodes. The system determines the evidence addressing propagation direction along the direction of the directed edges. In this embodiment, the evidence addressing propagation direction is a unidirectional path from h1 to h5. Starting from the initial node h1, the system sequentially extracts each vertex along the direction of the directed edges and arranges them into a reasoning node sequence H. seq =[h1,h2,h3,h4,h5].

[0046] III. Establish a channel for passing the set of trustworthy reference edges between adjacent nodes. The system establishes a transmission channel for the set of admissible reference edges between adjacent inference nodes, forming an evidence inference chain. The function of the transmission channel is as follows: after the upstream node completes the verification, it outputs a set of admissible reference edges consisting of the verified reference edges; this set is directly mapped to the evidence addressing constraints of the downstream node through the transmission channel, that is, when the downstream node starts the verification, it determines the set of admissible reference edges as the addressing range of the entity association pairs of the current node.

[0047] Specifically, upstream node h i After verification, the association confidence and verification conclusion are extracted from the reference edges that pass the verification. Reference edges with an association confidence not lower than the preset reference edge confidence threshold are grouped into a set L of admissible reference edges. i ⊆E (where E is the complete set of reference edges in the evidence graph structure). Downstream node h i+1 When starting the verification, L i The addressing range for the entity association pair determined to be itself is only along L. i The source entity identifier and target entity identifier of each reference edge are used to locate the corresponding charging entity and diagnostic entity from the entity nodes of the evidence sequence, forming the entity association pair to be verified. (Not included in L) i The reference edge in the middle, whose corresponding entity association pair is not within the addressing range of the downstream node.

[0048] The transmission channels between adjacent nodes constitute the propagation chain of evidence addressing constraints: L1 output by h1 is passed to h2 as an addressing constraint, L2 output by h2 is passed to h3, L3 output by h3 is passed to h4, and L4 output by h4 is passed to h5. As the inference chain advances, the set of admissible reference edges narrows layer by layer—the verification conclusion of the upstream node directly limits which reference edges the downstream node can use to address entity association pairs, realizing the step-by-step propagation and dynamic contraction of the evidence addressing range.

[0049] like Figures 3 to 5 As shown, in this embodiment, the system pre-configures product information, pharmacy information, and review process node information. This configuration information serves as the rule source for evidence addressing constraints. The system loads this configuration information into the inference chain, which then acts as the basis for judgment during verification at each node.

[0050] Through the above steps, the system has completed the construction of the evidence reasoning chain, established a complete reasoning framework from material integrity verification to compensation calculation, and established evidence addressing constraints based on the transfer of credible reference edge sets between adjacent nodes, providing an execution basis for subsequent correlation pair location and adaptability verification.

[0051] Example 3: Correlation of Location and Adaptability Verification and Verification Path Generation Step 3: Input the medical audit evidence sequence into the evidence reasoning chain. The reasoning node locates the association pair between the diagnostic entity and the billing entity in the evidence sequence along the reference edge. Input the association pair into the pre-set clinical diagnosis and treatment standard library and medical insurance catalog to perform indication matching and payment scope verification, and output the verification path of each node.

[0052] I. Verification process of material integrity verification node h1 The starting node h1 is not subject to upstream evidence addressing constraints and can directly access the entire set of reference edges E in the evidence graph structure. This node performs a material integrity check on each reference edge, based on the pre-defined necessary evidence rule R. req The necessary evidence rule defines the minimum material coverage requirements for each medical fact (such as drug charges, examination charges, and surgery charges)—for example, drug charges require the presence of the prescription field from the medical records, the detailed field from the expense records, and the invoice field from the voucher records. Upon verification, M edges in the reference edge set E satisfy the material completeness requirement (M ≤ |E|), while the remaining (|E|-M) edges fail verification due to missing necessary materials. This node aggregates the M verified reference edges into a set of admissible reference edges L1 and outputs L1 as the evidence addressing constraint for the downstream node h2. The verification conclusion is: the reference edge completeness ratio is M / |E|. The confidence level of this node is C1 = M / |E|.

[0053] II. Verification process of reasonableness verification node h2 (constraint propagation) Reasonableness verification node h2 receives the set of admissible reference edges L1 (containing M reference edges) transmitted by h1. This node determines L1 as the current entity association pair addressing range, and only along the source entity identifier and target entity identifier of each reference edge in L1, locates the corresponding charging entity and diagnostic entity from the entity nodes of the evidence sequence to form the entity association pair to be verified.

[0054] Specifically, the node traverses each referencing edge e in L1. i ∈L1, based on the source entity identifier s of the reference edge ei id Extract the corresponding fee entity (e.g., "Osimertinib fee item: 4980 yuan") from the entity node set V, and then extract the fee entity based on the target entity identifier t. id Extract the corresponding diagnostic entity (e.g., "diagnosis of malignant lung tumor") and combine the two into an entity association pair. i =(Payment entity, Diagnosis entity).

[0055] This node is for Pair i A reasonableness verification is performed to determine whether the correspondence between the chargeable entity and the diagnostic entity is reasonable from the perspective of medical common sense. For example, for the association pair "Osimertinib cost → Lung cancer diagnosis", the system queries the built-in drug-disease common sense database and confirms that osimertinib is a standard targeted therapy drug for patients with EGFR mutation-positive lung cancer, so the correspondence is reasonable and the verification passes. However, for the association pair "Vitamin C injection cost → Lung cancer diagnosis", vitamin C injection does not constitute a targeted treatment for lung cancer, so the correspondence is unreasonable and the verification fails.

[0056] Upon verification, P out of the M reference edges passed the rationality check (P≤M), while (MP) failed. This node groups the P verified reference edges into a set of admissible reference edges L2 and outputs L2 as the evidence addressing constraint for the downstream node h3. The verification conclusion is: the rationality check pass rate is P / M. The confidence level of this node is C2 = P / M.

[0057] III. Verification process of indication matching node h3 (constrained transmission) The indication matching node h3 receives the set of acceptable reference edges L2 (containing P reference edges) passed by h2 and determines L2 as the addressing range. This node traverses each reference edge in L2, locates the corresponding billing entity and diagnosis entity, forms the entity association pair to be verified, and then inputs the association pair into the pre-set clinical diagnosis and treatment standard library to perform indication matching.

[0058] Specifically, the system compares the disease code of the diagnostic entity (such as the ICD-10 code "C34.9") with the indication mapping table in the clinical practice guidelines database. The clinical practice guidelines database includes indication information from clinical practice guidelines issued by the National Health Commission and drug instructions from various pharmaceutical companies, organized in key-value pairs of (disease code → set of compliant treatment item codes).

[0059] Using the reference edge e k Taking the corresponding association pair (osimertinib cost item, lung cancer diagnosis) as an example: the ICD-10 code of the diagnosis entity is "C34.9". The system retrieves the medical insurance drug code "A20B0023" for osimertinib from the set of compliant treatment item codes corresponding to this disease code in the indication mapping table. The system matches the medical insurance drug code of the chargeable entity with the code in the set. If the match is successful, the indication matching result is "passed", indicating that the chargeable entity belongs to the compliant treatment item under this diagnosis.

[0060] For association pairs that fail to match (e.g., the billing entity is "Coenzyme A for injection" and the diagnostic entity is "malignant lung tumor", where Coenzyme A is not in the compliant treatment item coding set for malignant lung tumor), the indication matching result is "failed".

[0061] After indication matching, Q out of P reference edges passed the matching (Q≤P), while (PQ) failed. This node aggregates the Q successful reference edges into a set of admissible reference edges L3 and outputs L3 as the evidence addressing constraint for the downstream node h4. The verification conclusion is: the indication matching success rate is Q / P. The confidence level of this node is C3 = Q / P.

[0062] IV. Verification process of payment scope verification node h4 (constraint propagation) The payment scope verification node h4 receives the set of acceptable reference edges L3 (containing Q reference edges) transmitted by h3 and determines L3 as the addressing scope. This node traverses each reference edge in L3, locates the corresponding charging entity and diagnosis entity, compares the medical insurance catalog code of the charging entity with the medical insurance payment scope rules, and performs payment scope verification.

[0063] Specifically, the system queries the current medical insurance payment catalog database to obtain the payment restrictions corresponding to the chargeable entity, including but not limited to: (1) whether it is in the current medical insurance catalog; (2) whether there are indication restrictions (payment is limited to specific diseases); (3) whether there are usage restrictions (such as the maximum number of boxes that can be paid per month); and (4) whether there are patient identity restrictions (such as outpatient / inpatient only). The system generates payment scope verification results after comparing each item.

[0064] Taking "Osimertinib Cost" as an example: The system confirms that osimertinib is included in the current National Reimbursement Drug List (No. A20B0023), and its payment restriction condition is "limited to second-line or higher treatment for patients with advanced non-small cell lung cancer who have epidermal growth factor receptor (EGFR) sensitive mutations." Combining the diagnostic entity "malignant lung tumor" and existing gene testing reports in the system (if available), it determines whether the patient meets the payment restriction condition. If yes, the payment scope verification result is "full payment available" or "proportional payment available" (depending on the specific insurance terms); if not yes (e.g., the patient has no EGFR mutation testing record), the verification result is "not payable."

[0065] After payment scope verification, R out of Q reference edges passed the verification (R≤Q), while (QR) failed. This node aggregates the R passing reference edges into a set of admissible reference edges L4 and outputs L4 as the evidence addressing constraint for downstream node h5. The verification conclusion is: the payment scope verification pass rate is R / Q. The confidence level of this node is C4 = R / Q.

[0066] The aforementioned indication matching and payment scope verification together constitute the adaptation verification of entity association pairs, which is a complete process of verifying the indication matching of diagnosis and chargeable items based on the medical knowledge base, and verifying the payment scope of chargeable items based on the medical insurance policy database.

[0067] V. Output of the verification path During the execution of the inference chain, each inference node h i After verification is completed, the system records the evidence addressing field A of that node. i (i.e., the set of acceptable reference edges L received) i-1 ), Verification conclusion Res i and confidence level C iWrite the trajectory sequence and append the verification records of each node to the trajectory sequence Trajectory in sequence according to the evidence addressing propagation direction to generate the case reasoning path: Trajectory=[(h1,A1,Res1,C1),(h2,L1,Res2,C2),...,(h4,L3,Res4,C4)] Among them, the evidence addressing constraints of adjacent nodes satisfy A i+1 =L i (The receiving set of the next node is equal to the output set of the previous node).

[0068] The case reasoning path reflects the entire process from material integrity verification to payment scope verification. The input constraints, verification basis, reasoning conclusions and confidence levels of each node are fully recorded, realizing complete traceability of the entire review process.

[0069] Example 4: Collaborative Review, Local Reasoning Reenactment, and Compensation Instruction Generation Step 4: Push the subset of verification paths that need to be collaboratively verified to the multi-role collaborative review end. Bind the rejection judgment of the review end with the credibility weight of the corresponding reference edge to update the evidence addressing constraints. Based on the updated reference edge, re-execute the location and adaptability verification of the association pair of the node and downstream node to generate a compensation instruction.

[0070] This embodiment builds upon the case reasoning path (Trajectory) generated in Embodiment 3. The system assigns the confidence level C of each node... i Compared with the preset confidence threshold θ conf (θ in this embodiment) conf Based on statistical analysis of node confidence and audit accuracy in historical audit cases, each node is compared to identify the abnormal node set Abnormal={h i |C i <θ conf}

[0071] I. Targeted push notifications and multi-role collaborative review of verification path subsets Let h j ∈Abnormal represents nodes with abnormal confidence levels (e.g., the h4 payment range verification node, whose confidence level C4 = R / Q; if C4 is low due to unclear payment restrictions on a certain reference edge, then h4 is marked as abnormal). The system will transfer nodes from h in the Trajectory... jThe initial subsequence (i.e., the relevant records of h4 and downstream h5, as well as the L3 received by h4, the verification conclusion Res4, and the confidence level C4) is designated as a "component of verification path requiring collaborative verification" and pushed to the multi-role collaborative review end. The pushed content includes: node identifier h4, confidence level C4, and complete verification records of each upstream node h1-h3, so that reviewers can fully understand the context.

[0072] After receiving the push notification, the multi-role collaborative review platform allows reviewers (such as medical review experts and medical insurance policy specialists) to re-examine nodes with abnormal confidence levels and their associated entity pairs. Reviewers do not need to review all materials from the beginning; they only need to focus on marking questionable pairs within the verification path subset. Upon completion of the review, reviewers input structured review comments into the collaborative review platform, including a "pass" or "reject" decision and the reasons for rejection.

[0073] II. Rejection Decision Binding to the Credibility Weight of Referenced Edges The system parses the structured review comments returned by the collaborative review end: Rev=(node id ,edge id ,reject type ), where node id To correct node identifiers (such as h4), edge id The identifier of the target reference edge corresponding to the rejected association pair (e.g., e) k (i.e., the reference to "Osimertinib cost item → diagnosis of lung malignancy"), reject type For rejection criteria (such as "payment restrictions not met" or "diagnosis is irrelevant to charges").

[0074] The system is based on edge id Locating the target reference edge e k And according to the rejection type Reduce the credibility weight of this reference edge: If the rejection type is "diagnosis is completely unrelated to billing", then e k The credibility weight was directly reduced to 0; If the rejection type is "Payment restrictions not met (e.g., lack of genetic testing report)," then e k The credibility weight is multiplied by the attenuation coefficient λ (in this embodiment, λ=0.2, that is, reduced to 20% of the original weight). If the rejection type is "partially compliant but requires supplementary explanation", then e k The credibility weight is multiplied by the attenuation coefficient μ (μ=0.6).

[0075] Use the target reference edge e kTaking osimertinib as an example, the initial credibility weight was 0.91, and the reviewer rejected the application under the category of "payment restrictions not met" (missing EGFR mutation detection report). The updated reference edge weight w k =0.91×0.2=0.182.

[0076] Next, the system reconstructs the evidence addressing constraints of downstream nodes based on the updated reference edge weights. The system iterates through each reference edge in the current set of admissible reference edges L3, assigning edges with credibility weights below a preset weight threshold θ. weight (θ in this embodiment) weight Edges with a reference value of 0.5 or less are removed from the set of acceptable reference edges. This is because w... k =0.182<0.5, referencing edge e k It is removed from the set. The updated set of admissible reference edges is denoted as L3', which contains fewer reference edges than Q to Q-1 (assuming only e). k (Removed).

[0077] III. Re-execute node and downstream node verification based on the updated reference edges. The system takes the node h4 that received the rejection decision as the replay starting node, and re-executes the correlation pair location and suitability verification of h4 and downstream node h5 according to the updated evidence addressing constraints (i.e., L3').

[0078] Specifically: h4 replay: h4 receives L3' as the new addressing range, locates the entity association pair only along the reference edges in L3', and re-executes the payment range check. Because e k The item has been removed. h4 no longer performs payment scope verification for the osimertinib expense item. Its verification conclusion is updated to "the remaining Q-1 reference edges have passed the verification". The updated set of acceptable reference edges L4' is output.

[0079] h5 execution: h5 receives the updated L4' as the addressing range, locates the charging entity and the diagnostic entity only along the reference edges in L4', and re-executes the compensation calculation (see below for details).

[0080] During the replay process, the system maintains the verification paths of all upstream nodes (h1, h2, h3) of the starting node h4 unchanged—that is, the records of h1, h2, and h3 in the Trajectory are preserved as is. The system then concatenates the upstream records with the updated records h4 and h5 obtained after the replay in sequence to form the updated complete case reasoning path: Trajectory'=[(h1,A1,Res1,C1),(h2,L1,Res2,C2),(h3,L2,Res3,C3),(h4,L3',Res4',C4')),(h5,L4',Res5',C5')].

[0081] The essential difference between local inference replay and global re-audit lies in the fact that global re-audit requires re-executing all five nodes starting from h1, while local replay only executes the verification of downstream nodes starting from h4. When the number of nodes in the inference chain is large, the computational complexity of local replay is reduced from O(|H|) to O(|H|-j+1), where j is the position index of the abnormal node in the inference chain, significantly saving the overhead of repeated verification of upstream nodes.

[0082] IV. Generation of Claim Instructions Within the compensation constraint domain, h5 performs compensation rule deduction: based on medical insurance payment policies (such as deductible, coverage limit, and reimbursement ratio) and insurance product terms (such as deductible, reimbursement ratio, and exclusions), it gradually eliminates related pairs with no compensation space. For example, if a referencing edge corresponds to a fee entity amount of 200 yuan, but this fee item is listed as an exclusion in the insurance terms, it will be removed from the compensation constraint domain. Through layer-by-layer rule deduction, the compensation space shrinks to the final compensation domain.

[0083] Within the final payout domain, h5 calculates the payout amount Pay. For example, if the final payout domain contains only one compliant infusion fee pair with an amount of 500 yuan, and the insurance policy stipulates a payout ratio of 80%, then Pay = 500 × 0.8 = 400 yuan.

[0084] h5 generates payment instruction Pay instruction It contains the following fields: Compensation Amount Field: Pay=400 yuan; Admissible reference edge list field: Lists the reference edge identifiers corresponding to each associated pair within the final compensation domain, used to trace the evidence objects on which compensation is based; Claims Rule Matching Identifier Field: Identifies the specific rule / clause number to which this claim applies; The compensation condition coding field indicates the prerequisites for compensation execution (such as "compensation can only be made after supplementing the genetic testing report" or "direct compensation").

[0085] like Figure 5 This demonstrates how nodes with abnormal confidence levels are presented in the review panel, and pushes the corresponding case reasoning path to the multi-role collaborative review end. Reviewers can then conduct a review based on the reasoning path, eliminating the need to review all materials from scratch.

[0086] At this point, the evidence chain-driven medical collaborative claims review method provided by this invention has been completed. The system outputs a structured claims instruction, completing the fully automated process from material acquisition to claims instruction generation. In the process, the structured injection of human review experience and continuous optimization of the reasoning chain are realized.

[0087] In summary, this solution unifies multi-source heterogeneous materials into an evidence graph structure by solidifying cross-material reference edges, establishes evidence addressing dependencies between review nodes through a reasoning chain topology graph, dynamically records and progressively converges the evidence addressing range through a set of admissible reference edges, updates evidence addressing constraints by binding the rejection judgment at the review end with the credibility weight of the corresponding reference edge, and achieves incremental reasoning correction starting from abnormal nodes by re-executing the association between nodes and downstream nodes based on the updated reference edges to verify positioning and adaptability. This forms a complete reference edge-driven evidence addressing constraint reasoning chain review architecture.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A medical collaborative claims review method based on evidence chain drive, characterized in that, include: Obtain medical treatment materials, expense materials and voucher materials for medical claims, collect the spatial adjacency fields within the same material, establish cross-material entity associations, solidify entity associations that meet preset conditions into cross-material reference edges, and generate a medical review evidence sequence; The reasoning nodes are constructed according to the preset review order, and the set of admissible reference edges output by the upstream nodes is used as the evidence addressing constraint for the downstream nodes, thus forming an evidence reasoning chain. The medical audit evidence sequence is input into the evidence reasoning chain. The reasoning nodes locate the association pairs between the diagnostic entity and the billing entity in the evidence sequence along the reference edge. The association pairs are input into the pre-set clinical diagnosis and treatment standard library and medical insurance catalog to perform indication matching and payment scope verification. The verification path of each node is output. The subset of verification paths requiring collaborative verification is pushed to the multi-role collaborative review end. The rejection judgment of the review end is bound to the credibility weight of the corresponding reference edge to update the evidence addressing constraints. Based on the updated reference edge, the location and adaptability verification of the association pair of the node and downstream node are re-executed to generate a compensation instruction.

2. The method according to claim 1, characterized in that, The process of forming the cross-material reference edge includes: For a text area within the same material, multiple fields describing the same medical fact are grouped into a set of spatially adjacent fields based on the spatial adjacency distance and layout relationship between fields. Extract entity attributes from different materials that point to the same medical fact, and establish cross-material entity associations when the similarity between entity attributes exceeds a preset entity matching threshold; Calculate the association confidence of the cross-material entity association. When the association confidence reaches a preset solidification threshold, solidify the cross-material entity association into a cross-material reference edge. The reference edge includes the source entity identifier, the target entity identifier, and the initial confidence weight.

3. The method according to claim 2, characterized in that, The cross-material reference edges are represented by directed edges, with the direction pointing from the charging entity to the diagnostic entity; multiple reference edges constitute a reference edge set, and the medical review evidence sequence includes an evidence graph structure composed of a set of entity nodes and a set of reference edges.

4. The method according to claim 1, characterized in that, The construction process of the inference node includes: A reasoning chain topology graph is constructed with each review node as a vertex and the evidence addressing dependency relationship between each review node as a directed edge. The evidence addressing propagation direction is determined based on the direction of the directed edge. Starting from the initial node in the propagation direction, each vertex is extracted sequentially along the direction of the directed edge and arranged into a sequence of reasoning nodes. A transmission channel for a set of credible reference edges is established between adjacent reasoning nodes to form an evidence reasoning chain.

5. The method according to claim 4, characterized in that, The evidence addressing constraints include: After the upstream node completes the verification, it outputs a set of acceptable reference edges consisting of the verified reference edges; When a downstream node initiates verification, it determines the set of credible reference edges as the addressing range of the entity association pair of the current node, and locates the corresponding billing entity and diagnosis entity from the medical review evidence sequence along the source entity identifier and target entity identifier of each reference edge in the set of credible reference edges, thus forming the entity association pair to be verified.

6. The method according to claim 5, characterized in that, The execution of the evidence addressing constraints includes: Based on the set of credible reference edges received by the current node, an admission determination is performed on the reference edges in the medical review evidence sequence, and the reference edges that pass the determination are loaded into the associated pair positioning buffer of the current node.

7. The method according to claim 1, characterized in that, The process of locating the association pair between the diagnostic entity and the billing entity includes: Traverse each reference edge in the set of credible reference edges for the current node, and extract the corresponding charging entity and diagnostic entity from the entity nodes of the evidence sequence according to the source entity identifier and target entity identifier of the reference edge. The extracted billing entity and the diagnostic entity are combined into an association pair, which includes drug-indication pair, surgery-indication pair or consumable-indication pair.

8. The method according to claim 7, characterized in that, The indication matching and reimbursement scope verification includes: The disease code of the diagnostic entity is compared with the indication mapping table in the clinical diagnosis and treatment guidelines library to determine whether the charging entity belongs to the compliant treatment item under the clinical diagnosis, and an indication matching result is generated. The medical insurance catalog code of the charging entity is compared with the medical insurance payment scope rules to determine whether the charging entity is in the current medical insurance payment catalog and whether it meets the payment restriction conditions, and a payment scope verification result is generated. Based on the combined indication matching results and payment scope verification results, the verification path of the current node is output. The verification path includes the associated pair identifier, matching result, verification result, and node confidence.

9. The method according to claim 1, characterized in that, The update process for the evidence addressing constraints includes: Analyze the rejection judgment returned by the collaborative review end and locate the target reference edge corresponding to the rejected association pair; Based on the type of rejection judgment, the credibility weight of the target reference edge is reduced, and an updated reference edge weight is generated; Based on the updated reference edge weights, the evidence addressing constraints of downstream nodes are reconstructed. Each reference edge in the set of admissible reference edges is traversed, and reference edges with credibility weights lower than a preset weight threshold are removed from the set of admissible reference edges.

10. The method according to claim 1 or 9, characterized in that, The association between the node and its downstream nodes includes location and compatibility verification, including: Using the node that received the rejection decision as the replay starting node, the location and adaptability verification of the associated pairs of the node and each downstream node are re-executed according to the updated evidence addressing constraints. During the replay process, the verification paths of each node upstream of the starting node remain unchanged, the verification paths of the starting node and downstream nodes are updated, and an updated complete case reasoning path is generated. The adaptability verification includes the process of verifying the matching of the diagnosis and the indications of the chargeable items based on the medical knowledge base, and verifying the payment scope of the chargeable items based on the medical insurance policy database.

11. The method according to claim 1, characterized in that, The generation of the compensation instruction includes: Based on the set of credible reference edges output by the final node, a compensation constraint domain is constructed, which includes the association pairs verified by all nodes and the corresponding chargeable entity amount data; Within the compensation constraint domain, compensation rule deduction is performed, and related pairs with no compensation space are gradually eliminated based on medical insurance payment policies and insurance product terms, thus shrinking the compensation space to the final compensation domain; The compensation amount is calculated within the final compensation domain, and a compensation instruction is generated that includes the compensation amount, a list of acceptable reference edges, a rule matching identifier, and a compensation condition code.

12. A medical collaborative claims review system based on evidence chain, comprising a processor for performing the method as described in any one of claims 1 to 11.

Images