Medical record intelligent auditing method and system based on versioned medical insurance rules and AI model

Abstract

The present application relates to the technical field of medical data processing, and more particularly to a medical record intelligent auditing method and system based on versioned medical insurance rules and AI models. The method obtains the medical record data to be audited and its timestamp information, parses the diagnosis and treatment event sequence and constructs a multi-version knowledge graph of medical insurance rules, traces the target rule version and its evolution path based on the timestamp, generates a version-aware vector through semantic tensor transformation and establishes a semantic alignment relationship with the diagnosis and treatment event sequence to generate a hybrid representation, feeds the hybrid representation into an adversarial dual-path auditing architecture, makes the dual paths constrain each other in the feature space and uses the divergent samples to guide the AI model to reconstruct the semantic space, and finally outputs the fused auditing conclusion. The present application effectively deals with the auditing differences brought by the iteration of medical insurance rule versions, and improves the auditing accuracy and adaptability.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method and system for intelligent review of medical records based on versioned medical insurance rules and AI models. Background Technology

[0002] In the healthcare field, medical insurance rule review systems typically employ a combination of static rule bases and machine learning models to conduct compliance checks on medical record data. The conventional approach involves first building a stable set of medical insurance rules, transforming them into machine-executable judgment conditions, and then training an AI model using historical labeled data to assist in identifying abnormal medical practices. During the review process, the system matches the medical events in the medical record text against the rule base one by one, while simultaneously using the AI ​​model to output risk probabilities. Finally, the results are aggregated manually or through a simple voting mechanism. This architecture provides a certain level of review efficiency when the rule versions are stable.

[0003] However, the above approach has two significant drawbacks. Firstly, frequent updates to medical insurance policies lead to rapid iterations of the rule base, but traditional systems lack the ability to explicitly model the rule evolution path. When a new rule takes effect, historical models often fail due to a mismatch between training data and the current rules, requiring the collection of new labeled data and significant computational resources for full-scale training, making a smooth transition during rule changes impossible. Secondly, there is a lack of effective coordination between rule-based logical reasoning and AI model semantic reasoning. They typically operate independently in a loosely coupled manner, failing to form mutual constraints within the feature space. When rules are updated, the AI ​​model adapts slowly to the new rule semantic space, causing significant fluctuations in review results during version transitions, reducing system credibility. Summary of the Invention

[0004] This invention provides a method and system for intelligent medical record review based on versioned medical insurance rules and AI models, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for intelligent medical record review based on versioned medical insurance rules and an AI model, comprising:

[0006] Obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information;

[0007] The rule change features in the evolution path are mapped to version-aware vectors through semantic tensor transformation, and a semantic alignment relationship is established between the version-aware vectors and the diagnosis and treatment event sequence to generate a hybrid representation carrying the semantics of rule evolution.

[0008] The hybrid representation is fed into an adversarial dual-path review architecture, where the first path generates a deterministic review result based on rule-based logical reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning. This allows the two paths to constrain each other in the feature space. When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under limited sample conditions. The final review conclusion is output based on the dual-path fusion result after adversarial convergence.

[0009] Obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information, including:

[0010] The historical medical insurance rule texts are parsed in a time sequence to extract the effective time interval, applicable scope boundary and constraint logic expression of each rule version. The constraint logic expression is decomposed into diagnosis and treatment behavior nodes, cost restriction nodes and condition judgment nodes. All nodes are connected by directed edges to represent the constraint transmission relationship between nodes. A single version rule subgraph is constructed based on the constraint transmission relationship.

[0011] The single-version rule subgraphs at different time points are topologically aligned on the time axis. By calculating the semantic similarity of nodes in the single-version rule subgraphs, the homologous nodes and mutated nodes across versions are identified. Version inheritance edges are established between homologous nodes, and version differentiation edges are established between mutated nodes. The version inheritance edges and version differentiation edges are used to connect each single-version rule subgraph into a temporal association structure, forming a multi-version knowledge graph of medical insurance rules.

[0012] Based on timestamp information, the effective time interval covering the timestamp is retrieved in the temporal association structure of the multi-version knowledge graph. The corresponding target rule version subgraph is located, and the version inheritance edge and version differentiation edge of the target rule version subgraph are traversed forward along the time axis to the initial rule version. The mutation nodes associated with the version differentiation edge during the traversal are recorded, and the constraint logic differences of the mutation nodes are extracted as the evolution path.

[0013] By mapping the rule change features in the evolutionary path to version-aware vectors through semantic tensor transformation, establishing a semantic alignment relationship between the version-aware vectors and the diagnosis and treatment event sequence, and generating a hybrid representation carrying the semantics of rule evolution, the following is achieved:

[0014] Semantically aggregate the mutated nodes associated with the differentiation edges of each version in the evolution path, form rule change features through constraint logic difference analysis, construct a third-order semantic tensor representation of the rule change features in the time dimension, constraint strength dimension and applicable scope dimension, perform tensor decomposition on the third-order semantic tensor representation along the time dimension to capture the dominant mode of rule evolution, and convert the dominant mode into a version-aware vector through low-dimensional manifold mapping.

[0015] The behavioral semantics, temporal relationships, and cost association attributes of each diagnosis and treatment event in the diagnosis and treatment event sequence are structured and encoded to construct a vector representation of the diagnosis and treatment event. The mutual information distribution between the version-aware vector and the vector representation of the diagnosis and treatment event in the semantic space is calculated. Based on the mutual information distribution, sensitive diagnosis and treatment events affected by rule evolution are screened, and a semantic alignment relationship is established between the sensitive diagnosis and treatment events and the version-aware vector.

[0016] The version-aware vector is dynamically modulated according to the semantic alignment relationship, so that the rule evolution pattern in the version-aware vector and the semantic features of the sensitive diagnosis and treatment event have a resonance enhancement effect. The modulated version-aware vector is fused with the overall semantic representation of the diagnosis and treatment event sequence by tensor product to generate a hybrid representation that simultaneously encodes the diagnosis and treatment behavior logic and the rule evolution trajectory.

[0017] Dynamically modulating the version-aware vector based on the semantic alignment relationship, so that the rule evolution pattern in the version-aware vector resonates and enhances the semantic features of the sensitive medical events, includes:

[0018] Based on semantic alignment, a coupled tensor field is constructed between the version-aware vector and the sensitive diagnosis and treatment event. In the coupled tensor field, the semantic response gradient of each dimension component of the version-aware vector to the sensitive diagnosis and treatment event is calculated. Based on the semantic response gradient, the important dimensions that need to be enhanced and the redundant dimensions that need to be suppressed are determined, and a dimension-selective modulation mask is generated.

[0019] The dimension-selective modulation mask is applied to the version-aware vector. The representation intensity of the important dimensions is amplified and the representation intensity of the redundant dimensions is attenuated through a gating mechanism. This causes the modulated version-aware vector to move in the semantic space toward the distribution area of ​​the sensitive diagnosis and treatment event. The local curvature matching degree between the moved version-aware vector and the semantic features of the sensitive diagnosis and treatment event on the feature manifold is calculated.

[0020] When the local curvature matching degree exceeds the coordination threshold, a resonance loop is established between the version perception vector and the sensitive diagnosis and treatment event in the coupled tensor field. Through the resonance loop, the rule evolution mode in the version perception vector and the semantic features of the sensitive diagnosis and treatment event generate synchronous oscillations in the corresponding dimensions of the feature space. The synchronous oscillations cause the semantic expressions of the two to reinforce each other in the corresponding dimensions, forming a resonance enhancement effect.

[0021] The hybrid representation is fed into an adversarial dual-path review architecture, wherein the first path generates a deterministic review result based on rule-based logical reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning, including:

[0022] The hybrid representation is decomposed into an explicit rule constraint subspace and an implicit semantic association subspace in the feature space. The projection vector of the explicit rule constraint subspace is fed into the first path. A constraint propagation network is constructed in the first path. The rule constraint conditions in the projection vector are forward reasoned and backward verified through the constraint propagation network. When the truth value determination of all constraint conditions converges, a binary deterministic audit result is generated.

[0023] The complete representation of the hybrid representation is fed into the second path, where the diagnosis and treatment behavior patterns and rule evolution patterns in the implicit semantic association subspace are jointly probabilistically modeled by an AI model. Based on the AI ​​model, the posterior probability of the hybrid representation under the anomaly detection distribution is calculated, and a probabilistic audit result with a continuous value range is generated based on the posterior probability.

[0024] In the second path, an AI model is used to jointly model the diagnostic behavior patterns and rule evolution patterns in the implicit semantic association subspace. Based on the AI ​​model, the posterior probability of the hybrid representation under the anomaly detection distribution is calculated, including:

[0025] In the implicit semantic association subspace, the behavioral sequence dependency structure of the diagnosis and treatment behavior pattern and the version migration trajectory of the rule evolution pattern are extracted. The behavioral sequence dependency structure and the version migration trajectory are geometrically aligned on the feature manifold through topological isomorphism mapping to generate a joint representation structure that encodes the dual-mode coupling relationship.

[0026] The AI ​​model performs conditional dependency learning on the treatment behavior patterns and rule evolution patterns in the joint representation structure, and establishes a conditional probability distribution of the treatment behavior patterns with the rule evolution patterns as conditional variables. The parameters of the conditional probability distribution are adaptively inferred by the AI ​​model from the joint representation structure, forming a joint probability modeling result that characterizes the impact of rule versions on treatment compliance.

[0027] The hybrid representation is input into the AI ​​model. Based on the joint probability modeling results, the probability density ratio of the hybrid representation under the normal diagnosis and treatment behavior distribution and the abnormal diagnosis and treatment behavior distribution is calculated. The anomaly confidence of the hybrid representation is calculated by combining the probability density ratio with the historical statistical prior of the rule evolution mode. The anomaly confidence is the posterior probability of the hybrid representation under the anomaly detection distribution.

[0028] When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under conditions of few samples. Based on the dual-path fusion result after adversarial convergence, the final review conclusion is output, including:

[0029] Disagreement samples are identified by calculating the decision vector distance between deterministic review results and probabilistic review results. The rule evolution path features associated with the disagreement samples are extracted, and the rule evolution path features are transformed into meta-learning signals containing semantic difference information between versions.

[0030] The meta-learning signal is fed into the meta-parameter update layer of the AI ​​model. The meta-learning signal guides the AI ​​model to construct a fast adaptation direction pointing to the semantics of the new version of the rules in the gradient space. The fast adaptation direction is used to make incremental adjustments to the semantic encoding parameters of the AI ​​model with few samples, so that the AI ​​model can complete the reconstruction of the semantic space of the new version of the rules in the local semantic space of the divergent samples.

[0031] In the adversarial dual-path review architecture, continuous iterative optimization is performed. When the consistency measure between the deterministic review result and the probabilistic review result on the divergent sample satisfies the adversarial convergence condition, the deterministic review result and the probabilistic review result are weighted by credibility to generate a dual-path fusion result, which is then used as the final review conclusion.

[0032] A second aspect of this invention provides a medical record intelligent review system based on versioned medical insurance rules and an AI model, comprising:

[0033] The version tracing unit is used to obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information.

[0034] A hybrid representation unit is used to map the rule change features in the evolution path into version-aware vectors through semantic tensor transformation, establish a semantic alignment relationship between the version-aware vectors and the diagnosis and treatment event sequence, and generate a hybrid representation carrying the semantics of rule evolution.

[0035] The dual-path review unit is used to feed the hybrid representation to the adversarial dual-path review architecture. The first path generates a deterministic review result based on rule logic reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning. This allows the two paths to constrain each other in the feature space. When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under limited sample conditions. The final review conclusion is output based on the dual-path fusion result after adversarial convergence.

[0036] A third aspect of the present invention provides an electronic device, comprising:

[0037] processor;

[0038] Memory used to store processor-executable instructions;

[0039] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0040] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0041] This method completely solves the problem of inconsistent review benchmarks caused by rule version changes by constructing a multi-version knowledge graph of medical insurance rules and combining it with timestamp information to trace the target rule version and its evolution path. Rule change features are fully mapped to version-aware vectors and semantically aligned with the sequence of medical events, thereby generating a hybrid representation carrying the semantics of rule evolution. This enables the system to accurately distinguish compliance requirements under rules from different periods, avoiding misjudgments due to old rules or omissions due to new rules, and significantly improving the accuracy and consistency of cross-version medical record review.

[0042] By feeding hybrid representations into an adversarial dual-path review architecture, the rule-based logical reasoning and AI semantic reasoning paths constrain and challenge each other in the feature space, integrating the advantages of both deterministic logic and probabilistic semantics. When the two paths reach the same conclusion, a high-confidence result is output; when they diverge, in-depth analysis of the branch samples is triggered. This mechanism effectively suppresses the bias of a single review path, significantly reducing false positives and false negatives, while retaining the ability to flexibly judge complex and marginal cases.

[0043] When the review process encounters disagreements due to rule version updates, the system transforms the rule evolution path features of the disagreeing samples into meta-learning signals, guiding the AI ​​model to quickly reconstruct the semantic space of the new version's rules under limited sample conditions. This capability allows the system to adapt to rule changes without relying on large-scale retraining with new data, significantly shortening the calibration cycle after version iteration. The dual-path fusion result after adversarial convergence outputs the final review conclusion, ensuring the stability and timeliness of the review results in the context of rule evolution. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the intelligent medical record review method based on versioned medical insurance rules and AI models, as described in an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of the intelligent review process for medical insurance records based on an adversarial dual-path architecture, as described in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0048] Figure 1 This is a flowchart illustrating the intelligent medical record review method based on versioned medical insurance rules and AI models, as described in an embodiment of the present invention.

[0049] Intelligent medical record review methods based on versioned medical insurance rules and AI models include:

[0050] Obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information;

[0051] The rule change features in the evolution path are mapped to version-aware vectors through semantic tensor transformation, and a semantic alignment relationship is established between the version-aware vectors and the diagnosis and treatment event sequence to generate a hybrid representation carrying the semantics of rule evolution.

[0052] The hybrid representation is fed into an adversarial dual-path review architecture, where the first path generates a deterministic review result based on rule-based logical reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning. This allows the two paths to constrain each other in the feature space. When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under limited sample conditions. The final review conclusion is output based on the dual-path fusion result after adversarial convergence.

[0053] In one optional implementation, the process involves obtaining medical record data to be reviewed and its timestamp information, parsing the sequence of medical events in the medical record data, constructing a multi-version knowledge graph of medical insurance rules, and tracing the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information, including:

[0054] The historical medical insurance rule texts are parsed in a time sequence to extract the effective time interval, applicable scope boundary and constraint logic expression of each rule version. The constraint logic expression is decomposed into diagnosis and treatment behavior nodes, cost restriction nodes and condition judgment nodes. All nodes are connected by directed edges to represent the constraint transmission relationship between nodes. A single version rule subgraph is constructed based on the constraint transmission relationship.

[0055] The single-version rule subgraphs at different time points are topologically aligned on the time axis. By calculating the semantic similarity of nodes in the single-version rule subgraphs, the homologous nodes and mutated nodes across versions are identified. Version inheritance edges are established between homologous nodes, and version differentiation edges are established between mutated nodes. The version inheritance edges and version differentiation edges are used to connect each single-version rule subgraph into a temporal association structure, forming a multi-version knowledge graph of medical insurance rules.

[0056] Based on timestamp information, the effective time interval covering the timestamp is retrieved in the temporal association structure of the multi-version knowledge graph. The corresponding target rule version subgraph is located, and the version inheritance edge and version differentiation edge of the target rule version subgraph are traversed forward along the time axis to the initial rule version. The mutation nodes associated with the version differentiation edge during the traversal are recorded, and the constraint logic differences of the mutation nodes are extracted as the evolution path.

[0057] When performing time-series analysis on historically published medical insurance rules, it is necessary to extract structured information from the original policy documents one by one. Each rule record contains three core attributes: effective time interval, applicable scope boundary, and constraint logic expression. The constraint logic expression usually appears in the form of condition-result, such as "If the number of days of hospitalization exceeds a certain threshold and the primary diagnosis belongs to a specific category, then the corresponding expenses will be included in the reimbursement scope at a specific ratio." This type of expression is further decomposed into three types of atomic nodes: treatment behavior nodes, expense restriction nodes, and condition judgment nodes. Nodes are connected by directed edges, and the direction of the directed edge indicates the direction of constraint propagation, i.e., the logical dependency relationship of "if condition A is met, constraint B is triggered." After integrating all nodes and their directed edges within the same version, a single-version rule subgraph is formed, which fully expresses the logical structure of the medical insurance rules in that version.

[0058] After constructing the single-version rule subgraph, it is necessary to perform topological alignment of the subgraphs at different time points along the timeline to discover cross-version node relationships. During the alignment process, the semantic similarity between nodes in the subgraphs of different versions is calculated. Specifically, the text description of each node is converted into a semantic embedding vector, and the node is denoted as... The semantic embedding vector is ,node The semantic embedding vector is Then the semantic similarity between the two nodes for:

[0059] ;

[0060] when Exceeding the preset homology threshold At that time, the node With nodes If nodes are identified as having the same origin, a version inheritance edge is established between them; when Below However, it is still within the mutation determination interval (i.e. ,in When this happens, both are identified as mutated nodes, and a version differentiation edge is established between them. This version differentiation edge records the differences in the constraint logic expressions of the two nodes. If... If two nodes belong to completely different rule entries, no cross-version association edge is established. Version inheritance edges and version differentiation edges are used to connect the various single-version rule subgraphs, forming a temporal association structure extending along the time axis, i.e., a multi-version knowledge graph of medical insurance rules. This graph not only preserves the rule logic of each version but also retains a clear evolutionary trajectory between versions, allowing rule changes between any two adjacent versions to be accurately traced through the attributes of the edges in the graph.

[0061] After constructing the multi-version knowledge graph, timestamp information is extracted for each medical record awaiting review. The timestamp typically corresponds to the actual date of the treatment event in the medical record, or the date of the medical insurance settlement application. Based on this timestamp, a range search is performed within the temporal association structure of the multi-version knowledge graph: the effective time range corresponding to each version subgraph is traversed to find the version that covers the timestamp, and the corresponding single-version rule subgraph is positioned as the target rule version subgraph. If the timestamp falls exactly at the boundary between two versions, the target version is determined according to the actual execution priority of the rule.

[0062] After locating the target rule version subgraph, a graph traversal is performed forward along the timeline along the version inheritance and version differentiation edges until the initial rule version is reached. During the traversal, all version differentiation edges traversed and their associated mutated node pairs are recorded. For each mutated node pair, the differences in its constraint logic expressions are extracted. Specific differences include: changes to the preconditions of condition judgment nodes, changes to the numerical constraints of cost limit nodes, and changes to the behavioral scope of treatment behavior nodes. These differences are arranged in chronological order to form an evolutionary path from the initial version to the target version. This evolutionary path, with the sequence of mutated nodes and their constraint logic differences as its core content, fully records every substantial change in the target rule from its initial form to its current form.

[0063] The extraction results of the evolution path directly affect the construction quality of subsequent version perception vectors. If there is a multi-hop propagation relationship between the mutated nodes associated with a version differentiation edge during the traversal, the indirect changes of downstream nodes must also be included in the evolution path record. This recursive propagation mechanism ensures that the evolution path can capture the complete scope of the rule change's impact, not just the directly modified nodes. For example, if the reimbursement limit of a certain expense restriction node changes, and this node constrains the applicable conditions of multiple medical behavior nodes through directed edges, then the indirect changes of these medical behavior nodes will also be recorded in the evolution path, thus providing more complete rule evolution information for subsequent semantic tensor transformations.

[0064] In practical engineering implementation, multi-version knowledge graphs need to support efficient time interval retrieval and graph traversal operations. An interval tree structure can be used to index the effective time intervals of each version, reducing the complexity of timestamp queries to logarithmic levels. For large-scale rule texts, the calculation of node semantic embedding vectors can be completed using a pre-trained medical domain language model to improve the accuracy of semantic similarity calculation between nodes. The constraint logic difference information recorded on version differentiation edges is stored in a structured differential format, facilitating direct reading and conversion into rule change feature vectors in subsequent steps. The entire multi-version knowledge graph construction process is updated incrementally when a new version of the medical insurance rule is released. Only topological alignment of the new version subgraph with the boundary version subgraph of the existing graph is required; the entire graph does not need to be reconstructed, thus ensuring sufficient response efficiency in real-world scenarios with frequent rule iterations.

[0065] In one optional implementation, the rule change features in the evolution path are mapped to version-aware vectors through semantic tensor transformation, and a semantic alignment relationship is established between the version-aware vectors and the diagnosis event sequence to generate a hybrid representation carrying rule evolution semantics, including:

[0066] Semantically aggregate the mutated nodes associated with the differentiation edges of each version in the evolution path, form rule change features through constraint logic difference analysis, construct a third-order semantic tensor representation of the rule change features in the time dimension, constraint strength dimension and applicable scope dimension, perform tensor decomposition on the third-order semantic tensor representation along the time dimension to capture the dominant mode of rule evolution, and convert the dominant mode into a version-aware vector through low-dimensional manifold mapping.

[0067] The behavioral semantics, temporal relationships, and cost association attributes of each diagnosis and treatment event in the diagnosis and treatment event sequence are structured and encoded to construct a vector representation of the diagnosis and treatment event. The mutual information distribution between the version-aware vector and the vector representation of the diagnosis and treatment event in the semantic space is calculated. Based on the mutual information distribution, sensitive diagnosis and treatment events affected by rule evolution are screened, and a semantic alignment relationship is established between the sensitive diagnosis and treatment events and the version-aware vector.

[0068] The version-aware vector is dynamically modulated according to the semantic alignment relationship, so that the rule evolution pattern in the version-aware vector and the semantic features of the sensitive diagnosis and treatment event have a resonance enhancement effect. The modulated version-aware vector is fused with the overall semantic representation of the diagnosis and treatment event sequence by tensor product to generate a hybrid representation that simultaneously encodes the diagnosis and treatment behavior logic and the rule evolution trajectory.

[0069] After obtaining the variant nodes associated with the differentiation edges of each version in the evolutionary path, semantic aggregation processing is performed on these variant nodes. A variant node refers to a knowledge graph node where the rule content has undergone substantial changes between adjacent versions, including nodes adjusting reimbursement ratios, expanding or narrowing the scope of applicable diseases, and changing cost limits. For each pair of variant nodes between adjacent versions, rule change features are extracted through constraint logic difference analysis: the constraints of the variant nodes in the old and new versions are compared item by item to identify the addition or deletion of constraints, threshold shifts, and changes in the boundaries of the applicable scope. This difference information is encoded into a structured rule change feature vector, which serves as the basic input for subsequent tensor construction.

[0070] Tensor-based representation of rule change features involves expanding these features across three independent dimensions. The time dimension corresponds to the evolution sequence of rule versions, recording snapshots of the rule's state at each time point. The constraint strength dimension characterizes the strictness of the rule conditions; for example, when the reimbursement condition changes from "insured for at least one year" to "insured for at least two years," the constraint strength increases accordingly. The scope of application dimension describes changes in the types of medical services covered by the rule, the insured population, and the geographical area covered. These three dimensions are organized into a third-order semantic tensor. ,in , Version time steps, Let the constraint strength feature dimension be , The dimension represents the scope of application. Each element of the third-order tensor comprehensively reflects the rule change intensity information under a specific combination of a specific time point, a specific constraint strength level, and a specific scope of application.

[0071] For third-order semantic tensors Tensor decomposition along the time dimension aims to extract dominant patterns from the complex changes in the evolution of multiple rule versions. Specifically, the Tucker decomposition method is employed. It is decomposed into the product of the kernel tensor and the factor matrices of each dimension, where the factor matrix of the time dimension is... ( The number of dominant modes, This captures the low-rank structure of rule evolution over time, i.e., the main trend direction of rule change. By preserving the previous... The process first identifies the dominant components, filtering out noisy fine-tuning during rule version iterations and focusing on rule evolution patterns that substantially impact the review results. Then, a low-dimensional manifold mapping is used to convert the extracted dominant patterns into version-aware vectors. ( (This refers to the dimension of the version-aware vector). Low-dimensional manifold mapping is implemented using a nonlinear projection network to compress high-dimensional dominant pattern features into a compact semantic space. This allows the version-aware vector to express the direction and magnitude information of rule evolution in a continuous semantic space, providing a unified representation basis for rule-side semantic alignment with subsequent diagnostic event sequences.

[0072] In the encoding phase of the medical event sequence, each medical event is structured and encoded across three attribute dimensions. Behavioral semantic attributes encompass descriptive information such as the type of medical procedure, medication category, and examination items, which are mapped into behavioral semantic vectors using a pre-trained medical semantic encoder. Temporal relation attributes record the relative position of the medical event within the treatment process, the time interval between it and preceding and following events, and the causal dependencies between events, encoded as temporal relation vectors. Cost association attributes include the single-session cost amount, cumulative cost percentage, and cost category label, encoded as cost association vectors. The three attribute vectors are concatenated and subjected to a linear transformation to obtain a comprehensive vector representation of each medical event. ( For sequence indexing of diagnostic and treatment events, (The dimension of the diagnostic event vector).

[0073] Calculate version-aware vectors Representation of each diagnosis and treatment event vector The mutual information distribution in the semantic space is used to measure the statistical dependence strength between the rule evolution signal and each diagnostic event. Mutual information estimation employs a neural network-based mutual information maximization method, which trains a discriminator to distinguish sample pairs from the product of the joint distribution and the marginal distribution, thereby obtaining... and Mutual information estimate between All diagnostic and treatment events are sorted according to their mutual information estimates, and those with mutual information values ​​exceeding a preset sensitivity threshold are selected. Medical events are marked as sensitive medical events. These events will face different review and judgment logic before and after the rule version change, and are the medical behavior nodes most concentratedly affected by the rule evolution. The selected subset of sensitive medical events is then compared with the version awareness vector. Establish explicit semantic alignment relationships, specifically through computation. With each sensitive diagnostic and treatment event vector ( Attention weights between (indexes of sensitive medical events) This enables a directional association between the semantics of rule evolution and the semantics of diagnosis and treatment events.

[0074] Based on the semantic alignment relationship described above, the version-aware vector Dynamic modulation is performed. The modulation process uses attention weights for each sensitive diagnostic and treatment event. As a gating signal, the semantic features of sensitive medical events are back-injected into the feature space of the version-aware vector, so that... The rule evolution pattern in the version awareness vector resonates with the semantic features of sensitive medical events, creating a resonant enhancement effect. The physical meaning of this resonant enhancement is that when a certain type of medical behavior is repeatedly involved in the rule evolution process, the corresponding rule evolution pattern component in the version awareness vector will be amplified, thereby imposing a stronger rule constraint signal on that type of medical behavior in subsequent review and reasoning. The modulated version awareness vector is denoted as... Its characteristic distribution is compared to the original It is closer to the semantic context of the diagnosis and treatment of the medical records currently awaiting review.

[0075] Modulated version-aware vector The overall semantic representation of the diagnosis and treatment event sequence ( The overall semantic representation dimension (obtained by sequence-level aggregation of all diagnostic and treatment event vectors) is then fused using tensor products to generate a hybrid representation. The calculation method for tensor product fusion is as follows: The resulting tensor Each element in the representation simultaneously encodes the interaction strength between a rule-evolutionary component of the version-aware vector and a semantic component of the overall diagnostic semantic representation. By... Flattened and dimensionality-reduced, the final hybrid representation vector is obtained. This vector carries two types of information in the same semantic space: the logic of diagnosis and treatment behavior and the trajectory of rule evolution. It provides a high-quality input representation for the subsequent adversarial dual-path review architecture, ensuring that the review reasoning process can perceive the substantial impact of rule version differences on the judgment of medical record compliance.

[0076] In one optional implementation, dynamically modulating the version-aware vector according to the semantic alignment relationship, so that the rule evolution pattern in the version-aware vector resonates and enhances the semantic features of the sensitive medical event, includes:

[0077] Based on semantic alignment, a coupled tensor field is constructed between the version-aware vector and the sensitive diagnosis and treatment event. In the coupled tensor field, the semantic response gradient of each dimension component of the version-aware vector to the sensitive diagnosis and treatment event is calculated. Based on the semantic response gradient, the important dimensions that need to be enhanced and the redundant dimensions that need to be suppressed are determined, and a dimension-selective modulation mask is generated.

[0078] The dimension-selective modulation mask is applied to the version-aware vector. The representation intensity of the important dimensions is amplified and the representation intensity of the redundant dimensions is attenuated through a gating mechanism. This causes the modulated version-aware vector to move in the semantic space toward the distribution area of ​​the sensitive diagnosis and treatment event. The local curvature matching degree between the moved version-aware vector and the semantic features of the sensitive diagnosis and treatment event on the feature manifold is calculated.

[0079] When the local curvature matching degree exceeds the coordination threshold, a resonance loop is established between the version perception vector and the sensitive diagnosis and treatment event in the coupled tensor field. Through the resonance loop, the rule evolution mode in the version perception vector and the semantic features of the sensitive diagnosis and treatment event generate synchronous oscillations in the corresponding dimensions of the feature space. The synchronous oscillations cause the semantic expressions of the two to reinforce each other in the corresponding dimensions, forming a resonance enhancement effect.

[0080] Obtaining version-aware vectors After establishing the set of sensitive medical events, it is necessary to further establish a deep coupling relationship between them, so that the semantic evolution of rules can accurately penetrate into the feature dimensions of the related medical events. To this end, based on the established semantic alignment relationship, a coupled tensor field is constructed on the Cartesian product of the version-aware vector space and the feature space of sensitive medical events. Specifically, for dimensions... The version-aware vector and dimension are The A vector of sensitive medical events, coupled to each field element of a tensor field. Represents the version-aware vector. Dimensional composition and sensitive diagnostic and treatment events The semantic coupling strength between dimensional components, where , The coupled tensor field is derived from the attention weights in the semantic alignment relation via a bilinear mapping. The weighted summation results in a greater contribution of sensitive medical events with higher attention weights to the coupled tensor field, thus ensuring that the construction of the field fully reflects the sensitivity of key medical nodes in the medical record to the evolution of rules.

[0081] In the coupled tensor field After establishment, the semantic response gradients of each dimension component of the version-aware vector to sensitive medical events are calculated. For the version-aware vector... dimensional components, their semantic response gradients Defined as the comprehensive partial derivative response of this dimensional component with respect to the semantic features of all sensitive diagnostic and treatment events in the coupled tensor field, i.e., along the ... The sensitivity metric obtained by integrating the coupling strength along the dimensional direction. Semantic response gradient. Reflects the version-aware vector of the first The influence of the rule evolution information carried by each dimension on the semantic space of sensitive medical events: Dimensions with larger absolute gradient values ​​indicate that the rule evolution pattern of that dimension is highly correlated with sensitive medical events and should be enhanced; dimensions with smaller absolute gradient values ​​indicate that the information carried by that dimension contributes little to sensitive medical events and are redundant, and should be suppressed. Based on the semantic response gradients of all dimensions... Set gradient quantile threshold The absolute value of the gradient exceeds The dimensions are marked as the set of important dimensions. The absolute value of the gradient is lower than The dimensions are marked as redundant dimension sets. Based on this, a dimension-selective modulation mask vector is generated. Important dimensions correspond to mask values ​​greater than 1, while redundant dimensions correspond to mask values ​​less than 1 but greater than 0, ensuring that the mask has a non-zero response for all dimensions to avoid hard truncation of information.

[0082] Dimension-selective modulation mask Create version-aware vectors using a gating machine. The version-aware vector after gating and modulation is obtained. The calculation method is to use and After element-wise multiplication, normalization is performed. ,in This represents element-wise multiplication, and the normalization operation ensures that the overall semantic scale of the vector remains stable before and after modulation. The core function of the gating mechanism is that the mask value of important dimensions is greater than 1, which amplifies the rule evolution features in these dimensions, resulting in stronger discriminative ability in subsequent semantic reasoning; the mask value of redundant dimensions is less than 1, which effectively suppresses noise information in these dimensions, reducing its interference with the final review conclusion. After gating modulation, the distribution center of the version-aware vector in the semantic space shifts, moving closer to the semantic distribution region of sensitive medical events, thus laying the foundation for subsequent curvature matching calculations.

[0083] In semantic space, both version-aware vectors and the feature distributions of sensitive medical events can be viewed as low-dimensional manifold structures embedded in a high-dimensional Euclidean space. Modulated version-aware vectors With the A vector of sensitive medical events Local curvature matching degree on characteristic manifolds It is measured by calculating approximate values ​​of the Riemann curvature tensor in the neighborhood of two points. Specifically, a local principal curvature estimation method is used: [the method is described in the original text]. and of Within a neighborhood, several nearest neighbor points are sampled, and the local tangent spaces of each neighborhood are estimated using principal component analysis. The alignment degree between these tangent spaces is then calculated as the local curvature matching degree. Approximate to . The range of values ​​is normalized to The closer the value is to 1, the more closely the modulated version perception vector matches the local geometric structure of the sensitive diagnosis and treatment event on the feature manifold, which means that the intrinsic structure between the rule evolution semantics and the semantic features of the diagnosis and treatment event is highly consistent.

[0084] When the local curvature matching degree Exceeding the preset collaboration threshold At that time, determine the version-aware vector and the first The resonance establishment conditions are met between several sensitive diagnostic and treatment events in the coupled tensor field. A resonant circuit is established for this pairing. The process of establishing the resonant circuit is as follows: In the dimension that satisfies the condition... (Right now And the corresponding Exceeding the field strength threshold On the above, a bidirectional feedback connection is introduced, enabling the version-aware vector to... Dimensional feature activation can positively enhance the sensitivity of diagnostic and therapeutic events through coupling fields. Semantic response of dimensions, and sensitive diagnosis and treatment events. Semantic activation of the dimensional vector can also inversely enhance the version-aware vector. The representation strength of the dimension. This bidirectional positive feedback structure produces a synchronous oscillation effect in the corresponding dimension of the feature space: after multiple rounds of iterative interaction, the activation values ​​of the two feature vectors in the resonance dimension gradually tend to be co-amplified, and finally form a stable high activation state in the corresponding dimension, that is, the resonance enhancement effect.

[0085] The convergence of the resonance enhancement effect is determined by monitoring the change in the activation value of the resonance dimension between two adjacent iterations. To achieve, when Below the convergence threshold When resonance is considered stable, the iteration stops and the final resonance-enhanced version-aware vector is output. With the updated sensitive medical events Version-aware vector enhanced by resonance Compared to before modulation Its rule evolution pattern was significantly strengthened in dimensions highly correlated with sensitive diagnostic and treatment events, enabling the hybrid representations generated by subsequent tensor product fusion. It can more accurately capture the deep correspondence between rule version changes and specific diagnosis and treatment behaviors, providing more distinctive input features for the adversarial dual-path review architecture, and ultimately improving the accuracy and interpretability of review conclusions.

[0086] In one optional implementation, the hybrid representation is fed into an adversarial dual-path review architecture, wherein the first path generates a deterministic review result based on rule-based logical reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning, including:

[0087] The hybrid representation is decomposed into an explicit rule constraint subspace and an implicit semantic association subspace in the feature space. The projection vector of the explicit rule constraint subspace is fed into the first path. A constraint propagation network is constructed in the first path. The rule constraint conditions in the projection vector are forward reasoned and backward verified through the constraint propagation network. When the truth value determination of all constraint conditions converges, a binary deterministic audit result is generated.

[0088] The complete representation of the hybrid representation is fed into the second path, where the diagnosis and treatment behavior patterns and rule evolution patterns in the implicit semantic association subspace are jointly probabilistically modeled by an AI model. Based on the AI ​​model, the posterior probability of the hybrid representation under the anomaly detection distribution is calculated, and a probabilistic audit result with a continuous value range is generated based on the posterior probability.

[0089] After the hybrid representation is incorporated into the adversarial dual-path review architecture, it first needs to undergo subspace decomposition in the feature space to distinguish between explicit rule-constrained components and implicit semantic association components. Let the hybrid representation tensor... A unified feature vector is obtained after linear projection. Through orthogonal projection matrix and Extract the projection vectors of the explicit rule-constrained subspaces respectively Projection vectors of subspaces associated with implicit semantics ,in To constrain the dimension of the subspace by rules, Let be the dimension of the semantic association subspace. During the training phase, the two projection matrices are jointly optimized through constrained orthogonality loss to ensure that the two subspaces are decoupled at the feature level. This allows the first path to focus on logically interpretable rule-constrained reasoning, while the second path focuses on data-driven deep semantic mining.

[0090] In the first path, the projection vector The constraints are parsed into a set of structured rules and constraints. The constraint propagation network organizes the logical dependencies between the rules and constraints in the form of a directed graph. Each node in the graph corresponds to an atomic rule constraint (such as diagnostic coding compliance, drug indication matching, duplicate detection of billing items, etc.), and edges represent the premise implication or mutual exclusion relationship between constraints. During the forward inference phase, constraint nodes are activated layer by layer along the topological order of the directed graph, and the current truth value of each constraint node is calculated. ,in This indicates that the constraint condition is met. This indicates a violation. Once the truth values ​​of all predecessor nodes of a given node have been determined, the truth value of that node is calculated according to the rule logic (conjunction, disjunction, or negation), and the result is propagated to subsequent nodes. The reverse verification phase starts from the final review conclusion node and traces back each triggering path to check for logical contradictions. If the truth value of a constraint node is inconsistent with the logical combination result of its predecessor nodes, a backtracking correction mechanism is triggered to re-evaluate the activation state of the relevant constraint nodes.

[0091] Convergence is determined by monitoring the total change in the truth value of all nodes in the constrained network between two adjacent iterations. To achieve, when Below the preset convergence threshold At this point, it is determined that constraint propagation has reached a stable state. The final truth values ​​of all constraint nodes are then aggregated. If no violation nodes are found, a binary deterministic audit result is output. (Indicates compliance), otherwise output (Indicating a violation), along with the set of constraint nodes that triggered the violation as interpretable evidence. This binary output ensures the certainty and traceability of the first-path review conclusion, directly corresponding to specific rule clauses and meeting the mandatory requirements for compliance interpretation in medical insurance review scenarios.

[0092] The second path receives the complete representation of the hybrid representation. (Include and (Complete information from both parts) is used to perform joint probabilistic modeling of diagnostic behavior patterns and rule evolution patterns through an AI model. The AI ​​model employs a variational inference framework, integrating hybrid representations. Encoding as latent variables The posterior distribution parameters, i.e., the mean vector. The logarithmic variance vector of the diagonal covariance matrix ,in Let this be the dimension of the latent variable space. Latent variables. At the same time, it captures the deep semantic patterns of diagnosis and treatment behavior (such as the rationality of diagnosis and treatment pathways and the coherence of medication logic) as well as the semantic shifts introduced by rule evolution (such as the compliance differences of the same diagnosis and treatment behavior under the old and new rule versions).

[0093] In terms of anomaly detection distribution modeling, a prior distribution of normal medical behavior is constructed using a set of compliant medical record samples. By optimizing the variational lower bound during the training phase, compliant samples in the latent variable space form a compact cluster structure. For samples to be reviewed, their latent variable posterior distributions are calculated. Compared with normal prior distribution KL divergence between And combined with reconstruction error Together, they constitute the anomaly detection score. Posterior probability. By normalizing the anomaly detection score to The interval is obtained. A value closer to 1 indicates that the medical record deviates more from the compliance distribution at the semantic level, and is more likely to contain hidden violations. (Probabilistic audit result) As a continuous value range output, it can provide fine-grained characterization of violation risks, and is especially suitable for audit scenarios where rule boundaries are vague or the semantics of new version rules have not yet fully converged.

[0094] The advantage of joint probabilistic modeling lies in the fact that the implicit semantic association subspace encodes the temporal dependencies between medical events, the semantic associations between diagnosis and medication, and the version-sensitive features introduced by rule evolution. This information cannot be directly expressed through explicit rule constraint networks, but it is of great value in identifying circumventing violations (such as circumventing single-treatment cost limits by splitting medical procedures). By training on a large amount of historical medical record data in an unsupervised or weakly supervised manner, the AI ​​model can automatically learn the semantic boundaries of normal medical treatment patterns and assign a higher anomalous posterior probability to samples that exceed these boundaries during the inference phase.

[0095] To ensure the mutual constraints between the first and second paths in the feature space, a consistency alignment loss is introduced during the training phase. The requirement is that the first path is deemed compliant. When ), the posterior probability output by the second path. It should be below the compliance confidence threshold. When the first path is determined to be in violation ( )hour, It should be higher than the violation detection threshold. Through this consistency constraint, the two paths form an implicit adversarial game relationship at the feature representation level: the deterministic reasoning of the first path provides anchoring supervision signals for the probability distribution of the second path, while the semantic perception capability of the second path supplements the constraint coverage of the first path. This dual-path mutual constraint mechanism not only improves the review accuracy of the single path, but also lays a structural foundation for subsequent divergence handling and meta-learning signal extraction.

[0096] In one optional implementation, in the second path, a joint probability model is performed on the diagnostic behavior patterns and rule evolution patterns in the implicit semantic association subspace using an AI model. The posterior probability of the hybrid representation under the anomaly detection distribution is calculated based on the AI ​​model, including:

[0097] In the implicit semantic association subspace, the behavioral sequence dependency structure of the diagnosis and treatment behavior pattern and the version migration trajectory of the rule evolution pattern are extracted. The behavioral sequence dependency structure and the version migration trajectory are geometrically aligned on the feature manifold through topological isomorphism mapping to generate a joint representation structure that encodes the dual-mode coupling relationship.

[0098] The AI ​​model performs conditional dependency learning on the treatment behavior patterns and rule evolution patterns in the joint representation structure, and establishes a conditional probability distribution of the treatment behavior patterns with the rule evolution patterns as conditional variables. The parameters of the conditional probability distribution are adaptively inferred by the AI ​​model from the joint representation structure, forming a joint probability modeling result that characterizes the impact of rule versions on treatment compliance.

[0099] The hybrid representation is input into the AI ​​model. Based on the joint probability modeling results, the probability density ratio of the hybrid representation under the normal diagnosis and treatment behavior distribution and the abnormal diagnosis and treatment behavior distribution is calculated. The anomaly confidence of the hybrid representation is calculated by combining the probability density ratio with the historical statistical prior of the rule evolution mode. The anomaly confidence is the posterior probability of the hybrid representation under the anomaly detection distribution.

[0100] like Figure 2 As shown, the method includes:

[0101] In the implicit semantic association subspace, diagnostic and treatment behavior patterns contain event-dependent structures spanning time steps. For example, a certain type of diagnostic event triggers a specific medication sequence, or a combination of operations repeatedly occurs within a specific time window. These behavioral sequence dependency structures are not simply the superposition of independent events, but rather reflect the inherent temporal causal relationships within the clinical diagnostic and treatment process. Simultaneously, rule evolution patterns record the trajectory of medical insurance rules migrating from old to new versions, including version migration information such as a certain type of medical item changing from reimbursement-allowed to reimbursement-restricted, or the reduction or expansion of the indications for a certain type of drug. These two types of patterns are modeled as geometric structures on a high-dimensional feature manifold. The behavioral sequence dependency structure is represented as the manifold curve of diagnostic and treatment events in the feature space, and the version migration trajectory is represented as the path of rule version nodes in the rule semantic space.

[0102] The core objective of topological isomorphism mapping is to find a structure-preserving mapping from a manifold with behavioral sequence dependencies to a manifold with version migration trajectories, ensuring that the topological relationships (such as connectivity, neighborhood structure, and path length) on both manifolds remain consistent before and after the mapping. Specifically, persistent cohomology features are extracted from behavioral sequence dependencies to capture the topological invariants of event dependencies at different time scales; similarly, persistent cohomology features are extracted from version migration trajectories to characterize the topological change patterns during rule evolution. Let the topological feature vector of the behavioral sequence dependency structure be... Dimension is The topological feature vector of the version migration trajectory is Dimension is By learning a bidirectional mapping function and its inverse mapping This results in a topological distance error before and after mapping. Minimize. After geometric alignment is complete, and By concatenating the structures, a joint representation structure is obtained that encodes the coupling relationship between the two modes. Dimension is This joint representation structure carries both temporal dependency information of the diagnosis and treatment behavior and evolutionary geometric information of the rule version.

[0103] Based on joint representation structure The AI ​​model learns the conditional dependencies between diagnostic behavior patterns and rule evolution patterns. The rule evolution pattern is treated as a conditional variable, and its topological characteristics are used to... As a conditional input, a conditional probability distribution of the diagnostic and treatment behavior pattern is established under this condition. The AI ​​model employs a conditional normalized flow structure, which incorporates condition variables... The scaling and offset parameters of each affine coupling transformation are injected into the parameters of the stream transformation, allowing the parameters to adaptively adjust according to the regular evolution pattern. Let the... The scaling function of the layer affine coupling transformation is The offset function is Both are based on For additional input, then the first The transformation output of the layer is ,in Represents element-wise product. For the first The layer outputs a segmentation vector. By maximizing the conditional log-likelihood of the training samples, the AI ​​model adaptively infers the parameters of the conditional probability distribution from the joint representation structure, forming a joint probability modeling result that characterizes the impact of rule versions on medical compliance. This modeling result can accurately reflect which types of medical behavior sequences belong to high-probability compliance patterns and which belong to low-probability abnormal patterns under specific rule version conditions.

[0104] The hybrid representation is linearly projected to obtain a unified feature vector. Its projection in the implicit semantic association subspace The data is input into an AI model to calculate the ratio of its probability density under normal and abnormal patient behavior distributions. (Normal patient behavior distribution) The distribution of abnormal medical behavior was obtained by fitting the results of joint probability modeling from historical compliant medical record samples. Then, the sample of historical violation medical records is fitted in the same way. The probability density ratio is defined as... ,when This indicates that the density of the current mixed representation is higher under the abnormal distribution, meaning that the diagnosis and treatment behavior pattern of this medical record is closer to the characteristic distribution of historical violation samples.

[0105] Historical statistical priors of rule evolution patterns This data comes from statistical analysis of rule change events across multiple versions of the knowledge graph, specifically including the historical frequency of violations triggered by a certain type of rule change and the growth rate of violation cases after a particular rule version took effect. Normalization The ratio of prior weights within the interval to probability density Combined, calculate the anomaly confidence of the mixed representation. The combination method uses a Bayesian update formula, with... As the prior probability of an anomalous event, As the likelihood ratio, we obtain the posterior probability:

[0106] ;

[0107] This formula shows that when the rule evolution history is prior... High (i.e., this type of rule change has historically frequently triggered violations) and probability density ratio When the posterior probability is larger (i.e., the current treatment behavior is closer to the abnormal distribution), This will significantly increase, thereby enhancing the sensitivity of the review process. Conversely, when lower and When it approaches 1, Approaching the prior level, avoiding misjudgment of low-risk medical records. This refers to the posterior probability of the hybrid representation under the anomaly detection distribution, serving as the probabilistic review result of the second path. The output is used for subsequent dual-path fusion.

[0108] In practice, the results of joint probability modeling need to be dynamically adjusted as the medical insurance rules are updated. Each time a new version of the rules takes effect, the corresponding version migration trajectory features are... By injecting conditional normalization streams into the conditional input, the AI ​​model does not need to retrain all parameters; it only needs to scale the conditional transformation layer using a small number of new compliant and non-compliant samples. With offset function Fine-tuning is all that's needed to adapt to the semantic space of the new rule version. This mechanism is integrated with the design of using learning signals in the sovereign term to guide the AI ​​model in rapidly reconstructing the semantic space of the new rule version, ensuring that the second path maintains its consistently effective anomaly detection capability in the frequently iterating medical insurance environment. Simultaneously, historical statistical priors are utilized. It is also updated on a rolling basis according to the actual execution data of the new version of the rules to ensure the timeliness and accuracy of the posterior probability calculation.

[0109] In one optional implementation, when the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under conditions of few samples. The final review conclusion is output based on the dual-path fusion result after adversarial convergence, including:

[0110] Disagreement samples are identified by calculating the decision vector distance between deterministic review results and probabilistic review results. The rule evolution path features associated with the disagreement samples are extracted, and the rule evolution path features are transformed into meta-learning signals containing semantic difference information between versions.

[0111] The meta-learning signal is fed into the meta-parameter update layer of the AI ​​model. The meta-learning signal guides the AI ​​model to construct a fast adaptation direction pointing to the semantics of the new version of the rules in the gradient space. The fast adaptation direction is used to make incremental adjustments to the semantic encoding parameters of the AI ​​model with few samples, so that the AI ​​model can complete the reconstruction of the semantic space of the new version of the rules in the local semantic space of the divergent samples.

[0112] In the adversarial dual-path review architecture, continuous iterative optimization is performed. When the consistency measure between the deterministic review result and the probabilistic review result on the divergent sample satisfies the adversarial convergence condition, the deterministic review result and the probabilistic review result are weighted by credibility to generate a dual-path fusion result, which is then used as the final review conclusion.

[0113] When discrepancies arise between the deterministic and probabilistic review results in an adversarial dual-path review architecture, it is necessary to first accurately locate the divergent samples. Specifically, for each medical record to be reviewed, the deterministic review result output by the first path is encoded into a decision vector. The probabilistic review results output by the second path are encoded into a judgment vector. Calculate the distance between the decision vectors of the two entities in the decision space. .when Exceeding the preset divergence identification threshold When a divergence occurs, the sample is marked as a divergence sample and included in the subsequent meta-learning process. The calculation of the decision vector distance comprehensively considers semantic direction differences and amplitude differences, which can effectively distinguish between slight deviations and substantial decision conflicts, avoiding misjudging minor fluctuations caused by numerical perturbations as genuine divergences. After divergence identification, a set of all divergence samples is collected as the core data source driving meta-learning.

[0114] Extracting the associated rule evolution path features from divergent samples is a crucial step in constructing the meta-learning signal. Each divergent sample carries index information pointing to a specific evolution path in the multi-version knowledge graph. Along this path, structural changes such as differences in rule nodes, variations in constraints, and expansions or contractions in the scope of application between adjacent versions are extracted. These changes are then integrated into a version-specific semantic difference description. This version-specific semantic difference not only records the differences in the literal expression of old and new rules but also captures the deeper evolution of rules in terms of applicable scenarios, constraint strength, and logical relationships. This version-specific semantic difference information is then encoded into a meta-learning signal vector. Its dimension is Each dimension component corresponds to the intensity or direction of a certain type of semantic change in the rule evolution path. During the construction process, each change node on the rule evolution path is weighted and aggregated in chronological order, with change nodes closer to the current version being given higher aggregation weights to ensure the sensitivity of the meta-learning signal to the latest rule semantics.

[0115] Meta-learning signal When feeding data to the metaparameter update layer of the AI ​​model, a fast adaptation direction pointing to the semantics of the new version of the rules needs to be constructed in the gradient space. The metaparameter update layer maintains a set of metaparameters. These meta-parameters encode the general adaptability for cross-version rule migration. (Received) Subsequently, the meta-parameter update layer calculates the gradient response between the meta-learned signal and the current model's semantic encoding parameters to obtain the fast-adaptive direction vector. . This indicates the direction along which, in the gradient space, a small adjustment to the semantic encoding parameters most effectively transfers the model's semantic understanding capabilities from the old version of the rule-based semantic space to the new version. This process borrows the core idea of ​​model-independent meta-learning, which uses gradient information from a small number of samples to quickly locate the parameter update direction without requiring large-scale retraining of the entire model. This allows for rapid direction adaptation. The computation uses meta-learning signals to impose directional constraints on the parameter matrices of the semantic coding layer, so that parameter updates are concentrated along directions that are conducive to the semantic reconstruction of the new version of the rules, rather than being diffused throughout the entire parameter space.

[0116] Utilize rapid adaptation direction When performing few-sample incremental adjustments to the semantic encoding parameters of an AI model, the parameters are updated using only a small amount of labeled data (typically on the order of single digits) from the divergent sample set. The step size for the incremental adjustment is determined by an adaptive step size controller. The controller dynamically adjusts the update step size based on the current degree of divergence and the confidence level of the meta-learning signal: when the divergence is high and the confidence level of the meta-learning signal is strong, a larger parameter update step size is allowed to accelerate the semantic space reconstruction speed; when the divergence is low or the confidence level of the meta-learning signal is insufficient, the step size is reduced to prevent overfitting. The range of parameter increment adjustment is strictly limited to a subset of coding layer parameters related to the local semantic space of the divergent samples, through parameter importance masking. By identifying the parameter dimensions most sensitive to the semantics of the new version of the rules, and applying incremental adjustments only to these dimensions, the model's auditing capabilities on the stable version of the rules are protected while completing a partial reconstruction of the semantic space of the new version of the rules.

[0117] In the adversarial dual-path review architecture, the aforementioned meta-learning-driven parameter adjustment and dual-path adversarial optimization processes alternate, forming a continuously iterative optimization loop. After each round of meta-learning signal-driven parameter increment adjustment, the probabilistic review result of the updated AI model on the divergent samples is recalculated, and a consistency measure is calculated between this result and the deterministic review result of the first path. (Consistency measure...) Defined as the weighted consistency rate of the decision conclusions of two paths on a set of divergent samples, when Exceeding the adversarial convergence threshold When the condition for adversarial convergence is met, the iteration stops and the process enters the fusion output stage. If If the convergence threshold is not reached, a new round of divergence samples is extracted, the meta-learning signal is updated, and parameter incremental adjustment is performed again until the convergence condition is met or the maximum number of iterations is reached. .

[0118] After adversarial convergence, the deterministic and probabilistic review results are weighted and fused based on their credibility to generate the final review conclusion. The credibility weights of the two paths are not fixed values ​​but are dynamically estimated based on the historical performance of each path on the current sample. (The credibility weights of the deterministic review path are described below.) The calculation of the integrity of the rule logic reasoning chain and the rule version matching degree is based on the following: when the timestamp of the medical record to be reviewed is highly consistent with the traced target rule version, and the reasoning chain is unbroken, Take the higher value; when there is ambiguity in the rule version or missing nodes in the inference chain. The credibility weight of the probabilistic review path will be reduced accordingly. Stability estimation of the confidence interval width of the posterior distribution and the posterior probability of anomaly detection based on the variational inference framework: When the posterior distribution is concentrated and the variance of the probabilistic review results is small after multiple samplings. Take the higher value. The weights of the two paths satisfy the normalization constraint. fusion results The output is obtained by weighted combination of the two paths. Output a compliance conclusion when the compliance confidence threshold is exceeded; when When the value is below the violation detection threshold, output a violation conclusion; when... When the sample falls between the two thresholds, it is marked as requiring manual review, and accompanied by details of the dual-path divergence and a description of the rule evolution path characteristics to assist reviewers in making manual judgments. The final review conclusion, along with the credibility weight allocation, meta-learning adjustment log, and rule version traceability information, is output to form a complete and interpretable review report.

[0119] The method further includes:

[0120] After acquiring the medical record data to be reviewed, the system connects to the medical institution's server via a file transfer protocol and performs a three-level hierarchical verification on the directory storing the medical record files. The root directory is divided into first-level subdirectories based on hospital codes, and each hospital directory is divided into second-level subdirectories based on inpatient or outpatient type. The type directories store PDF medical record files named after the patient. The system parses the relative path of each file item and calculates the hierarchical depth. When the level is 1, it checks if the entry is a directory type and extracts the hospital code. When the level is 2, it checks if the directory name matches the inpatient or outpatient enumeration value. When the level is 3, the entry must be a file type, the filename must conform to a regular expression pattern, and the file size must be non-zero. If the conditions are met, the patient name, medical record type, and hospital code are extracted from the path to construct a medical record object and added to the valid file list. When the hierarchical depth reaches 4 levels or higher, it is marked as a depth exceedance error. After verification, a structured report is generated containing a list of valid files grouped by hospital and medical record type, and a list of invalid entries categorized by error type.

[0121] In a single database transaction, batch creation of 4-level related records is performed. Review batch records are created to record the total number of tasks and global progress statistics. Review monitoring records are created by hospital code to independently track the review progress of each hospital. Review task records are created for each medical record to identify the current execution stage and status. Medical record entity records are created to store basic medical record information and extended inpatient or outpatient information. The primary key of all records uses the snowflake algorithm to generate a distributed unique identifier. In the batch record, the total task number field is set to the total number of valid files and the parsed number, completed number, triggered number and progress percentage fields are initialized to 0.

[0122] The system is designed based on an event-driven asynchronous task scheduler. The scheduler maintains a queue of tasks to be processed, scheduling event objects, a set of active tasks, and a maximum number of concurrent slots. The system includes a batch statistics dictionary and a historical queue of completed batches. The historical queue uses a sliding window mechanism to retain only the 10 most recent completed batch records. During the task addition process, the task identifier and payload data are encapsulated and added to the waiting queue. At the same time, under the protection of a mutex lock, the total number of tasks and the number of tasks to be processed in the corresponding batch in the batch statistics dictionary are atomically incremented. After the task is added, a scheduling event signal is immediately triggered to wake up the scheduling loop. When there are no tasks, the scheduling loop is permanently blocked by waiting for the scheduling event, thus achieving zero CPU overhead.

[0123] After the event is triggered, the scheduling loop resets the event state and enters a 3-stage processing flow. In the first stage, all completed tasks are removed from the active task set and the number of currently active tasks is counted. The second phase calculates the number of available concurrent slots. ,like Then return directly to the waiting event phase. In the third phase, when the number of available slots is greater than 0, tasks are dequeued sequentially from the waiting queue and an independent worker instance is created for each task. Under the protection of the mutex lock, the number of tasks to be processed in the corresponding batch in the batch statistics dictionary is atomically decremented and the number of active tasks is atomically incremented. Workers are created through the factory pattern and task execution functions are dispatched asynchronously. After the task execution function finishes running, the number of active tasks in the corresponding batch in the batch statistics dictionary is atomically decremented under the protection of the mutex lock. The number of successful or failed tasks is atomically incremented according to the task execution status. If the total number of tasks in the batch is equal to the sum of the number of successful and failed tasks, the batch is considered to be completed and the batch statistics data copy is moved into the completed batch history queue.

[0124] After the PDF file is downloaded, the system simultaneously starts two parallel processing channels. The first channel uploads the PDF file to the AI ​​platform to call the OCR document parsing service to extract structured text. If the service call fails, an automatic retry mechanism is executed, with the waiting time for the k-th retry being [time value missing]. The OCR-parsed medical record content is typically tens of thousands of characters long. The system defines an ordered set of medical keywords, sorted from highest to lowest clinical review value. The keyword sequence for inpatient medical records is as follows: inpatient medical record cover page, admission record, expense list, expense details, inpatient settlement expense summary list, discharge record, discharge summary, surgical record, death record, report, assessment scale, outpatient medical record, outpatient and emergency medical record, long-term medical orders, temporary medical orders, death discussion record, progress notes, initial progress notes, and daily progress notes. The system sets an input length limit. 22,000 bytes, lower limit of content validity The number of characters is 100, and the cropping algorithm determines whether the medical record type is inpatient.

[0125] If the type is inpatient, a paragraph extraction sub-algorithm is executed. This sub-algorithm scans the OCR-parsed Markdown text line by line, performing three pattern matches on each line: Pattern A matches heading lines starting with a hash symbol and containing the keyword; Pattern B matches independent lines whose entire text is exactly equal to the keyword; and Pattern C matches bold lines enclosed in asterisks and containing the keyword. When any pattern match is successful, the matched keyword is recorded, and the next heading line is located. All content between the current line and the next heading line is collected as one paragraph. All matched paragraphs are concatenated and returned as the extraction result. The extracted or retained text is processed by removing image references and then its UTF-8 encoded byte length is calculated. If the byte length exceeds... Then, remove the lowest priority keyword from the end of the ordered keyword set, re-execute the paragraph extraction sub-algorithm using the reduced keyword subset, and after the iteration, perform a validity check on the cropped text. If the character length is less than... The system will return an error indicating insufficient content.

[0126] The cropped medical record text, along with the task identifier, primary diagnosis name, and medical record type, is input into a versioning rule level 2 matching algorithm. The algorithm first obtains the project identifier based on the task identifier, queries the set of valid rule versions associated with the project, and filters the rules by medical record type to obtain an applicable rule version list V. Level 2 matching is divided into two sub-stages, A and B. Sub-stage A iterates through each rule version v in version list V, performing fuzzy matching on the rule entries of version v to find rule entries whose names contain the primary diagnosis name. If the query result is not empty, the set of matched rule entries is merged into the diagnosis-specific rule set. Add version v to the set of matched versions After sub-phase A is completed, the remaining version set is calculated. ,like If the value is not empty, proceed to sub-phase B, and then iterate through sub-phase B. For each rule version v, perform an exact match on the rule entries of version v that match the query name with the general rule entry that is equal to the comprehensive code, and merge the query results into the general rule set. ,merge and Obtain the final set of matching rules traversal For each rule r, a structured Prompt is constructed. An association record between the medical record entity and the rule entry is created in the database. In the association record, the rule version identifier field is set to the version identifier of rule r to realize version snapshot, and the trigger status field is set to the enumeration value to be judged.

[0127] The cropped medical record content, along with the structured Prompt, is input into a large language model for inference. The unstructured text returned by the model undergoes a three-layer structured parsing algorithm. The first layer extracts the thought chain content, locating the interval from the beginning to the end of the "think" tag in the original output text and extracting the content of that interval. The second layer performs JSON structured data extraction, locating the first left curly brace in the original output text through bidirectional lookup. With the position of the last right curly brace If any position is invalid or It will then return a JSON extraction failure error, and extract the data from the original output text. The substring is treated as a JSON string and parsed to obtain a data object named "data". The third layer performs Chinese semantic field mapping, extracting values ​​from the data object where the key is the model conclusion and assigning them to the conclusion variable, the key is the medical decision basis and assigning it to the basis variable, and the key is the confidence level and assigning it to the confidence level variable. Differential field extraction is performed based on the medical record type. If the medical record type is inpatient, values ​​with keys of department, discharge date, primary diagnosis, and related medical insurance costs are extracted to construct a patient information object. If the medical record type is outpatient, values ​​with keys of patient name, patient gender, patient number, and patient age are extracted to construct a patient information object, as are values ​​with keys of visit date, department, chief complaint, and cost details. The fourth layer performs conclusion classification judgment, checking whether the conclusion variable contains a "non-triggered" keyword. If it contains a "non-triggered" keyword, the trigger status is determined to be "non-triggered medical insurance rule" and the trigger increment value is set to 0. If it does not contain a "non-triggered" keyword, the trigger status is determined to be "triggered medical insurance rule" and the trigger increment value is set to 1.

[0128] After the structured audit results are generated, the system executes a Level 3 atomic progress tracking algorithm to update the batch, monitoring, and task-level statistical data. The algorithm performs the following operations within a single database transaction to ensure atomicity: Level 3 task-level update updates the execution stage field of the task record to the new execution stage, the status field to the new status, and merges the historical and new results in the result chain field using an incremental merging function. Level 2 monitoring-level update searches for the corresponding monitoring record based on the batch identifier and hospital code of the task record. If the parsed incremental value in the incremental indicator is not null, an atomic increment operation is performed on the parsed numerical segment of the monitoring record. If the completed incremental value in the incremental indicator is not null, an atomic increment operation is performed on the completed numerical segment of the monitoring record. The progress percentage field of the monitoring record is updated using a formula... Automatic calculations show that the Level 1 batch update performs atomic increment operations on the parsed, completed, and triggered digital segments of the batch record. The progress percentage field of the batch record is automatically calculated using the same formula. After the transaction is committed, all three levels of statistical data updates take effect. Each stage of the review process is encapsulated as an independent tool interface and exposed to the AI ​​agent through the model context protocol. The AI ​​agent calls the tools of each stage in a preset order to form a complete review pipeline.

[0129] A second aspect of this invention provides a medical record intelligent review system based on versioned medical insurance rules and an AI model, comprising:

[0130] The version tracing unit is used to obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information.

[0131] A hybrid representation unit is used to map the rule change features in the evolution path into version-aware vectors through semantic tensor transformation, establish a semantic alignment relationship between the version-aware vectors and the diagnosis and treatment event sequence, and generate a hybrid representation carrying the semantics of rule evolution.

[0132] The dual-path review unit is used to feed the hybrid representation to the adversarial dual-path review architecture. The first path generates a deterministic review result based on rule logic reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning. This allows the two paths to constrain each other in the feature space. When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under limited sample conditions. The final review conclusion is output based on the dual-path fusion result after adversarial convergence.

[0133] A third aspect of the present invention provides an electronic device, comprising:

[0134] processor;

[0135] Memory used to store processor-executable instructions;

[0136] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0137] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0138] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent medical record review based on versioned medical insurance rules and AI models, characterized in that: include: Obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information; The rule change features in the evolution path are mapped to version-aware vectors through semantic tensor transformation, and a semantic alignment relationship is established between the version-aware vectors and the diagnosis and treatment event sequence to generate a hybrid representation carrying the semantics of rule evolution. The hybrid representation is fed into an adversarial dual-path review architecture, where the first path generates a deterministic review result based on rule-based logical reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning. This allows the two paths to constrain each other in the feature space. When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under limited sample conditions. The final review conclusion is output based on the dual-path fusion result after adversarial convergence.

2. The method according to claim 1, characterized in that, Obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information, including: The historical medical insurance rule texts are parsed in a time sequence to extract the effective time interval, applicable scope boundary and constraint logic expression of each rule version. The constraint logic expression is decomposed into diagnosis and treatment behavior nodes, cost restriction nodes and condition judgment nodes. All nodes are connected by directed edges to represent the constraint transmission relationship between nodes. A single version rule subgraph is constructed based on the constraint transmission relationship. The single-version rule subgraphs at different time points are topologically aligned on the time axis. By calculating the semantic similarity of nodes in the single-version rule subgraphs, the homologous nodes and mutated nodes across versions are identified. Version inheritance edges are established between homologous nodes, and version differentiation edges are established between mutated nodes. The version inheritance edges and version differentiation edges are used to connect each single-version rule subgraph into a temporal association structure, forming a multi-version knowledge graph of medical insurance rules. Based on timestamp information, the effective time interval covering the timestamp is retrieved in the temporal association structure of the multi-version knowledge graph. The corresponding target rule version subgraph is located, and the version inheritance edge and version differentiation edge of the target rule version subgraph are traversed forward along the time axis to the initial rule version. The mutation nodes associated with the version differentiation edge during the traversal are recorded, and the constraint logic differences of the mutation nodes are extracted as the evolution path.

3. The method according to claim 1, characterized in that, By mapping the rule change features in the evolutionary path to version-aware vectors through semantic tensor transformation, establishing a semantic alignment relationship between the version-aware vectors and the diagnosis and treatment event sequence, and generating a hybrid representation carrying the semantics of rule evolution, the following is achieved: Semantically aggregate the mutated nodes associated with the differentiation edges of each version in the evolution path, form rule change features through constraint logic difference analysis, construct a third-order semantic tensor representation of the rule change features in the time dimension, constraint strength dimension and applicable scope dimension, perform tensor decomposition on the third-order semantic tensor representation along the time dimension to capture the dominant mode of rule evolution, and convert the dominant mode into a version-aware vector through low-dimensional manifold mapping. The behavioral semantics, temporal relationships, and cost association attributes of each diagnosis and treatment event in the diagnosis and treatment event sequence are structured and encoded to construct a vector representation of the diagnosis and treatment event. The mutual information distribution between the version-aware vector and the vector representation of the diagnosis and treatment event in the semantic space is calculated. Based on the mutual information distribution, sensitive diagnosis and treatment events affected by rule evolution are screened, and a semantic alignment relationship is established between the sensitive diagnosis and treatment events and the version-aware vector. The version-aware vector is dynamically modulated according to the semantic alignment relationship, so that the rule evolution pattern in the version-aware vector and the semantic features of the sensitive diagnosis and treatment event have a resonance enhancement effect. The modulated version-aware vector is fused with the overall semantic representation of the diagnosis and treatment event sequence by tensor product to generate a hybrid representation that simultaneously encodes the diagnosis and treatment behavior logic and the rule evolution trajectory.

4. The method according to claim 3, characterized in that, Dynamically modulating the version-aware vector based on the semantic alignment relationship, so that the rule evolution pattern in the version-aware vector resonates and enhances the semantic features of the sensitive medical events, includes: Based on semantic alignment, a coupled tensor field is constructed between the version-aware vector and the sensitive diagnosis and treatment event. In the coupled tensor field, the semantic response gradient of each dimension component of the version-aware vector to the sensitive diagnosis and treatment event is calculated. Based on the semantic response gradient, the important dimensions that need to be enhanced and the redundant dimensions that need to be suppressed are determined, and a dimension-selective modulation mask is generated. The dimension-selective modulation mask is applied to the version-aware vector. The representation intensity of the important dimensions is amplified and the representation intensity of the redundant dimensions is attenuated through a gating mechanism. This causes the modulated version-aware vector to move in the semantic space toward the distribution area of ​​the sensitive diagnosis and treatment event. The local curvature matching degree between the moved version-aware vector and the semantic features of the sensitive diagnosis and treatment event on the feature manifold is calculated. When the local curvature matching degree exceeds the coordination threshold, a resonance loop is established between the version perception vector and the sensitive diagnosis and treatment event in the coupled tensor field. Through the resonance loop, the rule evolution mode in the version perception vector and the semantic features of the sensitive diagnosis and treatment event generate synchronous oscillations in the corresponding dimensions of the feature space. The synchronous oscillations cause the semantic expressions of the two to reinforce each other in the corresponding dimensions, forming a resonance enhancement effect.

5. The method according to claim 1, characterized in that, The hybrid representation is fed into an adversarial dual-path review architecture, wherein the first path generates a deterministic review result based on rule-based logical reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning, including: The hybrid representation is decomposed into an explicit rule constraint subspace and an implicit semantic association subspace in the feature space. The projection vector of the explicit rule constraint subspace is fed into the first path. A constraint propagation network is constructed in the first path. The rule constraint conditions in the projection vector are forward reasoned and backward verified through the constraint propagation network. When the truth value determination of all constraint conditions converges, a binary deterministic audit result is generated. The complete representation of the hybrid representation is fed into the second path, where the diagnosis and treatment behavior patterns and rule evolution patterns in the implicit semantic association subspace are jointly probabilistically modeled by an AI model. Based on the AI ​​model, the posterior probability of the hybrid representation under the anomaly detection distribution is calculated, and a probabilistic audit result with a continuous value range is generated based on the posterior probability.

6. The method according to claim 5, characterized in that, In the second path, an AI model is used to jointly model the diagnostic behavior patterns and rule evolution patterns in the implicit semantic association subspace. Based on the AI ​​model, the posterior probability of the hybrid representation under the anomaly detection distribution is calculated, including: In the implicit semantic association subspace, the behavioral sequence dependency structure of the diagnosis and treatment behavior pattern and the version migration trajectory of the rule evolution pattern are extracted. The behavioral sequence dependency structure and the version migration trajectory are geometrically aligned on the feature manifold through topological isomorphism mapping to generate a joint representation structure that encodes the dual-mode coupling relationship. The AI ​​model performs conditional dependency learning on the treatment behavior patterns and rule evolution patterns in the joint representation structure, and establishes a conditional probability distribution of the treatment behavior patterns with the rule evolution patterns as conditional variables. The parameters of the conditional probability distribution are adaptively inferred by the AI ​​model from the joint representation structure, forming a joint probability modeling result that characterizes the impact of rule versions on treatment compliance. The hybrid representation is input into the AI ​​model. Based on the joint probability modeling results, the probability density ratio of the hybrid representation under the normal diagnosis and treatment behavior distribution and the abnormal diagnosis and treatment behavior distribution is calculated. The anomaly confidence of the hybrid representation is calculated by combining the probability density ratio with the historical statistical prior of the rule evolution mode. The anomaly confidence is the posterior probability of the hybrid representation under the anomaly detection distribution.

7. The method according to claim 1, characterized in that, When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under conditions of few samples. Based on the dual-path fusion result after adversarial convergence, the final review conclusion is output, including: Disagreement samples are identified by calculating the decision vector distance between deterministic review results and probabilistic review results. The rule evolution path features associated with the disagreement samples are extracted, and the rule evolution path features are transformed into meta-learning signals containing semantic difference information between versions. The meta-learning signal is fed into the meta-parameter update layer of the AI ​​model. The meta-learning signal guides the AI ​​model to construct a fast adaptation direction pointing to the semantics of the new version of the rules in the gradient space. The fast adaptation direction is used to make incremental adjustments to the semantic encoding parameters of the AI ​​model with few samples, so that the AI ​​model can complete the reconstruction of the semantic space of the new version of the rules in the local semantic space of the divergent samples. In the adversarial dual-path review architecture, continuous iterative optimization is performed. When the consistency measure between the deterministic review result and the probabilistic review result on the divergent sample satisfies the adversarial convergence condition, the deterministic review result and the probabilistic review result are weighted by credibility to generate a dual-path fusion result, which is then used as the final review conclusion.

8. A medical record intelligent review system based on versioned medical insurance rules and AI models, used to implement the method as described in any one of claims 1-7, characterized in that, include: The version tracing unit is used to obtain the medical record data to be reviewed and its timestamp information, parse the diagnosis and treatment event sequence in the medical record data to be reviewed, construct a multi-version knowledge graph of medical insurance rules, and trace the target rule version and its evolution path in the multi-version knowledge graph based on the timestamp information. A hybrid representation unit is used to map the rule change features in the evolution path into version-aware vectors through semantic tensor transformation, establish a semantic alignment relationship between the version-aware vectors and the diagnosis and treatment event sequence, and generate a hybrid representation carrying the semantics of rule evolution. The dual-path review unit is used to feed the hybrid representation to the adversarial dual-path review architecture. The first path generates a deterministic review result based on rule logic reasoning, and the second path generates a probabilistic review result based on AI model semantic reasoning. This allows the two paths to constrain each other in the feature space. When the conclusions of the two paths diverge, the rule evolution path features of the divergent samples are transformed into meta-learning signals to guide the AI ​​model to quickly reconstruct the semantic space of the new version of the rules under limited sample conditions. The final review conclusion is output based on the dual-path fusion result after adversarial convergence.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images