A Medical Data Field Mapping Relationship Prediction Method Based on World Model Learning

Abstract

This invention provides a method for predicting mapping relationships of medical data fields based on world model learning, belonging to the field of model prediction technology. The method includes acquiring a medical data object containing a source field set and a target field set, performing causal state configuration based on a generative world model to generate intermediate world states, performing alignment inference, and combining consistency policy constraints to generate mapping relationship prediction results. Simultaneously, it utilizes backfilling feedback information for closed-loop optimization. This invention improves the accuracy and interpretability of medical data field mapping and reduces the mapping error rate.

Description

Technical Field

[0001] This invention relates to model prediction technology, and more particularly to a method for predicting the mapping relationship of medical data fields based on world model learning. Background Technology

[0002] With the rapid development of medical informatization, the interconnection and integration of medical data has become an important direction for industry development. Data exchange and integration between different medical institutions and different medical information systems require solving the problem of data field mapping, that is, determining the correspondence between fields with the same meaning in different data sources. Currently, medical data field mapping mainly relies on manual methods, where professionals need to manually establish the correspondence between source data fields and target data fields based on their experience and understanding of medical data standards. However, with the continuous enrichment of medical data standards and the increasing complexity of medical systems, traditional manual mapping methods can no longer meet the needs of large-scale medical data fusion.

[0003] Machine learning-based data mapping methods are increasingly being applied in the medical field, including rule-based mapping methods, string similarity-based mapping methods, and deep learning-based mapping methods. These methods attempt to automatically establish mapping relationships between fields by analyzing features such as field names, data types, and value ranges.

[0004] Existing field mapping methods lack a deep understanding of the unique semantics and structure of medical data. Medical data is highly specialized and has complex internal relationships. Existing methods often only focus on surface feature matching and fail to capture the semantic relevance between fields and the constraints of domain knowledge, resulting in low accuracy when faced with ambiguous terms and synonymous expressions.

[0005] Traditional mapping methods struggle to handle cross-table constraints and causal dependencies. Medical data is often distributed across multiple related table structures, with complex references and constraints between fields. Existing technologies mostly process data on a single table or single field, lacking the ability to perform structured mapping from a global perspective, making it difficult to guarantee the overall consistency of the mapping results.

[0006] Existing technologies lack interpretability and adaptive optimization mechanisms for the mapping process. Medical data processing demands extremely high accuracy; mapping errors can lead to serious consequences. However, existing methods typically fail to provide a basis for mapping decisions and confidence assessments, and also lack effective feedback mechanisms to continuously optimize the mapping algorithm, making it difficult for the system to adapt to the specific needs of different medical scenarios. Summary of the Invention

[0007] This invention provides a method for predicting the mapping relationship of medical data fields based on world model learning, which can solve the problems in the prior art.

[0008] A first aspect of the present invention provides a method for predicting the mapping relationship of medical data fields based on world model learning, comprising: acquiring a target medical data object, wherein the target medical data object includes a set of source fields with structural identifiers and a set of target fields to be aligned; uniformly extracting the source field set and the target field set to form an original description containing field semantic symbols and contextual relationships;

[0009] Based on an interpretable generative world model, the original description is configured with causal states, and intermediate world states representing the semantic causal dependencies and cross-table constraints of the fields are output. Based on the structural association identifiers and semantic constraint identifiers in the intermediate world states, the source field set and the target field set are aligned and inferred to generate alignment results.

[0010] Based on the alignment results and consistency strategy constraints, each candidate mapping pair is given a deterministic judgment and confidence basis under the conditions of satisfying structural consistency and cross-table constraint consistency, and the mapping relationship prediction result is obtained.

[0011] Based on the prediction results of the mapping relationship, field mapping is performed on the target medical data object to generate backfill feedback information. The backfill feedback information is used to adaptively update the results of the causal state configuration and the consistency strategy constraints. The updated causal state configuration results and consistency strategy constraints are used for a new round of alignment inference and decision generation, forming a closed-loop optimization driven by the backfill feedback information.

[0012] The original description is causally configured based on an interpretable generative world model, and the intermediate world state of the semantic causal dependencies and cross-table constraints of the representation fields is output. The process includes: the original description information contains multiple representation fields, the multiple representation fields are input into the interpretable generative world model, a semantic dependency graph of the representation fields is constructed based on multi-agent collaborative reasoning; based on the semantic dependency graph, a random walk algorithm is used to extract important causal paths of the representation fields, a feature mapping matrix of the representation fields is constructed according to the important causal paths, and the state of the original description information is reconstructed based on the feature mapping matrix.

[0013] The reconstructed state space is clustered according to the semantic similarity of the fields to obtain multiple semantic subspaces. Local constraints and global constraints are set for each semantic subspace. The state is iteratively updated based on the local constraints and the global constraints. During each iteration, the state that violates the constraints is adjusted by gradient descent until the multiple semantic subspaces meet the preset cross-table consistency threshold, and the intermediate world state is obtained.

[0014] Based on the semantic dependency graph, a random walk algorithm is used to extract important causal paths of the representation fields. A feature mapping matrix of the representation fields is constructed based on the important causal paths. The state reconstruction of the original description information based on the feature mapping matrix includes: performing a random walk process based on the semantic dependency graph, setting the transition probability at each node according to the outgoing edge weight, generating multiple walk paths through Monte Carlo sampling, calculating the cumulative value of information gain between nodes for each path, and marking the path as an important causal path when the cumulative value exceeds a preset information threshold.

[0015] Node sequences are extracted from the important causal path, and the transition probabilities between adjacent node pairs in the node sequence are calculated to generate a node transition matrix. For each source node in the node transition matrix, the correlation score between it and all target nodes is calculated. The correlation score is normalized to obtain a weight coefficient. The weight coefficient is used to linearly combine the original representation fields to generate a new state vector.

[0016] Based on the state vector, a state mapping rule is constructed, and the original description information is reconstructed according to the state mapping rule to obtain a reconstructed state containing causal dependencies.

[0017] Based on the structural association identifier and semantic constraint identifier in the intermediate world state, the source field set and the target field set are aligned and inferred to generate an alignment result including: the intermediate world state contains the source field set, the target field set, the structural association identifier, and the semantic constraint identifier; the structural information of each field in the source field set is extracted to generate a source field value vector; the structural information of each field in the target field set is extracted to generate a target field value vector.

[0018] Calculate the cosine distance between the source field value vector and the target field value vector to obtain an initial similarity matrix, and construct a field association graph based on the initial similarity matrix; use a graph embedding method to transform the field association graph into a low-dimensional representation space, execute a random walk algorithm in the low-dimensional representation space to generate a walk sequence between fields, and calculate the transition probability between field pairs based on the walk sequence;

[0019] Semantic rules for the source and target fields are extracted based on the semantic constraint identifiers. The semantic compatibility of the field pairs is calculated based on the semantic rules. The transition probability and the semantic compatibility are weighted and combined to obtain the matching probability of the field pairs. A greedy search strategy is used to perform iterative matching of the fields based on the matching probability. The consistency of the results of the iterative matching of the fields is verified, and the field mapping relationship that passes the consistency verification is output as the final alignment result.

[0020] Based on the alignment results and consistency policy constraints, each candidate mapping pair is given a deterministic judgment and confidence basis under the condition of satisfying structural consistency and cross-table constraint consistency. The mapping relationship prediction results are obtained by: obtaining the alignment results and consistency policy constraints, wherein the alignment results include a set of candidate mapping pairs, and the consistency policy constraints include structural consistency constraints and cross-table constraints.

[0021] A constraint propagation network is constructed based on the set of candidate mapping pairs. The nodes of the constraint propagation network represent candidate mapping pairs, and the edges represent the association between structural consistency constraints and cross-table constraints.

[0022] The belief propagation algorithm is executed on the constraint propagation network to iteratively calculate and update the confidence state of the nodes; the confidence state is probabilistically modeled, and the Monte Carlo tree search method is used to sample the constraint-satisfied paths. Based on the sampling results, the frequency of each candidate mapping pair appearing in the effective path is counted, and the frequency is used as the confidence index for deterministic judgment.

[0023] Candidate mapping pairs are selected based on the confidence index, and mapping pairs with a confidence level higher than a preset confidence threshold and their path evidence in the constraint propagation network are output as mapping relationship prediction results.

[0024] The belief propagation algorithm is executed on the constraint propagation network to iteratively calculate and update the confidence state of the nodes; probabilistic modeling of the confidence state is performed, and the Monte Carlo tree search method is used to sample the constraint satisfaction path, including: constructing a factor graph on the constraint propagation network, executing the belief propagation algorithm to message the factor graph, updating the probability information of variable nodes and function nodes alternately in each iteration, obtaining the marginal probability distribution of the nodes after multiple iterations, and calculating the confidence state of the nodes based on the marginal probability distribution;

[0025] A Monte Carlo search tree is constructed by probabilistically modeling the confidence state. Hierarchical sampling is performed on the Monte Carlo search tree. During the sampling process, the sampling probability is dynamically adjusted according to the confidence state of the node. Branches with confidence scores higher than the preset confidence threshold are selected for expansion first. The trajectory of node state changes and the degree of constraint satisfaction on each sampling path are recorded, and the common characteristics of high-quality paths are statistically analyzed.

[0026] The confidence state of the nodes is corrected based on the common characteristics, and the corrected state is output as the optimization result.

[0027] Based on the mapping relationship prediction results, field mapping is performed on the target medical data object to generate backfill feedback information. The adaptive update of the causal state configuration result and the consistency strategy constraints using the backfill feedback information includes: performing mapping transformation on the fields in the target medical data object according to the mapping relationship prediction results, recording the mapping success identifier and data quality index of each field; constructing a backfill feedback vector based on the mapping success identifier and the data quality index; performing time series analysis on the backfill feedback vector to identify the changing trend of the mapping effect over time, and generating a feedback feature matrix.

[0028] The causal state configuration is updated using the feedback feature matrix. By adjusting the field dependency strength and influence direction, the association structure between fields is optimized. Based on the rule compliance in the feedback feature matrix, the threshold parameters and constraint weights in the consistency strategy constraints are dynamically adjusted to improve the adaptability of the constraints to the data features.

[0029] A second aspect of the present invention provides an electronic device, comprising: a processor;

[0030] Memory used to store processor-executable instructions;

[0031] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0032] A third 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.

[0033] The beneficial effects of this application are as follows: It adopts an interpretable generative world model for causal state configuration, effectively captures the semantic causal dependencies and cross-table constraints between fields, breaks through the limitations of traditional methods in understanding complex medical terms, and improves the accuracy and reliability of mapping, especially when facing professional medical terms and non-standardized expressions.

[0034] By employing a dual verification mechanism of structural consistency and cross-table constraint consistency, deterministic judgment and confidence basis are provided for each mapping pair, resolving the ambiguity problem in medical data mapping, ensuring the professionalism and reliability of the mapping results, and meeting the high-precision requirements of the medical field for data processing.

[0035] By introducing a closed-loop optimization mechanism driven by backfill feedback information, the system achieves adaptive updates of causal state configuration results and consistency strategy constraints. The system can continuously learn and improve from practical applications, forming a continuously evolving mapping capability, which significantly enhances its adaptability and scalability in complex and ever-changing medical data environments. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the medical data field mapping relationship prediction method based on world model learning, as described in an embodiment of the present invention.

[0037] Figure 2 This is a flowchart illustrating the optimization of medical data field mapping status and the construction of dependency relationships in an embodiment of the present invention. Detailed Implementation

[0038] 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.

[0039] 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.

[0040] Figure 1 This is a flowchart illustrating the medical data field mapping relationship prediction method based on world model learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: obtaining a target medical data object, the target medical data object including a source field set with structural identifiers and a target field set to be aligned; uniformly extracting the source field set and the target field set to form an original description containing field semantic symbols and contextual relationships;

[0041] Based on an interpretable generative world model, the original description is configured with causal states, and intermediate world states representing the semantic causal dependencies and cross-table constraints of the fields are output. Based on the structural association identifiers and semantic constraint identifiers in the intermediate world states, the source field set and the target field set are aligned and inferred to generate alignment results.

[0042] Based on the alignment results and consistency strategy constraints, each candidate mapping pair is given a deterministic judgment and confidence basis under the conditions of satisfying structural consistency and cross-table constraint consistency, and the mapping relationship prediction result is obtained.

[0043] Based on the prediction results of the mapping relationship, field mapping is performed on the target medical data object to generate backfill feedback information. The backfill feedback information is used to adaptively update the results of the causal state configuration and the consistency strategy constraints. The updated causal state configuration results and consistency strategy constraints are used for a new round of alignment inference and decision generation, forming a closed-loop optimization driven by the backfill feedback information.

[0044] In one optional implementation, the original description is configured with causal states based on an interpretable generative world model, and the intermediate world states of semantic causal dependencies and cross-table constraints of the representation fields are output. This includes: the original description information contains multiple representation fields; the multiple representation fields are input into the interpretable generative world model; a semantic dependency graph of the representation fields is constructed based on multi-agent collaborative reasoning; important causal paths of the representation fields are extracted using a random walk algorithm based on the semantic dependency graph; a feature mapping matrix of the representation fields is constructed based on the important causal paths; and the original description information is reconstructed based on the feature mapping matrix.

[0045] The reconstructed state space is clustered according to the semantic similarity of the fields to obtain multiple semantic subspaces. Local constraints and global constraints are set for each semantic subspace. The state is iteratively updated based on the local constraints and the global constraints. During each iteration, the state that violates the constraints is adjusted by gradient descent until the multiple semantic subspaces meet the preset cross-table consistency threshold, and the intermediate world state is obtained.

[0046] like Figure 2 As shown, the method includes: a parsing process of the original descriptive information to extract multiple representation fields from the medical data. Each representation field contains attribute information such as field name, data type, value range, business meaning, and correlation. Preprocessing of the representation fields unifies the field representation format through standardized operations. Field names adopt a naming convention using lowercase letters and underscores, data types are mapped to a unified type code, value ranges are converted to numerical interval representations, and business meanings are semantically annotated using a medical domain dictionary. Vectorization of field attributes converts text information into numerical vectors, facilitating subsequent calculations and processing.

[0047] The interpretable generative world model is constructed using a layered architecture: a bottom layer is the field representation layer, a middle layer is the relation inference layer, and a top layer is the state generation layer. The field representation layer converts the input representation fields into high-dimensional vector representations, using a pre-trained medical domain word vector model to encode field names and business meanings. The relation inference layer identifies potential relationships between fields based on an attention mechanism, with attention weights reflecting the strength of dependencies between fields. The state generation layer generates state-space representations of the representation fields through a variational autoencoder. The encoder compresses the field vectors into low-dimensional latent representations, and the decoder reconstructs the field's state information from these latent representations.

[0048] A multi-agent collaborative reasoning mechanism constructs a semantic dependency graph representing fields, with each agent specializing in the semantic analysis of a specific type of field. The patient information agent handles basic personal information fields, the diagnosis information agent handles disease diagnosis-related fields, the treatment information agent handles treatment plan-related fields, and the laboratory information agent handles laboratory test-related fields. Agents interact via a message passing mechanism, with messages including the semantic labels of fields, association strength, and constraints. The collaborative reasoning process is iterative; in each iteration, agents update their semantic understanding of the fields they are responsible for and broadcast the updated results to other agents.

[0049] The construction of a semantic dependency graph uses representation fields as graph nodes and semantic dependencies between fields as graph edges. Node attributes include the semantic type, importance weight, and constraints of the field, while edge attributes include dependency type, dependency strength, and directionality. Dependency types are categorized into three types: causal dependency, correlational dependency, and constraint dependency. Causal dependency indicates that the value of one field directly affects the value of another field; correlational dependency indicates that there is a statistical correlation between fields but no explicit causal relationship; and constraint dependency indicates that there are business rule constraints between fields. Dependency strength is quantified using mutual information; field pairs with high dependency strength are represented by thick edges in the graph, while field pairs with low dependency strength are represented by thin edges.

[0050] The random walk algorithm performs path sampling on a semantically dependent graph, starting from each node and performing multiple random walks to explore the local structure of the graph. The walk step size is set to 15 steps, and the number of walks is set to 500 to ensure sufficient coverage of the connected regions of the graph. During the walk, transition probabilities are calculated based on edge weights, with edges with higher weights receiving a greater selection probability. The walk strategy employs biased random walks, introducing backtracking and exploration parameters to balance the search biases of depth-first and breadth-first search. The backtracking parameter controls the probability of a walk returning to a visited node, while the exploration parameter controls the probability of a walk visiting distant nodes.

[0051] The extraction of important causal paths is based on path frequency statistics and importance scoring according to the random walk results. Path frequency statistics calculate the number of times each path appears in all walk sequences; paths with higher frequencies indicate greater importance in the graph structure. The importance score is calculated by combining path length, node importance, and edge weight. Paths with moderate length, containing important nodes, and high edge weights receive higher scores. The path filtering mechanism removes invalid paths that are too short or too long, retaining valid paths with lengths between 3 and 10. Path clustering groups similar paths into path clusters, and the most representative path in each cluster is selected as the important causal path.

[0052] The feature mapping matrix is ​​constructed by filling the matrix with the co-occurrence relationships of nodes in important causal paths. Matrix rows correspond to source fields, and matrix columns correspond to target fields. Matrix element values ​​reflect the strength of the influence of the source field on the target field. The calculation of influence strength considers three factors: path length, path weight, and node importance. Shorter paths indicate more direct influence, higher path weights indicate more reliable influence, and higher node importance indicates more significant influence. Matrix symmetry is handled through averaging to ensure consistency of influence relationships, and matrix sparsity is addressed through threshold filtering to ensure the effectiveness of the mapping relationships.

[0053] The state reconstruction process of the original description information utilizes a feature mapping matrix to perform a linear transformation on the state representation of the fields. The reconstruction operation multiplies the original field vector with the mapping matrix to obtain a new state vector. The new state vector retains the original information while incorporating causal dependencies. The dimensionality of the reconstructed state space remains consistent with the original space to ensure the integrity of the information. Normalization of the state vector prevents numerical overflow and gradient vanishing problems. The normalization method uses L2 norm normalization to ensure that the vector magnitude is one.

[0054] Clustering analysis for reconstructing the state space is based on hierarchical clustering using semantic similarity of fields. Semantic similarity is calculated using the cosine similarity of word vectors, and fields with high similarity are grouped into the same semantic subspace. The clustering algorithm employs agglomerative hierarchical clustering, starting with each field as an independent cluster and gradually merging the clusters with the highest similarity until a preset number of clusters or a similarity threshold is reached. The quality of the clustering results is evaluated using the silhouette coefficient; a high silhouette coefficient indicates good clustering performance, with high intra-cluster similarity and large inter-cluster differences.

[0055] The local constraints of the semantic subspace define constraints for specific business rules of fields within each subspace. Local constraints include three categories: data type constraints, value range constraints, and format specification constraints. Data type constraints ensure the consistency of field types within the subspace; value range constraints limit field values ​​to a reasonable range; and format specification constraints unify the representation format of fields. Constraint strength is graded according to business importance, with strong constraints applied to core business fields and weak constraints applied to auxiliary business fields. The penalty function for constraint violation uses a squared loss form, with heavier penalties for greater violations.

[0056] Global constraints are set considering consistency requirements across multiple semantic subspaces and cross-table relationships. Cross-table constraints ensure data consistency of related fields in different tables, referential integrity constraints guarantee the validity of foreign key relationships, and business logic constraints maintain the correctness of domain knowledge. Global constraints have higher weights than local constraints, reflecting the importance of overall consistency. The constraint conflict detection mechanism identifies contradictions between local and global constraints, and the conflict resolution strategy prioritizes satisfying the requirements of global constraints.

[0057] The state iterative update process employs gradient descent optimization to minimize the constraint violation loss function. This loss function comprises two parts: local constraint loss and global constraint loss. The local constraint loss measures the degree to which constraints are satisfied within a subspace, while the global constraint loss measures the degree to which constraints are satisfied across the entire subspace. Gradient calculation is achieved through automatic differentiation; the gradient direction indicates the direction of state adjustment, and the gradient magnitude reflects the magnitude of the adjustment. The adaptive adjustment of the learning rate is based on the trend of the loss function's change; increasing the learning rate accelerates convergence when the loss decreases rapidly, and decreasing it stabilizes training when the loss oscillates.

[0058] The adjustment strategy for constraint violations employs different correction methods depending on the constraint type. For hard constraint violations, a projection method is used to forcibly adjust the state to within the feasible region. For soft constraint violations, a penalty method is used, guiding the state to gradually adjust through a loss function. The adjustment magnitude is controlled by a step size limit to avoid instability caused by over-adjustment; the upper limit of the step size is dynamically set based on the constraint sensitivity. The priority of state adjustments is ranked based on the importance of the constraints, with important constraints adjusted before less important ones.

[0059] Iterative convergence is determined by monitoring the satisfaction of a cross-table consistency threshold, which is set as the weighted average of all constraint violations. Meeting the threshold indicates that the expected level of consistency has been achieved among the semantic subspaces. The convergence criteria also include monitoring the magnitude of state changes between consecutive iterations; iteration stops when the magnitude of the change falls below a preset threshold. An upper limit on the number of iterations is set to prevent infinite loops; if the maximum number of iterations is exceeded, the iteration is forcibly terminated and the current optimal state is output.

[0060] The generation of intermediate world states integrates the semantic subspace after iterative convergence, forming a unified state representation that includes field semantic causal dependencies and cross-table constraints. The intermediate state comprises four core components: a source field set, a target field set, a structural association identifier, and a semantic constraint identifier. The source field set records the field information of the original data, the target field set records the field information of the target schema, the structural association identifier records the dependencies between fields, and the semantic constraint identifier records business rules and constraints. The serialization of the state representation uses a standard format to ensure compatibility with subsequent processing, and the quality of the state is quantified through constraint satisfaction and semantic consistency.

[0061] In one optional implementation, the important causal paths of the representation fields are extracted using a random walk algorithm based on the semantic dependency graph, a feature mapping matrix of the representation fields is constructed based on the important causal paths, and the state reconstruction of the original description information based on the feature mapping matrix includes: performing a random walk process based on the semantic dependency graph, setting the transition probability at each node according to the outgoing edge weight, generating multiple walk paths through Monte Carlo sampling, calculating the cumulative value of information gain between nodes for each path, and marking the path as an important causal path when the cumulative value exceeds a preset information threshold;

[0062] Node sequences are extracted from the important causal path, and the transition probabilities between adjacent node pairs in the node sequence are calculated to generate a node transition matrix. For each source node in the node transition matrix, the correlation score between it and all target nodes is calculated. The correlation score is normalized to obtain a weight coefficient. The weight coefficient is used to linearly combine the original representation fields to generate a new state vector.

[0063] Based on the state vector, a state mapping rule is constructed, and the original description information is reconstructed according to the state mapping rule to obtain a reconstructed state containing causal dependencies.

[0064] A random walk process is performed based on a constructed semantic dependency graph. In the semantic dependency graph, nodes represent key elements in the descriptive information, edges represent semantic relationships between nodes, and edge weights represent the strength of the relationship. For each node in the graph, the system sets the transition probability based on the weights of its outgoing edges. Specifically, if the current node is vi, the set of nodes connected to it is Neighbors(vi), and the weight of the connecting edge is wij, then the probability P(vi→vj) of transitioning from node vi to node vj is calculated as the edge weight divided by the sum of the weights of all outgoing edges. For example, if node A connects to nodes B, C, and D, with edge weights of 0.5, 0.3, and 0.2 respectively, then the probability of transitioning from A to B is 0.5, the probability of transitioning to C is 0.3, and the probability of transitioning to D is 0.2.

[0065] Multiple walk paths are generated using the Monte Carlo sampling method. Specifically, starting from a key node in the semantic dependency graph (such as a node corresponding to a representation field), the next node is randomly selected based on the aforementioned transition probabilities, and the sequence of visited nodes is recorded until a preset maximum step size (e.g., 20 steps) is reached or the same node is visited repeatedly. To ensure sufficient sampling, the system typically generates a large number of paths (e.g., 10,000) to cover important connections in the graph.

[0066] For each generated walk path, the cumulative information gain between nodes is calculated. The information gain IG(vi,vj) between adjacent nodes vi and vj on the path is determined by calculating the mutual information of the node feature vectors. For a path p=(v1,v2,...,vn) containing n nodes, its cumulative information gain value IG(p) is the sum of the information gains of all adjacent nodes on the path. When the cumulative information gain value of path p exceeds a preset information threshold τ (e.g., τ=0.65), the path is marked as an important causal path and saved. For example, if a path contains the nodes "device-fault-temperature-overheat-shutdown" and has a cumulative information gain value of 0.78, exceeding the threshold of 0.65, then the path is identified as an important causal path.

[0067] From the identified important causal paths, node sequences are extracted and the transition probabilities between nodes are calculated to generate a node transition matrix M. Specifically, for the set P of all important causal paths, the frequency (vi→vj) of node vi transitioning to node vj is counted. The element Mij of the node transition matrix is ​​calculated as this transition frequency divided by the sum of the transition frequencies of all outgoing edges of node vi. For example, if the frequency of node "fault" transitioning to "temperature" is 85 times and the frequency of transitioning to "pressure" is 15 times in all important paths, then the transition probability from "fault" to "temperature" is 0.85, and the transition probability to "pressure" is 0.15.

[0068] For each source node vi in ​​the node transition matrix, calculate the correlation score Score(vi,vj) between it and all target nodes vj. This score comprehensively considers the direct transition probability Mij and the indirect transition probability through intermediate nodes, using a decay coefficient α (usually set to 0.85) to control the influence of indirect paths. Specifically, the first-order correlation is the direct transition probability, the second-order correlation is the transition probability through an intermediate node multiplied by the decay coefficient, and so on. The final correlation score is the sum of all order correlations, considering a maximum of k-order correlations (e.g., k=3). For example, the correlation calculation from node "fault" to "shutdown": direct transition probability 0.2, indirect transition probability through "temperature" 0.85×0.75×0.6=0.38, indirect transition probability through "pressure" 0.85×0.15×0.4=0.05, and the total correlation score is 0.63.

[0069] The correlation scores are normalized using the softmax function to obtain the weight coefficients W(vi,vj). For the source node vi and all target nodes vj∈V, the weight coefficients represent the strength of the source node's influence on each target node. For example, the weight coefficients of the node "fault" on the three nodes "temperature", "pressure" and "shutdown" are 0.58, 0.12 and 0.30, respectively, and the sum is 1.

[0070] Using the calculated weight coefficients, the original representation fields are linearly combined to generate a new state vector. If the feature vector corresponding to the original representation field is f(vi), then the updated state vector s(vi) is the weighted sum of its feature vectors and those of all associated nodes. For example, the original feature vector of the node "fault" is [0.5, 0.6, 0.3, 0.2], and the feature vectors of the associated nodes "temperature", "pressure" and "shutdown" are [0.7, 0.3, 0.5, 0.4], [0.2, 0.8, 0.3, 0.6] and [0.4, 0.5, 0.7, 0.3], respectively. According to the above weight coefficients, the updated state vector is [0.58, 0.44, 0.52, 0.35].

[0071] State mapping rules are constructed based on the generated state vectors. These rules define how each element in the original descriptive information is mapped to a new state containing causal dependencies. The mapping rules include node replacement rules and relationship enhancement rules. The former determines which original nodes need to be replaced with the updated state vectors, while the latter determines the causal relationships that need to be explicitly expressed in the reconstruction result. For example, the mapping rule specifies that the "fault" node is replaced with the updated state vector representation, and explicit causal links are established with the "overheating" and "shutdown" nodes.

[0072] The original description information is reconstructed according to state mapping rules to generate a reconstructed state containing causal dependencies. The reconstruction process preserves the core semantics of the original information while enhancing the expression of the relationships between nodes on key causal paths. For example, the original description "the equipment malfunctioned and shut down" is reconstructed as "the equipment malfunctioned due to excessive temperature, which triggered shutdown protection," where the causal chain "excessive temperature → malfunction → shutdown" is clearly reflected in the reconstruction result.

[0073] By employing the aforementioned techniques, implicit causal dependencies can be extracted from the original descriptive information, and these dependencies can be explicitly expressed during the state reconstruction process, thereby improving the completeness and interpretability of information expression.

[0074] In one optional implementation, based on the structural association identifier and semantic constraint identifier in the intermediate world state, the source field set and the target field set are aligned and inferred to generate an alignment result including: the intermediate world state contains a source field set, a target field set, a structural association identifier, and a semantic constraint identifier; the structural information of each field in the source field set is extracted to generate a source field value vector; the structural information of each field in the target field set is extracted to generate a target field value vector.

[0075] Calculate the cosine distance between the source field value vector and the target field value vector to obtain an initial similarity matrix, and construct a field association graph based on the initial similarity matrix; use a graph embedding method to transform the field association graph into a low-dimensional representation space, execute a random walk algorithm in the low-dimensional representation space to generate a walk sequence between fields, and calculate the transition probability between field pairs based on the walk sequence;

[0076] Semantic rules for the source and target fields are extracted based on the semantic constraint identifiers. The semantic compatibility of the field pairs is calculated based on the semantic rules. The transition probability and the semantic compatibility are weighted and combined to obtain the matching probability of the field pairs. A greedy search strategy is used to perform iterative matching of the fields based on the matching probability. The consistency of the results of the iterative matching of the fields is verified, and the field mapping relationship that passes the consistency verification is output as the final alignment result.

[0077] The intermediate world state analysis extracts four core components from the causal state configuration results: a source field set, a target field set, structural association identifiers, and semantic constraint identifiers. The source field set contains the original medical data fields to be mapped, each carrying attribute information such as field name, data type, length limit, value range, and business meaning. The target field set contains the field structure of the mapping target, with field attribute formats consistent with the source fields to facilitate subsequent comparison and analysis. The structural association identifier records the hierarchical, dependency, and constraint relationships between fields, including parent-child relationship identifiers, sibling relationship identifiers, and cross-table relationship identifiers. The semantic constraint identifier contains semantic type markers for fields, business rule descriptions, and constraint definitions, used to guide semantic-level matching judgments.

[0078] The extraction of source field structure information employs a multi-dimensional feature extraction mechanism to convert field attributes into computable numerical representations. Field names are extracted using lexical analysis and semantic decomposition to extract keywords, which are then converted into word frequency vectors representing the field's naming characteristics. Data types are mapped to type encoding vectors: integers are encoded as 001, floating-point numbers as 010, character types as 100, date types as 111, and boolean types as 101. Field lengths and value ranges are standardized to values ​​between zero and one; lengths are normalized by dividing by a preset maximum length, and value ranges are normalized by the difference between the minimum and maximum values. Business meanings are mapped to semantic encoding vectors using a domain dictionary: patient information fields are encoded as 001, diagnostic information fields as 010, and treatment information fields as 100.

[0079] The generation of the source field numerical vector involves concatenating and weighting the extracted multidimensional features to form a unified vector representation. The vector dimensions include naming feature dimension, type feature dimension, numerical feature dimension, and semantic feature dimension. The weight of each dimension is set according to its importance to field matching. The naming feature dimension has a weight of 0.3, reflecting the significant impact of field names on matching. The type feature dimension has a weight of 0.4, emphasizing the crucial role of data type compatibility. The numerical feature dimension has a weight of 0.1, supplementing field size information. The semantic feature dimension has a weight of 0.2, reflecting the matching requirements of business semantics. During vector generation, the features of each dimension are weighted and summed to obtain a fixed-length source field numerical vector.

[0080] The extraction process for the target field's structural information is consistent with that of the source field, ensuring that both types of fields use the same feature extraction standards and encoding methods. The target field's naming features, type features, numerical features, and semantic features are processed using the same lexical analysis method, type encoding method, normalization method, and semantic mapping method, respectively. The generation of the target field's numerical vector also employs a weighted concatenation method, with weight coefficients consistent with those of the source field, ensuring the comparability of the source field vector and the target field vector.

[0081] Cosine distance is calculated by comparing the directional similarity between the source and target field numerical vectors to measure the structural similarity between them. The calculation involves vector dot product and vector magnitude calculation. The dot product multiplies the corresponding dimensions of the two vectors and then sums the results. The magnitude calculation takes the square root of the sum of the squares of each dimension. The cosine distance equals one minus the cosine similarity, and the cosine similarity equals the vector dot product divided by the product of the magnitudes of the two vectors. The distance value ranges from zero to one, where zero indicates that the vectors are identical and one indicates that the vectors are completely opposite.

[0082] The initial similarity matrix is ​​constructed by organizing the cosine distances between all source and target fields into a matrix form. Matrix rows correspond to source fields, matrix columns correspond to target fields, and matrix elements are the cosine distance values ​​of corresponding field pairs. The similarity matrix is ​​obtained by transforming the inverse of the distance values; the similarity is equal to one minus the normalized distance value, ensuring that the similarity value is within the range of zero to one. The symmetry of the matrix is ​​checked by comparing the difference between the transpose matrix and the original matrix; for non-square matrices, a pseudo-symmetry test method is used.

[0083] The field association graph is constructed based on establishing connections between highly similar field pairs in the initial similarity matrix. The graph employs a bipartite structure, with left-hand nodes representing source fields and right-hand nodes representing target fields. Edges represent the similarity relationships between fields. Edges are established only if the similarity value exceeds a preset threshold of 0.5 to filter out irrelevant connections with low similarity. Edge weights are directly represented by similarity values; higher weights indicate a stronger association between fields. The association graph also includes node attribute information, recording the structural and semantic features of the fields for subsequent graph embedding processing.

[0084] The graph embedding method uses a node-to-vector mapping technique to transform the field association graph into a low-dimensional vector representation space. The embedding process involves randomly initializing node vectors and iteratively optimizing the vector representation using graph structure and node attribute information. The optimization objective is to place nodes with high association levels closer together in the embedding space and nodes with low association levels further apart. The embedding dimension is set to 128 dimensions to balance representational power and computational efficiency. The iterative process uses gradient descent to update node vectors, with a learning rate of 0.01 and 1000 iterations to ensure convergence.

[0085] The random walk algorithm in the low-dimensional representation space starts from any node in the graph and selects the next node to visit according to the probability distribution of edge weights. The walk step size is set to 10 steps, and the number of walks is set to 1000 to obtain sufficient sampling data. Each walk generates a node sequence, recording all nodes visited along the walk path. The walk probability is calculated based on the node's out-degree and edge weights; nodes with higher out-degrees and larger edge weights receive a higher visit probability. To avoid getting trapped in local loops during the walk, a visit history record and a revisit penalty mechanism are used to ensure the diversity of the walks.

[0086] The generation of inter-field walk sequences involves summarizing all random walk paths to form a complete sequence set. Sequence preprocessing includes three steps: deduplication, filtering, and normalization. Deduplication removes identical walk sequences, filtering removes invalid sequences that are too short, and normalization standardizes the sequence representation format. Sequence analysis identifies implicit relationships between fields by statistically analyzing the co-occurrence frequency of node pairs in the sequence; field pairs with high co-occurrence frequencies indicate a strong association.

[0087] The transition probability is estimated based on the statistical frequency of node transitions in the walk sequence. The transition probability represents the likelihood of transitioning from a source field node to a target field node, calculated by dividing the number of transitions from the source field to the target field by the total number of transitions from the source field. The transition probability matrix is ​​constructed by organizing the transition probabilities of all field pairs into a matrix form. Row normalization ensures that the sum of elements in each row equals one. The smoothing of the transition probabilities uses Laplace smoothing to avoid zero-probability issues, adding a small constant term to all transitions before re-normalizing.

[0088] The semantic constraint identifier parsing extracts the semantic rule definitions for source and target fields. These semantic rules include three aspects: the business meaning of the field, its application scenario, and constraints. The business meaning is semantically annotated using a medical domain ontology library; the application scenario is determined through contextual analysis to define the field's usage environment; and the constraints are resolved by a rule engine that parses the field's value restrictions. The standardized processing of semantic rules transforms natural language descriptions into structured rule representations, facilitating computer processing and matching.

[0089] Semantic compatibility is calculated to evaluate the semantic matching degree between the source and target fields. The compatibility calculation considers three dimensions: semantic similarity, application compatibility, and constraint consistency. Semantic similarity is calculated using the cosine similarity of word vectors; application compatibility is analyzed through scenario overlap; and constraint consistency is determined through rule conflict detection. The weights of the three dimensions are set to 0.5, 0.3, and 0.2, respectively, with semantic similarity having the highest weight, reflecting the core role of semantic matching. The final compatibility value is calculated using a weighted average, ranging from zero to one.

[0090] The weighted combination of transition probability and semantic compatibility uses a linear fusion method to generate the matching probability of field pairs. The combination weights are allocated according to the importance of structural and semantic information, with the transition probability weight set to 0.6 and the semantic compatibility weight set to 0.4. The matching probability is calculated by multiplying the transition probability by 0.6 and adding the semantic compatibility weight by 0.4, ensuring that the matching probability comprehensively considers both structural similarity and semantic compatibility. Post-processing of the matching probability includes normalization and threshold filtering. Normalization ensures the validity of the probability values, while threshold filtering eliminates invalid matches with low probabilities.

[0091] The greedy search strategy iteratively selects fields based on matching probabilities. The search process starts with the field pair with the highest matching probability and progressively selects the highest-probability matching pairs from the remaining fields until all fields are matched or no matching pairs are available. The search strategy employs the principle of local optimum selection, choosing the currently optimal matching pair in each iteration to avoid the computational complexity of a global search. A conflict resolution mechanism ensures that each source field matches at most one target field, and each target field is matched at most one source field, maintaining the one-to-one mapping relationship.

[0092] The consistency verification of field iteration matching results checks whether the matching relationships meet the preset consistency constraints. The verification includes three aspects: data type compatibility, business logic consistency, and relational integrity. Data type compatibility verifies the feasibility of type conversion between the source and target fields; business logic consistency verifies the semantic rationality after matching; and relational integrity verifies the satisfaction of cross-field constraints. Matching pairs that fail verification are marked as conflicting matches and require re-selection or manual intervention.

[0093] The final alignment output includes a list of field mappings that have passed consistency verification. Each mapping records four pieces of information: source field identifier, target field identifier, matching probability, and verification status. The mappings are sorted from highest to lowest matching probability to facilitate user identification of high-quality matches. The alignment result also includes a list of unmatched fields and a list of conflicting matches, providing reference information for subsequent manual review and adjustments.

[0094] In one optional implementation, based on the alignment result and consistency strategy constraints, each candidate mapping pair is given a deterministic judgment and confidence basis under the condition of satisfying structural consistency and cross-table constraint consistency, and the mapping relationship prediction result is obtained by: obtaining the alignment result and consistency strategy constraints, wherein the alignment result includes a set of candidate mapping pairs, and the consistency strategy constraints include structural consistency constraints and cross-table constraints.

[0095] A constraint propagation network is constructed based on the set of candidate mapping pairs. The nodes of the constraint propagation network represent candidate mapping pairs, and the edges represent the association between structural consistency constraints and cross-table constraints.

[0096] The belief propagation algorithm is executed on the constraint propagation network to iteratively calculate and update the confidence state of the nodes; the confidence state is probabilistically modeled, and the Monte Carlo tree search method is used to sample the constraint-satisfied paths. Based on the sampling results, the frequency of each candidate mapping pair appearing in the effective path is counted, and the frequency is used as the confidence index for deterministic judgment.

[0097] Candidate mapping pairs are selected based on the confidence index, and mapping pairs with a confidence level higher than a preset confidence threshold and their path evidence in the constraint propagation network are output as mapping relationship prediction results.

[0098] The alignment results are obtained by parsing the candidate mapping pair set output by the semantic matching engine. Each candidate mapping pair includes a source field identifier, a target field identifier, a semantic similarity score, and a matching type tag. The source field identifier uses a fully qualified name format, including the complete path information of the database name, table name, and field name. The semantic similarity score reflects the degree of semantic matching between the source field and the target field, with the score range set between zero and one. The matching type tag distinguishes between three types: exact match, approximate match, and inferential match.

[0099] The construction of consistency policy constraints involves two levels of rule definitions: structural consistency constraints and cross-table constraints. Structural consistency constraints ensure that the mapped data maintains its original structural characteristics, including data type consistency, field length consistency, and precision consistency. Cross-table constraints ensure the integrity of relationships during multi-table mapping, including maintaining primary and foreign key relationships, referential integrity, and consistency of association paths.

[0100] The preprocessing of the candidate mapping pair set optimizes its quality through similarity threshold filtering and conflict detection mechanisms. Similarity threshold filtering removes candidate mapping pairs with semantic similarity below a minimum threshold. The conflict detection mechanism identifies and handles one-to-many and many-to-one mapping conflicts. Conflict handling employs a priority ranking mechanism, calculating a comprehensive priority score based on semantic similarity, structural matching, and business importance.

[0101] The constraint propagation network is constructed by converting candidate mapping pairs into network nodes. Each node contains a mapping pair identifier, state information, and a set of attributes. The mapping pair identifier uses a combination of source and target field identifiers to ensure the uniqueness of the node. Network edges are constructed by assigning weights based on the type and strength of the constraints. The edge weights corresponding to structural consistency constraints reflect the cost of constraint violation, while the edge weights corresponding to cross-table constraints reflect the importance of the association.

[0102] The topology optimization of the constraint propagation network reduces network complexity through clustering analysis and graph partitioning techniques. Clustering analysis groups semantically similar candidate mapping pairs into subnetworks, while graph partitioning divides the large-scale constraint network into multiple relatively independent subgraphs. The hierarchical structure of the network facilitates the parallel execution of the constraint propagation algorithm.

[0103] The belief propagation algorithm on the constraint propagation network employs an asynchronous message passing mechanism to iteratively update node states. Message passing includes two types: confidence messages and constraint messages. Confidence messages carry deterministic information about the node's current state, while constraint messages carry feedback information about constraint satisfaction. After receiving messages from neighboring nodes, a node updates its own confidence state based on the message content and local constraints. The confidence state update uses a weighted averaging mechanism, weighting and fusing multiple received messages according to the confidence level of the sending node.

[0104] Probabilistic modeling of confidence states transforms the discrete states of nodes into continuous probability distributions. This probabilistic modeling employs a Bayesian framework, unifying the prior probability, likelihood function, and posterior probability of a node. The prior probability is calculated based on the semantic similarity of nodes and the success rate of historical mappings. The likelihood function is calculated based on constraint satisfaction and the confidence states of neighboring nodes. The posterior probability is derived through Bayesian inference, resulting in the final probability distribution of the node's state.

[0105] Monte Carlo tree search constructs a search tree structure on a probabilistic constraint propagation network. The root node corresponds to the initial state of the network, the branches correspond to different node choices, and the leaf nodes correspond to the complete mapping scheme. The expansion of the search tree employs an upper confidence boundary strategy to balance the exploration of new paths with the utilization of known good paths. Each tree node contains three statistics: the number of visits, the cumulative reward, and the average reward.

[0106] The sampling process for constraint-satisfied paths begins at the root node of the search tree, with nodes randomly selected based on their confidence probabilities. The quality assessment of the sampled paths is based on a comprehensive calculation of the satisfaction of all constraints along the path. The path quality calculation considers the different importance of hard and soft constraints; violations of hard constraints result in the path being marked as invalid, while violations of soft constraints affect the path's score but not its validity.

[0107] The frequency of candidate mapping pairs in valid paths is statistically analyzed using two methods: frequency counting and weighted statistics. Frequency counting directly counts the number of times a candidate mapping pair appears in all valid paths, while weighted statistics consider the impact of path quality on the frequency count. The statistical results are then normalized to form a confidence index for the candidate mapping pairs.

[0108] The confidence index for deterministic decisions is calculated by combining frequency statistics and the probability distribution of confidence states for a comprehensive evaluation. The calculation of the confidence index considers both the absolute and relative frequencies of the mapping pairs, as well as the stability of the mapping pairs under different constraints. The dynamic adjustment of the confidence index is based on incremental updates using new sampling results.

[0109] The pre-set confidence threshold is determined based on the success rate analysis and risk assessment of historical mapping data. Setting the confidence threshold requires balancing the accuracy and coverage of the mapping results. The adaptive threshold adjustment mechanism optimizes parameters based on feedback information from the mapping results; the threshold is lowered when the mapping success rate is lower than the expected value, and raised when the mapping error rate is higher than the tolerance value.

[0110] The candidate mapping pair screening process compares the confidence index with a preset threshold, selecting high-confidence mapping candidates as the final recommendation. The screening process also considers the mutual exclusion and dependency relationships between mapping pairs to avoid selecting conflicting mapping combinations. Path evidence extraction includes the supporting paths of the mapping pairs in the constraint propagation network, constraint satisfaction trajectories, and decision-making basis. The output format of the mapping relationship prediction results includes four parts: a list of mapping pairs, confidence scores, evidence paths, and quality assessment.

[0111] In one optional implementation, a belief propagation algorithm is executed on the constraint propagation network to iteratively calculate and update the confidence state of nodes; probabilistic modeling of the confidence state and sampling of constraint-satisfied paths using a Monte Carlo tree search method include: constructing a factor graph on the constraint propagation network, executing the belief propagation algorithm to message-pass the factor graph, with variable nodes and function nodes alternately updating probability information in each iteration, obtaining the marginal probability distribution of nodes after multiple iterations, and calculating the confidence state of nodes based on the marginal probability distribution;

[0112] A Monte Carlo search tree is constructed by probabilistically modeling the confidence state. Hierarchical sampling is performed on the Monte Carlo search tree. During the sampling process, the sampling probability is dynamically adjusted according to the confidence state of the node. Branches with confidence scores higher than the preset confidence threshold are selected for expansion first. The trajectory of node state changes and the degree of constraint satisfaction on each sampling path are recorded, and the common characteristics of high-quality paths are statistically analyzed.

[0113] The confidence state of the nodes is corrected based on the common characteristics, and the corrected state is output as the optimization result.

[0114] The factor graph construction on the constraint propagation network categorizes network nodes into two types: variable nodes and function nodes. Variable nodes represent the value states of medical data fields, while function nodes represent the constraint relationships between fields. During factor graph construction, each medical data field corresponds to a variable node, and the field values ​​constitute the node's state space. Mapping constraints, semantic constraints, and structural constraints between fields correspond to different types of function nodes. The connection between variable nodes and function nodes is determined based on the scope of the constraint relationship. When a constraint involves multiple fields, the corresponding function nodes establish connections with the variable nodes corresponding to those fields. The factor graph's topology maintains sparsity to avoid increased computational complexity due to over-connection, and reduces direct dependencies between distant nodes through localized constraint design.

[0115] The message passing mechanism of the belief propagation algorithm establishes a bidirectional information flow between variable nodes and function nodes. Messages passed from a variable node to a function node contain the probability distribution information of the variable's current state, reflecting the likelihood of the variable taking different values. Messages passed from a function node to a variable node contain information about the impact of constraints on the variable's values, reflecting the correction effect of constraint satisfaction on the variable's state probability. Message passing employs an asynchronous update mechanism: variable nodes update their own state probabilities after receiving messages from all neighboring function nodes, while function nodes calculate constraint satisfaction and generate feedback messages after receiving messages from all neighboring variable nodes.

[0116] During iterative computation, the probability updates of variable nodes are based on Bayesian inference principles. Variable nodes use received function node messages as observational evidence and combine them with the node's prior probability distribution to calculate the posterior probability distribution. Probability updates employ normalization to ensure the validity of the probability distribution; the normalization operation adjusts the sum of probabilities for all states to one. Function node message generation is based on the definition of constraint functions, which map the state combinations of input variables to constraint satisfaction values. Higher satisfaction values ​​indicate that the constraints are more easily satisfied. The convergence of message propagation is determined by monitoring the magnitude of message changes between consecutive iterations; iteration stops when the magnitude of change falls below a preset convergence threshold.

[0117] The calculation of the marginal probability distribution integrates message information from multiple iterations into a final probabilistic representation of the node state. The marginal probability distribution reflects the possible distribution of each state of the variable node under all constraints. During the calculation, a probability product operation is performed on all function node messages received by the variable node, and the product result is normalized to obtain the marginal probability distribution. The quality of the marginal probability distribution is evaluated using the entropy index. A lower entropy value indicates a higher probability distribution concentration and less uncertainty in the node state, while a higher entropy value indicates a more even probability distribution and greater uncertainty in the node state.

[0118] The calculation of node confidence states is quantified based on the statistical characteristics of marginal probability distributions. The confidence state uses the probability value of the most probable state as the primary indicator, reflecting the degree of certainty of the node's state. The confidence state also considers the concentration of the probability distribution, measuring the certainty of the state choice by calculating the ratio of the highest probability to the second-highest probability; a larger ratio indicates a more certain state choice and higher confidence. The dynamic range of the confidence state is set between zero and one, where zero represents complete uncertainty and one represents complete certainty. The calculation of confidence states also incorporates a time decay factor, where the contribution of probability information from earlier iterations gradually decreases to the final confidence state, ensuring that the confidence state reflects the latest inference results.

[0119] The Monte Carlo search tree is constructed based on probabilistic modeling of confidence states for tree structure initialization. The root node of the search tree corresponds to the initial state of the current medical data field mapping problem, the branches of the tree correspond to different field value choices, and the depth of the tree corresponds to the length of the complete path for the field mapping decision. Tree nodes contain three types of data: state information, confidence information, and access statistics. State information records the field value combination corresponding to the current node, confidence information records the credibility of the state combination, and access statistics record the number of times the node is accessed and the cumulative benefit during the search process. The search tree employs a lazy expansion strategy, creating child nodes only when a node is accessed, avoiding the memory overhead caused by pre-constructing the complete tree structure.

[0120] The hierarchical sampling mechanism explores paths on the Monte Carlo search tree in a depth-first manner. The sampling process starts from the root node, calculating the selection probability at each node based on the confidence state of its child nodes. Child nodes with higher confidence levels receive a greater selection probability. The selection probability calculation combines confidence values ​​and exploration rewards, which encourage nodes with fewer visits to maintain search diversity. During the generation of the sampling path, the state transitions and constraint satisfaction of each node are recorded. Constraint satisfaction is determined by checking whether the current state combination violates preset constraints. Sampling termination conditions include reaching a leaf node, violating critical constraints, or reaching the maximum search depth.

[0121] The dynamic sampling probability adjustment mechanism updates the node selection strategy in real time based on feedback information obtained during the search process. This mechanism monitors the quality metrics of the sampled paths, including constraint satisfaction rate, path completeness, and goal achievement rate. The constraint satisfaction rate is calculated as the ratio of constraint violations on the path to the total number of constraint checks; a higher satisfaction rate indicates better path quality. Path completeness is calculated by comparing the actual path length with the expected path length; a higher completeness indicates a deeper search. Goal achievement rate is calculated by evaluating the degree of matching between the path's endpoint state and the target state; a higher achievement rate indicates better search performance.

[0122] The preset confidence threshold is adaptively adjusted based on statistical analysis of historical search data. This threshold serves as a filtering condition for branch selection; only branches with a confidence level exceeding the threshold are prioritized for expansion. The initial threshold value is estimated based on problem complexity and data characteristics. Higher complexity problems use lower initial thresholds to ensure search coverage, while lower complexity problems use higher initial thresholds to improve search efficiency. The threshold is dynamically adjusted based on the search success rate. When the search success rate is lower than expected, the threshold is lowered to increase the exploration scope; when the search success rate exceeds expected, the threshold is raised to focus on high-quality paths.

[0123] The path change trajectory record includes three types of information: node state sequence, state transition reason, and transition time. The node state sequence records the complete state information of each node on the sampling path, including field values, constraint satisfaction status, and confidence scores. The state transition reason records the specific constraints or decision rules that lead to the state change, helping to analyze the logical basis of state evolution. The transition time records the relative position of the state change during the search process, used to analyze the temporal characteristics of state evolution. The trajectory record uses a compressed storage format to reduce memory usage, retaining only detailed information of key state transition points, while intermediate stable states are represented in a simplified manner.

[0124] The statistical evaluation of constraint satisfaction employs a multi-dimensional assessment mechanism to comprehensively evaluate path quality. The assessment dimensions include hard constraint satisfaction, soft constraint satisfaction, and constraint conflict. Hard constraint satisfaction is calculated by checking whether the path violates mandatory constraints; paths that violate hard constraints are marked as invalid. Soft constraint satisfaction is calculated by assessing the path's compliance with optional constraints; the satisfaction of soft constraints affects the path's quality score but not its effectiveness. Constraint conflict is calculated by analyzing the contradictions between different constraints on the path; paths with high conflict levels indicate inconsistencies in the constraint design.

[0125] The extraction of common features of high-quality paths is achieved through pattern mining of successful paths. Feature extraction includes analysis at three levels: state patterns, transition patterns, and constraint patterns. State pattern analysis identifies frequently occurring node state combinations on successful paths, recognizing key state configurations conducive to search success. Transition pattern analysis identifies common state transition sequences on successful paths, recognizing effective decision chains and state evolution patterns. Constraint pattern analysis analyzes how successful paths handle different constraints, identifying constraint priorities and constraint coordination strategies. The feature extraction results form a rule base, providing heuristic guidance for subsequent searches.

[0126] The confidence state correction mechanism uses extracted common features to fine-tune the confidence of nodes. The correction process calculates the matching degree between the node's current confidence state and high-quality path features. Nodes with high matching degrees receive a confidence increase, while nodes with low matching degrees face a confidence decrease. The correction magnitude is determined based on the significance of feature matching; features with higher significance have a greater impact on confidence. The corrected confidence state is output as the optimization result, providing more accurate prior information for the next round of search, forming a closed-loop mechanism for iterative optimization.

[0127] In one optional implementation, performing field mapping on the target medical data object based on the mapping relationship prediction result to generate backfill feedback information, and using the backfill feedback information to adaptively update the result of the causal state configuration and the consistency strategy constraints includes: performing mapping transformation on the fields in the target medical data object according to the mapping relationship prediction result, recording the mapping success identifier and data quality index of each field; constructing a backfill feedback vector based on the mapping success identifier and the data quality index, performing time series analysis on the backfill feedback vector, identifying the trend of mapping effect over time, and generating a feedback feature matrix;

[0128] The causal state configuration is updated using the feedback feature matrix. By adjusting the field dependency strength and influence direction, the association structure between fields is optimized. Based on the rule compliance in the feedback feature matrix, the threshold parameters and constraint weights in the consistency strategy constraints are dynamically adjusted to improve the adaptability of the constraints to the data features.

[0129] When performing field mapping transformation on the target medical data object based on the mapping relationship prediction results, a field mapping execution engine is established. This engine receives the mapping relationship prediction results as input and transforms the source field data into the target field structure according to the predicted mapping relationship. The mapping execution engine establishes an independent transformation channel for each field. For numeric fields, a data type conversion mechanism is used; for text fields, a semantic alignment conversion mechanism is used; and for enumerated fields, a value range mapping conversion mechanism is used. During the mapping transformation process, the engine monitors the transformation status of each field in real time. When a field transformation is successful, a positive value is recorded as a success marker; when a field transformation fails, a negative value is recorded as a success marker; and when a field is partially transformed, an intermediate value is recorded as a success marker.

[0130] The construction of data quality indicators encompasses three dimensions: completeness, accuracy, and consistency. Completeness is measured by the proportion of non-empty fields after transformation, calculated as the ratio of valid data entries in the target field to the total number of data entries. Accuracy is measured by comparing the semantic consistency of data before and after transformation, establishing a semantic similarity assessment mechanism to analyze the semantic matching degree between the source field data before transformation and the target field data after transformation. Consistency is measured by verifying whether the transformed data meets the constraints of the target data model, including validation of data format constraints, value range constraints, and association constraints.

[0131] The construction of the backfill feedback vector integrates the mapping success identifier and data quality indicators into a unified feedback representation. The feedback vector includes mapping success, completeness, accuracy, and consistency dimensions. Each dimension undergoes standardization to normalize its numerical range to the zero-to-one interval. The mapping success dimension reflects the transformation effect at the field level, the completeness dimension reflects the data coverage, the accuracy dimension reflects the degree of semantic preservation, and the consistency dimension reflects the degree of constraint satisfaction. The feedback vector uses a weighted fusion method, assigning weight coefficients based on the characteristics of different medical data types: clinical diagnostic data emphasizes accuracy, patient basic information data emphasizes completeness, and laboratory test data emphasizes consistency.

[0132] Time series analysis identifies trends in the historical records of the backfilled feedback vector and establishes a sliding window mechanism to capture the temporal variation patterns of the mapping effect. The sliding window size is dynamically adjusted according to the data update frequency; a smaller window is used for frequently updated data to capture short-term changes, while a larger window is used for infrequently updated data to identify long-term trends. Time series analysis calculates the rate of change for each dimension of the feedback vector, identifying three basic patterns: upward trend, downward trend, and stationary trend. The rate of change is measured by the difference between the current time point value and the historical time point value, and combined with the direction and magnitude of change to form trend characteristics.

[0133] The generation of the feedback feature matrix transforms the time-series analysis results into a structured feature representation. Matrix rows correspond to different field mapping pairs, and matrix columns correspond to different feedback feature dimensions. Feature dimensions include current mapping effect, historical average effect, trend direction, magnitude of change, and volatility stability. The current mapping effect is represented by the mean of the feedback vectors within the most recent time window; the historical average effect is represented by the mean of the feedback vectors from all historical data; the trend direction is represented by the sign of the linear fitting slope; the magnitude of change is represented by the absolute value of the linear fitting slope; and volatility stability is represented by the variance of the feedback vectors.

[0134] The update of the causal state configuration utilizes the feedback feature matrix to adjust the strength of dependencies between fields. The dependency strength adjustment mechanism makes directional corrections based on the positive or negative nature of the feedback effect. When the field mapping effect is good, the dependency strength between related fields is strengthened; when the field mapping effect is poor, the dependency strength between related fields is weakened. The numerical update of dependency strength uses an adaptive learning rate, the magnitude of which is proportional to the magnitude of change in the feedback effect. A larger magnitude of change indicates a larger adjustment space, and the learning rate is increased accordingly to accelerate convergence. The adjustment of the influence direction is optimized based on the causal logic between fields. When the mapping effect of the source field to the target field remains good, the positive influence relationship is strengthened; when the mapping effect remains poor, a negative suppression relationship is introduced to avoid error propagation.

[0135] The optimization of the inter-field relationship structure is achieved by dynamically adjusting the connection weights of the causal graph. The update of connection weights considers both direct and indirect associations. Direct associations are adjusted based on the mapping feedback between field pairs, while indirect associations are adjusted based on the influence of field pairs transmitted through intermediate fields. The relationship structure optimization employs a local search strategy, prioritizing adjustments to field connections with the most significant feedback changes to avoid structural instability caused by simultaneous global adjustments. During the optimization process, the directed acyclic property of the causal graph is maintained, and the structural rationality is verified through topological sorting to prevent the formation of circular dependencies.

[0136] The dynamic adjustment of consistency strategy constraints is based on parameter optimization using the rule compliance index in the feedback feature matrix. Rule compliance is measured by the proportion of data that satisfies various constraints after statistical transformation, including data type constraint compliance, value range constraint compliance, and association constraint compliance. Threshold parameter adjustment employs a feedback-driven adaptive mechanism: when constraints are too strict, causing a large amount of data to fail validation, the threshold is appropriately relaxed; when constraints are too lenient, leading to a decline in data quality, the threshold is appropriately tightened. Constraint weights are adjusted based on the contribution of different constraint types to the mapping effect. The contribution is measured by the correlation between constraint satisfaction and the final mapping quality; constraint types with higher correlation are assigned greater weight.

[0137] The improved adaptability of constraints to data features is achieved through a self-learning mechanism for constraint rules. This mechanism analyzes the distribution characteristics of historical mapping data, identifies inherent patterns and anomalies, and dynamically generates new constraint rules or modifies existing ones based on the identification results. Constraint rule generation employs rule mining techniques, extracting common features from successful mapping cases as positive constraints and extracting exclusionary features from failed mapping cases as negative constraints. Constraint rule validation utilizes a cross-validation mechanism, dividing historical data into training and test sets. Constraint rules are generated on the training set and validated on the test set to ensure the newly generated constraints have good generalization ability.

[0138] A second aspect of the present invention provides an electronic device, comprising: a processor;

[0139] Memory used to store processor-executable instructions;

[0140] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0141] A third 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.

[0142] 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.

[0143] 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 predicting mapping relationships of medical data fields based on world model learning, characterized in that, include: Obtain the target medical data object, which includes a set of source fields with structural identifiers and a set of target fields to be aligned. Extract the source field set and the target field set in a unified manner to form an original description containing field semantic symbols and contextual relationships. Perform causal state configuration on the original description based on an interpretable generative world model. The original description information contains multiple representation fields. These multiple representation fields are input into an interpretable generative world model, and a semantic dependency graph of the representation fields is constructed based on multi-agent collaborative reasoning. Based on the semantic dependency graph, a random walk algorithm is used to extract important causal paths of the representation fields. Based on the important causal paths, a feature mapping matrix of the representation fields is constructed. Based on the feature mapping matrix, the original description information is reconstructed. The reconstructed state space is clustered according to the semantic similarity of the fields to obtain multiple semantic subspaces. Local and global constraints are set for each semantic subspace. The state is iteratively updated based on the local and global constraints. In each iteration, the state that violates the constraints is adjusted by gradient descent until the multiple semantic subspaces meet the preset cross-table consistency threshold, and an intermediate world state is obtained. The intermediate world state includes a source field set, a target field set, a structural association identifier, and a semantic constraint identifier. The structural information of each field in the source field set is extracted to generate a source field numerical vector. The structural information of each field in the target field set is extracted to generate a target field numerical vector. Calculate the cosine distance between the source field value vector and the target field value vector to obtain an initial similarity matrix, and construct a field association graph based on the initial similarity matrix; use a graph embedding method to transform the field association graph into a low-dimensional representation space, execute a random walk algorithm in the low-dimensional representation space to generate a walk sequence between fields, and calculate the transition probability between field pairs based on the walk sequence; Semantic rules for the source and target fields are extracted based on the semantic constraint identifiers. The semantic compatibility of the field pairs is calculated based on the semantic rules. The transition probability and the semantic compatibility are weighted and combined to obtain the matching probability of the field pairs. A greedy search strategy is used to perform iterative matching of the fields based on the matching probability. The consistency of the results of the iterative matching of the fields is verified, and the field mapping relationship that passes the consistency verification is output as the alignment result. Based on the alignment results and consistency strategy constraints, each candidate mapping pair is given a deterministic judgment and confidence basis under the conditions of satisfying structural consistency and cross-table constraint consistency, and the mapping relationship prediction result is obtained. Based on the prediction results of the mapping relationship, field mapping is performed on the target medical data object to generate backfill feedback information. The backfill feedback information is used to adaptively update the results of the causal state configuration and the consistency strategy constraints. The updated causal state configuration results and consistency strategy constraints are used for a new round of alignment inference and decision generation, forming a closed-loop optimization driven by the backfill feedback information.

2. The method according to claim 1, characterized in that, Based on the semantic dependency graph, a random walk algorithm is used to extract important causal paths of the representation fields. A feature mapping matrix of the representation fields is constructed based on the important causal paths. The state reconstruction of the original description information based on the feature mapping matrix includes: performing a random walk process based on the semantic dependency graph, setting the transition probability at each node according to the outgoing edge weight, generating multiple walk paths through Monte Carlo sampling, calculating the cumulative value of information gain between nodes for each path, and marking the path as an important causal path when the cumulative value exceeds a preset information threshold. Node sequences are extracted from the important causal path, and the transition probabilities between adjacent node pairs in the node sequence are calculated to generate a node transition matrix. For each source node in the node transition matrix, the correlation score between it and all target nodes is calculated. The correlation score is normalized to obtain a weight coefficient. The weight coefficient is used to linearly combine the original representation fields to generate a new state vector. Based on the state vector, a state mapping rule is constructed, and the original description information is reconstructed according to the state mapping rule to obtain a reconstructed state containing causal dependencies.

3. The method according to claim 1, characterized in that, Based on the alignment results and consistency policy constraints, each candidate mapping pair is given a deterministic judgment and confidence basis under the condition of satisfying structural consistency and cross-table constraint consistency. The mapping relationship prediction results are obtained by: obtaining the alignment results and consistency policy constraints, wherein the alignment results include a set of candidate mapping pairs, and the consistency policy constraints include structural consistency constraints and cross-table constraints. A constraint propagation network is constructed based on the set of candidate mapping pairs. The nodes of the constraint propagation network represent candidate mapping pairs, and the edges represent the association between structural consistency constraints and cross-table constraints. The belief propagation algorithm is executed on the constraint propagation network to iteratively calculate and update the confidence state of the nodes; the confidence state is probabilistically modeled, and the Monte Carlo tree search method is used to sample the constraint-satisfied paths. Based on the sampling results, the frequency of each candidate mapping pair appearing in the effective path is counted, and the frequency is used as the confidence index for deterministic judgment. Candidate mapping pairs are selected based on the confidence index, and mapping pairs with a confidence level higher than a preset confidence threshold and their path evidence in the constraint propagation network are output as mapping relationship prediction results.

4. The method according to claim 3, characterized in that, The belief propagation algorithm is executed on the constraint propagation network to iteratively calculate and update the confidence state of the nodes; probabilistic modeling of the confidence state is performed, and the Monte Carlo tree search method is used to sample the constraint satisfaction path, including: constructing a factor graph on the constraint propagation network, executing the belief propagation algorithm to message the factor graph, updating the probability information of variable nodes and function nodes alternately in each iteration, obtaining the marginal probability distribution of the nodes after multiple iterations, and calculating the confidence state of the nodes based on the marginal probability distribution; A Monte Carlo search tree is constructed by probabilistically modeling the confidence state. Hierarchical sampling is performed on the Monte Carlo search tree. During the sampling process, the sampling probability is dynamically adjusted according to the confidence state of the node. Branches with confidence scores higher than the preset confidence threshold are selected for expansion first. The trajectory of node state changes and the degree of constraint satisfaction on each sampling path are recorded, and the common characteristics of high-quality paths are statistically analyzed. The confidence state of the nodes is corrected based on the common characteristics, and the corrected state is output as the optimization result.

5. The method according to claim 1, characterized in that, Based on the mapping relationship prediction results, field mapping is performed on the target medical data object to generate backfill feedback information. The adaptive update of the causal state configuration result and the consistency strategy constraints using the backfill feedback information includes: performing mapping transformation on the fields in the target medical data object according to the mapping relationship prediction results, recording the mapping success identifier and data quality index of each field; constructing a backfill feedback vector based on the mapping success identifier and the data quality index; performing time series analysis on the backfill feedback vector to identify the changing trend of the mapping effect over time, and generating a feedback feature matrix. The causal state configuration is updated using the feedback feature matrix. By adjusting the field dependency strength and influence direction, the association structure between fields is optimized. Based on the rule compliance in the feedback feature matrix, the threshold parameters and constraint weights in the consistency strategy constraints are dynamically adjusted to improve the adaptability of the constraints to the data features.

6. A medical data field mapping relationship prediction system based on world model learning, used to implement the method of any one of claims 1-5, characterized in that, include: The first unit is used to acquire the target medical data object, which includes a set of source fields with structural identifiers and a set of target fields to be aligned. The source field set and the target field set are uniformly extracted to form an original description containing field semantic symbols and contextual relationships. The second unit is used to perform causal state configuration on the original description based on an interpretable generative world model, and output intermediate world states that represent the semantic causal dependencies and cross-table constraints of the fields; based on the structural association identifiers and semantic constraint identifiers in the intermediate world states, the source field set and the target field set are aligned and inferred to generate alignment results. The third unit is used to provide a deterministic judgment and confidence basis for each candidate mapping pair based on the alignment result and consistency strategy constraints, under the condition of satisfying structural consistency and cross-table constraint consistency, and to obtain the mapping relationship prediction result. The fourth unit is used to perform field mapping on the target medical data object based on the prediction results of the mapping relationship, generate backfill feedback information, use the backfill feedback information to adaptively update the results of the causal state configuration and the consistency strategy constraints, and use the updated causal state configuration results and consistency strategy constraints for a new round of alignment inference and decision generation, forming a closed-loop optimization driven by the backfill feedback information.

7. 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 5.

8. 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 5.

Images