A medical scientific research intelligent topic selection method and system based on multi-mode data analysis

Abstract

This invention provides a method and system for intelligent medical research topic selection based on multimodal data analysis. The invention extracts and standardizes subdivided research directions using a hierarchical departmental navigation tree, and combines UMLS and MeSH medical ontologies to assign semantic, weighted, and policy-oriented attributes to knowledge nodes, organizing them into a structured domain knowledge ontology. After fusing multi-source heterogeneous data, a hierarchical causal graph is constructed to automatically identify three types of causal edges: macro-policy, meso-academic collaboration, and micro-clinical evidence, ensuring cross-layer semantic consistency. Through multi-hop causal propagation driven by multi-attribute indicators, attribute activation paths are obtained based on dimensions of innovation, clinical need, technological frontier, and impact. Path consistency is evaluated using semantic aggregation degree, and finally, back-attribution is applied to feasibility constraints, forming a topic ranking with a traceable evidence chain. The hierarchical causal graph construction and multi-attribute independent propagation mechanism of this invention significantly improve the scientific rigor and personalized matching accuracy of research topic generation.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence knowledge reasoning and multi-attribute medical topic recommendation technology, and in particular to a method and system for intelligent selection of medical research topics based on multi-modal data analysis. Background Technology

[0002] Currently, intelligent topic selection and recommendation systems for medical research projects are constantly evolving, particularly in areas such as multi-modal data fusion, multi-attribute decision models, and the construction of medical knowledge ontology, where significant achievements have been made. Mainstream methods for evaluating and recommending medical research projects generally employ multi-attribute scoring models, combined with bibliometric analysis, clinical needs assessment, and cutting-edge technology mining. Through comprehensive scoring or weighted ranking, these models provide users with priority suggestions for proposed research directions. Commonly used implementation methods in the industry include rule engines based on expert experience knowledge bases, deep learning models employing attention weighting mechanisms, and recommendation algorithms integrating social evaluation systems. These technologies generally rely on multi-dimensional statistical analysis of historical data, expert opinion annotation, and deep modeling of user behavior feedback data, continuously promoting intelligent management of medical research information, enhancing information transparency, and improving the scientific rigor of decision-making.

[0003] However, the interpretability and causal traceability of existing recommendation systems still have significant limitations. Most mainstream technologies employ top-down scoring mechanisms or black-box machine learning models. While some systems introduce interpretive modules such as feature importance descriptions and weight visualizations, their outputs primarily remain at the level of "which attributes score highly and which indicators receive attention," failing to reveal the causal chains and logical flow between core evaluation factors. For example, ranking methods based on neural network attention mechanisms typically only provide the importance weights of input features or local sensitivity analysis, making it difficult to present the causal order and structured reasoning process between attributes. Meanwhile, experience models based on expert rules focus on static constraints and matching, lacking flexible causal chain modeling and traceable evidence. Furthermore, the interpretability of many current systems still heavily relies on user preference modeling or historical behavior feedback; without sufficient user interaction data, the transparency of the recommendation logic and the reliability of the system are affected.

[0004] Some existing publicly available solutions have the following shortcomings: First, the recommended output is usually presented in the form of a weighted comprehensive score or a score ranking form, which fails to provide users with a causal chain explanation, making it impossible to trace the logical deduction path of the final decision suggestion, and only remaining at the explanation of the result rather than the explanation of the process.

[0005] Secondly, mainstream models focus on attribute correlation or weight attribution, failing to combine medical knowledge ontology (such as structured semantic networks like UMLS and MeSH) with hierarchical causal reasoning mechanisms. This makes it difficult to fully utilize the rich professional terminology system and clinical logic framework in the medical field to achieve hierarchical and multi-granular causal path modeling for the multiple decision factors of topic recommendations.

[0006] Furthermore, existing explanation methods are susceptible to problems such as missing data dimensions and user feedback bias. They typically only address single attributes or fragmented explanations, lacking the ability to support multi-attribute collaborative decision-making and cross-dimensional causal tracing. This limits the flexibility and reliability of recommendation systems in diverse medical research application scenarios.

[0007] Furthermore, while some counterfactual explanations and attention heatmaps improve model interpretability, they all involve structural perturbations to the model itself and artificial construction of data distributions, posing significant challenges to the rigor of medical reasoning and the verifiability of results. For highly specialized scenarios requiring the tracing of medical facts, policy basis, and multi-source heterogeneous evidence, these methods struggle to meet the demands for high reliability and full observability.

[0008] Therefore, there is an urgent need for a new recommendation explanation mechanism that is free from user preference dependence, does not resort to complex neural hidden variables, and does not rely on external social trust networks, and can rely on multi-level medical knowledge ontology and Structured CausalGraph technology. Summary of the Invention

[0009] This invention provides a method and system for intelligent selection of medical research topics based on multimodal data analysis, aiming to solve the problems of the existing technology mentioned in the background section.

[0010] In a first aspect, the present invention provides a method for intelligent selection of medical research topics based on multimodal data analysis, comprising the following steps: S1: Based on the hierarchical departmental navigation tree already constructed in the intelligent medical research topic selection system, extract the subdivided research direction nodes under the target research field and construct a structured domain knowledge node set; S2: Construct a hierarchical causal graph containing macro, meso, and micro layers based on structured domain knowledge node sets, clinical datasets, funding project databases, and patent intelligence databases; S3: Map the evaluation index in the multi-attribute decision modeling module to the initial excitation signal of the corresponding subgraph, and input it into the corresponding subgraph structure in the hierarchical causal graph to trigger the multi-hop causal propagation process in each attribute dimension and generate the attribute activation path sequence. S4: For each attribute activation path sequence, identify the endpoint node that reaches the maximum activation depth during the propagation process, and calculate the semantic aggregation degree value between each path based on the preset path semantic consistency function. S5: Based on the semantic aggregation degree value, integrate attribute activation paths to generate a comprehensive causal support score for candidate topics. Then, through the path back projection algorithm, map the comprehensive causal support score back to the original feasibility dimension space in the constraint filter to obtain a causal attribution mapping result with traceable intervals. S6: Based on the causal attribution mapping results, generate ranking criteria texts corresponding to each candidate topic, forming a structured causal explanation evidence package; S7: Synchronize and associate the causal explanation evidence package with the original topic ranking list output by the multi-attribute decision model to construct an aligned topic-explanation data structure.

[0011] Preferably, in step S1, the navigation tree organizes clinical departments and their subordinate research directions in a tree structure, wherein each non-leaf node represents a clinical department and each leaf node represents a specific research direction branch.

[0012] Preferably, in step S2, based on the structured domain knowledge node set, subdivided research direction nodes under the target research domain are obtained as semantic anchors; the semantic anchors are used to match the epidemiological trends, diagnosis and treatment behavior sequences and patient outcome indicators of the corresponding diseases in real-world clinical datasets, retrieve the relevant national key research and development program project topics in the funding project database, extract the high-frequency citation paths of related technical fields in the patent intelligence database, and perform timestamp alignment and entity normalization processing to generate a cross-domain aligned original evidence tuple set; A three-layer causal edge identification algorithm is executed on the original evidence tuple set to obtain a set of macroscopic layer strongly constrained causal edges, a set of mesoscopic layer weakly supervised causal edges, and a set of microscopic layer probabilistic causal edges, and outputs a weighted directed edge list of the three types of causal edges. Inject the weighted directed edge lists of the three types of causal edges into the corresponding layer spaces to generate an initial version of the hierarchical causal graph. A consistency check and topology optimization are performed on the initial version of the hierarchical causal graph with a multi-level semantic structure to obtain a stable version of the hierarchical causal graph.

[0013] Preferably, in step S2, strong constraint causal edges from the macro level are injected into the macro level graph structure to form a top-level causal framework driven by policy guidance and clinical norms; weak supervision causal edges from the meso level are injected into the meso level graph structure to construct a knowledge evolution network driven by academic cooperation and technological evolution; probabilistic causal edges from the micro level are injected into the micro level graph structure to establish an individualized response map driven by real-world efficacy feedback; and cross-layer semantic bridging is achieved between layers through shared UMLS concept nodes.

[0014] Preferably, in step S3, based on a preset semantic mapping rule table, the evaluation indicators such as the innovation level score, the clinical need urgency score, the technological frontier index value, and the expected impact prediction value are bound to the corresponding causal propagation subgraph entry node to generate an initial incentive signal vector. The initial excitation signal vector is processed by a multi-hop causal propagation algorithm. When a node is activated, it inherits the excitation source identifier of the upstream node and adds the UMLS semantic relation type of the current edge to generate an intermediate activation state with a source tracing mark. This process is continued until the maximum propagation depth is reached or the activation energy is lower than a preset threshold. For each completed multi-hop propagation process, the complete node sequence from the initial excitation node to the final activation endpoint is extracted to form an attribute activation path. UMLS semantic type annotation is parsed for all intermediate nodes in the path to generate an attribute activation path sequence.

[0015] Preferably, in step S4, the UMLS semantic annotation node sequence arranged in the multi-hop causal propagation order in each path is traversed and the activation intensity numerical sequence of each node in the propagation process is extracted. The activation intensity is generated by accumulating the product of the edge weight and the node confidence in the hierarchical causal graph after being corrected by the time decay factor, so as to obtain the dynamic activation trajectory data on each path. For each attribute activation path, an extreme value detection operation is performed on the dynamic activation trajectory data. The node whose activation intensity reaches the peak and meets the persistence threshold condition is identified as the maximum activation depth endpoint node of the path. If there are multiple candidate nodes, the node that is farthest from the initial excitation signal and meets the significance test is selected as the final endpoint node. Based on the endpoint node and its path context information, a set of input parameters for the path semantic consistency function is constructed, and the semantic similarity matrix between each pair of paths is calculated using the standardized parameter vector; The semantic similarity matrix is ​​input into a predefined path semantic consistency function to generate semantic consistency scores between pairs of attribute-activated paths. The mean of elements outside the main diagonal is aggregated, and the average consistency score of all non-self-comparison path pairs is calculated as the final semantic aggregation degree value.

[0016] Preferably, in step S4, the set of input parameters includes the MeSH tree structure path code corresponding to each path endpoint node, the UMLS semantic type vector, and the co-occurrence document frequency statistics.

[0017] Preferably, in step S4, the path semantic consistency function is based on a linear weighted fusion model to fuse the MeSH tree structure distance score and the UMLS semantic type matching score. The weight coefficients are trained based on the consistency rating results of historical expert-annotated path pairs to generate semantic consistency scores between pairs of attribute-activated paths.

[0018] Preferably, in step S5, the semantic aggregation degree value calculated based on the attribute activation path sequence with semantic labels and its corresponding path semantic consistency function is used to integrate each activation path across dimensions using a weighted fusion strategy to obtain an initial score. Based on the initial score execution path back projection algorithm, a back propagation rule is established according to the directed causal edges between nodes in the hierarchical causal graph. The algorithm backtracks from the final activated node along the causal chain to the root node of the original feasibility dimension defined by the constraint filter, forming a preliminary attribution heat map. Threshold segmentation and interval quantization are performed on the attribution heatmap. A dynamic clustering algorithm is used to identify the value intervals of significant contributions on each feasibility dimension. The contribution intensity obtained by back projection is then mapped to the most matching discrete interval to generate attribution interval identifiers. Each attribution interval identifier is bound to its corresponding data source path to construct a structured causal attribution mapping table; the structured causal attribution mapping table is output as the causal attribution mapping result.

[0019] Secondly, the present invention provides a medical research intelligent topic selection system based on multimodal data analysis, wherein the system utilizes the above-mentioned method to achieve intelligent medical research topic selection.

[0020] The beneficial technical effects of this invention are as follows: 1. The hierarchical causal graph construction and multi-attribute independent propagation mechanism of the present invention significantly improve the scientific nature and personalized matching accuracy of scientific research proposition generation; the present invention realizes the structured expression of medical knowledge from the institutional layer to the data layer by constructing a three-layer hierarchical causal graph that includes macro-level normative constraints, meso-level knowledge evolution and micro-level empirical support. 2. This invention introduces a reversible attribution path tracing mechanism and a causal path folding panel output mode, which greatly improves the transparency of the system output and the clinical adoption intention. By performing semantic consistency analysis on the node with the maximum activation depth in each attribute activation path, using UMLS semantic type matching and MeSH tree structure distance weighting to calculate the path aggregation degree, and mapping it back to the original five-dimensional resource constraint space through a back projection algorithm, a closed-loop mapping from high-order causal reasoning to specific resource allocation suggestions is realized, generating ranking basis text with clear causal anchors. Attached Figure Description

[0021] Figure 1This is a flowchart illustrating the method of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the process for generating attribute activation path sequences in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of generating semantic aggregation degree values ​​for embodiments of the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown in the figure, this embodiment provides a method for intelligent selection of medical research topics based on multimodal data analysis, which specifically includes the following steps: S1: Based on the hierarchical departmental navigation tree already constructed in the intelligent medical research topic selection system, extract the subdivided research direction nodes under the target research field and construct a structured domain knowledge node set; S2: Construct a hierarchical causal graph containing macro, meso, and micro layers based on structured domain knowledge node sets, clinical datasets, funding project databases, and patent intelligence databases; S3: Map the evaluation index in the multi-attribute decision modeling module to the initial excitation signal of the corresponding subgraph, and input it into the corresponding subgraph structure in the hierarchical causal graph to trigger the multi-hop causal propagation process in each attribute dimension and generate the attribute activation path sequence. S4: For each attribute activation path sequence, identify the endpoint node that reaches the maximum activation depth during the propagation process, and calculate the semantic aggregation degree value between each path based on the preset path semantic consistency function. S5: Based on the semantic aggregation degree value, integrate attribute activation paths to generate a comprehensive causal support score for candidate topics. Then, through the path back projection algorithm, map the comprehensive causal support score back to the original feasibility dimension space in the constraint filter to obtain a causal attribution mapping result with traceable intervals. S6: Based on the causal attribution mapping results, generate ranking criteria texts corresponding to each candidate topic, forming a structured causal explanation evidence package; S7: Synchronize and associate the causal explanation evidence package with the original topic ranking list output by the multi-attribute decision model to construct an aligned topic-explanation data structure; S8: When a user accesses the topic sorting suggestions, the topic recommendation display format with a traceable logical chain is automatically loaded. If the user expands the causal path collapse panel of a certain topic, a visual causal chain graph containing the complete path chain, evidence sources and attribution weights is output.

[0025] In this embodiment, step S1 involves extracting subdivided research direction nodes under the target research field based on the hierarchical departmental navigation tree already constructed in the intelligent medical research topic selection system, and constructing a structured domain knowledge node set; including the following steps: S1.1: Obtain the predefined hierarchical departmental navigation tree in the intelligent medical research topic selection system. This navigation tree organizes clinical departments and their subordinate research directions in a tree structure. Each non-leaf node represents a clinical department (e.g., 'Endocrinology Department'), and each leaf node represents a specific research direction branch (e.g., 'Diabetic Foot Microcirculation Disorder'). Based on this tree structure, perform a depth-first traversal operation to identify and extract the semantic labels of the subdivided research directions corresponding to all leaf nodes, and generate an initial set of research direction nodes. S1.2: Perform standardized mapping processing on each semantic label in the initial research direction node set, use the metalexicon of the Unified Medical Language System (UMLS) as the basis for terminology standardization, convert non-standard research direction names into standard medical concepts with unique concept identifiers (CUI), ensure that each research direction node has cross-data source consistency and comparability at the semantic level, and output the standard research direction concept set after terminology standardization. S1.3: Based on each UMLS-annotated concept node in the standard research direction concept set, retrieve its corresponding parent-level topic path (Tree Number) from the MeSH tree structure, and combine the priority development direction tags in the GCP guidelines and the National Key Clinical Specialty Construction Catalog to add domain importance weight factors and policy orientation attributes to each research direction node, generating an enhanced research direction node set with semantic hierarchy and strategic value annotations; S1.4: Perform structured coding processing on the enhanced research direction node set, generate unique node identifiers using a three-level coding rule of department code-research direction type-policy level, and organize them into a knowledge representation structure in JSON-LD format, which includes fields such as node name, CUI identifier, MeSH path, policy label and weight value, forming a structured domain knowledge node set that can be called by subsequent modules.

[0026] S1.5: Persistently store the structured domain knowledge node set in the system's built-in domain ontology knowledge base, and simultaneously establish an inverted index to support rapid retrieval based on keywords, departmental classifications, or policy attributes; provide standardized input services to the subsequent module for building a hierarchical causal graph through the API interface to ensure that step S2 can carry out multi-source heterogeneous data fusion and causal edge inference based on consistent and reliable semantic anchors.

[0027] In this embodiment, step S2) involves constructing a hierarchical causal graph comprising macro, meso, and micro layers based on a structured domain knowledge node set, clinical dataset, funding project database, and patent information database. This includes the following steps: S2.1: Based on the structured domain knowledge node set output by the hierarchical departmental navigation tree already constructed in the medical research intelligent topic selection system, obtain the subdivided research direction nodes under the target research domain as semantic anchors; use these semantic anchors to match the epidemiological trends, diagnosis and treatment behavior sequences and patient outcome indicators of the corresponding diseases in real-world clinical datasets, retrieve the relevant national key research and development program project topics in the funding project database, and extract the high-frequency citation paths of related technical fields in the patent intelligence database; perform timestamp alignment and entity normalization processing on the above three types of heterogeneous data sources to uniformly map them to the standard concept identifiers under the UMLS semantic network, and generate a cross-domain aligned original evidence tuple set; S2.2: Based on the original evidence tuple set obtained through cross-domain alignment, a three-layer causal edge identification algorithm is executed: First, for GCP guidelines, clinical pathway specifications issued by the National Health Commission, and guidelines of the National Key Research and Development Program, node pairs with mandatory logical order are extracted using rule matching and keyword triggering mechanisms to generate a macro-level strongly constrained causal edge set; Second, for high-frequency co-cited research topic pairs in the bibliometric database, technical association items of joint applications of multiple projects in the funded project database, and forward and backward citation relationships on the technical citation chain in the patent information database, co-occurrence intensity threshold filtering and Jaccard similarity weighting are applied to generate a meso-level weakly supervised causal edge set; Finally, based on the Granger causality test results between centralized treatment interventions and dynamic changes in biomarkers in real-world clinical datasets, combined with the significance coefficient (p<0.05) output by the Cox proportional hazards model, a micro-level probabilistic causal edge set is constructed; a weighted directed edge list of the three types of causal edges is output; S2.3: Inject the weighted directed edge lists of the three types of causal edges into the corresponding layer spaces: inject the macro-level strongly constrained causal edges into the macro-level graph structure to form a top-level causal framework driven by policy guidance and clinical norms; inject the meso-level weakly supervised causal edges into the meso-level graph structure to construct a knowledge evolution network driven by academic cooperation and technological evolution; inject the micro-level probabilistic causal edges into the micro-level graph structure to establish an individualized response map driven by real-world efficacy feedback; cross-layer semantic bridging is achieved between layers through shared UMLS concept nodes to generate an initial version of a hierarchical causal graph with multi-level semantic structures; S2.4: Perform consistency verification and topology optimization on the initial version of the hierarchical causal graph with multi-level semantic structure: calculate the semantic distance between synonymous nodes across layers based on the MeSH tree structure, and trigger manual review and marking if the distance exceeds a preset threshold; evaluate the simplicity of causal paths within each layer using the Bayesian Information Criterion (BIC) and remove redundant edges to prevent overfitting; introduce a negative sample suppression strategy to exclude false association edges that have no significant difference in the control group data; finally output a stable version of the hierarchical causal graph after semantic consistency correction and structural sparsification. S2.5: Persistently store the stable version of the hierarchical causal graph to the graph database engine and configure a queryable interface for it, supporting index access by research direction node, UMLS semantic type or MeSH classification code; at the same time, generate the metadata description file of the graph, including the construction timestamp, data source version number, causal edge confidence distribution statistics table and coverage indicators of each layer, as the input basis and reliable audit basis for the subsequent multi-hop causal propagation process.

[0028] In this embodiment, in step S3, the evaluation index in the multi-attribute decision modeling module is mapped to the initial excitation signal of the corresponding subgraph and input into the corresponding subgraph structure in the hierarchical causal graph, triggering a multi-hop causal propagation process on each attribute dimension and generating an attribute activation path sequence, such as... Figure 2 As shown, the specific steps include the following: S3.1: Based on the hierarchical causal graph with multi-level semantic structure output by the previous step S2, identify the connected subgraph topology structure composed of macro-level strongly constrained causal edges, meso-level weakly supervised causal edges, and micro-level probabilistic causal edges. According to the medical semantic categories of the subdivided research direction nodes in the department navigation tree, divide the subgraph partitions corresponding to the four thematic domains of innovation, clinical needs, technological evolution, and social impact, and generate four sets of causal propagation subgraphs with clear functional boundaries as a structured carrier for attribute incentive signal injection. S3.2: Obtain the quantitative values ​​of the four core evaluation indicators output by the multi-attribute decision modeling module—innovation level score, clinical need urgency score, technological frontier index value, and expected impact prediction value. Standardize them into dimensionless incentive intensity parameters with unified dimensions. Based on the preset semantic mapping rule table, bind each indicator to the corresponding causal propagation subgraph entry node to generate four initial incentive signal vectors with clear medical semantic anchoring positions to start the directional causal propagation process. S3.3: Perform multi-hop causal propagation algorithm processing on the initial excitation signal vector. This algorithm is based on the causal edge weights, evidence strength levels (L1-L4) and path decay coefficients between nodes in the hierarchical causal graph. In each hop of propagation, it dynamically calculates the activation probability of downstream adjacent nodes. When a node is activated, it inherits the excitation source identifier of the upstream node and appends the UMLS semantic relationship type of the current edge (such as 'causes', 'associated_with') to generate an intermediate activation state with a source tracing mark. It continues to iterate until the maximum propagation depth is reached or the activation energy is lower than a preset threshold. For four initial excitation signal vectors with medical semantic anchoring locations, a hierarchical causal graph multi-hop propagation algorithm is used to calculate the activation probability of downstream nodes based on the graph structure. At each hop, the multi-hop propagation algorithm multiplies the current node's excitation intensity p with the causal edge weight w, and combines this with a correction coefficient for the evidence strength level L. The result is then input into a decay function to obtain the activation probability value of the downstream node.

[0029] Furthermore, energy attenuation control during propagation is achieved through the path attenuation function calculation formula: in, The activation probability of downstream nodes. For edge weights, ( ) is the correction function for the strength of evidence level. This is the path attenuation coefficient. To increase the depth of dissemination.

[0030] Furthermore, based on the UMLS semantic relation type matching table, a semantic tag appending algorithm is used to generate node source tags. Each activated node, while updating its state, inherits the incentive source identifier of its upstream node and appends the semantic relation type R of the current edge, thereby forming an intermediate activated state record with causal semantics.

[0031] Furthermore, by activating a state iteration mechanism, continuous propagation is achieved, and the termination of propagation is dynamically determined. This occurs when the propagation depth reaches the maximum propagation depth. Or the current node's excitation energy is lower than the excitation energy threshold. When this happens, the iteration on that propagation path is stopped to prevent invalid diffusion and waste of computational resources.

[0032] S3.4: For each completed multi-hop propagation process, extract the complete node sequence from the initial activation node to the final activation endpoint to form an attribute activation path; perform UMLS semantic type annotation parsing on all intermediate nodes in the path to obtain a serialized label stream including standard semantic categories such as 'Disease or Syndrome', 'Therapeutic or Preventive Procedure', and 'Biologic Function', and generate a structured attribute activation path sequence with semantic labels as an explicit expression of the causal reasoning trajectory under this attribute dimension; Furthermore, the semantic parsing module identifies the semantic type of each intermediate node in the path array, and maps the nodes to type codes such as 'Disease or Syndrome', 'Therapeuticor Preventive Procedure', and 'Biologic Function' according to the UMLS standard semantic type enumeration table, thus obtaining the node semantic label sequence.

[0033] Furthermore, a serialized tag stream generator is used to concatenate the aforementioned semantic tag sequences into an ordered tag stream according to the propagation order, while maintaining an index consistent with the node chain, so as to facilitate subsequent binding of traceability identifiers.

[0034] Furthermore, a structured path construction method is adopted to pair each node with its semantic label as a key-value structure and encapsulate it into an attribute activation path object with path meta-information (incentive source identifier, propagation depth, and sum of edge weights), thereby generating a structured attribute activation path sequence with semantic labels.

[0035] S3.5: Independently encapsulate the four attribute activation path sequences derived from innovation, clinical needs, technological frontiers, and influence, respectively, retaining their original source identifiers, propagation path chains, UMLS semantic type annotations for each node, and path termination condition information. Output four sets of attribute activation path sequence data packages with uniform format and complete semantics, which can be used in subsequent steps for path consistency analysis and cross-dimensional semantic aggregation degree calculation.

[0036] In this embodiment, in step S4, for each attribute activation path sequence, the endpoint node that reaches the maximum activation depth during propagation is identified, and the semantic aggregation degree value between each path is calculated based on a preset path semantic consistency function, such as... Figure 3 As shown, the specific steps include the following: S4.1: Based on the four sets of attribute activation path sequences with semantic labels, the UMLS semantic label node sequence arranged in the multi-hop causal propagation order in each path is traversed and the activation intensity numerical sequence of each node in the propagation process is extracted. The activation intensity is generated by the product of the edge weight and the node confidence in the hierarchical causal graph and then accumulated after being corrected by the time decay factor, so as to obtain the dynamic activation trajectory data on each path. S4.2: Perform extreme value detection on the dynamic activation trajectory data of each attribute activation path, identify the node whose activation intensity reaches the peak and meets the persistence threshold condition as the maximum activation depth endpoint node of the path. If there are multiple candidate nodes, select the node that is farthest from the initial excitation signal and meets the significance test (p<0.05) as the final endpoint node, and output four sets of endpoint node sets corresponding to the dimensions of innovation, clinical need urgency, technological frontier index and expected impact, respectively. S4.3: Based on the four sets of endpoint nodes and their path context information, construct the input parameter set of the path semantic consistency function. This parameter set includes the MeSH tree structure path code, UMLS semantic type vector, and co-occurrence document frequency statistics corresponding to each path endpoint node. Calculate the semantic similarity matrix between each pair of paths using the standardized parameter vectors. The MeSH structure distance is calculated using the tree edit distance algorithm, and the UMLS semantic matching degree is measured by cosine similarity. S4.4: Input the semantic similarity matrix into the predefined path semantic consistency function. This function is based on a linear weighted fusion model to fuse the MeSH tree structure distance score and the UMLS semantic type matching score. The weight coefficients are trained based on the consistency rating results of historical expert-annotated path pairs to generate semantic consistency scores between pairs of attribute-activated paths and output a 4×4 symmetric score matrix. Furthermore, the weight values ​​of each fusion factor are adjusted using the minimum mean square error criterion through the weight coefficient training algorithm, and the optimal weighted parameter vector is obtained to ensure the robustness and generalizability of the fusion model on different path types.

[0037] Furthermore, a weighted summation function is used to perform a fusion operation on each off-diagonal element of the semantic similarity matrix. The specific calculation formula is as follows: ; in, The consistency score is assigned to the merged path. Distance score for MeSH tree structure, The UMLS semantic type matching score is given. and These are the optimal weight coefficients obtained during training.

[0038] Furthermore, the above calculation steps are performed on all pairwise combinations of the four attribute activation paths to generate a 4×4 semantic consistency scoring matrix with symmetry constraints. The main diagonal element of the matrix is ​​fixed to 1, and the remaining elements are calculated according to the fusion formula.

[0039] S4.5: Perform mean aggregation on the elements outside the main diagonal of the 4×4 semantic consistency scoring matrix, calculate the average consistency score of all non-self-comparison path pairs as the final semantic aggregation degree value. This value represents the overall logical correlation strength between the attribute activation paths triggered by the four core evaluation indicators, and outputs a semantic aggregation degree value in the form of a single scalar for subsequent cross-dimensional integration and comprehensive causal support modeling.

[0040] In this embodiment, step S5 involves integrating attribute activation paths based on semantic aggregation degree values ​​to generate a comprehensive causal support score for candidate topics. The comprehensive causal support score is then mapped back to the original feasibility dimension space in the constraint filter using a path back projection algorithm, resulting in a causal attribution mapping result with traceable intervals. Specifically, this includes the following steps: S5.1: Based on the semantic aggregation degree value calculated by the four sets of attribute activation path sequences with semantic labels output from the previous steps and their corresponding path semantic consistency functions, a weighted fusion strategy is used to integrate the activation paths of the four dimensions of innovation level, clinical need urgency, technological frontier index and expected impact across dimensions. Specifically, the semantic aggregation degree value of each path is used as a weight coefficient to accumulate the importance of the key nodes covered in each path, generating an initial score of comprehensive causal support that reflects the synergistic incentive effect of multiple attributes, which serves as the input source for subsequent back projection. S5.2: Based on the initial score of the comprehensive causal support, execute the path back projection algorithm. This algorithm establishes back propagation rules based on the directed causal edges between nodes in the hierarchical causal graph, and backtracks from the final activated node along the causal chain to the root node of the original feasibility dimension defined by the five-dimensional constraint filter. In this process, maintain a differentiable semantic gradient flow to track the contribution path of each feasibility dimension (such as resource support strength, experimental platform conditions) to the current score, and form a preliminary attribution heat map. S5.3: The attribution heat map is segmented by threshold and quantized by interval. A dynamic clustering algorithm is used to identify the value intervals of significant contributions in each feasibility dimension. For example, in the dimension of 'resource support intensity', three support level intervals of low (0-3), medium (4-6) and high (7-10) are marked according to historical project funding data. The contribution intensity obtained by back projection is mapped to the most matching discrete interval to generate attribution interval identifiers with semantic readability. S5.4: Bind the attribution interval identifier of each feasibility dimension to its corresponding data tracing path, and construct a structured causal attribution mapping table containing dimension name, interval range, evidence source (such as fund project database ID, clinical data batch number) and attribution weight percentage; this table serves as a bridge connecting high-order causal reasoning and low-level feasibility constraints, ensuring that every attribution conclusion can be traced back to the original data source and knowledge ontology anchor point. S5.5: Output the structured causal attribution mapping table as a causal attribution mapping result with traceable intervals, which can be used in subsequent steps to generate ranking basis text and visual explanation panel; this result not only reflects the implementation adaptability distribution of candidate topics in the five-dimensional constraint space, but also explicitly reveals the logical transmission chain driven by the hierarchical causal graph behind it, providing technical support for improving system transparency and user trust.

[0041] In this embodiment, step S6 involves generating ranking criteria text corresponding to each candidate topic based on the causal attribution mapping results, forming a structured causal explanation evidence package, specifically including the following steps: S6.1: Based on the causal attribution mapping results with traceable intervals output by S5, the distribution of attribution intervals of candidate topics in each dimension of the five-dimensional constraint filter is extracted as input conditions. Context matching is performed using a predefined medical semantic template library to generate an initial explanation framework. Each value in the attribution interval distribution (such as 'Resource support intensity: medium to high') serves as a keyword that triggers a specific semantic template. The corresponding natural language fragment is generated through a rule mapping mechanism, and a structured explanation skeleton is output.

[0042] S6.2: Enhance the structured explanation skeleton generated in S6.1. Input the key path node chain in the four attribute activation paths. Based on UMLS semantic type annotation and MeSH tree structure path information, use the semantic distance weighted algorithm to calculate the association confidence between nodes, and merge the node pairs with high confidence (>0.85) into composite causal units. Use the composite causal units to replace the isolated node descriptions in the original skeleton to generate a semantically coherent causal chain description and output the enhanced explanation text intermediate. S6.3: Based on the metadata index of real-world clinical datasets, funding project databases, and patent information databases, data source tracing and binding processing is performed on each causal chain description involved in S6.2. A multi-source evidence alignment algorithm is used to identify the corresponding data source entries (such as 'Section 4.2 of the 2023 National Chronic Disease Surveillance Report'), and L1-L4 evidence strength levels are automatically assigned according to the evidence type (statistical correlation, intervention study, expert consensus, technical verification), and annotated explanatory text with source labeling and evidence classification is output. S6.4: Perform attribution quantification embedding processing on the annotated explanatory text output by S6.3. Input the attribution weight (%) of each attribute activation path in the semantic aggregation degree calculation, and use a dynamic interpolation method to embed this value into the end of the corresponding causal chain in the form of 'contribution percentage' to form the standard expression format of '[causal chain] (contribution XX%)'. At the same time, retain the original comprehensive causal support score as a global confidence index, and output a complete explanatory instance integrating quantitative attribution information. S6.5: Performs consistency checks and format standardization on multiple complete explanation instances generated by S6.4, defines a unified terminology based on medical research writing standards (e.g., uniformly using 'end-stage diabetic nephropathy' instead of 'uremia stage'), uses a terminology consistency replacement algorithm to eliminate ambiguity, and encapsulates the final text into a structured causal explanation evidence package in JSON-LD format, which includes fields such as 'pathway_chain', 'evidence_source', 'confidence_level', and 'attribution_weight', and outputs standardized explanation data units that can be called by the visualization interface.

[0043] In this embodiment, step S7 involves synchronously associating the causal explanation evidence package with the original topic ranking list output by the multi-attribute decision model to construct an aligned topic-explanation data structure. This specifically includes the following steps: S7.1: Based on the original topic ranking list output by the multi-attribute decision modeling module, obtain the unique identifier of each candidate topic and its corresponding comprehensive score, sub-scores of each dimension and ranking position, as the primary key input for subsequent structured association, so as to establish the data anchor point for topic recommendation results; S7.2: Serialize and parse the structured causal explanation evidence package generated in S6, extract the key path node chain, data source label, evidence strength level (L1-L4) and quantitative attribution weight of each path contribution contained in each evidence package, and form a set of structured explanation tuples indexed by the unique identifier of the topic to support cross-modal data alignment. S7.3: Perform primary key matching operation based on the unique identifier of the topic, and associate each record in the original topic ranking list with its corresponding set of structured explanation tuples one-to-one to build a topic-explanation aligned data structure. Each topic item is bound to its exclusive causal path semantic chain and attribution evidence set to ensure semantic consistency between the recommendation results and the explanation content. S7.4: Using the component-based rendering engine in the front-end visualization framework, the topic-explanation aligned data structure is embedded into the recommendation display module of the web interface to generate a causal path collapse panel with hierarchical expansion function. The initial state of the panel only displays the core scoring indicators and a brief path summary. The full path is loaded through user interaction to optimize the information presentation density and interface response efficiency. S7.5: Deploy a dynamic graph rendering unit inside the folded panel. When the user clicks the expand command, based on the preloaded hierarchical causal graph topology and the attribution interval distribution output by the path back projection algorithm, generate a visualized causal chain graph in real time, which includes complete path chains, causal directions between nodes, data source labels, and attribution weight percentages. Output an interactive explanation interface with medical semantic annotation and evidence tracing capabilities, thereby achieving full observability, verifiability, and traceability of the recommendation logic.

[0044] In this embodiment, in step S8, when a user accesses the topic ranking suggestions, the topic recommendation display format with a traceable logical chain is automatically loaded. If it is detected that the user expands the causal path collapse panel of a certain topic, a visualized causal chain graph containing the complete path chain, evidence sources, and attribution weights is output, so as to realize the full observability, verifiability, and traceability of the recommendation results. Specifically, it includes the following steps: S8.1: Based on the aligned topic-explanation data structure embedded in the front-end visual interaction interface, obtain the identifier of the candidate topic currently accessed by the user, and trigger the recommendation display format loading request to initialize the topic recommendation interface component with a traceable logical chain.

[0045] S8.2: Perform context-aware parsing on the loading request, extract browsing preferences and historical interaction trajectories from the user's session state, combine them with the structured causal explanation evidence package cached on the backend, and generate a personalized display view template adapted to the current user role (such as a research manager or clinical researcher).

[0046] S8.3: Based on the personalized display view template, render a causal path collapsible panel that supports click interaction, embed an expandable area in the panel and bind an event listener to monitor in real time whether the user triggers the 'expand' operation behavior of a specific topic.

[0047] S8.4: When the user expands the causal path collapse panel of a certain topic, the corresponding four attribute activation path sequences, the maximum activation depth nodes of each path and their semantic aggregation degree values ​​are extracted from the structured causal explanation evidence package, and the attribution interval distribution output by the path back projection algorithm is used as the basis for explanation.

[0048] S8.5: Utilize a graph visualization engine to lay out and render the extracted attribute activation path sequence, generating a visualized causal chain graph that includes complete path chains, UMLS semantic type annotations, MeSH tree structure distance weighting information, data source annotations (such as national chronic disease monitoring reports), evidence strength levels (L1-L4), and attribution weights (%) of each path's contribution, thus achieving full observability, verifiability, and traceability of the recommendation logic.

[0049] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0050] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0051] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent selection of medical research topics based on multi-modal data analysis, characterized in that, Includes the following steps: S1: Based on the hierarchical departmental navigation tree already constructed in the intelligent medical research topic selection system, extract the subdivided research direction nodes under the target research field and construct a structured domain knowledge node set; S2: Construct a hierarchical causal graph containing macro, meso, and micro layers based on structured domain knowledge node sets, clinical datasets, funding project databases, and patent intelligence databases; S3: Map the evaluation index in the multi-attribute decision modeling module to the initial excitation signal of the corresponding subgraph, and input it into the corresponding subgraph structure in the hierarchical causal graph to trigger the multi-hop causal propagation process in each attribute dimension and generate the attribute activation path sequence. S4: For each attribute activation path sequence, identify the endpoint node that reaches the maximum activation depth during the propagation process, and calculate the semantic aggregation degree value between each path based on the preset path semantic consistency function. S5: Based on the semantic aggregation degree value, integrate attribute activation paths to generate a comprehensive causal support score for candidate topics. Then, through the path back projection algorithm, map the comprehensive causal support score back to the original feasibility dimension space in the constraint filter to obtain a causal attribution mapping result with traceable intervals. S6: Based on the causal attribution mapping results, generate ranking criteria texts corresponding to each candidate topic, forming a structured causal explanation evidence package; S7: Synchronize and associate the causal explanation evidence package with the original topic ranking list output by the multi-attribute decision model to construct an aligned topic-explanation data structure.

2. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 1, characterized in that, In step S1, the navigation tree organizes clinical departments and their subordinate research directions in a tree structure, where each non-leaf node represents a clinical department and each leaf node represents a specific research direction branch.

3. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 1, characterized in that, In step S2, based on the structured domain knowledge node set, subdivided research direction nodes under the target research domain are obtained as semantic anchors; the semantic anchors are used to match the epidemiological trends, diagnosis and treatment behavior sequences and patient outcome indicators of the corresponding diseases in real-world clinical datasets, retrieve the relevant national key research and development program project topics in the funding project database, extract the high-frequency citation paths of related technical fields in the patent intelligence database, and perform timestamp alignment and entity normalization processing to generate a cross-domain aligned original evidence tuple set; A three-layer causal edge identification algorithm is executed on the original evidence tuple set to obtain a set of macroscopic layer strongly constrained causal edges, a set of mesoscopic layer weakly supervised causal edges, and a set of microscopic layer probabilistic causal edges, and outputs a weighted directed edge list of the three types of causal edges. Inject the weighted directed edge lists of the three types of causal edges into the corresponding layer spaces to generate an initial version of the hierarchical causal graph. A consistency check and topology optimization are performed on the initial version of the hierarchical causal graph with a multi-level semantic structure to obtain a stable version of the hierarchical causal graph.

4. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 3, characterized in that, In step S2, strong constraint causal edges at the macro level are injected into the macro level graph structure to form a top-level causal framework driven by policy guidance and clinical norms; weak supervision causal edges at the meso level are injected into the meso level graph structure to construct a knowledge evolution network driven by academic cooperation and technological evolution. By injecting probabilistic causal edges into the micro-layer graph structure, an individualized response map driven by real-world therapeutic feedback is established; cross-layer semantic bridging is achieved between layers through shared UMLS concept nodes.

5. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 4, characterized in that, In step S3, based on the preset semantic mapping rule table, the evaluation indicators such as the innovation level score, the clinical need urgency score, the technological frontier index value, and the expected impact prediction value are bound to the corresponding causal propagation subgraph entry node to generate the initial incentive signal vector. The initial excitation signal vector is processed by a multi-hop causal propagation algorithm. When a node is activated, it inherits the excitation source identifier of the upstream node and adds the UMLS semantic relation type of the current edge to generate an intermediate activation state with a source tracing mark. This process is continued until the maximum propagation depth is reached or the activation energy is lower than a preset threshold. For each completed multi-hop propagation process, the complete node sequence from the initial excitation node to the final activation endpoint is extracted to form an attribute activation path. UMLS semantic type annotation is parsed for all intermediate nodes in the path to generate an attribute activation path sequence.

6. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 1, characterized in that, In step S4, the UMLS semantic annotation node sequence arranged in the multi-hop causal propagation order in each path is traversed and the activation intensity numerical sequence of each node in the propagation process is extracted. The activation intensity is generated by accumulating the product of the edge weight and the node confidence in the hierarchical causal graph after being corrected by the time decay factor, so as to obtain the dynamic activation trajectory data on each path. For each attribute activation path, an extreme value detection operation is performed on the dynamic activation trajectory data. The node whose activation intensity reaches the peak and meets the persistence threshold condition is identified as the maximum activation depth endpoint node of the path. If there are multiple candidate nodes, the node that is farthest from the initial excitation signal and meets the significance test is selected as the final endpoint node. Based on the endpoint node and its path context information, a set of input parameters for the path semantic consistency function is constructed, and the semantic similarity matrix between each pair of paths is calculated using the standardized parameter vector; The semantic similarity matrix is ​​input into a predefined path semantic consistency function to generate semantic consistency scores between pairs of attribute-activated paths. The mean of elements outside the main diagonal is aggregated, and the average consistency score of all non-self-comparison path pairs is calculated as the final semantic aggregation degree value.

7. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 6, characterized in that, In step S4, the input parameter set includes the MeSH tree structure path code corresponding to each path endpoint node, the UMLS semantic type vector, and the co-occurrence document frequency statistics.

8. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 7, characterized in that, In step S4, the path semantic consistency function is based on a linear weighted fusion model to fuse the MeSH tree structure distance score and the UMLS semantic type matching score. The weight coefficients are trained based on the consistency rating results of historical expert-annotated path pairs to generate semantic consistency scores between pairs of attribute-activated paths.

9. The intelligent medical research topic selection method based on multi-modal data analysis according to claim 1, characterized in that, In step S5, the semantic aggregation degree value calculated based on the attribute activation path sequence with semantic labels and its corresponding path semantic consistency function is used to integrate each activation path across dimensions using a weighted fusion strategy to obtain an initial score. Based on the initial score execution path back projection algorithm, a back propagation rule is established according to the directed causal edges between nodes in the hierarchical causal graph. The algorithm backtracks from the final activated node along the causal chain to the root node of the original feasibility dimension defined by the constraint filter, forming a preliminary attribution heat map. Threshold segmentation and interval quantization are performed on the attribution heatmap. A dynamic clustering algorithm is used to identify the value intervals of significant contributions on each feasibility dimension. The contribution intensity obtained by back projection is then mapped to the most matching discrete interval to generate attribution interval identifiers. Each attribution interval identifier is bound to its corresponding data source path to construct a structured causal attribution mapping table; the structured causal attribution mapping table is output as the causal attribution mapping result.

10. A medical research intelligent topic selection system based on multi-modal data analysis, characterized in that, The system described herein utilizes the method described in any one of claims 1-9 to achieve intelligent topic selection for medical research.

Images