Intelligent data processing method for clinical research

Abstract

The invention relates to the technical field of electric digital data processing, and discloses an intelligent data processing method for clinical research. The method comprises the following steps: constructing a basic clinical knowledge graph fused with an international medical ontology; the method comprises the following steps: performing deep learning driven entity recognition and knowledge graph linking on original data from heterogeneous sources such as an electronic medical record and an inspection system; executing cross-source entity alignment and knowledge fusion based on the graph attention network; performing automatic data quality verification and restoration according to clinical logic rules embedded in the atlas; and finally, extracting and generating a standardized analysis ready data set from the enhanced atlas according to research requirements. According to the technical scheme, high-quality, automatic and semantic integration and flexible delivery of clinical research data are realized, and the data processing efficiency and research reliability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric digital data processing, and particularly relates to an intelligent data processing method for clinical research. BACKGROUND

[0002] With the increasing role of clinical research in precision medicine, drug development and public health decision-making, the demand for high-quality and high-consistency data is increasingly urgent. Modern clinical research usually relies on multi-source heterogeneous data, including electronic health records, laboratory information systems, image archiving systems and patient self-reported mobile health data. These data differ in structure, coding standards, terminology systems and update frequency, making it difficult for traditional data integration methods to achieve effective alignment and fusion.

[0003] Especially in large-scale multi-center studies, the lack of a unified semantic framework makes the data cleaning, mapping and verification process highly dependent on manual intervention, which not only is inefficient, but also easily introduces subjective bias, seriously affecting the scientificity, repeatability and regulatory compliance of the research results.

[0004] Intelligent data governance technology based on natural language processing and knowledge representation has become a key path to break through data silos. This direction aims to automatically analyze unstructured or semi-structured clinical text through computational means, identify medical entities and their mutual relationships, and map them to standardized terminology systems, thereby building integrated data sets with semantic consistency.

[0005] The core of such technology is to achieve semantic interoperability across source data, rather than just simple conversion at the field or format level. Its goal is to support subsequent advanced analysis tasks, such as real-world evidence generation, cohort identification and causal inference.

[0006] Existing technologies still face multiple bottlenecks in processing clinical multi-source data: on the one hand, traditional ETL processes rely on pre-defined rules and static mapping tables, which cannot adapt to the dynamic evolution of medical terminology and clinical practice;

[0007] On the other hand, although some studies have attempted to introduce ontologies or dictionaries for terminology standardization, they lack a deep understanding of contextual semantics, making it difficult to accurately parse negations, hypotheses or temporal relationships in clinical notes. In addition, existing systems generally do not establish a dynamically scalable knowledge graph architecture, which cannot continuously capture new medical concepts and their associations, limiting the data fusion capabilities.

[0008] In practical applications, these problems directly manifest as high data missing rates, frequent entity alignment errors, and non-comparability of cross-institutional data, which severely restrict the transition of clinical research from data availability to data credibility, computability and traceability, and an intelligent data processing method that integrates deep semantic understanding and dynamic knowledge modeling is urgently needed. SUMMARY

[0009] The technical problem to be solved by the present application is to provide an intelligent data processing method for clinical research to overcome the defects of data integration difficulty, semantic ambiguity, low quality and poor interoperability caused by heterogeneous data sources, different formats and standards in the prior art.

[0010] To solve the above technical problems, the present application provides an intelligent data processing method for clinical research, which constructs a clinical ontology knowledge graph as a semantic hub, uses a deep learning model to perform entity recognition and relationship extraction on heterogeneous source data, and realizes semantic mapping and fusion of data to knowledge graph nodes.

[0011] Further, logical consistency and clinical path constraint rules are defined in the fused knowledge graph, and automatic data quality checking and repair based on the graph structure are performed.

[0012] Finally, according to the preset research analysis requirements, a structured and standardized unified data set is generated through graph query and traversal technology, thereby realizing high-quality and automatic integration and processing of multi-source heterogeneous clinical research data.

[0013] The present application provides an intelligent data processing method for clinical research, which comprises the following steps:

[0014] Constructing a basic clinical knowledge graph, which presets core concept nodes and semantic relationship edges between them covering diseases, symptoms, drugs, test items, treatment plans and anatomical structures;

[0015] Receiving and preprocessing clinical raw data from multiple heterogeneous data sources, including hospital information systems, laboratory information management systems, electronic medical record systems and patient self-report data interfaces;

[0016] Performing semantic annotation and entity linking on the clinical raw data, identifying unstructured text and semi-structured field content in the data as clinical entities, and linking them to unique concept nodes in the basic clinical knowledge graph;

[0017] Performing cross-source entity alignment and knowledge fusion, merging entity information from different data sources linked to the same concept node, and creating instance nodes and attribute information in the basic clinical knowledge graph to form an enhanced clinical instance knowledge graph;

[0018] Based on the enhanced clinical instance knowledge graph, performing graph-constrained data quality verification and cleaning, identifying and correcting logical contradictions, numerical abnormalities and integrity defects in the data;

[0019] Based on the pre-defined clinical research data specifications, a standardized analysis-ready dataset is extracted and generated from the enhanced clinical instance knowledge graph.

[0020] As one embodiment of the present invention, the construction of the basic clinical knowledge graph specifically includes:

[0021] Integrate at least two publicly available, standardized medical ontology databases: one is the International Classification of Diseases, Injuries and Causes of Death, and the other is a logical naming and coding system for clinical terminology. Extract the core conceptual entities and their hierarchical structure and relationships from these databases to serve as the initial skeleton of the knowledge graph.

[0022] Define the schema layer of the knowledge graph, including entity types such as patient, diagnosis, drug, test, surgery, and relation types such as suffer from, receive, take, indicator is, execute, and define a set of attributes for each entity type, such as the age and gender attributes of the patient entity, and the dosage and usage attributes of the drug entity.

[0023] Using a medical literature corpus, natural language processing technology is used to automatically mine implicit relationships between conceptual entities that are not included in standard ontology databases, such as the association between a specific drug and a rare adverse reaction. The newly mined relationships are then added as relation edges to the basic clinical knowledge graph to enrich its semantic expressive power.

[0024] As one embodiment of the present invention, the semantic annotation and entity linking of the original clinical data specifically includes:

[0025] For unstructured text data, such as medical records and discharge summaries, a pre-trained biomedical language model based on a transformer architecture is used. The model is fine-tuned for a clinical named entity recognition task and is used to identify entity mentions such as diseases, symptoms, drugs, tests, and surgeries in the text.

[0026] For semi-structured data, such as test reports, key-value pairs are extracted using preset regular expressions and field mapping rules, where the key is the test item name and the value is the test result and unit.

[0027] The identified entity mentions or test item names are used to calculate the cosine similarity between the entity mentions and test item names and the vector representations of all concept nodes in the KOMO spectrum of the basic clinical knowledge graph through an entity linking model based on a Siamese neural network.

[0028] Concept nodes with cosine similarity scores higher than a preset linking threshold are selected as linking targets, thereby mapping strings in the original data to deterministic concept entities in the knowledge graph. The preset linking threshold ranges from 0.90 to 0.98.

[0029] As one embodiment of the present invention, the execution of cross-source entity alignment and knowledge fusion specifically includes:

[0030] For instance nodes representing the same real-world object, such as the same patient from different systems, a multimodal feature vector is constructed. The feature vector contains a hash code of its demographic information, a bag-of-words representation of the International Classification of Diseases (ICD) codes of the main diagnoses in the medical records, and statistical features of the distribution of its adjacent nodes in the knowledge graph.

[0031] An alignment model based on graph attention network is adopted. The input is the feature vectors of two instance nodes to be judged, and the output is a probability value indicating that the two are the same object.

[0032] When the probability value exceeds a preset alignment threshold, the two instance nodes are merged. The merging strategy includes migrating all unique attributes and relationship edges to one retained node, deleting the other node, and updating all links pointing to the deleted node. The preset alignment threshold is set to 0.95.

[0033] In the knowledge graph, a link prediction algorithm based on restricted Boltzmann machine is used to infer and complete the missing patient-drug relationship edges based on the existing patient-diagnosis-drug triplet relationship, thereby enhancing the integrity of the graph data.

[0034] As one embodiment of the present invention, the data quality verification and cleaning based on the enhanced clinical instance knowledge graph, performing graph constraint verification and cleaning specifically includes:

[0035] A set of data quality rules based on shape constraint language is predefined. These rules describe the logical constraints that data should follow in the form of a graph schema. The logical constraints include value type constraints, i.e., the data type of a specific attribute value must be an integer or a floating-point number; numerical range constraints, i.e., the numerical value of a specific test item must be within the range of clinical normal values; cardinality constraints, i.e., a patient entity can only be associated with one live state node at any given time; and association logic constraints, i.e., a patient entity diagnosed with a specific male disease cannot have a gender attribute value of female.

[0036] Launch a validation engine that executes the data quality rules on the enhanced clinical instance knowledge graph, traversing all instance nodes and their attributes and relationships, and outputting all data nodes that violate the constraints and their violation details;

[0037] For the identified non-compliant data, an automated cleaning process is performed using a preset correction strategy. The correction strategy includes: for numerical anomalies, if the data deviates from three standard deviations of the distribution of similar data in the patient group, it is marked as a null value; for logical contradictions, such as gender not matching the diagnosis, the confidence attribute value of the diagnostic relationship is reduced, and the entry is recorded in the data audit log for manual review.

[0038] As one embodiment of the present invention, the step of extracting and generating a standardized analysis-ready dataset from the enhanced clinical instance knowledge graph according to a preset clinical research data specification specifically includes:

[0039] Receive a clinical research data specification defined in the form of a structured query language, which specifies the inclusion and exclusion criteria for the research cohort, the list of data variables to be observed, and the data types and coding standards of the variables;

[0040] The clinical research data specifications are automatically translated into a set of query statements in a graph query protocol language, and the inclusion and exclusion criteria are converted into pattern matching conditions for nodes and edges in the graph.

[0041] Execute the query statement on the enhanced clinical instance knowledge graph to filter out a set of patient instance nodes that meet all inclusion criteria and do not meet any exclusion criteria;

[0042] For each selected patient instance node, traverse along its relation edges, extract the specified attribute values ​​or associated node information from the data variable list, and convert them according to the coding standards in the specification, such as converting the diagnosis node into the International Classification of Diseases (ICD) code string.

[0043] All extracted and transformed data are organized into a two-dimensional table structure, with patient instances as behaviors and data variables as columns, and exported as a comma-separated value file format or a columnar storage file format as the final analysis-ready dataset to be delivered.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. This invention constructs a clinical knowledge graph as a unified semantic model, mapping data from different sources and with different structures to a unified conceptual space, fundamentally solving the semantic ambiguity and integration barriers caused by data heterogeneity, and achieving seamless integration of cross-source data;

[0046] 2. This invention utilizes deep learning-based natural language processing and entity linking technology to achieve automated, high-precision extraction and structuring of information contained in clinical text data, thereby improving the automation level of data processing and the depth of information utilization.

[0047] 3. This invention proposes a data quality verification method based on graph constraints, which embeds clinical knowledge and logical rules into the data model, realizing in-depth, context-aware verification and cleaning of data quality, and ensuring the accuracy and reliability of the final generated dataset.

[0048] 4. By converting data query requirements into graph traversal operations, this invention provides flexible and efficient data extraction and reconstruction capabilities, enabling rapid response to diverse clinical research needs and generating customized and standardized analysis-ready datasets, which greatly improves the efficiency of data preparation and the reproducibility of research results in clinical research. Attached Figure Description

[0049] Fig. 1 This is a schematic diagram of the overall technical architecture of an intelligent data processing method for clinical research proposed in this invention.

[0050] Fig. 2 This is a schematic diagram of the core principle framework of semantic mapping and fusion based on clinical ontology knowledge graph in this invention;

[0051] Fig. 3 This is a logical flow diagram of the semantic annotation and entity linking of heterogeneous clinical raw data in this invention;

[0052] Fig. 4 This is a logical flowchart of the data quality verification of cross-source entity alignment, knowledge fusion and graph constraints in this invention;

[0053] Fig. 5 This is a logical flowchart of the [Generation of Standardized Analysis-Ready Dataset Based on Enhanced Clinical Instance Knowledge Graph] in this invention; Detailed Implementation

[0054] Please refer to the attached document. Figs. 1 to 5 This invention provides an intelligent data processing method for clinical research, aiming to solve the technical challenges of integrating multi-source heterogeneous clinical data in terms of semantic consistency, quality control, and standardized output. The method centers on constructing and continuously enhancing a clinical knowledge graph, and through steps such as deep learning-driven semantic mapping, cross-source entity alignment, quality verification under graph constraints, and on-demand structured extraction, achieves fully automated processing from raw heterogeneous data to high-quality, analysis-ready datasets.

[0055] The method includes the following steps:

[0056] S1, Constructing a basic clinical knowledge graph;

[0057] S2 receives and preprocesses raw clinical data from multiple heterogeneous data sources;

[0058] S3, semantic annotation and entity linking of the raw clinical data;

[0059] S4 performs cross-source entity alignment and knowledge fusion;

[0060] S5, based on an enhanced clinical instance knowledge graph, performs data quality validation and cleaning based on graph constraints;

[0061] S6. Based on the preset clinical research data specifications, extract and generate a standardized analysis-ready dataset from the enhanced clinical instance knowledge graph.

[0062] In step S1, a basic clinical knowledge graph is constructed. This knowledge graph serves as the semantic hub of the entire data processing flow. Its construction process first integrates at least two publicly available, standardized medical ontology databases. One is the International Classification of Diseases, Injuries and Causes of Death (ICD-1), and the other is a logical naming and coding system for clinical terminology. Core conceptual entities, their hierarchical structures, and relationships are extracted from these two authoritative resources to form the initial skeleton of the knowledge graph. These conceptual entities cover key elements in clinical research, such as diseases, symptoms, drugs, laboratory tests, treatment plans, and anatomical structures.

[0063] Building upon this, a schema layer for the knowledge graph is defined. The schema layer explicitly specifies entity types, including patient, diagnosis, drug, test, and surgery, and defines relation types, such as having, receiving, taking, indicator being, and executing. Simultaneously, a set of attributes is assigned to each entity type. For example, the "patient" entity includes age and gender attributes; the drug entity includes dosage and usage attributes; and the test entity includes result value, unit, and reference range attributes.

[0064] To further enrich the semantic expressive power of the graph, a large-scale medical literature corpus is utilized, and natural language processing techniques are employed to automatically mine implicit relationships between conceptual entities that are not included in standard ontology databases. Examples include the association between specific drugs and rare adverse reactions, and the dynamic trends of certain test indicators under specific disease states.

[0065] These new relationships obtained through text mining are added as relation edges to the basic clinical knowledge graph, thereby constructing a semantic model that is both authoritative and extensible.

[0066] In step S2, raw clinical data from multiple heterogeneous data sources are received and preprocessed. These heterogeneous data sources include hospital information systems, laboratory information management systems, electronic medical record systems, and patient self-report data interfaces. The data generated by these systems have different formats, including structured tables, semi-structured key-value pairs, and unstructured free text. The preprocessing stage first performs unified timestamp alignment on all data to ensure that data records from different sources can be correctly correlated in time sequence.

[0067] For structured data, the field content is directly parsed; for semi-structured data, such as test reports, key-value pairs are extracted using preset regular expressions and field mapping rules, where the key is the test item name and the value is the test result and unit; for unstructured text data, such as medical records and discharge summaries, the original text form is retained for subsequent semantic processing.

[0068] All preprocessed data is encapsulated into data units with metadata tags. The metadata tags record information such as data source, data type, collection time, and patient identifier, providing contextual support for subsequent entity linking and alignment.

[0069] In step S3, semantic annotation and entity linking are performed on the raw clinical data. The core task of this step is to map the string representations in the raw data to unique, standardized concept nodes in the basic clinical knowledge graph.

[0070] For unstructured text data, a pre-trained biomedical language model based on a transformer architecture is employed. This model has been pre-trained on a large-scale clinical text corpus and further fine-tuned on a clinical named entity recognition task, enabling it to recognize entity mentions such as diseases, symptoms, drugs, tests, and surgeries in the text.

[0071] The model outputs each identified entity mention along with its location and type in the text. For semi-structured data, as mentioned earlier, key-value pairs have been extracted using rules, where the key represents the entity mention to be linked. Subsequently, all identified entity mentions or check item names are input into an entity linking model based on a Siamese neural network.

[0072] The model generates high-dimensional vector representations for each entity mention and each concept node in the knowledge graph. A similarity score is obtained by calculating the cosine similarity between the two vectors. Concept nodes with cosine similarity scores higher than a preset linking threshold are selected as linking targets.

[0073] The preset link threshold ranges from 0.90 to 0.98. If the highest score is less than this threshold, the entity mention is marked as unlinked and recorded in the log for subsequent manual review or model iteration optimization.

[0074] Through this process, ambiguous, redundant, or synonymous expressions in the original data are accurately mapped to deterministic conceptual entities in the knowledge graph, thus achieving semantic standardization of the data.

[0075] In step S4, cross-source entity alignment and knowledge fusion are performed. After step S3, information about the same real-world object from different data sources is represented as multiple instance nodes in the knowledge graph. The goal of this step is to identify and merge these instance nodes representing the same object, thereby constructing a coherent and non-redundant augmented clinical instance knowledge graph.

[0076] First, a multimodal feature vector is constructed for each instance node. This feature vector integrates multiple information sources: its demographic information is hashed to protect privacy; the main diagnostic information in its medical records is converted into a bag-of-words representation of the International Classification of Diseases (ICD); in addition, it also includes the statistical features of the distribution of the node's neighboring nodes in the knowledge graph, that is, the number and type distribution of various entity nodes directly connected to it.

[0077] Next, an alignment model based on a graph attention network is adopted. This model takes the feature vectors of two instance nodes to be judged as input, learns their deep association in the feature space through a multi-layer graph attention mechanism, and finally outputs a probability value representing the likelihood that the two are the same real-world object. When this probability value is greater than a preset alignment threshold, the two are determined to be the same object. The preset alignment threshold is set to 0.95. Once determined to be the same object, a node merging operation is performed.

[0078] The merging strategy is as follows: Select one node as the retained node, migrate all unique attributes and relation edges of the other node to the retained node, and delete the merged node. Simultaneously, update all links in the knowledge graph pointing to the deleted node, redirecting them to the retained node.

[0079] After entity alignment, to further enhance the integrity of the graph data, a link prediction algorithm based on a restricted Boltzmann machine is used. This algorithm infers and completes missing direct relation edges based on existing triplet relationships in the graph.

[0080] Through this knowledge fusion process, the originally scattered and isolated data fragments are integrated into a tightly structured and semantically rich clinical instance knowledge graph.

[0081] In step S5, based on the enhanced clinical instance knowledge graph, data quality verification and cleaning of the graph constraints are performed. This step embeds clinical knowledge and logical rules into the data model, achieving a deeper level of data quality assurance.

[0082] First, a set of data quality rules based on a shape constraint language is predefined. These rules describe the logical constraints that the data should follow in the form of a graph schema, specifically including:

[0083] Value type constraints stipulate that the data type of a specific attribute value must be an integer or a floating-point number. For example, the "age" attribute cannot be a string.

[0084] Numerical range constraints stipulate that the values ​​of specific test items must be within the clinically recognized normal range. For example, blood glucose values ​​that exceed the reasonable physiological range are considered abnormal.

[0085] Cardinality constraint stipulates that a patient entity can only be associated with one live state node at any given time to avoid state conflicts; and association logic constraint stipulates that a patient entity diagnosed with a specific male disease cannot have a female gender attribute value.

[0086] A validation engine is then activated. This engine traverses all instance nodes, attributes, and relationship edges in the augmented clinical instance knowledge graph, checking each one to ensure it conforms to the predefined data quality rules. For any violation, the validation engine generates a detailed violation report, recording the violating node, violating attribute or relationship, the specific violation type, and the expected compliance status.

[0087] For identified non-compliant data, a pre-defined correction strategy is applied for automated cleaning. The correction strategy includes: for numerical anomalies, if the data deviates from three standard deviations of the distribution of similar data in the patient group, it is marked as a null value and the reason for the anomaly is recorded; for logical contradictions, such as gender not matching the diagnosis, the data is not directly deleted, but the confidence attribute value of the diagnostic relationship is reduced, and the entry is recorded in the data audit log for manual review.

[0088] This quality verification method based on graph structure can make accurate judgments by utilizing rich contextual information, far exceeding the traditional simple threshold verification based on isolated fields.

[0089] In step S6, a standardized analysis-ready dataset is extracted and generated from the enhanced clinical instance knowledge graph according to preset clinical research data specifications. This step enables on-demand conversion from a complex graph structure to a concise two-dimensional table.

[0090] First, we receive a clinical research data specification defined in Structured Query Language (SMQ). This specification, provided by the researchers, clearly specifies the inclusion and exclusion criteria for the study cohort, the list of data variables to be observed, and the data types and coding standards for the variables. For example, inclusion criteria might be age over 50 and having type 2 diabetes, while exclusion criteria might be a history of severe hepatic or renal insufficiency. Required variables might include glycated hemoglobin levels, types of medications used, and incidence of complications.

[0091] Next, the system will automatically translate the clinical research data specifications into a set of query statements in a graph query protocol language.

[0092] The inclusion and exclusion criteria are precisely translated into pattern matching conditions between nodes and edges in the graph. For example, "having type 2 diabetes" is translated into finding nodes with the relationship "patient-having-type 2 diabetes".

[0093] These queries are then executed on the enhanced clinical instance knowledge graph to filter out the set of patient instance nodes that meet all inclusion criteria and do not meet any exclusion criteria, which is the target study cohort.

[0094] For each patient instance node in the queue, the system performs a depth-first traversal along its relational edges, extracting the specified attribute values ​​or associated node information from the data variable list. For example, to obtain the glycated hemoglobin level, the system searches for all child nodes of the test types under that patient node, filters out the node whose test item is glycated hemoglobin, and extracts its result value attribute.

[0095] Finally, the extracted data is transformed according to the coding standards in the specification. For example, diagnostic nodes are converted into corresponding International Classification of Disease (ICD) code strings, and drug names are converted into generic names. All extracted and transformed data is organized into a two-dimensional table structure, where each row corresponds to a patient instance and each column corresponds to a data variable.

[0096] The table is ultimately exported as a comma-separated value file or a columnar storage file, and delivered to researchers as an analysis-ready dataset that can be directly used in statistical analysis software.

[0097] Through the collaborative work of the above six steps, the method of this invention constructs a closed-loop, intelligent clinical data processing pipeline. This pipeline uses a knowledge graph as a unified semantic model, eliminating obstacles caused by data heterogeneity; it utilizes advanced deep learning technology to achieve high-precision automated information extraction; it ensures data quality through clinical logic rules embedded in the knowledge graph structure; and it can flexibly respond to diverse research needs, efficiently generate standardized datasets, and comprehensively improve the efficiency, quality, and reliability of clinical research data preparation.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent data processing method for clinical research, characterized in that, include: A basic clinical knowledge graph is constructed, which pre-defines core concept nodes covering diseases, symptoms, drugs, test items, treatment plans, and anatomical structures, as well as semantic relationship edges between them; Receive and preprocess raw clinical data from multiple heterogeneous data sources, including hospital information systems, laboratory information management systems, electronic medical record systems, and patient self-report data interfaces; Semantic annotation and entity linking are performed on the raw clinical data. Unstructured text and semi-structured field content in the data are identified as clinical entities and linked to the unique concept node in the basic clinical knowledge graph. Cross-source entity alignment and knowledge fusion are performed to merge entity information from different data sources that are linked to the same concept node, and instance nodes and attribute information are created in the basic clinical knowledge graph to form an enhanced clinical instance knowledge graph. Based on the enhanced clinical case knowledge graph, data quality verification and cleaning are performed under graph constraints to identify and correct logical contradictions, numerical anomalies and integrity defects in the data. Based on the pre-defined clinical research data specifications, a standardized analysis-ready dataset is extracted and generated from the enhanced clinical instance knowledge graph.

2. The intelligent data processing method for clinical research according to claim 1, characterized in that, Constructing a basic clinical knowledge graph, including: Integrate at least two publicly available, standardized medical ontology databases: one is the International Classification of Diseases, Injuries and Causes of Death, and the other is a logical naming and coding system for clinical terminology. Extract the core conceptual entities and their hierarchical structure and relationships from these databases to serve as the initial skeleton of the knowledge graph. Define the schema layer of the knowledge graph, including entity types and relation types, and define a set of attributes for each entity type; Using a medical literature corpus, implicit relationships between conceptual entities that are not included in standard ontology databases are automatically mined through natural language processing technology. The newly mined relationships are then added as relation edges to the basic clinical knowledge graph to enrich its semantic expressive power.

3. The intelligent data processing method for clinical research according to claim 1, characterized in that, Semantic annotation and entity linking of the raw clinical data, including: For unstructured text data, such as medical records and discharge summaries, a pre-trained biomedical language model based on a transformer architecture is used. The model is fine-tuned for a clinical named entity recognition task and is used to identify entity mentions such as diseases, symptoms, drugs, tests, and surgeries in the text. For semi-structured data, such as test reports, key-value pairs are extracted using preset regular expressions and field mapping rules, where the key is the test item name and the value is the test result and unit. The identified entity mentions or test item names are used to calculate the cosine similarity between the entity and the vector representation of all concept nodes in the basic clinical knowledge graph through an entity linking model based on a Siamese neural network. Concept nodes with cosine similarity scores higher than a preset link threshold are selected as link targets, thereby mapping strings in the original data to a deterministic concept entity in the knowledge graph.

4. The intelligent data processing method for clinical research according to claim 1, characterized in that, The execution of cross-source entity alignment and knowledge fusion includes: For instance nodes representing the same real-world object, such as the same patient from different systems, a multimodal feature vector is constructed. The feature vector contains a hash code of its demographic information, a bag-of-words representation of the International Classification of Diseases (ICD) codes of the main diagnoses in the medical records, and statistical features of the distribution of its adjacent nodes in the knowledge graph. An alignment model based on graph attention network is adopted. The input is the feature vectors of two instance nodes to be judged, and the output is a probability value indicating that the two are the same object. When the probability value is greater than the preset alignment threshold, the two instance nodes are merged. The merging strategy includes migrating all unique attributes and relationship edges to a retained node, deleting the other node, and updating all links pointing to the deleted node. In the knowledge graph, a link prediction algorithm based on restricted Boltzmann machine is used to infer and complete the missing patient-drug relationship edges based on the existing patient-diagnosis-drug triplet relationship, thereby enhancing the integrity of the graph data.

5. The intelligent data processing method for clinical research according to claim 4, characterized in that, The construction of its multimodal feature vector includes: Their demographic information is hashed to generate privacy-preserving identifier features; Its main diagnostic information is converted into a bag-of-words representation of the International Classification of Diseases (ICD) code to generate diagnostic semantic features; The number and type distribution of various entity nodes directly connected in the knowledge graph are statistically analyzed to generate graph topological features; The privacy-preserving identifier features, diagnostic semantic features, and graph topology features are concatenated to form the multimodal feature vector.

6. The intelligent data processing method for clinical research according to claim 4, characterized in that, The missing patient-drug relationship edges are inferred and completed using a link prediction algorithm based on a restricted Boltzmann machine, including: The existing triplet relationships in the knowledge graph are used as training samples and input into the restricted Boltzmann machine model; By optimizing model parameters using a comparative divergence algorithm, the potential joint probability distribution among entities is learned. For any patient node and drug node pair, calculate their joint activation probability; When the joint activation probability is higher than the preset completion threshold, a new take-up relationship edge is created between the two, and the confidence attribute value of the relationship edge is set to the joint activation probability.

7. The intelligent data processing method for clinical research according to claim 1, characterized in that, The data quality verification and cleaning for performing the graph constraints includes: A set of data quality rules based on shape constraint language is predefined. These rules describe the logical constraints that the data should follow in the form of a graph pattern. The logical constraints include value type constraints, numerical range constraints, cardinality constraints, and association logical constraints. Launch a validation engine that executes the data quality rules on the enhanced clinical instance knowledge graph, traversing all instance nodes and their attributes and relationships, and outputting all data nodes that violate the constraints and their violation details; For the identified non-compliant data, an automated cleaning process is performed using a preset correction strategy. The correction strategy includes: for numerical anomalies, if the data deviates from three standard deviations of the distribution of similar data in the patient group, it is marked as a null value; for logical contradictions, such as gender not matching the diagnosis, the confidence attribute value of the diagnostic relationship is reduced, and the entry is recorded in the data audit log for manual review.

8. The intelligent data processing method for clinical research according to claim 7, characterized in that, The value type constraint stipulates that the data type of a specific attribute value must be an integer or a floating-point number; the numerical range constraint stipulates that the numerical value of a specific test item must be within the clinical normal range; the cardinality constraint stipulates that a patient entity can only be associated with one live state node at any given time; the association logic constraint stipulates that a patient entity diagnosed with a specific male disease cannot have a female gender attribute value.

9. The intelligent data processing method for clinical research according to claim 1, characterized in that, Based on pre-defined clinical research data specifications, a standardized analysis-ready dataset is extracted and generated from the enhanced clinical instance knowledge graph, including: Receive a clinical research data specification defined in the form of a structured query language, which specifies the inclusion and exclusion criteria for the research cohort, the list of data variables to be observed, and the data types and coding standards of the variables; The clinical research data specifications are automatically translated into a set of query statements in a graph query protocol language, and the inclusion and exclusion criteria are converted into pattern matching conditions for nodes and edges in the graph. Execute the query statement on the enhanced clinical instance knowledge graph to filter out a set of patient instance nodes that meet all inclusion criteria and do not meet any exclusion criteria; For each selected patient instance node, traverse along its relation edges, extract the specified attribute values ​​or associated node information from the data variable list, and convert them according to the coding standards in the specification. All extracted and transformed data are organized into a two-dimensional table structure, with patient instances as behaviors and data variables as columns, and exported as a comma-separated value file format or a columnar storage file format as the final analysis-ready dataset to be delivered.

10. The intelligent data processing method for clinical research according to claim 9, characterized in that, The clinical research data specifications are automatically translated into a set of query statements in a graph query protocol language, including: Parse the logical condition expressions in the structured query language; Each condition is mapped to a corresponding node type, attribute constraint, or relationship path in the knowledge graph. Multiple conditions are combined to generate a composite graph pattern query statement, which supports existence quantifiers and Boolean logic operators.

Images