A medical record data structuring processing method and system

Abstract

This invention relates to the field of medical information technology, and more particularly to a method and system for structuring medical record data. The method includes: constructing a preprocessing pipeline for text cleaning and temporal reorganization; utilizing a deep bidirectional representation parser to perform semantic parsing and logical nesting deconstruction; implementing context-aware medical entity recognition and attribute association; standardizing and normalizing terminology based on a dynamic knowledge graph; and establishing a logical consistency verification and associative reasoning mechanism. This invention's deep parsing and logical deconstruction of complex medical semantics significantly improves the accuracy and logical coherence of unstructured medical data transformation.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method and system for structuring medical record data. Background Technology

[0002] Electronic medical records are not only a digital mapping of clinical diagnosis and treatment activities, but also a core data carrier for medical knowledge discovery, clinical decision support, and regional health resource allocation. With the development of natural language processing technology, existing medical record data processing technologies, such as statistical machine learning methods like Hidden Markov Models and Conditional Random Fields, are widely used in named entity recognition tasks. These methods learn the feature vectors of the text to annotate key elements such as symptoms, diagnoses, and surgeries.

[0003] As medical applications demand higher data accuracy and logical consistency, the limitations of existing technologies become increasingly apparent when dealing with complex medical texts. Specifically, a complete medical record is not merely a collection of isolated entities, but rather encompasses the evolution of a doctor's diagnostic and treatment thought process. The same medical entity often carries entirely different clinical weights at different stages of diagnosis and treatment. Existing structured models tend to focus on local feature extraction, lacking an understanding of the overall logical chain of the medical record. This makes it difficult for the models to capture implicit causal relationships across paragraphs and time sequences. For example, when processing descriptive text, the inability to establish a dynamic mapping between symptom evolution and subsequent examinations and interventions easily leads to the loss of logical continuity in disease progression after the data is structured. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for structuring medical record data, in order to solve the technical problems of insufficient structuring accuracy and logical breaks caused by the strong context dependence and complex logical nesting of existing medical record texts.

[0005] One aspect of the present invention provides a method for structuring medical record data, comprising:

[0006] S1. Construct a medical record data preprocessing pipeline and use the preprocessing module to perform standardized cleaning and temporal reorganization of the original unstructured medical record text;

[0007] S2. Perform deep semantic parsing and logical nesting deconstruction, and use the semantic parsing module to extract deep semantic features from the medical record text; among them, multi-level semantic features are captured through a deep bidirectional representation parser, and syntactic tree units are constructed to locate core components and long-distance dependencies, and the logical judgment unit is activated to deconstruct negation logic, uncertainty logic and transition logic to determine the definite state of medical concepts.

[0008] S3. Implement context-aware medical entity recognition and attribute association, and use the multi-task learning recognition framework in the entity recognition module to achieve accurate localization and classification of medical entities; extract the modifier attributes related to each entity synchronously through the attribute extraction submodule, and use the position encoder to maintain the physical position association between the entity and its attributes, and combine the self-attention mechanism to capture the attribute supplementary information across sentences;

[0009] S4. Based on dynamic knowledge graphs, medical terminology is standardized and normalized, and a multi-level medical knowledge graph is mounted using a normalization module. Among them, the similarity is calculated by combining character overlap, semantic distance and departmental context through a semantic similarity matching engine, and a disambiguation processor is used to process abbreviations.

[0010] S5. Establish a logical self-consistency verification mechanism and use the quality control module to perform quality control on the structured results. Specifically, the temporal logic verification engine checks for temporal logic conflicts, the clinical rule matching engine performs entity dependency verification, and the coreference resolution unit handles pronoun reference problems. Finally, the verified data is encapsulated and output through the data encapsulation interface according to the preset medical data model.

[0011] In some embodiments, S1 further includes: performing a multimodal data alignment process:

[0012] When the original medical record includes semi-structured tabular data or discrete text after image text recognition, the multimodal data alignment module is used to strongly correlate the indicator items in the table with their corresponding values, units and reference ranges through spatial location mapping and keyword anchoring technology.

[0013] Cross-validate the associated table data with the text description to identify logical deviations between the indicator values ​​and the text descriptions, and record these logical deviations as key inconsistencies for subsequent processing by the quality control module.

[0014] In some embodiments, in S2, the deep bidirectional representation parser execution process includes at least:

[0015] During the training phase, a layered masking mechanism is adopted, which forces the model to learn the internal structural features of entities and external contextual dependencies by randomly masking individual characters and selectively masking complete medical entity phrases.

[0016] During the parsing phase, a gating mechanism integrated within the parser is used to dynamically adjust the attention given to contextual information at different distances based on the semantic complexity of the current parsed content.

[0017] The syntax tree construction unit implements a long word priority segmentation strategy based on the characteristics of Chinese medical language. Before constructing the syntax tree, word segmentation intervention is performed through a medical lexicon to prevent medical terms from being segmented into fragmented word groups.

[0018] In some embodiments, in S3, the context-aware medical entity recognition process further includes:

[0019] An integrated dynamic negative sampling module automatically generates similar-looking, similar-sounding, and ambiguous interference items as negative samples by analyzing historical false alarm data, thereby improving the model's ability to distinguish medical concepts during training.

[0020] For newly emerging or rare medical terms, an open-domain identification strategy is adopted, which combines the generation capabilities of language models to predict the part-of-speech and semantic category of the terms.

[0021] The attribute extraction submodule employs a multi-head attention mechanism, assigning independent attention heads to different types of entities. Specifically, the attention head for drug entities specifically searches for surrounding dosage and frequency words; the attention head for symptom entities specifically perceives location and degree words.

[0022] In some embodiments, in S4, the normalization process further includes:

[0023] The normalization module supports the coexistence of multiple versions of the terminology set and cross-standard mapping. By establishing a mapping matrix between different versions of the International Classification of Diseases and the medical terminology set, it enables real-time reorganization of structured data.

[0024] When performing term alignment, the semantic similarity matching engine introduces a topological structure reference to analyze the hierarchical position of candidate nodes in the knowledge graph. If the classification level of a node has a closer topological distance to the department ontology to which the medical record belongs, then the node is given a higher recommendation weight.

[0025] It provides an expert rule feedback mechanism, allowing clinical experts to correct mapping results online through a knowledge alignment interface, and adjust the weight parameters of the mapping strategy in real time based on the corrected samples.

[0026] In some embodiments, S5 further includes establishing a multidimensional associative reasoning mechanism, which includes:

[0027] A clinical decision support mapping layer is set up to compare structured medical record data with standard clinical pathways, automatically identify missing or abnormal items in the diagnosis and treatment process, and generate prompt information.

[0028] Construct a traceable evidence chain display interface, and label each generated structured field with its starting position in the original medical record, parsing path, matching knowledge graph node, and verification rule pass status;

[0029] The time-series logic verification engine performs a global sorting of all extracted time nodes. When a medical logic contradiction is found, it automatically corrects the error by retrieving the credibility score of the original text recognition, or marks the original text as having a questionable logic.

[0030] In some embodiments, the feature is that it further includes end-to-end privacy protection and robustness maintenance:

[0031] The privacy engine automatically identifies and masks patients' sensitive privacy information throughout the entire data structuring process.

[0032] The asynchronous message queue mechanism is used to manage the data flow between each step, and timestamp-based version control and global transaction management are implemented during the processing. When an exception occurs in any processing link, a rollback mechanism is executed to ensure the atomicity of structured processing.

[0033] Through an online learning loop, isolated words that cannot be directly mapped by a knowledge graph are periodically clustered and pushed to the expert review end. The annotation results are used as incremental training data and fed back to the entity recognition and normalization process.

[0034] Another aspect of the present invention provides a medical record data structuring processing system, comprising:

[0035] The preprocessing module is used to build a medical record data preprocessing pipeline to perform standardized cleaning and temporal reconstruction of the original unstructured medical record text; the preprocessing module integrates an encoding verifier, an interference information filter and a text segmentation engine.

[0036] The semantic parsing module, connected to the preprocessing module, is used to perform deep semantic parsing and logical nesting deconstruction to extract deep semantic features from medical record texts; the semantic parsing module is equipped with a deep bidirectional representation parser, a word vector mapping unit, a syntax tree construction unit, and a logical decision unit;

[0037] An entity recognition module, connected to the semantic parsing module, is used to perform context-aware medical entity recognition and attribute association; the entity recognition module includes a multi-task learning recognition framework, which includes a boundary determination submodule, an entity classification submodule, and an attribute extraction submodule;

[0038] The normalization module, connected to the entity recognition module, is used for the standardization processing of medical terms based on a dynamic knowledge graph. The normalization module is equipped with a distributed medical knowledge graph library and internally encapsulates a semantic similarity matching engine and an ambiguity disambiguation processor.

[0039] The quality control module, connected to the normalization module, is used to establish a logical self-consistency verification and multi-dimensional correlation reasoning mechanism; the quality control module consists of a temporal logic verification engine, a clinical rule matching engine, a coreference resolution unit, and a data encapsulation interface.

[0040] In other embodiments, the interference information filter in the preprocessing module adopts a dynamic feedback learning mechanism, which supports online updates and real-time distribution of filtering rules through the administrator terminal;

[0041] The logic decision unit internally stores a negation lexicon and uncertainty operators, which are used to limit the scope of logic on the dependency syntax tree and realize the measurement of the certainty of medical concepts.

[0042] The entity recognition module uses a position encoder to strongly couple the entity's modification attributes with the entity body, and uses a multi-head attention mechanism to allocate independent attention heads to improve the accuracy of attribute association.

[0043] The system also includes a privacy-de-identification engine, which automatically identifies and masks sensitive privacy information before data flows to an insecure domain.

[0044] In other embodiments, the preprocessing module and the semantic parsing module are deployed on a high-performance computing cluster.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention no longer stops at the extraction of surface-level words, but instead delves into the analysis of negation logic, uncertainty and nested clauses in medical records by constructing a deep bidirectional representation parser and a logic judgment module;

[0047] 2. By mounting a dynamically updated medical knowledge graph and combining semantic similarity and context disambiguation technology, this invention can achieve accurate mapping from extremely non-standard clinical colloquialisms and abbreviations to standardized terminology;

[0048] 3. This invention utilizes clinical medical logic to perform secondary verification of the processing results, eliminating logical noise and contradictions in the structured process through timing verification, dependency verification, and coreference resolution;

[0049] 4. The layered architecture, distributed computing, and asynchronous processing mode adopted in this invention enable it to support the structuring needs of tens of thousands of medical records per day in large hospitals. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the medical record data structuring method of the present invention;

[0052] Figure 2 This is a block diagram of the medical record data structuring processing system of the present invention. Detailed Implementation

[0053] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] Application Overview

[0055] Existing methods for structuring medical records have not fundamentally resolved the imbalance between surface-level semantic extraction and deep diagnostic logic reconstruction. Structuring medical record data is not merely an information extraction task, but a complex knowledge engineering process involving semantic disambiguation, evidence reasoning, and dynamic injection of knowledge graphs. In practical applications, the lack of a comprehensive processing framework that can simultaneously handle text grammar parsing, medical logic verification, and clinical reasoning simulation often results in fragmented information, untraceable logic, and distorted semantic weights in the structured data.

[0056] Therefore, how to construct a structured processing method and system with deep semantic perception, self-consistency of diagnosis and treatment logic, and dynamic knowledge alignment capabilities to address the implicit logical coupling in medical record texts, while ensuring extraction accuracy and realizing the transformation from unstructured text to structured knowledge with deep clinical logic, is an urgent problem that needs to be solved in the field of medical informatization.

[0057] Based on this, the method and system of the present invention are proposed, and the details are as follows.

[0058] Example 1

[0059] This invention provides a method for structuring medical record data. In practical engineering deployment, the method first constructs a medical record data preprocessing pipeline. The main function of this pipeline is to perform high-precision cleaning and logical temporal reorganization of the original unstructured medical record text. Specifically, after receiving the original text stream from a heterogeneous system in a medical institution, the pipeline first calls its internal encoding verifier. This verifier performs traversal verification of character-level encodings. Its working mechanism is to identify the Byte Order Mark (BOM) in the original stream and uniformly convert non-standard encodings, including but not limited to GBK, Big5, EUC-KR, or ISO-8859-1, into UTF-8 standard encoding. This step eliminates the risk of garbled characters in the subsequent semantic parsing process caused by encoding conflicts from the bottom layer.

[0060] With standardized coding, the interference filter in the pipeline begins to function. This filter is loaded with a pre-built cluster of regular expressions that can accurately anchor and remove redundant system information from medical record text that is irrelevant to the medical diagnosis logic. This redundant information includes, but is not limited to, system log tags generated by the database at the beginning and end of the document, illegal special escape characters generated by front-end page layout, whitespace placeholders commonly found in medical document systems, and redundant print control characters.

[0061] To ensure that the cleaning process does not inadvertently damage valid medical text, the filter introduces a contextual verification operator. Before determining a character as redundant information, its surrounding semantic environment is searched. If the character is located within a specific medical term, it is retained through conflict resolution logic. The cleaned text is further processed by a segmentation engine based on preset medical domain syntax rules. This engine does not simply segment punctuation marks; instead, it scans specific business keyword anchors and complex pragmatic patterns to segment the continuous text stream into sections with clear business attributes. These sections specifically cover core content such as chief complaint, present illness, past medical history, personal history, family history, physical examination, auxiliary examinations, diagnostic recommendations, and treatment plans. During the section segmentation process, not only are primary headings identified, but also substructures nested within sections are recognized through hierarchical perception logic, such as the specialist examination subsection under physical examination.

[0062] Based on the segmentation, refined time-series marking is applied to statements within each business segment. The recognition logic captures time words (e.g., May 10, 2023), relatively offset time words (e.g., three days ago), and their modified clinical action verb phrases (e.g., "appeared") in the text, establishing a linear time coordinate axis in the background based on admission time or first visit time. To handle common imprecise time descriptions in medical descriptions (e.g., "recently"), fuzzy time interval modeling technology is introduced, assigning a confidence-based time range to each discrete medical event, ensuring that all medical events can find their precise or reasonable location in the time series.

[0063] After basic preprocessing, deep semantic parsing and logical nesting deconstruction are performed. The deep bidirectional representation parser uses large-scale pre-trained contextual modeling techniques (such as the improved Transformer architecture) to perform multi-level semantic feature mapping on the text. In the lexical dimension, through high-dimensional vector space mapping, each medical term is transformed into a dense vector containing rich contextual relationships. These vectors not only represent the literal meaning of the words, but also encapsulate their medical part-of-speech features, positional offset information, and morphological features within the word in the medical context.

[0064] At the syntactic level, the parser analyzes the skeletal structure of sentences by constructing a dependency syntactic tree. The algorithm identifies the dominance relationships between words, locating the core predicate and its associated subject, object, modifier, and adverbial components. For complex, long sentences prevalent in medical records, the parser can penetrate multiple layers of modifiers to capture the dependency logic between the central word and its distantly defining components. For example, in a long sentence stating that a patient underwent internal fixation surgery at another hospital three years ago due to a fall that resulted in a left femoral neck fracture, the parser can accurately establish the causal logic between the surgery and the femoral neck fracture.

[0065] The logic determination module is launched simultaneously at this stage. It is responsible for deconstructing the negation logic, uncertainty logic, and transition logic in the medical record. This module scans a specific negation word library (including explicit negation such as "not seen" and implicit negation such as "deny") and limits the scope of negation words on the constructed dependency syntax tree, thereby accurately determining whether a certain symptom or sign belongs to the positive discovery reported by the patient or is clearly excluded.

[0066] For uncertain statements in the description (such as "suspected"), the logic judgment module will mark them as specific states to be verified and associate them with corresponding confidence weights. In addition, this deep bidirectional representation parser also integrates a hierarchical masking mechanism. This mechanism forces the model to learn the aggregate features inside the entity by selectively masking complete medical entity phrases during the training phase. Combined with the internal gating mechanism, it can dynamically adjust the attention weight to long-distance contextual information, effectively preventing semantic baseline drift during the parsing of long sentences.

[0067] Subsequently, this invention implements context-aware medical entity recognition and attribute association. In this process, a multi-task learning framework is used to achieve precise localization of medical entities. The scope of recognized entities is broad, including but not limited to various symptoms, disease names, names of complex surgeries, specific drug names, anatomical locations, laboratory test indicators, imaging results, microbial names, and genetic markers. While locking the entity boundaries, the entity recognition module simultaneously extracts strongly correlated modifier attributes through an attribute extraction head.

[0068] In practice, for symptom-related entities, the system automatically scans and extracts the exact location of the symptom, its severity, frequency, duration, and related triggering or relieving factors. For drug-related entities, the extraction focus shifts to the drug's single dose, frequency of administration, specific usage, route of administration, and treatment course information. The recognition process employs an advanced positional encoder to maintain a sense of physical distance between entities and attributes, and utilizes a self-attention mechanism to capture supplementary attribute information spanning multiple sentences or even paragraphs.

[0069] To further reduce the false alarm rate, the entity recognition process also integrates a dynamic negative sampling module. This module automatically generates similar-looking or similar-sounding medical concepts as interference items for comparative training by clustering analysis of historical false alarm cases, significantly improving the system's accuracy in distinguishing easily confused medical terms (such as atrial fibrillation and atrial flutter). For newly emerging long-tail medical terms, an open-domain recognition strategy is adopted, using semantic similarity to predict their possible semantic categories, ensuring dynamic coverage of cutting-edge medical descriptions.

[0070] After extracting the original entities, this invention employs a dynamic knowledge graph-based standardization and normalization process for medical terminology. Since clinical medical records contain numerous abbreviations, colloquial expressions, and non-standard abbreviations, the normalization module, by mounting a multi-level medical knowledge graph, achieves a precise mapping from the original expressions to a standardized terminology set. This knowledge graph deeply integrates ICD-10 (disease classification), ICD-11, SNOMED CT (medical terminology set), LOINC (laboratory indicators), and the domestically used surgical procedure classification codes (ICD-9-CM-3) and generic drug names.

[0071] Normalization logic determines the mapping relationship by calculating the multi-dimensional semantic similarity between the original entity and the candidate concept of the knowledge graph. This similarity calculation not only considers the literal overlap, but also takes into account the context of the department to which the patient belongs.

[0072] Finally, this invention establishes a logical self-consistency verification and multi-dimensional correlation reasoning mechanism to strictly control the quality of the generated structured data. The quality control module first performs temporal logic verification to check whether there are contradictions in the order of the past medical history, present medical history, and surgical history recorded in the medical records on the logical timeline. For example, the description of the present medical history of an acute attack of cholecystitis should not occur in the same timeline after cholecystectomy.

[0073] Subsequently, an inter-entity dependency check is performed, examining the rationality between the clinical medical diagnosis and the adopted drug treatment plan and auxiliary examination results. For example, if the diagnosis is bacterial pneumonia, logically it should be associated with a treatment plan using antibiotics and supporting auxiliary examinations such as elevated white blood cell count or chest X-ray infiltrates.

[0074] If the system detects a significant logical deviation, it will automatically trigger a correction procedure. This involves searching the global medical record context to try and find the entry error, or marking the entry as a high-risk anomaly for manual review. Coreference resolution technology is used at this stage to handle pronoun reference issues within the document, ensuring that the same entity can be accurately merged into the same logical node in different narrative contexts, avoiding entity redundancy.

[0075] Finally, the validated structured data is packaged according to a pre-set medical data model that conforms to the HL7FHIR standard or the domestic electronic medical record sharing document specification, generating a high-value dataset containing entity relationship graphs (RDF / JSON-LD), time-series evolution trajectories, and detailed attribute lists.

[0076] It should be noted that, at the hardware implementation level, the medical record data structuring processing system of the present invention adopts a highly scalable distributed computing architecture. The preprocessing module 100, semantic parsing module 200, entity recognition module 300, normalization module 400 and quality control module 500 are deployed as independent service components on a high-performance computing cluster, such as a Kubernetes container cloud.

[0077] To better understand this invention, the inventors provide a specific application example.

[0078] I received an original medical record from the gastroenterology department of a top-tier hospital: The patient underwent laparotomy (LC) at our hospital three years ago. Recently, he experienced severe upper abdominal colic, accompanied by nausea and vomiting, which was slightly relieved after self-administering scopolamine. Physical examination revealed mild scleral icterus. Preliminary diagnosis: bile duct stones? Admission for further examination is recommended.

[0079] Based on the foregoing information, the system performs the following tasks:

[0080] The preprocessing module 100 runs. The encoding verifier confirms that the text is UTF-8 encoded. The interference information filter identifies and deletes database printing tags at the beginning of the text. The text segmentation engine segments the text into four sections: past medical history, present medical history, physical examination, and diagnostic recommendations, by recognizing anchor points such as "three years ago," "recently," "physical examination," and "preliminary diagnosis." The time-series marking unit establishes a timeline: T1 (three years ago, surgery performed), T2 (recently, symptoms appeared), T3 (symptoms relieved after self-medication), T4 (time of physical examination and diagnosis).

[0081] The semantic parsing module 200 intervenes, and the deep bidirectional representation parser, through syntactic tree analysis, identifies the semantic center of "LC procedure" as "laparoscopic cholecystectomy" and recognizes that "accompanied by nausea and vomiting" is a symptom modification of "upper abdominal colic". The logic decision unit captures the question mark "?" uncertainty operator in the diagnostic item, and, combined with the context of "preliminary diagnosis", marks "bile duct stones" as "highly suspected / pending confirmation".

[0082] The entity recognition module 300 begins operation, identifying "LC procedure" as the surgical entity, "upper abdomen" as the anatomical location entity, "colic" as the symptom entity, "scopolamine" as the drug entity, and "scleral icterus" as the physical sign entity. The attribute extraction submodule 340 utilizes a multi-head attention mechanism to associate "severe" with the severity attribute of "colic," "slight relief" with the therapeutic attribute of "scopolamine," and "mild" with the degree attribute of "icterus."

[0083] The normalization module 400 started working, and the distributed medical knowledge graph database retrieved the standard term for "LC procedure" as "laparoscopic cholecystectomy," matching the ICD-9-CM-3 code 51.2300. "Bile duct stones" was mapped to the standard term "bile duct stones, not specifically identified," corresponding to the ICD-10 code K80.500. During the mapping process, the system identified "scopolamine" as an antispasmodic drug, consistent with common gastroenterology medications, confirming the accuracy of the mapping.

[0084] The quality control module 500 begins self-consistency verification. The timing logic verification confirms that the surgery (cholecystectomy) three years ago and the recent symptoms do not conflict on the timeline. Since the patient has undergone cholecystectomy, the "bile duct stones" in the current medical history is deemed highly medically plausible by the logic decision engine (possible common bile duct stones). The coreference resolution unit directs "its" or implied pronouns to this patient.

[0085] Finally, the system outputs a standardized JSON data packet, including a complete surgical history, symptom evolution trajectory, physical examination findings, preliminary diagnosis, and associated standard codes.

[0086] To further verify the technical superiority of this invention, a rigorous comparative example is also set up in this embodiment. The comparative example adopts a currently mainstream medical record processing system based on traditional rule engines, keyword matching, and a simple CRF (random field) model. The experimental sample consists of 200,000 real unstructured medical records accumulated by multiple departments of a tertiary hospital. These medical record samples have significant characteristics: many nested long sentences (average sentence length 45 characters), dense non-standard abbreviations (15 abbreviations per thousand characters), complex negation logic, and vague temporal descriptions.

[0087] Table 1. In-depth comparison of performance indicators between the embodiments of the present invention and the comparative examples.

[0088] Evaluation Dimensions Comparative model (rules + simple model) This invention (multi-dimensional framework) Increase Entity recognition accuracy 78.4% 96.7% +18.3% Entity recognition recall rate 75.1% 94.2% +19.1% Attribute association accuracy 65.8% 93.5% +27.7% Negation logic recognition accuracy 62.3% 98.1% +35.8% Uncertainty semantic annotation accuracy 58.9% 95.6% +36.7% Sensitivity to timing conflict detection 41.5% 92.8% +51.3% Term normalization matching degree 69.2% 97.4% +28.2% Cross-paragraph reference resolution accuracy 33.7% 91.5% +57.8% Redundant information filtering cleanliness 88.6% 99.5% +10.9% Data processing concurrency throughput 120 850 +608% Logical reasoning consistency score 55.4 94.6 +39.2

[0089] The quantitative data above clearly demonstrates that this invention achieves a qualitative leap in all core technical indicators by introducing deep semantic parsing and knowledge graph alignment. Particularly in negation logic recognition and uncertainty semantic annotation, this invention achieves extremely high accuracy rates of 98.1% and 95.6%, respectively. This directly addresses the industry pain point where traditional methods easily and incorrectly extract descriptions such as "no fever," "no cough," or "cannot rule out malignancy" as positive entities for fever, cough, or malignant tumors.

[0090] Meanwhile, the sensitivity of temporal conflict detection increased from 41.5% to 92.8%, meaning that the present invention can effectively capture logical errors or false records that may exist in medical records.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0092] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A medical record data structuring processing method and system, characterized in that, include: S1. Construct a medical record data preprocessing pipeline and use the preprocessing module to perform standardized cleaning and temporal reorganization of the original unstructured medical record text; S2. Perform deep semantic parsing and logical nesting deconstruction, and use the semantic parsing module to extract deep semantic features from the medical record text; among them, multi-level semantic features are captured through a deep bidirectional representation parser, and syntactic tree units are constructed to locate core components and long-distance dependencies, and the logical judgment unit is activated to deconstruct negation logic, uncertainty logic and transition logic to determine the definite state of medical concepts. S3. Implement context-aware medical entity recognition and attribute association, and use the multi-task learning recognition framework in the entity recognition module to achieve accurate localization and classification of medical entities; extract the modifier attributes related to each entity synchronously through the attribute extraction submodule, and use the position encoder to maintain the physical position association between the entity and its attributes, and combine the self-attention mechanism to capture the attribute supplementary information across sentences; S4. Based on dynamic knowledge graphs, medical terminology is standardized and normalized, and a multi-level medical knowledge graph is mounted using a normalization module. Among them, the similarity is calculated by combining character overlap, semantic distance and departmental context through a semantic similarity matching engine, and a disambiguation processor is used to process abbreviations. S5. Establish a logical self-consistency verification mechanism and use the quality control module to perform quality control on the structured results. Specifically, the temporal logic verification engine checks for temporal logic conflicts, the clinical rule matching engine performs entity dependency verification, and the coreference resolution unit handles pronoun reference problems. Finally, the verified data is encapsulated and output through the data encapsulation interface according to the preset medical data model.

2. The method of claim 1, wherein, S1 also includes: performing a multimodal data alignment process: When the original medical record includes semi-structured tabular data or discrete text after image text recognition, the multimodal data alignment module is used to strongly correlate the indicator items in the table with their corresponding values, units and reference ranges through spatial location mapping and keyword anchoring technology. Cross-validate the associated table data with the text description to identify logical deviations between the indicator values ​​and the text descriptions, and record these logical deviations as key inconsistencies for subsequent processing by the quality control module.

3. The method of claim 1, wherein, In S2, the execution flow of the deep bidirectional representation parser includes at least the following: During the training phase, a layered masking mechanism is adopted, which forces the model to learn the internal structural features of entities and external contextual dependencies by randomly masking individual characters and selectively masking complete medical entity phrases. During the parsing phase, a gating mechanism integrated within the parser is used to dynamically adjust the attention given to contextual information at different distances based on the semantic complexity of the current parsed content. The syntax tree construction unit implements a long word priority segmentation strategy based on the characteristics of Chinese medical language. Before constructing the syntax tree, word segmentation intervention is performed through a medical lexicon to prevent medical terms from being segmented into fragmented word groups.

4. The method of claim 1, wherein, In S3, the context-aware medical entity recognition process further includes: An integrated dynamic negative sampling module automatically generates similar-looking, similar-sounding, and ambiguous interference items as negative samples by analyzing historical false alarm data, thereby improving the model's ability to distinguish medical concepts during training. For newly emerging or rare medical terms, an open-domain identification strategy is adopted, which combines the generation capabilities of language models to predict the part-of-speech and semantic category of the terms. The attribute extraction submodule employs a multi-head attention mechanism, assigning independent attention heads to different types of entities. Specifically, the attention head for drug entities specifically searches for surrounding dosage and frequency words; the attention head for symptom entities specifically perceives location and degree words.

5. The method of claim 1, wherein, In S4, the normalization process further includes: The normalization module supports the coexistence of multiple versions of the terminology set and cross-standard mapping. By establishing a mapping matrix between different versions of the International Classification of Diseases and the medical terminology set, it enables real-time reorganization of structured data. When performing term alignment, the semantic similarity matching engine introduces a topological structure reference to analyze the hierarchical position of candidate nodes in the knowledge graph. If the classification level of a node has a closer topological distance to the department ontology to which the medical record belongs, then the node is given a higher recommendation weight. It provides an expert rule feedback mechanism, allowing clinical experts to correct mapping results online through a knowledge alignment interface, and adjust the weight parameters of the mapping strategy in real time based on the corrected samples.

6. The method of claim 1, wherein, S5 also includes establishing a multidimensional associative reasoning mechanism, which includes: A clinical decision support mapping layer is set up to compare structured medical record data with standard clinical pathways, automatically identify missing or abnormal items in the diagnosis and treatment process, and generate prompt information. Construct a traceable evidence chain display interface, and label each generated structured field with its starting position in the original medical record, parsing path, matching knowledge graph node, and verification rule pass status; The time-series logic verification engine performs a global sorting of all extracted time nodes. When a medical logic contradiction is found, it automatically corrects the error by retrieving the credibility score of the original text recognition, or marks the original text as having a questionable logic.

7. The method of claim 1, wherein, It also includes end-to-end privacy protection and robustness maintenance: The privacy engine automatically identifies and masks patients' sensitive privacy information throughout the entire data structuring process. An asynchronous message queue mechanism is used to manage the data flow between each step, and timestamp-based version control and global transaction management are implemented during the processing. When an exception occurs in any processing link, a rollback mechanism is executed to ensure the atomicity of structured processing. Through an online learning loop, isolated words that cannot be directly mapped by a knowledge graph are periodically clustered and pushed to the expert review end. The annotation results are used as incremental training data and fed back to the entity recognition and normalization process.

8. A medical record data structured processing system, implementing the method described in any one of claims 1 to 7, characterized in that, include: The preprocessing module (100) is used to build a medical record data preprocessing pipeline to perform standardized cleaning and temporal reorganization on the original unstructured medical record text; The preprocessing module (100) integrates an encoding verifier (110), an interference information filter (120), and a text segmentation engine (130). The semantic parsing module (200), connected to the preprocessing module (100), is used to perform deep semantic parsing and logical nesting deconstruction to extract deep semantic features from the medical record text; the semantic parsing module (200) is equipped with a deep bidirectional representation parser (210), a word vector mapping unit (220), a syntax tree construction unit (230), and a logical judgment unit (240). An entity recognition module (300) is connected to the semantic parsing module (200) and is used to implement context-aware medical entity recognition and attribute association. The entity recognition module (300) includes a multi-task learning recognition framework (310), which includes a boundary determination submodule (320), an entity classification submodule (330), and an attribute extraction submodule (340). The normalization module (400) is connected to the entity recognition module (300) and is used for the standardization processing of medical terms based on dynamic knowledge graph. The normalization module (400) is equipped with a distributed medical knowledge graph library (410) and internally encapsulates a semantic similarity matching engine (420) and an ambiguity disambiguation processor (430). The quality control module (500), connected to the normalization module (400), is used to establish a logical self-consistency verification and multi-dimensional association reasoning mechanism; the quality control module (500) consists of a temporal logic verification engine (510), a clinical rule matching engine (520), a coreference resolution unit (530), and a data encapsulation interface (540).

9. The system according to claim 8, characterized in that, The interference information filter (120) in the preprocessing module (100) adopts a dynamic feedback learning mechanism, which supports online updates and real-time distribution of filtering rules through the administrator terminal; The logic decision unit (240) internally stores a negation lexicon and an uncertainty operator, which are used to limit the scope of logic on the dependency syntax tree and realize the measurement of the certainty of medical concepts; The entity recognition module (300) uses a position encoder to strongly couple the entity's modification attributes with the entity body, and allocates independent attention heads through a multi-head attention mechanism to improve the accuracy of attribute association. The system also includes a de-privacy engine (600) for automatically identifying and masking sensitive privacy information before data flows to an insecure domain.

10. The system of claim 8 or 9, characterized in that, The preprocessing module (100) and the semantic parsing module (200) are deployed in a high-performance computing cluster.

Images