Deep learning large model-based medical record information extraction and analysis method and system

Abstract

The invention provides a medical record information extraction and analysis method and system based on a deep learning large model, and the method comprises the steps: carrying out the semantic understanding of a medical record text through a deep learning model, and recognizing key clinical features; extracting the key clinical features, and carrying out vectorization processing on the extracted key clinical features; performing sentence segmentation and word segmentation on the case text based on a rule engine and the key clinical features; adjusting the segmented words based on a symptom classification rule; and carrying out word frequency statistics on the adjusted segmented words. According to the method, accurate recognition and structured conversion of key clinical features can be ensured through a semantic understanding technology, knowledge graph construction and machine learning model optimization are supported, and a new view angle and a new tool are provided for exploration and research of diseases.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of artificial intelligence technology and medical big data analysis, and in particular to a method and system for extracting and analyzing medical record information based on a deep learning large model. Background Art

[0002] With the development of medical information technology, massive amounts of medical record data have become a valuable resource for medical research. However, traditional manual analysis methods are inefficient and time-consuming, making them difficult to meet the needs of large-scale data analysis. In the field of psychiatry, medical records are complex and subjective, and lack unified quantitative standards, making the extraction of clinical features and scientific analysis fraught with significant challenges.

[0003] While some existing natural language processing (NLP) tools can perform basic word segmentation and information extraction from medical records, these methods often rely on specific rules or small-scale data training, making them difficult to adapt to the complex scenarios of psychiatry. Furthermore, there is still a lack of research on the quantitative scoring and standardization of clinical characteristics, which seriously hinders the accurate diagnosis and scientific research of diseases.

[0004] In the field of psychiatric medical record processing, existing NLP tools have limitations due to their reliance on specific rules or small-scale data training. This is mainly due to the following four reasons:

[0005] 1. Domain complexity leads to insufficient rule coverage

[0006] Diverse terminology: Psychiatry encompasses over 30,000 specialized terms, such as "feeling of having one's thoughts read" and "emotional perversion," with numerous synonyms (e.g., "persecution delusion" and "delusion of relationship") and ambiguous expressions (e.g., "depression" and "anhedonia"). Traditional rule-based systems struggle to exhaustively capture all terminology variations and semantic associations. For example, a rule might fail to identify the entity "auditory hallucination" in the sentence "patient complains of hearing persistent condemning voices."

[0007] Semantic logical complexity: Medical records often contain multiple layers of nested logic (e.g., the negation "no obvious hallucinations" or the conditional statement "if extrapyramidal reactions occur, reduce medication"), making traditional rules difficult to parse. For example, a rule system might mistakenly ignore the negation in "deny suicidal ideation" and incorrectly extract the "suicidal ideation" entity.

[0008] 2. Data Labeling Dilemma Limits Model Training

[0009] High annotation costs: Annotation of psychiatric medical records requires professionals with a medical background (such as attending physicians). Since it is difficult to enumerate all terminology variants and semantic associations in medical records, repeated verification is required.

[0010] The quality of annotation varies: Due to the lack of uniformity in terminology, there is poor consistency between different annotators.

[0011] Data privacy restrictions: Psychiatric data involves sensitive information and is strictly regulated by regulations such as HIPAA, so research is often limited to a very small scope.

[0012] 3. Insufficient model generalization ability

[0013] Limitations of shallow models: Existing tools often use shallow models such as rule-based CRF or BiLSTM, which lack deep semantic understanding capabilities. For example, they cannot capture the multi-level relationships in "The patient exhibits significant catatonic stupor, meeting the diagnostic criteria for schizophrenia." Performance degrades significantly in long-distance dependency scenarios (such as symptom descriptions spanning multiple paragraphs).

[0014] Poor domain adaptability: Due to the uniqueness of the medical profession, traditional models perform poorly in medical scenarios.

[0015] Differences in medical record structure Most hospitals use a narrative structure with diverse content and poor consistency.

[0016] 4. Lack of Standardization of Quantitative Scoring

[0017] Commonly used clinical scales only score specific symptoms, such as anxiety and depression. These scales often fail to reflect the patient's overall medical history and disease progression. To understand the patient's overall condition, a diverse system of assessment criteria is needed.

[0018] Mental illnesses have a fluctuating course, such as: symptom changes in time series data (such as "depressive symptoms have gradually worsened in the past three months"), and the causal relationship between medication adjustments and symptom improvement (such as "the frequency of manic episodes decreased after combined use of sodium valproate"). Existing tools have difficulty capturing these changing information and are even more difficult to standardize.

[0019] Currently, the Deepseek large model has great advantages in semantic recognition and is expected to become an important tool for medical word segmentation, extraction and analysis. Related applications and research are also gradually being carried out, but compared with the highly specialized field of psychiatry, there is very little research. Summary of the Invention

[0020] In view of the above problems, the present invention is proposed to provide a technical solution to overcome the above problems or at least partially solve the above problems.

[0021] The present invention provides a medical record information extraction and analysis method based on a deep learning large model, the method comprising: using a deep learning model to perform semantic understanding of medical record text and identify key clinical features; extracting the key clinical features and vectorizing the extracted key clinical features, wherein the key clinical features include at least one of symptom descriptions, laboratory indicators, and medication records; performing sentence segmentation and word segmentation on the case text based on a rule engine and the key clinical features; adjusting the word segmentations based on symptom classification rules; and performing word frequency statistics on the adjusted word segmentations.

[0022] Optionally, the method further includes identifying the correlation between the quantified key clinical features, and constructing a knowledge graph based on the correlation, wherein the knowledge graph includes nodes and edges, wherein the nodes represent clinical features, and the edges represent the relationships between clinical features.

[0023] Optionally, the vectorization processing includes converting the extracted clinical features into a low-dimensional dense vector representation.

[0024] Optionally, the method further includes: fusing low-dimensional dense vectors of different types of clinical features to generate a unified feature vector representation.

[0025] Optionally, identifying the correlation between the quantified key clinical features includes: training a relationship extraction algorithm using a labeled medical data set; and using the trained relationship extraction algorithm to identify the correlation between the quantified key clinical features.

[0026] Optionally, the annotated medical data set includes medical records classified and annotated according to diseases, symptoms and / or drugs.

[0027] Optionally, the correlation between the quantified key clinical features is identified, including extracting clear regular relationships between clinical features directly through rule matching; for unclear regular relationships between clinical features, a predetermined machine algorithm is trained and then prediction is performed using the trained machine algorithm.

[0028] Optionally, the association between the key clinical features includes using a quadruple representation, i.e., {disease, symptom, sign, score}, or using a triple representation, i.e., {disease, symptom, treatment}, or using a triple representation, i.e., {disease, symptom, drug}.

[0029] Optionally, the method further includes: processing new medical record data regularly or in real time, identifying new key clinical features and new correlations between key clinical features, and updating the constructed knowledge graph.

[0030] The present invention also provides a medical record information extraction and analysis system based on a deep learning large model, which includes: a semantic understanding module, which is used to use a deep learning model to perform semantic understanding of medical record text and identify key clinical features; extract the key clinical features and vectorize the extracted key clinical features, wherein the key clinical features include at least one of symptom descriptions, laboratory indicators, and medication records; a word segmentation extraction module, which is used to perform sentence segmentation and word segmentation on the case text based on a rule engine and the key clinical features; a word segmentation adjustment module, which is used to adjust the word segmentation based on symptom classification rules; and a word frequency statistics module, which is used to perform word frequency statistics on the adjusted word segmentation.

[0031] The present invention has the following significant advantages:

[0032] Efficiency: Based on the deep learning big model, automatic word segmentation extraction and data mining of medical record information are realized, greatly improving processing efficiency.

[0033] Accuracy: Through semantic understanding technology, it ensures accurate identification and structured conversion of key clinical features.

[0034] Standardization: Design a unified scoring system to provide quantitative standards for clinical feature analysis to facilitate scientific research and clinical application.

[0035] Intelligence: Supports knowledge graph construction and machine learning model optimization, providing new perspectives and tools for disease exploration and research.

[0036] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above technical solution of the present invention and its objectives, features and advantages more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0038] Figure 1 shows a technical module architecture diagram of the present invention;

[0039] Figure 2 The workflow of the knowledge graph construction engine is shown;

[0040] Figure 3 The flowchart of the medical record information extraction and analysis method based on deep learning large model proposed by the present invention is shown

[0041] Figure 4 The process of optimizing word segmentation in the present invention is shown. DETAILED DESCRIPTION

[0042] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0043] Psychiatric medical records are characterized by over 85% unstructured text (chief complaint, current medical history, psychiatric examination, etc.), complex terminology (such as "clouded thoughts" and "emotional inversion"), and a large number of subjective descriptions (such as "the patient reported feeling fear"). Clinical assessment criteria have subjective differences (such as HAMD scale scores), which have always been difficult to automate, standardize, and quantify. Current medical data analysis models do not involve psychiatric medical record data.

[0044] In view of the unstructured characteristics of psychiatric medical records, such as Figure 1 As shown, the underlying technology of the present invention is based on the locally deployed Deepseek large language model 10, which has 32 billion parameters and can perform very detailed semantic analysis of text. Using Deepseek R132B can break through the terminology barriers of medical record data. Through pre-training learning combined with a dynamic domain dictionary mechanism, it can accurately identify ambiguous expressions, polysemous words, etc., and can perform in-depth thinking analysis, and can also automatically discover new words. The Deepseek R132B model captures long-distance dependencies through the Transformer architecture (supporting 4096 tokens context), and can implement complex logical analysis, such as processing nested negation "no clear..., but there is..., deny..." and causal relationship "because..., therefore, use..." and other logical analysis. It can also intelligently analyze time series and establish the logic of medical history development based on time series. The present invention can also parse other unstructured data besides text, which requires the underlying deep learning model to be able to perform multimodal fusion, that is, to integrate unstructured data such as voice and image.

[0045] In this paper, Deepseek's large language model was trained using 300GB of psychiatric literature (including DSM-5 / ICD-10 terminology), enabling accurate semantic understanding and reasoning of complex professional data in medical records. In local deployments, the local A100 single card throughput reaches 280 tokens / s and supports dynamic switching between mixed-precision INT4 and FP16, achieving a balance between deep optimization capabilities and efficiency in the medical field.

[0046] In order to solve the technical problems of extracting psychiatric medical record information and quantifying the scores based on the extracted information, it is necessary to process and adjust the output of the underlying DeepSeek R132B large model, such as Figure 1 As shown, the underlying DeepSeek R132B large model of the domain adaptation layer 20 of the present invention performs domain adaptation, including professional terminology adaptation, context understanding enhancement, and task-oriented fine-tuning.

[0047] Terminology adaptation: The DeepSeek R132B large model integrates a dictionary of psychiatric terminology, covering symptoms, medications, diagnoses, examination methods, and other aspects. For example, descriptions of psychiatric-specific symptoms such as "racing thoughts" and "obsessive-compulsive ideas," as well as the names of commonly used psychiatric drugs such as "clozapine" and "quetiapine." This dictionary will be updated through fine-tuning training to incorporate newly emerging professional terms, such as the names of new antidepressants. Terminology mapping and normalization: Since there may be multiple ways of expressing the same concept in psychiatric medical records, the domain adaptation layer also needs to perform terminology mapping and normalization. For example, expressions such as "depression," "frustration," and "bad mood" will be uniformly mapped to the standard term "depression" to facilitate subsequent information extraction and analysis.

[0048] Enhanced contextual understanding: Clinical context modeling is performed within the DeepSeek R132B large model. This model leverages prior knowledge in the field of psychiatry and contextual information from medical records to construct a clinical context model. This model helps understand the meaning of words and sentences in a text within a specific clinical context. For example, when describing a patient's hallucination symptoms, context can accurately determine whether the symptoms are auditory, visual, or other types of hallucinations, as well as information such as the frequency and severity of the hallucinations.

[0049] Task-oriented fine-tuning: Based on the specific tasks of extracting and quantitatively scoring psychiatric medical record information, the parameters of the DeepSeek R132B large model are fine-tuned to make the model perform better in tasks such as entity recognition, relationship extraction, and score prediction. For example, during the training process, additional labeled data for specific psychiatric entities and relationships is added to train the model to improve the model's ability to recognize and process this key information. Feedback mechanism and continuous optimization: A feedback mechanism is established to collect feedback from doctors during actual use and continuously optimize the domain adaptation layer. Doctors can evaluate and correct the output results of the model, and this feedback data will be used to update the model parameters and adjust the adaptation strategy to improve the accuracy and practicality of the model.

[0050] like Figure 1As shown in Figure 3, a knowledge graph construction engine 30 is constructed based on the domain-adapted large model. The knowledge graph construction engine 30 is responsible for converting the information extracted from the medical record text into a structured knowledge graph, providing strong support for subsequent quantitative scoring and scientific research analysis. Figure 2 As shown, the knowledge graph construction engine completes entity recognition and extraction, entity relationship extraction and representation, and dynamic update and maintenance of the knowledge graph. Detailed descriptions are given below.

[0051] Entity Recognition and Extraction 31

[0052] By leveraging information processed by the domain adaptation layer, combined with deep learning and rule matching, high-precision entity recognition can accurately identify entities such as diseases, symptoms, medications, examination items, and treatment plans in medical records. For example, entities such as "schizophrenia," "delusion," "risperidone," and "EEG" can be identified from medical record text.

[0053] Classify and label the identified entities, assigning each entity a corresponding type label. For example, label "schizophrenia" as a "disease" type, "delusion" as a "symptom" type, and "risperidone" as a "drug" type. This classification and labeling facilitates subsequent knowledge graph construction and querying.

[0054] Relation Extraction and Representation 32

[0055] Define the logical relationships between disease and symptoms, disease and signs, etc., and gradually achieve a comprehensive description of the relationships between entities.

[0056] Relationship extraction is performed using a combination of rule-based and deep learning. For clear rule-based relationships, such as "using a certain drug to treat a certain disease," they can be directly extracted through rule matching. For more complex relationships, machine learning algorithms are used for training and prediction.

[0057] The extracted relationships are represented as quadruple pairs, such as {disease, symptom, sign, score}. These relationships are stored in a graph database, forming the underlying structure of the knowledge graph. For example, in the knowledge graph, there is a "disease-symptom" relationship between "schizophrenia" and "delusion."

[0058] Dynamic update and maintenance of knowledge graph33

[0059] As new medical records are constantly generated and medical knowledge is constantly updated, the knowledge graph needs to be updated. The knowledge graph construction engine processes new medical record data regularly or in real time, extracts new entities and relationships from it, and adds them to the knowledge graph.

[0060] Knowledge graphs can be used for reasoning and querying, providing support for clinical diagnosis and treatment decisions. For example, doctors can query knowledge graphs to learn about common symptoms of a disease. They can also use knowledge graphs to reason and predict the progression of a patient's condition.

[0061] Knowledge graphs provide a rich data resource for scientific research and analysis. By mining and analyzing the entities and relationships within knowledge graphs, researchers can explore the pathogenesis, symptoms, and influencing factors of mental illness. For example, analyzing the relationship between different symptom combinations and disease subtypes can provide a basis for accurate diagnosis and personalized treatment.

[0062] Based on the construction of knowledge graph, Figure 1 As shown, the framework application layer implements the quantitative scoring module 40.

[0063] Information extraction, analysis, and quantitative visualization of psychiatric medical records are achieved based on the architecture of the above four modules, namely the locally deployed Deepseek large language model, domain adaptation layer, knowledge graph construction engine, and quantitative scoring module.

[0064] One aspect of the present invention, as Figure 3 As shown, a method for extracting and analyzing medical record information based on a deep learning large model is provided, which includes:

[0065] S1. Use deep learning models to understand the semantics of medical record text and identify key clinical features;

[0066] S2. Extracting the key clinical features and performing vector quantization on the extracted key clinical features, wherein the key clinical features include at least one of symptom description, laboratory index, and medication record;

[0067] S3. Segment the case text into sentences and words based on the rule engine and the key clinical features;

[0068] S4. Adjust the segmentation based on the symptom classification rules;

[0069] S5. Perform word frequency statistics on the adjusted word segmentation.

[0070] The present invention achieves standardization of clinical characteristics and standardized scoring by performing automated word segmentation extraction, data mining, and knowledge graph construction on medical record texts, especially psychiatric medical record texts, and can support scientific research and exploratory research on diseases.

[0071] In this method, this technical solution is mainly divided into four main steps: deep learning-assisted entity information extraction, rule engine optimization of word segmentation, word segmentation adjustment based on symptom classification, and word frequency statistics and result output. The deep learning model can perform semantic understanding of the text and provide a basis for the extraction of key information; and the rule engine can combine professional knowledge and established rules in the medical field to optimize the word segmentation results. The two complement each other and can improve the accuracy and efficiency of information extraction.

[0072] Before using a deep learning model to perform semantic understanding on medical record text, the medical record data is read into memory and initially processed. Generally, medical record data is stored as text files in a designated folder. As a specific implementation, the present invention iterates through all text files in the folder and uses Python's OS module and file reading functions to load the medical record text into memory. This approach is simple and direct, and can easily process large numbers of medical record files.

[0073] According to the structural characteristics of the medical records, a series of key parts that need to be extracted are defined, such as name, admission time, gender, discharge time, etc. As a specific embodiment, the present invention uses regular expressions to match each medical record text and extract the content of the corresponding part. Regular expressions have powerful text matching capabilities and can quickly and accurately locate and extract information with a specific format. For example, for the "name" part, the regular expression r"name: (.*?)(?=\n|$)" can be used to match the name information. In the extracted text information, information such as date, time, blood pressure, heart rate, and body temperature is not directly helpful for the extraction of key clinical features, but will increase the complexity of the processing. Therefore, regular expressions are used to remove this information from the text. This can reduce the impact of noise data on subsequent processing and improve processing efficiency and accuracy.

[0074] After denoising, a deep learning model is used to perform semantic understanding of psychiatric medical records, identifying key clinical features (such as symptom descriptions, laboratory indicators, medication records, etc.). The pre-trained deep learning model is then used to extract entity information from the medical records, and the rule engine is used to further optimize the word segmentation results.

[0075] The rule engine is implemented based on the aforementioned domain adaptation layer, dynamically updates the domain dictionary, integrates the DSM-5 / ICD-10 terminology library (containing 12,000+ professional words), and can use professional knowledge and established rules in the medical field to fine-tune the word segmentation results. Combined with the output of the deep learning model, the advantages of both can be fully utilized, utilizing both the semantic understanding ability of deep learning and the professionalism and flexibility of the rule engine to achieve more accurate word segmentation and information extraction. After actual testing, the present invention has an F1 value of 92.3% in context-aware entity recognition (15% better than the traditional CRF model), and in semantic similarity calculation, the cosine similarity threshold is 0.85 (supports fuzzy matching).

[0076] like Figure 4 As shown, the process of optimizing word segmentation includes: S21. Sentence segmentation and negation word processing, S22. Word segmentation and classification replacement, S23. Optimization combined with deep learning results, etc. Each step is explained in detail below.

[0077] During the sentence segmentation and negation processing steps, the extracted content is segmented, and the text is divided into multiple sentences according to delimiters such as commas, periods, and semicolons. At the same time, considering the importance of negation in medical records, the logical patterns of its occurrence are summarized and corresponding processing rules are determined. For sentences containing negation, the sentences are processed according to specific rules to avoid semantic errors during word segmentation. For example, if a negation is followed by multiple parallel symptom descriptions, the negation is added before each symptom description to accurately express the semantics.

[0078] During the word segmentation and category replacement steps, symptom classification rules are stored in JSON files. The JSON format is highly readable and extensible, making it convenient for storing and managing complex classification rules. The JSON file defines various symptoms and their corresponding categories, providing a basis for category replacement in the word segmentation results.

[0079] For each segmented word, check whether it exists in the classification rules defined in the JSON file. If so, replace the segmented word with the corresponding category; if not, retain the original segmented word. This classification replacement method can unify and standardize segmentation results, reduce vocabulary diversity, and facilitate subsequent statistics and analysis.

[0080] During the optimization step, the entity information extracted by the deep learning model is combined with the word segmentation results from the rule engine. Key entities identified by the deep learning model undergo special processing during the word segmentation process to ensure their accurate extraction and classification. Furthermore, the rule engine, drawing on medical expertise and experience, can correct and supplement potential errors in the deep learning model, further improving word segmentation accuracy.

[0081] After the optimization analysis, word frequency statistics can be performed. As a specific implementation method, Python's Counter class is used to perform word frequency statistics on the word segmentation results. The total word frequency of each word in all files, the word frequency of each word in each part, and the average word frequency of each word in each part can be counted separately. Word frequency statistics can help us understand the frequency of occurrence of key information in medical records and provide data support for subsequent analysis and research. The statistical results can be organized into a Pandas DataFrame object and sorted in descending order by average word frequency. Finally, the results of each part are saved to different worksheets in the Excel file, which makes it easy to view and analyze the frequency of occurrence of each word in different parts, providing intuitive information for doctors and researchers.

[0082] Based on the optimization analysis, data mining and knowledge graph construction are carried out. First, the extracted clinical features are vectorized to generate a low-dimensional dense representation. Then, a relation extraction algorithm (such as RE-BERT) is used to identify the correlation between clinical features and construct a clinical knowledge graph.

[0083] The extracted clinical features are vectorized, such as Figure 4 As shown, specifically including:

[0084] 1. Feature extraction: Use deep learning models to extract key clinical features from medical record text after semantic understanding and word segmentation optimization.

[0085] 2. Vectorized representation: Convert the extracted clinical features into a low-dimensional, dense vector representation. Dimensionality reduction algorithms (such as principal component analysis (PCA) and linear discriminant analysis (LDA)) can be used to reduce the high-dimensional vectors while retaining important information.

[0086] 3. Feature fusion: The vectors of different types of clinical features (such as symptom descriptions, physical signs, medication records, etc.) are fused to form a unified feature vector representation.

[0087] Identify the correlation between clinical features and build a clinical knowledge graph, including:

[0088] 1. Relationship Extraction Algorithm Selection: Select an appropriate relationship extraction algorithm (such as RE-BERT) to identify associations between clinical features. RE-BERT is a relationship extraction algorithm based on the BERT model. It leverages the BERT model's powerful semantic understanding capabilities to accurately identify relationships between entities. Other relationship extraction algorithms can also be selected, and this invention is not limited to this.

[0089] 2. Relationship extraction training: Use the labeled dataset to train the relationship extraction algorithm. During the training process, adjust the algorithm parameters to enable it to more accurately identify the relationship between clinical features.

[0090] 3. Knowledge graph construction: Based on the results of relationship extraction, a clinical knowledge graph is constructed. The knowledge graph consists of nodes and edges. Nodes represent clinical features (entities), and edges represent relationships between entities.

[0091] Identifying the correlations between the quantified key clinical features includes: training a relationship extraction algorithm using a labeled medical dataset; and identifying the correlations between the quantified key clinical features using the trained relationship extraction algorithm. The labeled medical dataset includes medical records categorized and labeled by disease, symptom, and / or medication.

[0092] For clear regular relationships between clinical features, they are directly extracted through rule matching; for unclear regular relationships between clinical features, a predetermined machine algorithm is trained and then used to make predictions.

[0093] The associations between the key clinical features can be represented using a quadruple, i.e., {disease, symptom, sign, score}, a triple, i.e., {disease, symptom, treatment}, or a triple, i.e., {disease, symptom, drug}. In the present invention, quadruple representation is preferred. The present invention supports 10 types of logical relationship reasoning, which are not listed here one by one.

[0094] The present invention has a dynamic update mechanism for the knowledge graph, and the method also includes: processing new medical record data regularly or in real time, identifying new key clinical features and new correlations between key clinical features, and updating the constructed knowledge graph.

[0095] In the present invention, combined with the opinions of clinical experts, a multidimensional scoring standard system is formulated (such as the severity of symptoms is divided into "mild", "moderate" and "severe" and scored), and current medical history and mental examination evaluation standards are formulated. The multidimensional scoring standard includes 11 major dimensions such as cognition / emotion / willpower in the basic dimension, and indicators are subdivided under each dimension. For example, there are 13 core symptom scores under the dimension of self-awareness. The scoring standard system is provided with a comprehensive index, namely the disease severity index (DSI value), and quantitative scores are calculated according to each dimension and its subdivided indicators. The quantitative scoring model (based on the above-mentioned multidimensional scoring standard system) is trained using expert-annotated data. After training, it is corrected in real time based on the difference feedback between the doctor's score and the system score. Generally speaking, if the difference between the doctor's score and the system score is less than 5%, it is considered that the system does not need to be corrected. The quantitative scoring model uses a machine learning module to deeply learn the results of multiple automated scorings before, so as to optimize and calibrate the scoring model to ensure the accuracy and consistency of the scoring.

[0096] The present invention also provides a medical record information extraction and analysis system based on a deep learning large model, which includes: a semantic understanding module, which is used to use a deep learning model to perform semantic understanding of medical record text and identify key clinical features; extract the key clinical features and vectorize the extracted key clinical features, wherein the key clinical features include at least one of symptom descriptions, laboratory indicators, and medication records; a word segmentation extraction module, which is used to perform sentence segmentation and word segmentation on the case text based on a rule engine and the key clinical features; a word segmentation adjustment module, which is used to adjust the word segmentation based on symptom classification rules; and a word frequency statistics module, which is used to perform word frequency statistics on the adjusted word segmentation.

[0097] The system also includes a knowledge graph construction module for identifying the correlation between quantified key clinical features and constructing a knowledge graph based on the correlation. The knowledge graph includes nodes and edges, wherein the nodes represent clinical features and the edges represent the relationships between clinical features.

[0098] The system also includes a quantitative scoring module for quantitatively evaluating the severity of the user's symptoms based on the segmented words and related knowledge graphs extracted from the user's disease data.

[0099] The system achieves efficient and accurate clinical feature extraction and standardized scoring by combining natural language processing, data mining and machine learning technologies.

[0100] The present invention avoids the risk of uploading data to the cloud by deploying the Deepseek 32B model locally, ensures the security and privacy of patient medical record data, and guarantees data security.

[0101] Large deep learning models were pre-trained using data containing 300GB of psychiatric literature (including DSM-5 / ICD-10 terminology). The local A1 card boasts a throughput of 280 tokens per second and supports dynamic switching between INT4 and FP16. This balances deep optimization capabilities and efficiency in the medical field, enabling high inference speeds for deep learning models, enabling rapid processing of large amounts of medical records and delivering high performance.

[0102] This paper solves complex logical parsing problems such as nested negation and long-distance dependencies in psychiatric medical record texts by utilizing the Transformer architecture of a deep learning model to capture long-distance dependencies. This paper improves the accuracy of medical record information extraction and quantitative scoring by improving the context-aware entity recognition and medical term mapping normalization of large models, fine-tuning the model, and optimizing the algorithm.

[0103] The system of the present invention has strong scalability, flexible system architecture design, is easy to expand and upgrade, and new functional modules can be added according to actual needs.

[0104] In terms of the above-mentioned innovative technical means and effective technical effects, the present invention has very good and wide application scenarios, such as being used by doctors for clinical diagnosis and formulation of treatment plans, and providing data and information support for scientific research. First, for doctors, it assists doctors in quickly and accurately diagnosing patients' conditions, provides quantitative scoring references, and improves the accuracy and efficiency of diagnosis. Doctors can more easily formulate personalized treatment plans based on the patients' quantitative scores and medical records. Secondly, for scientific researchers, the present invention supports large-scale psychiatric case analysis and research, which is conducive to exploring the pathogenesis and treatment rules of the disease, and provides strong data and information support for scientific research.

[0105] The above-mentioned system proposed by the present invention can realize mobile terminal applications in an offline environment, parse a large amount of medical record data and construct knowledge graphs in an offline manner, thereby enabling edge computing deployment.

[0106] In the description provided herein, numerous specific details are described. However, it is understood that embodiments of the present invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0107] Similarly, it should be understood that in order to streamline the present invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Accordingly, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of the present invention.

[0108] It should be noted that the above embodiments illustrate rather than limit the present invention, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between brackets shall not be construed as limiting the claims.

Claims

1. A method for extracting and analyzing medical record information based on a deep learning large model, characterized in that: The method includes: Using a deep learning model to perform semantic understanding of medical record text and identify key clinical features; extracting the key clinical features and performing vectorization on the extracted key clinical features, wherein the key clinical features include at least one of symptom descriptions, laboratory indicators, and medication records; Perform sentence segmentation and word segmentation on the case text based on the rule engine and the key clinical features; Adjust the segmentation based on symptom classification rules; Perform word frequency statistics on the adjusted word segmentations.

2. The medical record information extraction and analysis method based on deep learning large model according to claim 1 is characterized in that: The method also includes identifying the correlation between the quantified key clinical features, and constructing a knowledge graph based on the correlation, wherein the knowledge graph includes nodes and edges, wherein the nodes represent clinical features and the edges represent the relationships between clinical features.

3. The medical record information extraction and analysis method based on deep learning large model according to claim 1 is characterized in that: The vectorization process includes converting the extracted clinical features into low-dimensional dense vector representations.

4. The medical record information extraction and analysis method based on deep learning large model according to claim 3 is characterized in that: The method also includes: fusing low-dimensional dense vectors of different types of clinical features to generate a unified feature vector representation.

5. The medical record information extraction and analysis method based on deep learning large model according to claim 2 is characterized in that: Identifying the correlation between the quantified key clinical features, including: using a labeled medical data set to train a relationship extraction algorithm; and using the trained relationship extraction algorithm to identify the correlation between the quantified key clinical features.

6. The method for extracting and analyzing medical record information based on a deep learning large model according to claim 5 is characterized in that: The annotated medical data set includes medical records classified and annotated according to diseases, symptoms and / or drugs.

7. The medical record information extraction and analysis method based on deep learning large model according to claim 2 is characterized in that: Identify the correlation between quantified key clinical features, including extracting clear regular relationships between clinical features directly through rule matching; for unclear regular relationships between clinical features, train a predetermined machine algorithm and then use the trained machine algorithm to make predictions.

8. The medical record information extraction and analysis method based on deep learning large model according to claim 1 is characterized in that: The associations between the key clinical features include being represented by a quadruple, ie, {disease, symptom, sign, score}, or being represented by a triple, ie, {disease, symptom, treatment}, or being represented by a triple, ie, {disease, symptom, drug}.

9. The medical record information extraction and analysis method based on deep learning large model according to claim 2 is characterized in that: The method also includes: processing new medical record data regularly or in real time, identifying new key clinical features and new correlations between key clinical features, and updating the constructed knowledge graph.

10. A medical record information extraction and analysis system based on a deep learning large model, characterized in that: The system includes: A semantic understanding module is used to use a deep learning model to perform semantic understanding on medical record text and identify key clinical features; extract the key clinical features and perform vector processing on the extracted key clinical features, wherein the key clinical features include at least one of symptom description, laboratory index, and medication record; A word segmentation and extraction module, used to perform sentence segmentation and word segmentation on the case text based on the rule engine and the key clinical features; A word segmentation adjustment module, used to adjust word segmentations based on symptom classification rules; The word frequency statistics module is used to perform word frequency statistics on the adjusted word segmentations.

Images