Medical semantic ambiguity resolution method based on medical semantic network

Abstract

The invention relates to the field of medical artificial intelligence and natural language processing, in particular to a medical semantic ambiguity resolution method based on a medical semantic network. The method aims at solving the common problems of homophonic synonym, fuzzy expression, polysemy abbreviation, context uncertainty and the like in voice input of doctors or medical record texts. The method comprises the following steps: firstly, acquiring an initial medical record text through a medical special voice recognition module; secondly, constructing a multi-level and high-dimension medical semantic map by utilizing the comprehensive medical knowledge base; then, carrying out automatic detection on potential ambiguous vocabularies or expressions in the text, and generating a polysemy item candidate set from the semantic network; then, by integrating various modal data such as patient course information, historical medical records, examination reports, medical guidelines and current context logic, multi-dimensional feature vectors are constructed, and a deep semantic matching model is utilized to comprehensively score all candidate semantic items; finally, the system automatically selects the semantic item with the optimal score and replaces the semantic item into the original text, and the structured medical record which is clear in semantics, accurate and consistent with medical specifications is generated. According to the method, the accuracy, integrity and consistency of automatically generated medical records can be remarkably improved, clinical risks caused by ambiguity are effectively reduced, and a firm and reliable technical support is provided for applications such as intelligent medical record generation, clinical auxiliary decision making and medical scientific research data analysis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical information processing and natural language processing, and specifically relates to a medical semantic ambiguity resolution method based on a medical semantic network, aiming to improve the accuracy and reliability of automatic generation of electronic medical records. BACKGROUND

[0002] In the modern medical system, electronic medical records (EMR) have become a core component of clinical diagnosis and treatment. In order to improve the work efficiency of doctors, voice-driven electronic medical record generation systems have emerged, enabling doctors to quickly complete medical record recording through oral dictation. However, in actual clinical applications, the medical record text dictated or written by doctors often contains various forms of semantic ambiguity, which is due to the particularity of the medical field and is a pain point that existing general speech recognition and natural language processing technologies cannot solve.

[0003] Firstly, the confusion of homophonic words and near-homophonic words is a common problem in speech recognition. For example, in oral dictation, "renal failure" and "renal attenuation" are pronounced similarly, and the system may not be able to accurately distinguish between them, but the two represent completely different degrees of illness severity in clinical practice. "Inflammation" and "disorder" may also be misrecognized, the former is a medical term, and the latter may refer to "anorexia". This incorrect recognition may directly lead to serious deviations in medical record recording, and even affect subsequent diagnosis and treatment plans.

[0004] Secondly, ambiguous expressions and non-standard language are common in doctors' oral dictation. For example, a doctor may say "the index is a little high", which in clinical practice may refer to a variety of different examination items (such as blood sugar, blood pressure, white blood cell count, etc.), and the degree of its highness is not clear. "Heart function is not good" is also a typical ambiguous expression, which may refer to heart failure, heart failure, or even a specific stage (such as NYHA III). Traditional natural language processing methods often only perform literal matching and cannot understand the true clinical meaning behind these ambiguous expressions.

[0005] Furthermore, medical abbreviations and polysemous words are highly context-dependent. Many medical terms have their common abbreviations, such as "PT" can represent "Prothrombin Time" or "Physical Therapy". "CT" can refer to "Computed Tomography" or "Chemotherapy" in other contexts. The correct understanding of these abbreviations depends entirely on the context in which they are used, such as the patient's department, chief complaint, test results, medication history, etc. Existing technologies rely heavily on simple dictionary matching or local context analysis, lacking the ability to integrate multi-source information for global semantic reasoning, resulting in low accuracy of ambiguity resolution when facing complex and variable clinical scenarios.

[0006] In summary, the existing technology has the following main defects in dealing with medical semantic ambiguity: Dependence on a single data source: Most systems only rely on speech or text itself, failing to fully utilize key multi-modal information such as patient history, test reports, and imaging data, resulting in incomplete context understanding.

[0007] Lack of deep semantic understanding: Lack of in-depth application of medical knowledge graph, unable to understand the complex relationships between medical concepts (such as cause-symptom, drug-side effect, test-result), thus unable to conduct effective semantic reasoning.

[0008] Rule matching limitations: Over-reliance on pre-set rules or simple statistical models, making it difficult to cope with constantly changing medical terminology and individualized doctor expression habits. These problems directly affect the reliability and practicality of intelligent medical record systems, and may even pose potential risks to patient safety. Therefore, there is an urgent need for a new ambiguity resolution method that can effectively integrate multi-source data, deeply understand medical semantics, and adaptively learn. SUMMARY

[0009] TECHNICAL PROBLEM The present invention aims to solve the semantic ambiguity problem faced by existing technologies in medical speech recognition and electronic medical record automatic generation, specifically including: how to automatically identify and accurately resolve various forms of ambiguity such as homonymy, ambiguous expression, polysemous abbreviation, etc., so as to generate structured medical record text with clear semantics, accurate content and compliance with medical standards, ultimately improving the quality and efficiency of medical information processing. TECHNICAL SCHEME

[0010] To solve the above technical problems, the present application proposes a medical semantic ambiguity resolution method based on a medical semantic network. The method divides the entire process into multiple key steps, forming a complete ambiguity resolution workflow, which specifically includes: Step one: voice transcription and initial text acquisition The doctor uses a special microphone or mobile terminal to dictate the patient's medical history, diagnosis, treatment, etc.

[0011] After the system receives the voice signal, it uses a medical-specific speech recognition (ASR) module to transcribe it into an initial text. This ASR module is pre-trained using a large amount of medical field corpus (including different departments, different accents of doctor's voice) to ensure the preliminary recognition accuracy of professional terms.

[0012] At the same time, this module also needs to have real-time segmentation function, which can cut long dictation into meaningful short sentences or paragraphs according to voice pause, speed change and other characteristics, so as to facilitate subsequent text processing.

[0013] Step two: construction of medical semantic network The present application constructs a comprehensive medical semantic graph (Medical Knowledge Graph). This graph is based on a medical knowledge base and integrates multiple authoritative data sources, including but not limited to: Disease library: such as ICD-10 coding, common diseases, rare diseases, complications, etc.

[0014] Symptom library: pain, fever, cough, etc., and the association between symptoms is established.

[0015] Drug library: drug name, generic name, trade name, dosage form, dosage, usage and dosage, side effects, etc.

[0016] Laboratory examination library: laboratory examination items, normal value range, clinical significance of abnormal results, etc.

[0017] Medical guidelines and specifications: diagnosis and treatment guidelines, consensus, operation specifications, etc.

[0018] In this graph, the node (Node) represents various medical entities (such as "hypertension", "aspirin", "white blood cells", "angina pectoris", etc.), and the edge (Edge) represents the complex semantic relationship between entities (such as "causal relationship": hypertension leads to angina pectoris; "companion relationship": fever accompanied by cough; "treatment relationship": aspirin treats angina pectoris; "superior-inferior relationship": angina pectoris belongs to coronary heart disease). The construction of this semantic graph provides a strong foundation for subsequent semantic reasoning.

[0019] Step three: ambiguity detection and candidate generation The system performs natural language processing (NLP) on the initial medical record text, including word segmentation, part-of-speech tagging, etc.

[0020] A pre-trained ambiguity detection model is used to scan each word in the text. This model automatically identifies potential ambiguity points by comparing dictionaries, abbreviation libraries, and homonym libraries, mainly including: Homonymic polysemy: such as the ambiguity of "pneumonia" and "pneumonia".

[0021] Medical abbreviations: such as "PT", "CT", "MR", etc.

[0022] Ambiguous expressions: such as "a little high", "not too good", "obviously better", etc.

[0023] Once an ambiguous word is detected, the system extracts all possible candidate sets of polysemous items from the medical semantic network constructed in the second step based on the literal form or pronunciation of the ambiguous word. For example, after identifying "PT", the system generates the candidate set { "prothrombin time", "physical therapy"}.

[0024] Step four: Context modeling and multi-dimensional feature vector construction This is the most core step of the invention, aiming to provide sufficient context information for each candidate semantic item and perform accurate semantic matching. The system constructs a high-dimensional feature vector from the following multiple dimensions: Patient basic information: age, gender, occupation, etc. For example, if the patient is an elderly person, "PT" is more likely to refer to "physical therapy"; if the patient is a hemophilia patient, "PT" is more likely to refer to "prothrombin time".

[0025] Patient course information: chief complaint, history of present illness, past medical history, family history, etc. For example, if the patient complains of "uncontrolled bleeding after trauma", "PT" is more related to coagulation function.

[0026] Multi-source data fusion: This is the key innovation point of the invention. The system will real-time search and fuse the patient's test reports (such as blood routine, liver and kidney function), imaging reports (such as CT, X-ray), past medical orders, medication records, etc. structured data. For example, when the system identifies "PT", if it also finds "PT extension" or "INR value" indicators in the patient's test report, it greatly enhances the weight of the "prothrombin time" candidate.

[0027] Text context semantics: Use deep learning models (such as BERT, GPT, etc.) to model the text before and after the ambiguous word, capturing syntactic and lexical level semantic relationships.

[0028] The above multi-dimensional information is integrated into a unified feature vector as the input of the next step of deep semantic matching model.

[0029] Step five: deep semantic matching and scoring ranking The present application adopts a deep semantic matching model to evaluate the matching degree between each candidate meaning item and the multi-dimensional feature vector. The model can be an attention mechanism-based neural network that can automatically learn and weight information in different dimensions.

[0030] The model calculates the matching score of each candidate meaning item according to the feature vector. For example, in the case of "PT", if there is a related indicator in the test report, the model will give "prothrombin time" a higher score; on the contrary, if the medical record mentions words such as "rehabilitation training" and "physical therapy", the score of "physical therapy" will be higher.

[0031] Step six: semantic resolution and text generation According to the score given by the deep semantic matching model, the system selects the highest scoring semantic item as the final resolution result.

[0032] The system replaces the ambiguous expressions in the original text (such as "PT", "a little high") with standardized and accurate medical terms (such as "prothrombin time", "indicator higher than normal range", "heart function classification is NYHA Ⅲ grade").

[0033] Finally, the system formats the resolved text to generate a clear and complete structured electronic medical record for doctors to confirm and archive.

[0034] Advantages Compared with the prior art, the present application has the following significant advantages: 1. High-precision ambiguity resolution: By integrating multi-modal data, the present application can perform semantic reasoning from multiple dimensions, significantly improving the resolution accuracy of homonyms, ambiguous expressions and polysemous abbreviations, and fundamentally solving the difficulties faced by traditional methods.

[0035] 2. Enhancing the accuracy and consistency of medical records: The present application automatically converts non-standard and ambiguous spoken expressions into standard medical terms, ensuring the rigor and consistency of medical record content and reducing the risk of clinical misjudgment caused by inaccurate information.

[0036] 3. Improving clinical work efficiency: Doctors do not need to repeatedly speak or manually modify ambiguous content, and the system can automatically complete semantic clarification, greatly simplifying the medical record writing process and allowing doctors to devote more energy to diagnosis and treatment.

[0037] 4. Expanding application scenarios: The application not only can be applied to the automatic generation of electronic medical records, but also can be extended to medical research data extraction, clinical auxiliary decision-making, medical question and answer robots, and other fields due to its strong semantic understanding ability, and has a wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 Medical semantic ambiguity resolution method flow chart DETAILED DESCRIPTION

[0039] Example 1: Homonym resolution Scene: Internal medicine outpatient department, doctor dictating: "The patient has symptoms of renal failure and needs to be hospitalized immediately." Process: The ASR module recognizes "renal failure", but since it is similar in pronunciation to "renal attenuation", the system marks it as potential ambiguity.

[0040] Resolution: The system retrieves the patient's medical history and recent examination reports.

[0041] Context modeling: It is found that the patient has "chronic kidney disease" in the medical history, and the creatinine (Cr) and urea nitrogen (BUN) indicators in the latest examination report are significantly higher than the normal value.

[0042] Deep matching: In the semantic network, "renal failure" has strong association with "chronic kidney disease" and "elevated creatinine and urea nitrogen". "Renal attenuation" usually refers to mild decline in kidney function, which is not consistent with the current examination results.

[0043] Result: The system determines "renal failure" as the optimal semantics according to the strong association and outputs "the patient has symptoms of acute renal failure and needs to be hospitalized immediately".

[0044] Example 2: Fuzzy expression resolution Scene: Cardiology ward, doctor dictating: "The patient's heart function is not very good recently, and it is recommended to adjust the medication." Process: The system detects the fuzzy expression "heart function is not very good".

[0045] Resolution: Multi-source data fusion: The system retrieves the echocardiogram report of the patient and finds that the left ventricular ejection fraction (LVEF) is 35%. At the same time, combined with the patient's symptom description (such as shortness of breath and fatigue after exercise).

[0046] Semantic matching: The system uses the medical guideline atlas to match LVEF 35% with the NYHA classification of heart failure, and determines that the index and symptoms meet NYHA III.

[0047] Result: The system automatically replaces the ambiguous abbreviation with the standard medical term, and outputs "The patient's cardiac function is classified as NYHA Class III, and medication adjustment is recommended." Example 3: Ambiguous Abbreviation Resolution Scenario: Orthopedic clinic, doctor dictating: "The patient's PT and OT treatment plans have been determined." Process: The system recognizes the ambiguous abbreviation "PT" and the candidate set is { "Prothrombin Time", "Physical Therapy"}.

[0048] Resolution: Context Modeling: The system analyzes the context of the text and finds that "PT" appears simultaneously with "OT" (Occupational Therapy).

[0049] Multi-source Data Fusion: The patient's medical history is retrieved, and it is found that the patient's complaint is "post-fracture rehabilitation."

[0050] Deep Matching: In the semantic network, both "Physical Therapy" and "Occupational Therapy" are commonly used terms in the rehabilitation department, and their co-occurrence frequency is high. "Prothrombin Time" is a blood test item and is not related to the rehabilitation context.

[0051] Result: The system determines "Physical Therapy" as the optimal semantic and outputs "The patient's physical therapy (PT) and occupational therapy (OT) treatment plans have been determined." Example 4: Comprehensive Resolution in Complex Scenarios Scenario: Infectious disease ward, doctor dictating: "The patient's CRP value is very high, considering inflammation." Process: Ambiguity Detection: "CRP" is an ambiguous abbreviation (C-reactive protein, cardiopulmonary resuscitation); "inflammation" may be confused with "disgust" in speech recognition.

[0052] Resolution: Multi-source Data Fusion: The system retrieves the patient's blood test report and finds that the "C-reactive protein" indicator has abnormally high values; at the same time, it finds that the patient has symptoms of fever, white blood cell elevation, and other inflammation symptoms.

[0053] Semantic Matching: For "CRP": Combined with the test report, the system raises the weight of "C-reactive protein" to the highest.

[0054] For "inflammation": combined with the patient's fever and white blood cell elevation, the semantic network strengthens the concept of "inflammation" and excludes "disgust."

[0055] Result: The system generates "The patient's C-reactive protein (CRP) value is significantly elevated, combined with clinical symptoms, considering the presence of inflammation."

Claims

1. A medical semantic ambiguity resolution method, characterized in that, The method comprises the following steps: Obtaining doctor's voice input and transcribing it into initial medical record text; Building a medical semantic network based on a medical knowledge base, which includes medical entities and semantic relationships between them; Ambiguity detection on the initial text, identifying homophonic disambiguation, ambiguous expressions or polysemous abbreviations, and generating a candidate concept set for each ambiguous word; Combining the context semantics of the initial text with the patient's multi-source medical data (including but not limited to basic information, medical history, examination results, medical orders, etc.) to build a multi-dimensional feature vector for each candidate concept; Using a deep semantic matching model to score the multi-dimensional feature vector to rank the candidate concepts; Selecting the optimal semantics according to the scoring results and replacing the original ambiguous expression to generate the final medical text with clear semantics.

2. The method of claim 1, wherein the medical semantic network includes but is not limited to disease entities, symptom entities, drug entities, test index entities and diagnosis and treatment operation entities.

3. The method of claim 1, wherein the semantic relationships include cause-and-effect relationships, concomitant relationships, treatment relationships, hierarchical relationships and synonymous relationships.

4. The method of claim 1, wherein the multi-source medical data is obtained by real-time retrieval of electronic medical record system (EHR) and laboratory information system (LIS).

5. The method of claim 1, wherein the deep semantic matching model is a neural network model based on attention mechanism.

6. The method of claim 1, wherein the ambiguity detection includes methods based on dictionary matching, abbreviation library query and homophone library comparison.

7. The method of claim 1, wherein the generated medical text is structured electronic medical record text, and the structured text contains fields such as diagnosis, treatment and examination.

8. The method of claim 1, wherein the method further comprises a feedback learning module capable of continuously optimizing the ambiguity resolution model according to the doctor's subsequent modifications.

Images