Intelligent hypertension diagnosis system and method based on information driving

Abstract

The invention discloses an intelligent hypertension diagnosis system and method based on information driving, and relates to the technical field of medical artificial intelligence. Comprising an entropy minimization retrieval enhancement generation module used for constructing and optimizing a hypertension knowledge base and realizing efficient retrieval enhancement, and a zero-sample task-oriented dialogue module used for realizing high-quality multi-round information collection without a large amount of dialogue training data, the multi-stage training module is used for realizing structured hypertension medical field knowledge training; the hypertension professional decision logic module is used for realizing deep clinical decision based on hypertension field professional knowledge; the method has the advantages that the diagnosis accuracy is high, and the method is superior to human doctors and other models; the knowledge base is optimized, so that the storage and retrieval time is greatly shortened, and the performance is improved; the dialogue ability is strong, multi-round interaction can be realized, and the information acquisition rate and consistency are high; the method is professional in clinical output, has a zero-sample dialogue capability, is flexible in application, can optimize medical resources, and can explain decisions.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to an information-driven intelligent diagnostic system and method for hypertension. Background Technology

[0002] The intelligent hypertension diagnostic system is applied in medical consultation scenarios, suitable for hypertension diagnosis, classification, risk stratification, and treatment plan recommendations. Given the global shortage of medical resources, this system can serve as a self-service medical consultation assistant, helping to address the problem of hypertension as a "silent killer" and providing professional diagnostic services to hypertension patients.

[0003] Currently, artificial intelligence diagnostic systems for chronic diseases such as hypertension mainly have the following problems:

[0004] 1. Knowledge Acquisition Challenges: Existing AI diagnostic systems primarily rely on two types of knowledge injection methods: dynamic methods (such as retrieval-enhanced RAG generation) and static knowledge embedding (such as fine-tuning). Dynamic injection methods are highly dependent on retrieval quality, prone to conflicts between retrieved knowledge and pre-trained representations, and suffer from insufficient information extraction efficiency. The large amount of redundancy and document similarity in medical knowledge bases also affects the similarity search mechanism in retrieval systems.

[0005] 2. Data Dependence on Multi-Turn Dialogues: Guiding patients through multiple rounds of dialogue to collect personalized information requires exponentially increasing training data—if a single-turn question-and-answer session requires n samples, a k-turn dialogue would require n^k samples to achieve comparable accuracy. This is particularly challenging in a medical setting because patients have diverse conditions, varying levels of health literacy, and may struggle to accurately express their symptoms.

[0006] 3. Lack of clinical reasoning ability: Existing systems often lack diagnostic logic chains and structured reasoning ability, and cannot conduct systematic clinical thinking like professional doctors, thus affecting the accuracy of diagnosis and the reliability of treatment recommendations;

[0007] 4. Limited Interaction Capabilities: Existing medical AI systems mostly serve as assistant tools for doctors, only able to process pre-existing data and unable to autonomously interact with patients in a human-like manner. Traditional systems typically only maintain 4-6 rounds of dialogue, resulting in low information acquisition rates and failing to meet the in-depth information collection needs required for complex hypertension diagnosis.

[0008] 5. Lack of depth in specific fields: General-purpose large language models lack in-depth expertise in specific disease areas such as hypertension, and are insufficient in handling professional medical problems, making it difficult to achieve the diagnostic capabilities of specialist doctors. Summary of the Invention

[0009] The purpose of this invention is to provide an information-driven intelligent diagnostic system and method for hypertension in order to solve the above-mentioned problems.

[0010] The present invention achieves the above objectives through the following technical solutions:

[0011] An information-driven intelligent hypertension diagnostic system includes an entropy minimization retrieval enhancement generation module, a zero-shot task-oriented dialogue module, a multi-stage training module, and a hypertension professional decision-making logic module. The entropy minimization retrieval enhancement generation module is used to construct and optimize a hypertension knowledge base and achieve efficient retrieval enhancement, including information-knowledge matrix construction, information entropy quantification, greedy selection algorithm execution, vector encoding and index construction, and retrieval and enhancement units. The zero-shot task-oriented dialogue module is used to achieve multi-round information collection without requiring a large amount of dialogue training data, including information state tracking, intelligent node selection, information extraction and reliability assessment, and final diagnosis synthesis units. The multi-stage training module is used for structured hypertension medical knowledge training, including reinforcement learning, supervised fine-tuning, and MinRAG ensemble units. The hypertension professional decision-making logic module is used for deep clinical decision-making, including hypertension diagnosis classification, cardiovascular risk stratification, treatment plan decision-making, and special population handling units.

[0012] Preferably, the entropy minimization retrieval enhancement generation module also includes an importance calculation unit, which defines node importance as the product of knowledge point weight and uncertainty, with the weight reflecting the clinical relevance of the knowledge point; the vector encoding and index construction unit adopts a pre-trained model adapted to the semantic features of medical text; and the minimum set determination unit selects knowledge records to form a knowledge base through normalized entropy curve inflection point analysis to achieve high knowledge point coverage.

[0013] Preferably: the information status tracking unit of the zero-sample task-oriented dialogue module assigns clinical importance weights, uncertainty values ​​and status identifiers to each node; the intelligent node selection unit selects query nodes based on clinical importance scores, logical relevance and information acquisition priority, and the clinical importance score is a weighted combination of weights and uncertainty.

[0014] Preferably, the supervised fine-tuning unit of the multi-stage training module uses chain-thinking reasoning-enhanced question-and-answer data; the reinforcement learning unit uses a group preference optimization strategy to process hypertension-related multiple-choice questions and implements format and accuracy reward functions.

[0015] Preferably, the hypertension professional decision-making logic module includes a hypertension diagnosis classification unit that assesses blood pressure levels; a cardiovascular risk stratification unit that assesses target organ damage and analyzes associated risk factors; and a treatment plan decision-making unit that provides lifestyle interventions, drug treatment recommendations, and follow-up plan development.

[0016] An information-driven intelligent diagnostic method for hypertension includes:

[0017] Step S1: Construct and optimize the hypertension knowledge base and enhance retrieval;

[0018] Step S2: Conduct zero-sample, multi-round dialogues to collect patient information;

[0019] Step S3: Perform multi-stage training on the model;

[0020] Step S4: Make clinical decisions based on hypertension expertise.

[0021] Preferably, step S1 includes the steps of mapping knowledge records to form a binary matrix, calculating information entropy, defining importance, greedy selection, and vector encoding. The vector encoding adopts a pre-trained model adapted to the semantic features of medical text. The importance is defined as the product of the knowledge point weight and the uncertainty.

[0022] Preferably, step S2 includes constructing an information status tracking structure, selecting query nodes based on multi-dimensional indicators such as clinical importance scores, extracting information and attaching reliability indicators and diagnostic synthesis steps.

[0023] Preferably, step S3 includes steps such as using a grouping preference optimization strategy and a dual reward function to process multiple-choice questions, using chain-thinking reasoning to enhance question-and-answer data for supervised fine-tuning, and MinRAG integration optimization of the knowledge base.

[0024] Preferably, step S4 includes diagnostic classification, risk stratification, treatment plan development, and special population management steps, with the treatment plan including lifestyle interventions and drug therapy recommendations.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. Diagnostic accuracy: This invention achieves an accuracy rate of 91.50% in Chinese hypertension practice and 88.90% in international hypertension practice, significantly exceeding human hypertension physicians (72.58% and 67.88%, respectively) and other large-scale language models;

[0027] 2. Efficient Knowledge Management: The MinRAG method reduces the knowledge base from 32,898 records to 1,137, reducing storage requirements and retrieval processing time by 96.5%, while improving model performance. On smaller models, MinRAG improves accuracy by 60.5% compared to the basic model and by 13.2% compared to the full RAG model.

[0028] 3. Superior dialogue capabilities: The system can maintain an average of 26.41 rounds of dialogue, far exceeding the 4-6 rounds of the comparison model, achieving a clinical entity retrieval rate of over 95% and information consistency of 97.5%; the dialogue memory capacity and information retrieval capacity are significantly better than the comparison system;

[0029] 4. Professional Clinical Output: In the 25-point clinical assessment, this system achieved an excellent score of 19.2, performing well in multiple categories such as diagnostic assessment, lifestyle recommendations, medication management and follow-up protocols;

[0030] 5. Zero-sample dialogue capability: It does not require a large amount of task-specific dialogue data, and maintains diagnostic consistency through information status tracking while adapting to patient communication patterns, which greatly reduces deployment costs and data requirements;

[0031] 6. Application Flexibility: The method of this invention is not limited to hypertension, but can be extended to the management of other chronic diseases, such as diabetes and heart disease, and has broad application prospects;

[0032] 7. Medical resource optimization: It can serve as a self-medical consultation assistant to help solve the global shortage of doctors, especially in areas with limited medical resources, and provide patients with timely and accurate hypertension management consultation;

[0033] 8. Explainability: The system provides a structured reasoning process, including detailed explanations of diagnostic approaches, risk assessments, and treatment recommendations, which improves the transparency and credibility of clinical decision-making. Attached Figure Description

[0034] 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.

[0035] Figure 1 This is a flowchart of an information-driven intelligent diagnosis method for hypertension as described in this invention.

[0036] Figure 2 This refers to the model exam score of an information-driven intelligent hypertension diagnostic system as described in this invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings:

[0038] Example 1

[0039] like Figures 1-2As shown, an information-driven intelligent hypertension diagnostic system includes an entropy minimization retrieval enhancement generation module, a zero-shot task-oriented dialogue module, a multi-stage training module, and a hypertension professional decision-making logic module. The entropy minimization retrieval enhancement generation module is used to construct and optimize a hypertension knowledge base and achieve efficient retrieval enhancement, including information-knowledge matrix construction, information entropy quantification, greedy selection algorithm execution, vector encoding and index construction, and retrieval and enhancement units. The zero-shot task-oriented dialogue module is used to achieve multi-round information collection without requiring a large amount of dialogue training data, including information state tracking, intelligent node selection, information extraction and reliability assessment, and final diagnosis synthesis units. The multi-stage training module is used for structured hypertension medical knowledge training, including reinforcement learning, supervised fine-tuning, and MinRAG ensemble units. The hypertension professional decision-making logic module is used for deep clinical decision-making, including hypertension diagnosis classification, cardiovascular risk stratification, treatment plan decision-making, and special population handling units.

[0040] The construction and operation of the entropy minimization retrieval enhancement generation module includes information-knowledge matrix construction unit operation, information entropy quantification unit calculation, importance calculation unit execution, greedy selection algorithm unit operation, minimum set determination unit optimization, vector encoding and indexing unit processing, and retrieval and enhancement unit application.

[0041] The information-knowledge matrix construction unit operation involved collecting 32,898 knowledge records in the field of hypertension (covering clinical guidelines, case literature, pharmacological data, etc.) and identifying 234 core knowledge points related to hypertension (including blood pressure classification standards, complication-related factors, drug indications, etc.). Each knowledge record was matched with a knowledge point to form a binary information-knowledge matrix, where an element "1" indicates that the i-th knowledge record contains the j-th knowledge point, and "0" indicates that it does not.

[0042] The information entropy quantization unit is calculated using information entropy theory, through the formula... Calculate the information entropy of each knowledge point, where P(x_i) is the probability of the knowledge point appearing in all records. For example, the knowledge point "diagnostic criteria for essential hypertension" has a low information entropy (high information redundancy) because it appears in most records; the knowledge point "contraindications for special medications for hypertension during pregnancy" has a high information entropy (strong uniqueness) because it appears only in a few records.

[0043] The importance calculation unit is executed to define the importance of nodes. Where W represents the knowledge point weight (set based on clinical diagnostic relevance, such as setting the weight of "systolic / diastolic blood pressure range" to 0.9 and "patient family medical history" to 0.6), and U represents the knowledge point uncertainty (converted from the information entropy quantification result; the higher the entropy value, the larger U). This formula calculates the importance scores for 234 knowledge points, which serve as the basis for subsequent knowledge base optimization.

[0044] The greedy selection algorithm unit operates as a greedy selection algorithm that performs bi-objective optimization. In each iteration, it selects information blocks from 32,898 records that "contribute the maximum marginal entropy gain" and "cover uncovered knowledge points". For example, the first round selects records covering "diagnostic criteria for grade 3 hypertension" (a highly important knowledge point), the second round selects records covering "blood pressure control targets for patients with diabetes" (an uncovered knowledge point with high entropy), and so on, iterating step by step to filter out a core subset of records.

[0045] The optimization of the minimum set unit involved plotting a normalized entropy curve (the horizontal axis represents the number of selected records, and the vertical axis represents the normalized value of knowledge point coverage and average information entropy). The optimal subset size was determined through inflection point analysis of the curve. When the number of records reached 1,137, the curve showed a clear inflection point, at which point the knowledge point coverage reached 96.8% (only 8 low-importance knowledge points were not covered; the uncovered content included non-universal knowledge such as "diagnosis and treatment of rare complications" and "hypertension of special gene types," which did not affect the core diagnosis). Therefore, 1,137 records were selected as the final optimized knowledge base.

[0046] The vector encoding and indexing unit processing employs a pre-trained model (such as the BioBERT model) adapted to the semantic features of medical text, encoding 1,137 knowledge base records into 768-dimensional dense semantic vectors. A FAISS vector database is constructed, and the encoded vectors are stored in the database with an index for subsequent fast similarity searches.

[0047] The retrieval and enhancement unit is used during the inference phase. When a user query is received (e.g., "How should a 55-year-old male patient with a systolic blood pressure of 165 mmHg and a diastolic blood pressure of 105 mmHg be diagnosed?"), the unit retrieves the top 10 most relevant paragraphs (including records related to "diagnostic criteria for grade 2 hypertension" and "risk factors for hypertension in middle-aged men") from the vector database based on the cosine similarity algorithm. These 10 paragraphs are then concatenated with the original query to form enhanced prompts (with a total length controlled within 512 tokens) and input into the subsequent diagnostic module.

[0048] The aforementioned entropy minimization retrieval enhancement generation module's construction logic, multi-stage training module's training process (including reinforcement learning, supervised fine-tuning, and MinRAG ensemble), and the entire process of subsequent dialogue module's information collection and decision module's diagnostic output all correspond to... Figure 1 The technical framework shown: Figure 1The steps “Step1: Data Preprocessing, Training” clearly define the connection between knowledge base construction and model training; “Step2: Conversation Framework” corresponds to the interaction logic of the zero-shot dialogue module; and “Step3: Conversation, Diagnosis & Recommendation Evaluation” covers the entire process of generating and evaluating diagnostic results, providing process-oriented support for the system construction in this embodiment.

[0049] The implementation of the zero-sample task-oriented dialogue module includes the construction of the information state tracking unit, the operation of the intelligent node selection unit, the operation of the information extraction and reliability assessment unit, and the synthesis of state updates and final diagnosis.

[0050] The information status tracking unit is constructed to build an information status tracking structure, containing four core nodes: key clinical indicators (systolic blood pressure, diastolic blood pressure, heart rate, etc.), symptoms (headache, dizziness, chest tightness, etc.), risk factors (history of smoking, obesity, high-salt diet, etc.), and medical history (history of hypertension, history of diabetes / nephropathy, etc.). Parameters are assigned to each node: clinical importance weight (e.g., systolic blood pressure weight 0.95), initial uncertainty value (set to 1.0, representing the maximum uncertainty when no information is obtained), and status identifier ("not queried", "obtained", "pending verification").

[0051] The intelligent node selection unit operates by selecting the first query node based on node importance score, logical relevance, and information retrieval priority when a patient first enters the dialogue; for example, it prioritizes querying "systolic and diastolic blood pressure measurements" (clinical importance score). Logically, this is the basis for diagnosis and conforms to the interview sequence recommended by the clinical diagnosis guidelines for hypertension. If the patient answers "systolic blood pressure 150 mmHg, diastolic blood pressure 95 mmHg", the system calculates the semantic relevance of this node to the information already obtained (e.g., relevance to the knowledge point of "blood pressure classification" 0.8) and the clinical logical relevance (directly related to the diagnosis result, relevance 0.9), and then selects the next node "whether there is a history of diabetes" (because hypertension combined with diabetes will affect risk stratification, so it has a higher priority).

[0052] The information extraction and reliability assessment unit operates by employing Named Entity Recognition (NER) and semantic understanding models to extract medical information from patient responses. For example, if a patient answers, "I have had diabetes for 5 years, have been taking metformin, and occasionally experience dizziness," the system extracts three pieces of information: "5-year history of diabetes," "taking metformin," and "dizziness symptoms." Simultaneously, a reliability index is assigned to each piece of information through text credibility analysis (e.g., determining whether the answer is clear and consistent): "5-year history of diabetes" is clear and assigned a reliability of 0.9; "occasionally experience dizziness" is ambiguous and assigned a reliability of 0.6. Only information with a reliability ≥ 0.8 is used to update the diagnostic status, and nodes with clearly obtained information (such as "diabetes history") are marked as "obtained" to avoid repeated queries.

[0053] The training data for the reliability assessment module comes from two sources: ① publicly available medical dialogue datasets (such as the hypertension dialogue subset in MedDialog-CN, containing 12,000 sets of "patient statements - doctor annotation credibility" data, with annotation dimensions including statement clarity, consistency, and accuracy of medical terminology); ② a self-built annotation dataset (three cardiologists with the title of associate chief physician or above were invited to annotate the credibility of 5,000 sets of simulated hypertension patient dialogues, and the annotation results were tested for consistency using the Kappa coefficient (Kappa=0.89, P<0.01) to ensure annotation quality); the model is trained using binary classification (high credibility ≥0.8 / low credibility <0.8) based on the above data, and fine-tuned using the BERT-Base model. The final test set F1 score reached 0.92, ensuring the accuracy of the reliability assessment;

[0054] State updates and final diagnoses are combined after each interaction. The system integrates newly acquired high-reliability information and updates the uncertainty and weights of the corresponding nodes based on the Bayesian update formula. The formula for calculating node uncertainty U is: Initial uncertainty (When no information is available), the uncertainty after the update occurs once new information is available. Where S is the information similarity (the semantic similarity between the newly acquired information and the standard representation of the node, calculated by cosine similarity, ranging from 0 to 1), and C is the information credibility (i.e. the reliability index output by the reliability assessment unit, ranging from 0 to 1).

[0055] For example, if a patient answers "systolic blood pressure 150 mmHg" (similarity to the standard expression for "systolic blood pressure" S=0.98, confidence level C=0.9), then the updated uncertainty... If asked again later whether the patient had a history of smoking, the patient answered "no history of smoking" (S=0.95, C=0.98). Uncertainty after the update At this point, the uncertainty reduction is 0.118 - 0.008 = 0.11 > 0.05, so further questioning is needed; if the next round of questioning asks "Do you have a history of kidney disease?", the patient answers "No" (S=0.96, C=0.99). Uncertainty after the update The uncertainty reduction is 0.008 - 0.0003 = 0.0077 < 0.05, indicating that the system has completed information collection and synthesized the final diagnostic result.

[0056] Patient Information Summary: Record all collected clinical indicators, including systolic blood pressure 150 mmHg, diastolic blood pressure 95 mmHg, 5-year history of diabetes, no history of smoking, no history of kidney disease, and occasional dizziness.

[0057] Formal medical diagnosis: Grade 2 hypertension (primary), cardiovascular risk stratification is "medium risk" (due to comorbid diabetes);

[0058] Comprehensive recommendations: Lifestyle modifications (daily sodium intake ≤5g, adopt the DASH diet plan: 4-5 servings of vegetables, 4 servings of fruit, and 1-2 servings of whole grains daily, reducing saturated fat intake), medication recommendations (add valsartan 80mg / day to the existing metformin), follow-up plan (followed by the first follow-up 1 month later, and then every 3 months thereafter, monitoring indicators including blood pressure (measured 3 times a week and the average value is taken), blood glucose (fasting blood glucose), and renal function (serum creatinine). If blood pressure remains ≥140 / 90mmHg, adjust the medication dose to 160mg / day).

[0059] The training process of the multi-stage training module includes reinforcement learning unit training, supervised fine-tuning unit training, and MinRAG ensemble unit operations.

[0060] The reinforcement learning unit training involved collecting 2,641 hypertension-related multiple-choice questions (divided into three groups based on knowledge points: "Diagnosis," "Treatment," and "Pharmacology," with 892, 956, and 793 questions per group, respectively), and employing a group preference optimization strategy. Two reward functions were set up:

[0061] Format bonus: If the model's answer correctly contains the "thinking label" (such as "Step 1: Determine the blood pressure level; Step 2: Analyze comorbidities..."), it will receive a format bonus of 1.0; otherwise, it will receive 0 points.

[0062] Accuracy Bonus: A perfect answer earns 1.0 accuracy bonus (e.g., for the question "Standard for Grade 3 Hypertension Systolic Blood Pressure", if the correct answer is "≥180mmHg", the model's answer is consistent and earns 1.0 bonus); partially correct answers are awarded proportionally (e.g., if the model answers "170-180mmHg", 0.5 bonus is awarded); the bonus weight is adjusted for different question groups, with the accuracy bonus weight for "Pharmacology" questions increased to 1.2 to ensure the model's learning effect on drug knowledge;

[0063] Supervised fine-tuning unit training prepares 5000 chain-reasoning enhanced question-and-answer data (e.g., the question "A 60-year-old patient with a systolic blood pressure of 170 mmHg and a history of coronary heart disease, how should he / she be treated?", with the reasoning process "Step 1: Determine the blood pressure classification as grade 2; Step 2: Considering the poor vascular elasticity of elderly patients, avoid lowering blood pressure too quickly; Step 3: Given the presence of coronary heart disease, prioritize beta-blockers; Step 4: Determine the drug dosage as metoprolol 25 mg / day"). This data is input into the model for supervised fine-tuning, with 10 training rounds and a learning rate of 2e-5, enabling the model to output explicit intermediate reasoning steps reflecting the clinical decision-making process and improving the structured reasoning ability for complex queries.

[0064] The MinRAG integrated unit operation applies the aforementioned entropy minimization retrieval enhancement generation module technique to optimize the initial 32,898 knowledge records to 1,137 (knowledge point coverage rate of 96.8%), and integrates the optimized knowledge base with the trained model to achieve synergy between "retrieval-reasoning-generation". When the model answers, relevant information in the knowledge base can be called in real time to improve diagnostic accuracy.

[0065] The functional implementation of the hypertension professional decision-making logic module includes the operation of the hypertension diagnosis and classification unit, the analysis of the cardiovascular risk stratification unit, the output of the treatment plan decision-making unit, and the adaptation of the special population processing unit.

[0066] The hypertension diagnosis and classification unit operates by receiving patient information collected by the dialogue module. First, it differentiates between primary and secondary hypertension (e.g., by judging whether there is a history of renal artery stenosis or endocrine abnormalities; if no relevant abnormalities are found, it is classified as primary). Then, it classifies hypertension into grades 1-3 based on systolic / diastolic blood pressure values ​​(grade 1: systolic blood pressure 140-159 mmHg / diastolic blood pressure 90-99 mmHg, grade 2: 160-179 mmHg / 100-109 mmHg, grade 3: ≥180 mmHg / ≥110 mmHg). At the same time, it records the blood pressure level assessment results (e.g., "systolic blood pressure 165 mmHg, diastolic blood pressure 105 mmHg, meeting the criteria for grade 2 hypertension").

[0067] Cardiovascular risk stratification unit analysis includes risk stratification based on patient age, comorbidities, and target organ damage:

[0068] Low risk: No comorbidities, no target organ damage (e.g., patients under 50 years of age with grade 1 hypertension, no diabetes, no left ventricular hypertrophy).

[0069] Medium risk: Grade 1-2 hypertension with one risk factor (e.g., a 55-year-old male with grade 2 hypertension and diabetes).

[0070] High risk: Grade 3 hypertension or two or more risk factors (e.g., a 60-year-old female with grade 3 hypertension, combined with diabetes and renal dysfunction).

[0071] Very high risk: Combined target organ damage (such as left ventricular hypertrophy, proteinuria) or clinical complications (such as stroke, myocardial infarction).

[0072] In this embodiment, a 55-year-old male patient (with grade 2 hypertension and diabetes) was classified as medium risk.

[0073] The output of the treatment decision unit includes:

[0074] Lifestyle intervention: Define the daily sodium intake control standard (≤5g) and provide the specific content of the DASH diet plan (4-5 servings of vegetables, 4 servings of fruit, 1-2 servings of whole grains per day, and reduce saturated fat intake).

[0075] Drug treatment plan: Based on the patient's blood pressure classification (level 2), risk stratification (medium risk), comorbidities (diabetes) and absence of drug contraindications, valsartan (angiotensin II receptor antagonist, suitable for patients with diabetes) is recommended at a dose of 80 mg / day.

[0076] Follow-up plan: The first follow-up will be conducted 1 month after starting medication, followed by follow-up every 3 months thereafter. Monitoring indicators include blood pressure (measured 3 times a week and the average value is taken), blood glucose (fasting blood glucose), and renal function (serum creatinine). If blood pressure remains ≥140 / 90 mmHg, the medication dose will be adjusted to 160 mg / day.

[0077] Special population processing unit adaptation includes:

[0078] For elderly patients with hypertension (e.g., over 70 years old): Adjust the blood pressure target (systolic blood pressure <150 mmHg is sufficient to avoid excessively low blood pressure affecting cerebral blood supply), and prioritize long-acting calcium channel blockers (e.g., amlodipine).

[0079] For patients with diabetes: Angiotensin-converting enzyme inhibitors (ACEIs) or angiotensin II receptor blockers (ARBs) should be preferred, and beta-blockers that may raise blood sugar (such as propranolol) should be avoided.

[0080] For pregnant women with hypertension: ACEI / ARB drugs are contraindicated. Methyldopa or labetalol should be selected instead. At the same time, the follow-up interval should be shortened (once every 2 weeks) and fetal development should be closely monitored.

[0081] Through the synergy of the four modules mentioned above, this system demonstrates excellent accuracy in hypertension diagnosis, and its performance can be achieved through… Figure 2 (Model exam score chart) Verification: Figure 2 The scores of our system (DrHyper-7B / 32B) were compared with those of human doctors and other mainstream large models (such as Llama-3.3-70B-Instruct, Qwen2.5-32B-Instruct, etc.) in hypertension professional examinations (CHSPQ, IHSPQ). DrHyper-32B scored 91.50% in CHSPQ (Chinese Hypertension Professional Practice Questions) and 88.90% in IHSPQ (International Hypertension Professional Practice Questions), significantly higher than those of human doctors (72.58%, 67.88%) and other models. This fully demonstrates the advantages of our system in hypertension professional diagnostic capabilities, which is consistent with the technical goal of "prioritizing diagnostic accuracy" in this embodiment.

[0082] In addition, this embodiment also discloses an information-driven intelligent hypertension diagnostic system, including the following steps:

[0083] Step S1: Construct and optimize the hypertension knowledge base and enhance retrieval; perform information-knowledge matrix construction, matching 32,898 hypertension knowledge records with 234 hypertension-related knowledge points to obtain a binary matrix; use the formula... Calculate the information entropy of each knowledge point to determine redundancy and uniqueness; according to the formula The importance of nodes is calculated (W represents clinical relevance weight, and U represents uncertainty), and a greedy selection algorithm with bi-objective optimization is executed. Through inflection point analysis of the normalized entropy curve, 1,137 records are selected as the knowledge base (knowledge point coverage rate of 96.8%). The knowledge base records are encoded into 768-dimensional semantic vectors using the BioBERT model to construct the FAISS vector database. When receiving user queries, the top 10 relevant paragraphs are retrieved based on cosine similarity and concatenated with the original query to form enhanced prompts.

[0084] Step S2: Conduct zero-sample multi-turn dialogues to collect patient information; construct an information status tracking structure containing 4 types of core nodes, and assign clinical importance weights and initial uncertainty values ​​to each node. ) and status identifiers; based on node importance scores ( The system selects the first query node based on logical relevance (semantic relevance + clinical logical relevance) and priority according to hypertension clinical diagnostic guidelines. It then uses a Neural Error Analysis (NER) and semantic understanding model to extract patient response information. A reliability assessment model based on dual-source training data (12,000 publicly available sets + 5,000 self-built labeled sets) assigns reliability indices to the information (only information with a reliability ≥0.8 is used for updates). Finally, it updates the information according to the Bayesian update formula. (S represents semantic similarity, and C represents credibility) Update node uncertainty and weight, and activate relevant query paths; when the uncertainty reduction is less than 0.05 after 3 consecutive queries, synthesize the final result containing patient information summary, formal medical diagnosis (classification + risk stratification), and comprehensive recommendations (lifestyle + medication + follow-up);

[0085] Step S3: Perform multi-stage training on the model; use a group preference optimization strategy to process 2,641 multiple-choice questions and implement dual reward function reinforcement learning; use 5,000 chain-like reasoning data for supervised fine-tuning (10 rounds, learning rate 2e-5); integrate MinRAG technology to optimize the knowledge base to 1,137 items, and achieve retrieval-reasoning collaboration;

[0086] Step S4: Make clinical decisions based on hypertension expertise; differentiate between primary and secondary hypertension, complete grade 1-3 classification and blood pressure assessment; stratify low / medium / high / very high risk by combining risk factors and target organ damage; develop lifestyle intervention, drug treatment and follow-up plans; adjust diagnosis and treatment plans for special populations such as the elderly, those with diabetes, and pregnant women.

[0087] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are only illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the present invention, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. An information-driven based intelligent diagnosis system for hypertension, characterized in that, Comprise: An entropy minimization retrieval enhancement generation module for constructing an optimized hypertension knowledge base and implementing efficient retrieval enhancement, including information-knowledge matrix construction, information entropy quantification, greedy selection algorithm execution, vector encoding and index construction, and retrieval and enhancement units; A zero-shot task-oriented dialogue module for collecting multiple rounds of information without large amounts of dialogue training data, including information state tracking, intelligent node selection, information extraction and reliability evaluation, and final diagnosis synthesis units; A multi-stage training module for structured hypertension medical knowledge training, including reinforcement learning, supervised fine-tuning, and MinRAG integration units; And A hypertension professional decision-making logic module for deep clinical decision-making, including hypertension diagnosis classification, cardiovascular risk stratification, treatment plan decision-making, and special population handling units.

2. The system of claim 1, wherein, The entropy minimization retrieval enhancement generation module further includes an importance calculation unit that defines node importance as the product of knowledge point weight and uncertainty, with weight reflecting the clinical relevance of knowledge points; the vector encoding and index construction unit uses a pre-trained model adapted to the semantic features of medical text; and the minimum set determination unit selects knowledge records to form a knowledge base through normalized entropy curve inflection point analysis, achieving high knowledge point coverage.

3. The system of claim 1, wherein, The information state tracking unit of the zero-shot task-oriented dialogue module assigns each node a clinical importance weight, an uncertainty value, and a state identifier; the intelligent node selection unit selects inquiry nodes based on clinical importance scores, logical relevance, and information acquisition priority, with clinical importance scores being a weighted combination of weight and uncertainty.

4. The system of claim 1, wherein, The supervised fine-tuning unit of the multi-stage training module uses chain-of-thought reasoning-enhanced question-answer data; the reinforcement learning unit uses a grouping preference optimization strategy to handle hypertension-related multiple-choice questions, implementing format and accuracy reward functions.

5. The system of claim 1, wherein, The hypertension diagnosis classification unit of the hypertension professional decision-making logic module includes blood pressure level evaluation; the cardiovascular risk stratification unit includes target organ damage assessment and concomitant risk factor analysis; and the treatment plan decision-making unit provides lifestyle interventions, drug treatment recommendations, and follow-up plan development.

6. An information-driven intelligent diagnosis method for hypertension, characterized in that, Comprise: Step S1: Construct an optimized hypertension knowledge base and implement retrieval enhancement; Step S2: Conduct zero-shot multi-round dialogue to collect patient information; Step S3: Train the model in multiple stages; Step S4: Make clinical decisions based on hypertension professional knowledge.

7. The method of claim 6, wherein, The step S1 includes knowledge record mapping to form a binary matrix, information entropy calculation, importance definition, greedy selection, and vector encoding steps, with vector encoding using a pre-trained model adapted to the semantic features of medical text; The importance definition is the product of knowledge point weight and uncertainty.

8. The method of claim 6, wherein, Step S2 includes building an information state tracking structure, selecting inquiry nodes based on clinical importance scores and other multidimensional indicators, extracting information with reliability indicators, and synthesizing diagnoses.

9. The method of claim 6, wherein, Step S3 includes processing multiple-choice questions using a grouping preference optimization strategy and a double reward function, supervised fine-tuning with chain-of-thought reasoning-enhanced question-answer data, and MinRAG integration to optimize the knowledge base.

10. The method of claim 6, wherein, Step S4 includes diagnosis classification, risk stratification, treatment regimen development and special population handling steps, including lifestyle intervention and drug treatment recommendations.

Images