Intelligent medical inquiry auxiliary diagnosis system based on symptom reasoning and knowledge graph

Abstract

The invention discloses an intelligent medical inquiry auxiliary diagnosis system based on symptom reasoning and a knowledge graph. The intelligent medical inquiry auxiliary diagnosis system comprises a knowledge graph construction module, a common disease diagnosis module, a multi-round dialogue inquiry module, a multi-modal data fusion module, a treatment scheme generation module and a follow-up visit management module. The data layer is responsible for medical knowledge storage and case data management, the logic layer is used for realizing a reasoning algorithm and a dialogue strategy, and the application layer is used for providing a user interaction interface and an external system interface. Systematized collection of symptoms is achieved by constructing a large-scale medical knowledge graph and designing a multi-round dialogue inquiry strategy, the disease posterior probability is calculated by applying a Bayesian probabilistic reasoning model, multi-modal diagnosis is conducted by fusing medical images and inspection data, a personalized treatment scheme is generated, and medication safety is checked. And intelligent inquiry, accurate diagnosis, reasonable medication and whole-course management of the patient are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and medical information processing, in particular to an intelligent medical inquiry and auxiliary diagnosis system based on symptom reasoning and knowledge graph, which is suitable for application scenarios such as auxiliary diagnosis of primary medical institutions, online inquiry of internet medical treatment, family health management, emergency triage, and rare disease screening. BACKGROUND

[0002] Existing online inquiry systems mostly rely on manual services by doctors, which is difficult to meet the demand for 24-hour instant response, and lack standardized diagnosis and treatment processes, resulting in uneven service quality.

[0003] Traditional medical diagnosis relies on the clinical experience and medical knowledge accumulation of doctors, and for complex cases, rare diseases, and patients with atypical symptoms, even experienced doctors may make diagnosis errors. Existing medical auxiliary diagnosis systems have the following technical defects:

[0004] 1. Incomplete symptom collection: Most systems use a simple symptom check list, which cannot ask targeted questions based on patient complaints like real doctor inquiries, resulting in missing key symptom information, and inability to effectively obtain characteristic symptoms needed for differential diagnosis, affecting the accuracy of subsequent disease reasoning.

[0005] 2. Incomplete knowledge base construction: Existing medical knowledge bases mostly exist in the form of text data or rule bases, and fail to fully model the complex relationships between diseases, symptoms, signs, tests, drugs, and treatment plans, lack quantitative characterization of symptom-disease probability relationships, disease-test sensitivity relationships, and drug-disease indications, limiting the reasoning ability of the system.

[0006] 3. Single reasoning model: Traditional diagnosis systems mostly use simple rule matching or decision tree methods, which cannot handle the uncertainty of symptoms, the diversity of diseases, and the probability of diagnosis, and cannot consider multiple possible diseases and give differential diagnosis suggestions like clinical doctors, nor can they explain the supporting evidence and opposing evidence for each disease diagnosis conclusion.

[0007] 4. Lack of multi-modal data fusion: Existing systems mainly rely on text symptom descriptions for diagnosis, and fail to effectively integrate medical images (X-ray, CT, MRI), test reports (blood routine, biochemical indicators), physiological signals (ECG, blood pressure), and other multi-source heterogeneous medical data, which cannot fully utilize objective test results to improve diagnosis accuracy and reliability.

[0008] 5. Lack of personalized treatment plans: Most systems can only provide standard treatment plans for diseases without considering individual differences such as the patient's age, gender, medical history, allergy history, liver and kidney function, and medication history. They also do not perform intelligent checks on medication safety issues such as drug interactions, incompatibilities, and dosage adjustments, which poses medication risks.

[0009] 6. Weak ability to identify rare diseases: Existing diagnostic systems are mainly trained for common diseases and lack the ability to identify rare diseases with low incidence but complex symptoms, which can easily lead to misdiagnosis or missed diagnosis, delaying the best treatment time for patients. In particular, some serious diseases with atypical early symptoms cannot be detected in time.

[0010] 7. Lack of closed-loop management in diagnosis and treatment: Traditional systems lack follow-up patient education, medication guidance, follow-up reminders, and efficacy evaluation after providing diagnostic suggestions. This results in poor patient compliance, difficulty in guaranteeing treatment effectiveness, and an inability to form a complete closed loop of intelligent diagnosis and treatment services.

[0011] Therefore, there is an urgent need for an intelligent medical consultation and auxiliary diagnostic system capable of systematically collecting symptom information, constructing a complete medical knowledge graph, performing probabilistic disease reasoning, integrating multimodal medical data, generating personalized treatment plans, identifying rare diseases, and providing end-to-end management. This invention addresses these technical challenges by integrating artificial intelligence technologies such as natural language understanding, knowledge graphs, Bayesian reasoning, and deep learning, providing high-quality intelligent diagnostic assistance tools for primary healthcare and internet-based healthcare. Summary of the Invention

[0012] To address the aforementioned issues, this invention provides a real-time cognitive load assessment and adaptive early warning system based on multimodal physiological signals. It constructs a large-scale medical knowledge graph covering disease, symptoms, examinations, medications, and treatments; designs a multi-turn dialogue consultation strategy to achieve systematic symptom collection; applies a Bayesian probabilistic inference model to calculate the posterior probability of diseases; integrates medical imaging and laboratory data for multimodal diagnosis; generates personalized treatment plans; and checks medication safety, thereby achieving intelligent consultation, accurate diagnosis, rational drug use, and comprehensive management of patients.

[0013] The specific plan is as follows:

[0014] An intelligent medical consultation and auxiliary diagnostic system based on symptom reasoning and knowledge graphs includes the following steps:

[0015] Knowledge graph construction module: Constructs a large-scale medical knowledge graph containing disease entities, symptom entities, examination entities, drug entities, treatment plan entities and their relationships, and combines patient symptom information to achieve automated knowledge extraction through natural language processing technology, completing the structured representation of the initial symptom set;

[0016] Common disease diagnosis module: Construct a quantitative association model between symptoms and diseases, calculate the posterior probability of all candidate diseases based on the initial symptom set, and obtain a preliminary disease probability distribution;

[0017] Multi-turn dialogue consultation module: By asking about new symptoms, the posterior probability is updated, and the next optimal question is calculated based on the new probability distribution to update the symptom information;

[0018] Multimodal data fusion module: Integrates objective examination data from medical images and laboratory reports, and calculates the posterior probability of the disease in conjunction with symptom information to identify the target disease; For medical images, the system uses a deep convolutional neural network to extract lesion features and outputs image diagnosis results and confidence levels; For laboratory reports, the system compares the values ​​of each indicator with the normal reference range to identify abnormal items.

[0019] Treatment plan generation module: Based on the identified target disease, and using the disease-treatment relationship in the knowledge graph, combined with the individual characteristics of the patient, a personalized treatment plan is generated.

[0020] Follow-up management module: Automatically generates health education materials for patients, explaining the causes, pathogenesis, prognosis, daily care points, and timing of follow-up visits in plain language.

[0021] Furthermore, the knowledge graph construction module includes:

[0022] The knowledge graph uses triples to represent medical knowledge, with each triple recording as follows: ,For example , Knowledge graphs are constructed by extracting data from multiple sources, including medical textbooks, clinical guidelines, electronic medical records, drug instructions, and medical literature.

[0023] Furthermore, the common disease diagnosis module includes:

[0024] For symptom set and disease collection By statistically analyzing large-scale electronic medical record data, calculations can be made regarding diseases. Symptoms appear under certain conditions conditional probability and symptoms Prior probability in the general population These probability parameters are stored in the relational attributes of the knowledge graph, forming a quantitative association model between symptoms and diseases.

[0025] When patients exhibit a combination of symptoms At that time, the system calculates the posterior probability of each candidate disease based on Bayes' theorem; for each disease... The posterior probability is calculated as follows:

[0026]

[0027] in For disease The prevalence of disease in a population, i.e., the prior probability. In the disease Symptom combinations observed under certain conditions The likelihood probability, The marginal probability of the symptom combination;

[0028] Assuming that the symptoms are independent of each other under given disease conditions (Naive Bayes assumption), the likelihood probability is decomposed as follows:

[0029]

[0030] Marginal probabilities are calculated using the law of total probability:

[0031]

[0032] After calculating the posterior probability of all candidate diseases, the system sorts them from highest to lowest probability to generate a differential diagnosis list. For the top-ranked diseases, the system retrieves their typical symptoms, supporting evidence (characteristic symptoms exhibited by the patient), and opposing evidence (common symptoms not exhibited by the patient) from the knowledge graph, providing doctors with detailed explanations of the diagnostic basis. This probability-based reasoning method can handle the uncertainty of symptoms and the diversity of diseases, and has stronger robustness and interpretability compared to traditional rule-matching methods.

[0033] Furthermore, the multi-turn dialogue consultation module includes:

[0034] After initially calculating the disease probability distribution through the aforementioned Bayesian inference, this invention designs an active learning consultation strategy based on information entropy to maximize information gain and quickly identify the target disease during the dialogue. This strategy forms a closed loop with Bayesian inference: after each new symptom is asked, the Bayesian posterior probability is immediately updated, and the next optimal question is calculated based on the new probability distribution; during the consultation process, the system maintains the current disease probability distribution. ,in For the first The set of symptoms obtained through round-robin dialogue, Indicates the dialogue round number;

[0035] The system calculates the information entropy of the current disease probability distribution:

[0036]

[0037] in This indicates the uncertainty of the current diagnosis; the higher the entropy value, the greater the ambiguity in disease identification. To reduce uncertainty, the system needs to select the most valuable symptoms for inquiry.

[0038] For candidate symptoms (Symptoms not yet inquired about, subscript c indicates candidate), the system predicts the potential information gain after inquiring about these symptoms; patient's response to symptoms There are two possible answers: Existence (The superscript + indicates a positive result) or not present (Superscript - indicates negative), the probability of each answer is:

[0039]

[0040] Ask about symptoms The expected information entropy after that is:

[0041]

[0042] in and These are the updated probability distribution entropies when the patient answers yes and no, respectively;

[0043] Information gain is defined as the difference between the current entropy and the expected entropy:

[0044]

[0045] The system selects the symptom with the greatest information gain for inquiry, i.e. This strategy ensures that each question minimizes diagnostic uncertainty, quickly pinpointing the target disease within a limited number of dialogue rounds, and mimicking the diagnostic thought process of a clinician. Compared to random or sequential questioning, this information gain-based strategy can reduce the average number of consultation rounds by more than 40%, significantly improving consultation efficiency.

[0046] Furthermore, the multimodal data fusion module includes:

[0047] After obtaining a preliminary diagnosis through symptom inference and multiple rounds of consultation, this invention integrates objective examination data such as medical images and laboratory reports into a Bayesian inference framework to further improve the accuracy and reliability of the diagnosis. This multimodal fusion method is an important supplement and validation to the aforementioned symptom diagnosis. For medical images (such as chest X-rays and CT scans), the system uses a deep convolutional neural network to extract lesion features, identify abnormal manifestations such as pulmonary nodules, fractures, and tumors, and output image diagnosis results and confidence levels. The image analysis module is based on advanced network architectures such as ResNet or DenseNet and is trained on large-scale labeled image datasets, achieving a diagnostic level close to that of professional radiologists.

[0048] For test reports (such as complete blood count and biochemical indicators), the system compares each indicator value with the normal reference range to identify abnormal items (e.g., elevated white blood cell count suggests infection, elevated liver enzymes suggest liver damage). The test result analysis module also considers the correlation between indicators; for example, the simultaneous presence of anemia, thrombocytopenia, and leukopenia suggests abnormal bone marrow hematopoietic function. These objective test results are input as additional evidence into the Bayesian inference model, and are combined with symptom information to calculate the probability of disease.

[0049] Let the imaging diagnosis result be ,in For the identified image features, The total number of image features is used to detect anomalies. ,in As an anomaly detection indicator, This represents the total number of abnormal indicators.

[0050] The multimodal posterior probability, which combines symptoms, imaging, and laboratory tests, is calculated as follows:

[0051]

[0052] in and In the disease Imaging manifestations under certain conditions and test anomalies The conditional probabilities, whose parameters are also statistically derived from knowledge graphs and historical case data, are used to extend the aforementioned symptom reasoning to multimodal reasoning by multiplying the probabilities of evidence from different modalities using the Naive Bayes assumption. Multimodal data fusion significantly improves diagnostic accuracy, particularly for diseases requiring objective examinations (such as pneumonia, fractures, and tumors), increasing diagnostic accuracy by more than 25% compared to symptom-based diagnosis alone.

[0053] Furthermore, the treatment plan generation module includes:

[0054] After diagnosing the disease and identifying the target disease through the aforementioned medical knowledge graph reasoning, the system enters the treatment plan generation stage. This module generates personalized treatment plans based on the disease-treatment relationships in the knowledge graph and combined with the patient's individual characteristics. The treatment plan includes drug selection, dosage, course of treatment, precautions, and non-drug treatment suggestions. The system comprehensively considers individual factors such as the patient's age, gender, weight, past medical history, allergy history, liver and kidney function, and other medications currently in use to ensure the effectiveness and safety of the treatment plan. This module, together with the aforementioned diagnostic module, forms a complete diagnosis-treatment chain.

[0055] For diagnosing diseases (The subscript diag indicates diagnosed) The system retrieves the set of applicable drugs from the knowledge graph. and their preferred order, wherein As a candidate drug, This represents the total number of drugs; drug selection must comprehensively consider efficacy, safety, and individual patient factors. The system prioritizes first-line drugs from clinical practice guidelines, while also screening and ranking them based on the patient's contraindications, allergy history, and liver and kidney function.

[0056] The system adjusts drug dosage based on the patient's age, weight, and liver and kidney function indicators. For drugs requiring renal function adjustment, the dosage is adjusted based on creatinine clearance rate. Calculating and adjusting the dosage; Creatinine clearance reflects the kidneys' ability to filter and excrete metabolic waste products and is an important indicator for assessing kidney function. Many drugs metabolized by the kidneys require dosage adjustments based on kidney function to avoid drug accumulation and toxicity. This invention uses the widely used Cockcroft-Gault formula to estimate creatinine clearance. This formula establishes empirical relationships based on physiological parameters such as age, weight, serum creatinine concentration, and sex, enabling rapid estimation of kidney function without 24-hour urine collection.

[0057]

[0058] in Creatinine clearance rate (mL / min) The patient's age (in years). Weight (kg) This refers to serum creatinine concentration (mg / dL). The coefficient is for gender (1 for males, 0.85 for females, due to women having relatively less muscle mass); in the formula, 140 minus age reflects the decline in kidney function with age, body weight reflects creatinine production, and serum creatinine concentration reflects the kidney's clearance capacity; according to The system queries the drug dosage adjustment table to determine the safe dosage; for example, some antibiotics are used in patients with renal insufficiency (…). When the dosage needs to be reduced (to a certain level), the system automatically calculates the adjusted dose and notifies the doctor. For pediatric patients, the system calculates the pediatric dose based on weight or body surface area; for elderly patients, the system considers organ function decline and slowed drug metabolism, appropriately reducing the starting dose. This personalized dose adjustment mechanism effectively reduces the risk of adverse drug reactions and improves medication safety.

[0059] This invention constructs a drug-drug interaction database within a knowledge graph, recording the types and severity (mild, moderate, severe, contraindicated) of interactions between drug pairs. The interaction knowledge is derived from authoritative sources such as drug instructions, pharmacological literature, and clinical pharmacy guidelines. When the system recommends drug combinations or a patient has a history of medication use, it automatically performs interaction checks. For detected interactions, the system assesses the risk level and provides warning information. For example, the combination of warfarin and aspirin carries a serious risk of bleeding, and the system will mark it as "contraindicated" and suggest alternative treatments. The combined use of calcium and tetracycline may affect absorption, and the system will prompt "take them separately with a 2-hour interval." The combined use of digoxin and furosemide may increase digoxin toxicity due to low potassium, and the system will prompt "monitor blood potassium levels."

[0060] In addition, the system also checks for contraindications between drugs and patient conditions (e.g., beta-blockers are contraindicated in asthma patients, and anticholinergic drugs are contraindicated in glaucoma patients) and drug-allergy history (e.g., penicillin-allergic patients are contraindicated, and sulfonamide-allergic patients are contraindicated), comprehensively ensuring medication safety. It can identify over 90% of clinically important interactions, effectively avoiding medication errors.

[0061] Furthermore, it also includes a rare disease screening module. During the aforementioned diagnostic process, if the posterior probabilities of all common diseases are found to be below a set threshold (usually 0.6), indicating that the patient's symptom combination does not conform to common disease patterns, the system automatically activates the rare disease screening module. This module works in parallel with the common disease diagnosis module, forming a dual-guarantee mechanism for diagnosis. The rare disease knowledge base contains symptom spectra, diagnostic criteria, and characteristic examinations for thousands of rare diseases; rare disease data comes from the Orphanet rare disease database, medical literature, expert experience, etc.

[0062] The system calculates the matching degree between patient symptom combinations and various rare diseases, using weighted Jaccard similarity:

[0063]

[0064] in Indicates similarity. It is a candidate rare disease (the subscript rare indicates a rare disease). The set of symptoms observed by the patient (consistent with the aforementioned definition). This is a collection of typical symptoms of this rare disease. This represents the intersection operation of sets. This represents the union operation of sets. For symptoms The characteristic weights are calculated using the inverse document frequency (IDF) method. Rare and highly specific symptoms (such as "double vision + dysphagia + muscle weakness" indicating myasthenia gravis) are given higher weights, while common non-specific symptoms (such as fever and fatigue) are given lower weights.

[0065] For rare diseases with a high degree of matching, the system prompts doctors to pay attention and recommends specific tests (such as genetic testing, special biochemical indicators, and immunological tests). Simultaneously, the system assesses the severity and urgency of the patient's condition. For high-risk patients (such as those suspected of having myocardial infarction, stroke, acute infection, or visceral rupture and bleeding), it recommends emergency referral to the emergency department. For complex cases (such as those involving multiple systems, unclear diagnosis, or poor treatment response), it recommends referral to specialized hospitals (such as neurology, rheumatology, and hematology) to avoid misdiagnosis and delayed treatment by primary care physicians. It comprehensively considers factors such as the severity of the disease, the capacity of primary care facilities, and the patient's medical costs to provide reasonable referral recommendations.

[0066] Furthermore, the follow-up management module includes:

[0067] After completing disease diagnosis and treatment plan generation, to ensure treatment effectiveness and form a closed-loop diagnosis and treatment system, the system enters the patient education and follow-up management stage. This module is the final link in the diagnosis and treatment process, upgrading the system from a "diagnostic tool" to a "full-process management platform." The system automatically generates patient health education materials, explaining the etiology, pathogenesis, prognosis, daily care points, and follow-up appointments of the disease in plain language. The health education content is reviewed by medical experts to ensure scientific accuracy and easy understanding, avoiding the use of professional jargon that may cause difficulties for patients. For example, for pneumonia patients, the system explains that "pneumonia is an inflammatory infection of the lung tissue, mainly caused by bacteria or viruses, manifested as symptoms such as fever, cough, sputum production, and chest pain, and most cases can be cured with antibiotic treatment."

[0068] At the same time, personalized follow-up plans are developed, such as "re-examining blood routine tests 3 days after taking antibiotics to observe changes in infection indicators," "measuring blood pressure 2 weeks after taking antihypertensive drugs to assess efficacy and adjust dosage," and "monthly follow-up visits for patients with chronic diseases to monitor disease progression." The follow-up plan is customized according to the type of disease, treatment plan, and patient risk level.

[0069] The system pushes medication reminders, follow-up appointment reminders, and health advice to patients via a mobile application, and collects follow-up data on patient self-reported symptom changes, medication adherence, and adverse reactions. The mobile application features a simple and user-friendly interface, supporting multiple interaction methods such as voice input and image upload, making it convenient for elderly and less literate patients. Based on the follow-up data, the system evaluates the treatment effect. For patients with poor efficacy (such as no improvement or worsening of symptoms after 3 days of medication, or no improvement in test indicators), the system recommends adjusting the treatment plan or referring them to another specialist. For patients with good recovery, the system provides encouragement and suggestions for preventing relapse (such as vaccination, lifestyle improvement, and regular check-ups), forming a closed loop of diagnosis and treatment.

[0070] Through the above-mentioned technical solution, this invention realizes intelligent management of the entire process from symptom collection, disease inference, examination suggestions, treatment plan generation to follow-up management, providing high-quality diagnostic auxiliary tools for primary healthcare institutions and Internet healthcare platforms, effectively improving the level of diagnosis and treatment, reducing the rate of misdiagnosis and missed diagnosis, and improving the patient's medical experience.

[0071] The beneficial effects of this invention are as follows:

[0072] 1. By constructing a large-scale medical knowledge graph and applying Bayesian probabilistic reasoning, this invention can systematically analyze the association between symptom combinations and diseases, improving diagnostic accuracy by 20-30 percentage points compared to traditional rule-matching methods. The accuracy improvement is particularly significant for cases with complex symptoms requiring differential diagnosis. After integrating multimodal data, the diagnostic accuracy is further improved by over 25%.

[0073] 2. By adopting an information gain-based active learning consultation strategy, the system can intelligently select the most valuable symptoms to inquire about, reducing the average number of dialogue rounds from 15-20 rounds in traditional methods to 8-10 rounds, shortening the consultation time by more than 40%, significantly improving the communication efficiency between doctors and patients, and improving the medical experience.

[0074] 3. By integrating multimodal data such as symptoms, imaging, and laboratory tests, and making full use of objective examination results, the reliability of diagnosis is improved by more than 25%. In particular, for diseases that rely on imaging or laboratory results (such as pneumonia, fractures, tumors, and infections), multimodal fusion diagnosis is superior to simple symptom diagnosis, reducing the misdiagnosis rate by more than 30%.

[0075] 4. The system automatically checks for drug interactions, incompatibilities, allergy history, dosage adjustments, and other medication safety issues, identifying over 90% of potential medication risks, effectively avoiding medication errors and reducing the incidence of adverse drug reactions. Personalized dosage adjustments (based on age, weight, and liver and kidney function) make medication use more precise and safer.

[0076] 5. By constructing a rare disease knowledge base and employing intelligent matching algorithms, the system can identify thousands of rare diseases. The identification rate for rare diseases with atypical symptoms exceeds 70%, preventing primary care physicians from missing diagnoses and delaying treatment due to a lack of relevant knowledge. The rare disease identification module pays special attention to rare and highly specific symptom combinations, improving identification sensitivity.

[0077] 6. Provide expert-level diagnostic assistance to primary healthcare institutions, make up for the lack of experience of primary care physicians, enable the diagnostic level of primary healthcare institutions to approach that of tertiary hospitals, effectively alleviate the problem of uneven distribution of medical resources, promote the implementation of hierarchical medical treatment, and reduce the cost of medical treatment for patients.

[0078] 7. The system can be deployed on internet healthcare platforms to provide patients with 24-hour intelligent consultation services, conduct preliminary screenings of conditions, and offer medical advice, alleviating doctors' workload and improving patients' healthcare experience. Especially during peak medical resource periods such as nighttimes and holidays, the system can provide timely health consultation services.

[0079] 8. Through the follow-up management module, the system continuously tracks the patient's treatment progress, evaluates the efficacy, and adjusts the treatment plan in a timely manner, improving patient compliance and cure rate. Compared with traditional treatment models, the recovery period is shortened by 15-20%, the recurrence rate is reduced by 10-15%, and patient satisfaction is increased by more than 30%.

[0080] 9. The knowledge graph adopts a modular design, allowing for easy addition of new diseases, drugs, and treatment guidelines, supporting continuous learning and knowledge updates. By analyzing clinical data, the system can automatically optimize probability parameters and inference strategies, resulting in continuously improving diagnostic performance over time, and exhibiting excellent scalability and maintainability.

[0081] 10. This invention can be applied to multiple fields such as primary healthcare auxiliary diagnosis, online medical consultation, family health management, emergency triage, medical education and training, health check-up interpretation, chronic disease management, and rare disease screening. It has broad market application prospects and social value, and promotes the intelligent and inclusive development of healthcare.

[0082] 11. By improving the accuracy of primary care diagnoses, reducing misdiagnosis and mistreatment, and minimizing unnecessary examinations and medications, this system avoids increased treatment costs due to delayed diagnosis. It is estimated that applying this system can reduce the average medical cost per patient by 15-20%, and decrease the number of repeat visits, resulting in significant economic and social benefits.

[0083] 12. The treatment plans recommended by the system are all based on clinical practice guidelines and evidence-based medicine, avoiding the subjectivity and arbitrariness of empirical medicine, promoting the transformation of medicine from "empirical medicine" to "evidence-based medicine", improving the standardization and normalization of medical services, and ensuring medical quality and patient safety.

[0084] 13. The system not only provides diagnostic conclusions but also explains supporting evidence (characteristic symptoms exhibited by the patient) and opposing evidence (common symptoms not exhibited by the patient) for each disease, providing a transparent diagnostic reasoning process. This interpretability helps doctors understand the system's recommendations, enhances their trust in the system, and promotes the establishment of a human-machine collaborative diagnosis and treatment model. Attached Figure Description

[0085] Figure 1 This is a flowchart of the present invention.

[0086] Figure 2 This is a diagram of the architecture of the present invention. Detailed Implementation

[0087] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0088] As shown in the figure, this embodiment provides an intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph, with the application scenario of auxiliary diagnosis of respiratory diseases in primary medical institutions. The specific solution is as follows:

[0089] I. System Architecture and Knowledge Graph Construction

[0090] An intelligent medical consultation and auxiliary diagnostic system is constructed, comprising modules for knowledge graph construction, common disease diagnosis, multi-turn dialogue consultation, multimodal data fusion, treatment plan generation, and follow-up management. The system adopts a layered architecture: the data layer is responsible for medical knowledge storage and case data management; the logic layer implements reasoning algorithms and dialogue strategies; and the application layer provides the user interface and external system interfaces.

[0091] Medical knowledge graph construction and implementation:

[0092] The system constructs a medical knowledge graph encompassing 5,000 common diseases, 15,000 symptoms and signs, 8,000 examination items, 12,000 drugs, and 6,000 treatment plans. Knowledge sources include: medical textbooks such as *Practical Internal Medicine*, clinical practice guidelines from the Chinese Medical Association, the *Chinese Pharmacopoeia* and drug instructions, desensitized electronic medical record data (100,000 records), and medical literature from PubMed.

[0093] The knowledge graph is stored using the Neo4j graph database and contains entity nodes and relation edges. The main entity types include: Disease, Symptom, Examination, Medicine, and Treatment. The main relation types include: hasSymptom, requiresExam, treatedByMedicine, drugInteraction, and contraindication.

[0094] Each relationship comes with a probability or weight attribute. For example, the relationship Conditional probability attribute This indicates that 85% of pneumonia patients will experience fever. These probability parameters were obtained by analyzing 100,000 anonymized electronic medical records. The conditional probabilities were calculated using the maximum likelihood estimation method, and Laplace smoothing was used to handle zero-probability problems.

[0095] II. Implementation of Disease Diagnosis Based on Bayesian Inference

[0096] 1) Symptom Information Collection

[0097] The patient entered their chief complaint through the system interface: "I have had a persistent cough for the past three days, accompanied by a fever of around 38.5°C."

[0098] The system uses a pre-trained natural language understanding model in the medical field (based on BioBERT fine-tuning) to perform named entity recognition and relation extraction on the chief complaint, extracting symptom entities: and its attributes: The entity recognition module employs a Conditional Random Field (CRF) sequence annotation method, trained on a medical text corpus, achieving a recognition accuracy of 92%.

[0099] The system maps the extracted symptoms to standard symptom nodes in the knowledge graph, processes synonyms and hierarchical relationships (e.g., "fever" is mapped to "heat," "high fever" is identified as "heat" with a body temperature attribute >39℃), and constructs an initial symptom set. .

[0100] 2) Calculation of posterior probability of disease

[0101] Based on the initial symptom set, the system retrieves a set of candidate diseases associated with these symptoms from the knowledge graph. Candidate diseases are obtained by traversing the "hasSymptom" relationship. 20 diseases including.

[0102] The system uses Bayes' theorem to calculate the posterior probability of each candidate disease. Taking pneumonia as an example:

[0103]

[0104] Found from the knowledge graph: , , (Population prevalence, obtained from epidemiological data).

[0105] Assuming symptoms are independent, the likelihood probability is:

[0106]

[0107] Marginal probabilities are calculated by summing over all candidate diseases:

[0108]

[0109] The posterior probability of pneumonia is:

[0110]

[0111] Similarly, calculate the posterior probabilities for other diseases: , , , .

[0112] The most probable disease at present is pneumonia (31.9%), but the uncertainty is still high, and further consultation and symptom collection are needed to differentiate the symptoms.

[0113] III. Implementation of the Multi-Round Diagnosis Strategy

[0114] 1) Information gain calculation and question selection

[0115] The system calculates the information entropy of the current disease probability distribution:

[0116]

[0117]

[0118] A high entropy value indicates significant diagnostic uncertainty. The system needs to select the most valuable symptoms for inquiry.

[0119] Retrieve relevant symptoms of candidate diseases from the knowledge graph and filter out a candidate set of differential symptom candidates that have not yet been queried: There are 15 symptoms, including...

[0120] For each candidate symptom, calculate the expected information entropy after the query. Take the symptom "chest pain" as an example:

[0121] Calculate the probability of the patient experiencing chest pain:

[0122]

[0123] If the patient answers "chest pain", the updated disease probability is recalculated using Bayes' theorem to obtain a new probability distribution. Calculate its entropy bits; if the answer is "no chest pain", entropy bits.

[0124] The expected entropy is:

[0125]

[0126] Information gain is:

[0127]

[0128] Similarly, calculate the information gain for other candidate symptoms: bits bits bits bits.

[0129] The system selects "chest pain," the symptom with the greatest information gain, and asks: "Do you experience chest pain or discomfort?"

[0130] 2) Dialogue round iteration

[0131] The patient replied, "Yes, I feel a stinging sensation in my chest when I take a deep breath."

[0132] System extracts symptoms And its attributes "tingling sensation" and "breathing-related", update the symptom set. The conditional probabilities of chest pain in various diseases were obtained from a knowledge graph query. , , The disease probability was recalculated using the aforementioned Bayesian formula, and the updated probability was... , , The probability of pneumonia increases significantly because chest pain (especially pleural chest pain) is a characteristic symptom of pneumonia.

[0133] The system continues to calculate the information gain for the next question, selecting "Properties of Coughing" (information gain 0.31 bits) to ask: "Do you cough up phlegm? What color is your phlegm?"

[0134] The patient replied, "I have thick yellow phlegm."

[0135] Update symptom set Thick yellow sputum indicates a bacterial infection. A search of the knowledge graph revealed: , Apply Bayes' theorem again to update the probability, and then calculate... , .

[0136] The system determined that the probability of a pneumonia diagnosis had reached the threshold (set at 0.75) and could provide a preliminary diagnostic suggestion, but recommended further examination for confirmation. After three rounds of dialogue, the system successfully identified the target disease, demonstrating significantly higher consultation efficiency than traditional fixed questionnaire methods.

[0137] IV. Implementation of Multimodal Data Fusion Diagnosis

[0138] The system recommends that the patient undergo a chest X-ray and a complete blood count to confirm the diagnosis. The results are as follows:

[0139] Chest X-ray: shows patchy shadows in the right lower lung field with blurred edges, consistent with lung inflammation.

[0140] Complete blood count: White blood cell count (Normal range) The percentage of neutrophils was 85% (normal range 50%-70%), suggesting a bacterial infection.

[0141] The system uses a pre-trained convolutional neural network (based on the DenseNet-121 architecture, trained on the ChestX-ray14 dataset) to automatically analyze chest X-ray images, identify pulmonary infiltrative lesions, and output image diagnosis results. The confidence level was 0.92. The imaging analysis module also located the lesion (right lower lung field) and assessed the extent of the lesion (approximately 15% of the lung field area).

[0142] The blood routine analysis module identified abnormal indicators. Based on the elevated white blood cell count and the increased proportion of neutrophils, the system judges that a bacterial infection is highly likely.

[0143] The system applies the aforementioned multimodal fusion formula to integrate symptoms, imaging, and laboratory data to calculate the comprehensive posterior probability:

[0144]

[0145] The conditional probabilities of the images and tests obtained from the knowledge graph query are as follows: (88% of pneumonia patients showed infiltrating on X-ray). (82% of pneumonia patients have elevated white blood cell counts).

[0146] Symptom Collection It includes four symptoms: cough, fever, chest pain, and yellow phlegm. Applying the Naive Bayes hypothesis, the conditional probabilities of each symptom obtained from the previous query are multiplied together: .

[0147] Prior probability of disease .

[0148] After normalization, the calculation yields The diagnostic confidence increased significantly from 78% to 95% in symptom-based diagnoses.

[0149] The system outputs the following diagnostic conclusion: "Based on the symptoms (cough, fever, chest pain, yellow sputum), imaging examination (chest X-ray showing infiltrative lesions in the right lower lung field) and blood routine results (elevated white blood cell and neutrophil counts), the patient is highly likely to have community-acquired pneumonia (bacterial), and anti-infective treatment for pneumonia is recommended."

[0150] V. Generation and Implementation of Personalized Treatment Plans

[0151] 1) Drug selection and dosage adjustment

[0152] The system retrieves recommended treatment options for pneumonia from the knowledge graph. According to the "Guidelines for the Diagnosis and Treatment of Community-Acquired Pneumonia in Chinese Adults", first-line antibiotics include amoxicillin-clavulanic acid, levofloxacin, azithromycin, etc.

[0153] The system inquired about the patient's basic information: age 45, weight 70kg, no history of drug allergies, previously healthy, no abnormal liver or kidney function, and no other medications.

[0154] Based on the patient's age (adult) and absence of contraindications, amoxicillin-clavulanic acid was selected as the first-line drug (broad-spectrum antibacterial, effective, and safe). The standard dose is 1000 mg (875 mg amoxicillin + 125 mg clavulanic acid) twice daily (once in the morning and once in the evening) for 7-10 days.

[0155] The system queries the patient's renal function indicators: serum creatinine. mg / dL (normal range 0.7-1.3 mg / dL). Calculate creatinine clearance using the Cockcroft-Gault formula:

[0156]

[0157] If creatinine clearance is normal (>60 mL / min), no dose adjustment is required. If the patient has renal insufficiency ( (mL / min), the system will automatically adjust the dose to 500mg twice daily or 1000mg once daily.

[0158] 2) Medication safety inspection

[0159] A systematic review of the patient's past medication history revealed that the patient was taking the nonsteroidal anti-inflammatory drug ibuprofen for fever reduction and pain relief.

[0160] The system searched the drug interaction knowledge base and detected that "amoxicillin-clavulanic acid + ibuprofen" is a mild interaction (ibuprofen may slightly affect the renal excretion of antibiotics, but this is not clinically significant). They can be used together. It is recommended to drink plenty of water to promote drug excretion.

[0161] Further system checks for contraindications: The patient has no history of penicillin allergy (amoxicillin belongs to the penicillin class), no abnormal liver function (clavulanic acid may cause liver damage, so it should be used with caution in patients with abnormal liver function), and no infectious mononucleosis (patients with this disease may develop a rash when using amoxicillin), thus meeting the medication requirements.

[0162] The system generates personalized treatment plans:

[0163] Drug treatment: 1. Amoxicillin-clavulanic acid tablets 1000mg, twice daily (morning and evening), taken after meals to reduce gastrointestinal reactions, for 7-10 days. 2. Ibuprofen sustained-release capsules 300mg, taken when fever (body temperature >38.5℃) or significant pain, not exceeding 3 times daily, with an interval of at least 6 hours between doses.

[0164] Non-pharmacological treatment: 3. Drink plenty of water (2000-2500mL daily) to help dilute sputum and promote drug excretion. 4. Rest and avoid strenuous exercise and heavy physical labor. 5. Keep the room well-ventilated and avoid smoking and secondhand smoke.

[0165] Follow-up plan: 6. After 3 days of medication, repeat blood routine test to observe changes in white blood cells and neutrophils and assess infection control. 7. If symptoms do not improve or worsen after 3 days (such as persistent high fever, worsening dyspnea, increased purulent sputum), seek medical attention promptly to adjust the treatment plan. 8. After the course of treatment, repeat chest X-ray to confirm absorption of lung lesions.

[0166] VI. Follow-up Management and Implementation of Treatment Efficacy Evaluation

[0167] The system creates follow-up plans for patients and sends reminders and collects data via a mobile app.

[0168] Day 1 (treatment begins): The system sends a medication reminder: "Please take amoxicillin-clavulanic acid tablets as prescribed, once in the morning and once in the evening, after meals." Patients are required to record their daily body temperature (once in the morning and once in the evening) and changes in symptoms (frequency of cough, amount and color of sputum, degree of chest pain).

[0169] The patient reported that evening: temperature 37.8℃ (decreased), cough slightly reduced, still with yellow sputum, chest pain reduced. Systemic assessment indicated that treatment was beginning to show results.

[0170] Day 3 (First Follow-up): The system sends a follow-up reminder: "Please go to the hospital for a blood routine test to understand the infection control situation."

[0171] The patient uploaded the follow-up test results: the white blood cell count decreased. The proportion of neutrophils dropped to 72%, indicating that the infection was well controlled.

[0172] The system inquired about the patient's symptom changes, and the patient reported: "Body temperature returned to normal (36.5℃), cough significantly reduced, still a small amount of white phlegm (changed from yellow to white), chest pain basically disappeared."

[0173] System analysis: White blood cell and neutrophil counts decreased, symptoms improved, and sputum color changed from yellow to white (indicating bacterial clearance). Treatment was assessed as effective, and it is recommended to continue the original medication regimen until the end of the course of treatment. An encouraging message was sent: "Treatment is effective; please continue taking your medication as prescribed and do not stop taking it on your own."

[0174] Day 7 (End of Treatment): The system sends a notification that the treatment is complete and inquires about the patient's current symptoms. The patient reports: "No more fever or chest pain, occasional mild cough, no sputum, and good mental state."

[0175] The systemic assessment indicates good recovery; it is recommended to discontinue antibiotics (to avoid overuse) and continue observation for one week. If a mild cough persists, cough suppressants and expectorants can be taken for symptomatic relief.

[0176] Simultaneously, health education content is pushed out:

[0177] Pneumonia prevention measures: Get vaccinated against pneumococcus and influenza to reduce the risk of infection; avoid getting chilled and adjust clothing according to weather changes; quit smoking and limit alcohol consumption, as smoking damages the respiratory tract's defense function; strengthen exercise to improve the body's resistance;

[0178] Key points for rehabilitation care: Strengthen nutrition, eat more foods rich in protein and vitamins; exercise moderately, starting with light activities such as walking, and gradually restore physical strength; ensure sufficient sleep to promote body repair;

[0179] Relapse symptom identification: If fever, worsening cough, or difficulty breathing recurs, seek medical attention promptly; a mild cough may still be present within one month after pneumonia recovery, which is part of the normal recovery process;

[0180] Day 14 (Final Follow-up): The system conducted a final follow-up assessment, asking, "Do you currently have any discomfort symptoms?" The patient reported, "Completely recovered, no discomfort, and have returned to normal work."

[0181] The system records the diagnosis and treatment outcome: Community-acquired pneumonia (bacterial), clinically cured after 7 days of treatment with amoxicillin-clavulanic acid. The patient's health record is updated, completing the diagnosis and treatment loop.

[0182] This complete diagnosis and treatment process demonstrates the system's end-to-end intelligent capabilities, from symptom collection, multi-round consultation, disease reasoning, multimodal fusion diagnosis, personalized treatment to follow-up management. Through the detailed description of this embodiment, those skilled in the art can clearly understand the technical solution of this invention and can implement it in practical application scenarios based on the described methods and system configurations to achieve efficient and accurate intelligent medical consultation and assisted diagnosis.

[0183] In summary, this invention constructs a large-scale medical knowledge graph covering diseases, symptoms, examinations, drugs, and treatment plans. Medical knowledge is represented using triples, with each relation accompanied by a probability or weight attribute, supporting quantitative reasoning. Knowledge sources include multi-source heterogeneous data such as medical textbooks, clinical guidelines, electronic medical records, and medical literature. Natural language processing techniques (named entity recognition, relation extraction) are used for automated knowledge extraction and fusion. The knowledge graph supports efficient graph query and reasoning operations, providing a knowledge foundation for disease diagnosis and treatment plan generation.

[0184] This invention applies Bayes' theorem to calculate the posterior probability of a disease, comprehensively considering the conditional probability of symptoms and the prior probability of the disease, and performs probabilistic inference under conditions of uncertainty. It supports joint probability calculation of multiple symptoms, multiple examinations, and multimodal data, and can handle the uncertainty of symptoms and the diversity of diseases. The inference results not only provide the most probable disease but also generate a list of differential diagnoses and explain the supporting and opposing evidence for each disease, providing interpretable diagnostic suggestions.

[0185] This invention designs an active learning algorithm based on information gain, which dynamically selects the most valuable symptoms to inquire about during multi-turn dialogues, maximizing the diagnostic information gain brought by each question and rapidly reducing diagnostic uncertainty within a limited number of dialogue rounds. This strategy simulates the diagnostic thinking of a clinician, prioritizing the inquiry of key symptoms required for differential diagnosis, avoiding ineffective questions, and significantly improving consultation efficiency and patient experience.

[0186] This invention integrates heterogeneous data from multiple sources, including symptom text, medical images, and laboratory reports. It employs a deep learning model (convolutional neural network) to extract lesion features from images and identify abnormal patterns from laboratory indicators. The multimodal evidence is then fed into a Bayesian inference model to calculate the comprehensive posterior probability. This multimodal fusion fully utilizes the complementary information from subjective symptoms and objective examinations, significantly improving diagnostic accuracy and reliability.

[0187] This invention generates personalized treatment plans based on knowledge graphs, adjusting drug dosages according to factors such as patient age, weight, and liver and kidney function (using clinical calculation formulas such as Cockcroft-Gault). It also checks for drug interactions, contraindications, allergies, and other medication safety issues to ensure the effectiveness and safety of the treatment plan. The medication safety check module can identify over 90% of clinically important interactions and contraindications, effectively preventing medication errors.

[0188] This invention constructs a rare disease knowledge base and uses a weighted similarity algorithm (weighted Jaccard) to match patient symptoms with the symptom spectrum of rare diseases, identifying low-probability but high-risk rare diseases. It assesses the severity and urgency of the condition and intelligently decides whether referral to a specialist hospital or emergency department is necessary. The rare disease identification module places particular emphasis on rare and highly specific symptoms, assigning them higher weights to improve the sensitivity of rare disease identification.

[0189] This invention establishes a personalized follow-up plan, which uses mobile devices to push medication and follow-up appointment reminders, collects patient self-reported data (symptom changes, medication adherence, adverse reactions), assesses treatment effectiveness, adjusts treatment plans promptly for patients with poor efficacy, and provides relapse prevention advice for patients with good recovery, forming a complete closed-loop management system from diagnosis to treatment to rehabilitation. Follow-up management significantly improves patient adherence and cure rate, and reduces relapse rate.

[0190] The above embodiments are merely typical illustrative methods of the present invention, and the scope of protection of the present invention is not limited thereto. All equivalent substitutions and improvements made under the concept of the present invention should fall within the scope of protection. It should be emphasized that any modifications or minor adjustments made by those skilled in the art without departing from the basic principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph, characterized in that, include: Knowledge graph construction module: Constructs a large-scale medical knowledge graph containing disease entities, symptom entities, examination entities, drug entities, treatment plan entities and their relationships, and combines patient symptom information to achieve automated knowledge extraction through natural language processing technology, completing the structured representation of the initial symptom set; Common disease diagnosis module: Construct a quantitative association model between symptoms and diseases, calculate the posterior probability of all candidate diseases based on the initial symptom set, and obtain a preliminary disease probability distribution; Multi-turn dialogue consultation module: By asking about new symptoms, the posterior probability is updated, and the next optimal question is calculated based on the new probability distribution to update the symptom information; Multimodal data fusion module: Integrates objective examination data from medical images and laboratory reports, and calculates the posterior probability of the disease in conjunction with symptom information to identify the target disease; For medical images, the system uses a deep convolutional neural network to extract lesion features and outputs image diagnosis results and confidence levels; For laboratory reports, the system compares the values ​​of each indicator with the normal reference range to identify abnormal items. Treatment plan generation module: Based on the identified target disease, and using the disease-treatment relationship in the knowledge graph, combined with the individual characteristics of the patient, a personalized treatment plan is generated. Follow-up management module: Automatically generates health education materials for patients, explaining the causes, pathogenesis, prognosis, daily care points, and timing of follow-up visits in plain language.

2. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, The knowledge graph construction module includes: The knowledge graph uses triples to represent medical knowledge, with each triple recording as follows: The knowledge graph is constructed by extracting data from multiple sources, including medical textbooks, clinical practice guidelines, electronic medical records, drug instructions, and medical literature.

3. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, The common disease diagnosis module includes: For symptom set and disease collection By statistically analyzing large-scale electronic medical record data, calculations can be made regarding diseases. Symptoms appear under certain conditions conditional probability and symptoms Prior probability in the general population These probability parameters are stored in the relational attributes of the knowledge graph, forming a quantitative association model between symptoms and diseases. When patients exhibit a combination of symptoms At that time, the system calculates the posterior probability of each candidate disease based on Bayes' theorem; for each disease... The posterior probability is calculated as follows: in For disease The prevalence of disease in a population, i.e., the prior probability. In the disease Symptom combinations observed under certain conditions The likelihood probability, The marginal probability of the symptom combination; Assuming that the symptoms are independent under given disease conditions, the likelihood probability is decomposed as follows: Marginal probabilities are calculated using the law of total probability: After calculating the posterior probability of all candidate diseases, the system sorts them from high to low probability values ​​to generate a differential diagnosis list. For the top-ranked diseases, the system retrieves their typical symptoms, supporting evidence, and opposing evidence from the knowledge graph to provide doctors with detailed explanations of the diagnostic basis.

4. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, The multi-turn dialogue consultation module includes: After each new symptom is inquired about, the Bayesian posterior probability is immediately updated, and the next optimal question is calculated based on the new probability distribution. During the consultation process, the system maintains the current disease probability distribution. ,in For the first The set of symptoms obtained through round-robin dialogue, Indicates the dialogue round number; The system calculates the information entropy of the current disease probability distribution: in It indicates the uncertainty of the current diagnosis; the higher the entropy value, the greater the ambiguity in disease identification. For candidate symptoms The system predicts the potential information gain from inquiring about the symptom; the patient's response to the symptom... There are two possible answers: Existence Or does not exist The probability of each answer is: Ask about symptoms The expected information entropy after that is: in and These are the updated probability distribution entropies when the patient answers yes and no, respectively; Information gain is defined as the difference between the current entropy and the expected entropy: The system selects the symptom with the greatest information gain for inquiry, i.e. .

5. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, The multimodal data fusion module includes: Let the imaging diagnosis result be ,in For the identified image features, The total number of image features is used to detect anomalies. ,in As an anomaly detection indicator, This represents the total number of abnormal indicators. The multimodal posterior probability, which combines symptoms, imaging, and laboratory tests, is calculated as follows: in and In the disease Imaging manifestations under certain conditions and test anomalies The conditional probability.

6. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, The treatment plan generation module includes: Treatment plans include drug selection, dosage, course of treatment, precautions, and non-drug treatment recommendations; the system comprehensively considers individual factors such as the patient's age, gender, weight, medical history, allergy history, liver and kidney function, and other medications currently in use; For diagnosing diseases The system retrieves the set of applicable drugs from the knowledge graph. and their preferred order, wherein As a candidate drug, The total number of drugs is considered; the system prioritizes first-line drugs from clinical practice guidelines, while also screening and ranking them based on the patient's contraindications, allergy history, and liver and kidney function. The system adjusts drug dosage based on the patient's age, weight, and liver and kidney function indicators. For drugs requiring renal function adjustment, the dosage is adjusted based on creatinine clearance rate. Calculate the adjusted dose; estimate creatinine clearance using the Cockcroft-Gault formula: in Creatinine clearance rate For the patient's age, For weight, This refers to serum creatinine concentration. This is the sex coefficient; in the formula, 140 minus age reflects the decline in kidney function with age, body weight reflects creatinine production, and serum creatinine concentration reflects the kidney's clearance capacity; according to The system queries the drug dosage adjustment table to determine the safe dosage. A drug-drug interaction database was built in the knowledge graph to record the types and severity of interactions between drug pairs. When the system recommends drug combinations or when a patient has a history of medication use, the system automatically performs interaction checks. For detected interactions, the system assesses the risk level and provides warning information. In addition, the system also checks for contraindications between the drug and the patient's disease, and for the match between the drug and the patient's allergy history.

7. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, It also includes a rare disease screening module; if the posterior probability of all common diseases is found to be lower than the set threshold, it indicates that the patient's symptom combination does not conform to the common disease pattern, and the system will automatically start the rare disease screening module; the rare disease knowledge base contains symptom spectra, diagnostic criteria, and characteristic examinations for thousands of rare diseases; The system calculates the matching degree between patient symptom combinations and various rare diseases, using weighted Jaccard similarity: in Indicates similarity. As a candidate rare disease, A collection of symptoms observed by the patient. This is a collection of typical symptoms of this rare disease. This represents the intersection operation of sets. This represents the union operation of sets. For symptoms The characteristic weights; For rare diseases with a high degree of matching, the system prompts doctors to pay attention and suggests specific tests; at the same time, the system assesses the severity and urgency of the patient's condition, and suggests emergency referral to the emergency department for high-risk patients and referral to a specialist hospital for complex cases.

8. The intelligent medical consultation and auxiliary diagnosis system based on symptom reasoning and knowledge graph as described in claim 1, characterized in that, The follow-up management module includes: The system pushes medication reminders, follow-up appointment reminders, and health advice to patients through a mobile application, and collects follow-up data on patient self-reported symptom changes, medication adherence, and adverse reactions. Based on the follow-up data, the system evaluates the treatment effect, suggests adjusting the treatment plan or referring patients for those with poor efficacy, and provides encouragement and relapse prevention advice to patients with good recovery, thus forming a closed loop of diagnosis and treatment.

Images