Intelligent doctor auxiliary diagnosis and treatment system

Abstract

The invention relates to the technical field of auxiliary diagnosis and treatment systems, in particular to an intelligent auxiliary diagnosis and treatment system for doctors. Comprising a system initialization and role service registration module, a user input analysis and intention recognition module, an intelligent routing and professional tool set selection module, an SSE streaming response initialization and workflow starting module, a knowledge graph retrieval and reasoning module, a diagnosis and treatment logic reasoning and scheme generation module and a streaming result output and node state updating module. A dynamic knowledge update triggering and verifying module; according to the invention, on the system integration level, efficient integration of multi-professional medical resources is realized based on role-as-a-service architecture, and complete coverage of multi-professional knowledge modules of internal medicine, surgery, oncology and the like is ensured through a role service registration validity evaluation mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of auxiliary diagnosis and treatment systems, and in particular to an intelligent doctor auxiliary diagnosis and treatment system. BACKGROUND

[0002] In modern medical practice, with the intensification of population aging, the complication of disease spectrum and the explosive growth of medical knowledge, doctors have an increasingly urgent need for intelligent auxiliary diagnosis and treatment systems.

[0003] However, the existing medical auxiliary systems have many limitations in actual application, and it is difficult to meet the efficient and accurate needs of clinical diagnosis and treatment. The specific shortcomings are as follows: single professional coverage, difficulty in interdisciplinary cooperation, the existing systems are mostly focused on single disease field, and medical institutions need to deploy multiple independent systems to cover the needs of multiple departments, resulting in increased cost and serious data island, and cross-disciplinary case diagnosis and treatment need to be switched between multiple systems, which is cumbersome and easy to miss information. SUMMARY

[0004] In view of the technical problems of single professional coverage and difficulty in interdisciplinary cooperation of the medical auxiliary system in actual application, the present application provides an intelligent doctor auxiliary diagnosis and treatment system.

[0005] The technical scheme adopted by the present application is: an intelligent doctor auxiliary diagnosis and treatment system, comprising:

[0006] A system initialization and role service registration module, which abstracts different medical specialties into independent role services based on a role-as-a-service architecture, and realizes dynamic calling through a service registration mechanism. Each role service corresponds to a set of professional tool sets, including the knowledge graph, diagnosis and treatment rules and decision models in the field.

[0007] A user input analysis and intent recognition module, which receives diagnosis and treatment related information submitted by the user in natural language form, analyzes the input content and recognizes the user's intent through natural language processing technology.

[0008] An intelligent routing and professional tool set selection module, which automatically selects the most matching professional tool set for processing according to the user's intent and entity information.

[0009] An SSE streaming response initialization and workflow starting module, which starts the Server-SentEvents streaming response mechanism, sends a workflow start event to the user end, and establishes a real-time communication channel.

[0010] A knowledge graph retrieval and reasoning module that calls an integrated medical knowledge graph through a professional tool set, retrieves relevant medical knowledge, and performs reasoning;

[0011] A diagnosis and treatment logic reasoning and scheme generation module that performs diagnosis and treatment logic reasoning and generates preliminary assisted diagnosis and treatment suggestions based on the retrieved knowledge and clinical rules;

[0012] A streaming result output and node state update module that pushes the generated diagnosis and treatment results to users in blocks through an SSE mechanism and updates the node processing state in real time;

[0013] A dynamic knowledge update triggering and verification module that triggers a dynamic knowledge update mechanism according to new information or external medical progress during diagnosis and treatment, and verifies the effectiveness of the updated content;

[0014] A workflow completion and result summary module that pushes a workflow completion event and summarizes an assisted diagnosis report after all processing nodes are completed.

[0015] In one embodiment, the system initialization and role service registration module, the role service registration validity evaluation formula R valid is as follows:

[0016]

[0017] where R valid is the role service registration validity index; n is the number of professional core knowledge modules; w i is the weight of the i-th knowledge module; S ki is the i-th knowledge module content set registered by the system; S ri is the standard content set of the i-th professional knowledge module.

[0018] In one embodiment, the user input analysis and intent recognition module receives diagnosis and treatment-related information submitted by users in natural language form, analyzes the input content through natural language processing technology, and identifies user intent. The specific method is as follows:

[0019] A pre-trained language model based on Transformer is used in combination with field fine-tuning to realize the analysis of medical natural language. An intent recognition model is used to distinguish user demand types, including diagnosis suggestion generation, medical record writing assistance, and evidence-based medicine query;

[0020] User intent confidence calculation C j :

[0021]

[0022] wherein C j is the confidence of the jth intent; S j is the output score of the jth intent by the intent recognition model; and m is the total number of preset intent types.

[0023] The entity extraction accuracy P is evaluated as follows:

[0024]

[0025] wherein P is the entity extraction accuracy; T p is the number of correctly extracted entities; and F p is the number of incorrectly extracted entities.

[0026] In one embodiment, the intelligent routing and professional tool set selection module automatically selects the most matched professional tool set for processing according to the user intent and entity information, and the specific method is as follows:

[0027] Based on the entity correlation analysis and rule reasoning of the knowledge graph, the accurate mapping of user demand to professional tool sets is realized, and the intelligent routing mechanism considers the disease type, symptom characteristics, and patient population factors.

[0028] The tool set matching degree M t is calculated as follows:

[0029] M t = a x R et + b x R it + g x R dt

[0030] wherein M t is the matching degree of the tth tool set; a, b, and g are weight coefficients; R et is the entity matching rate; R it is the intent matching rate; and R dt is the disease field matching rate.

[0031] In one embodiment, the SSE streaming response initialization and workflow starting module starts the Server-SentEvents streaming response mechanism, sends a workflow start event to the user end, and establishes a real-time communication channel, and the specific method is as follows:

[0032] Firstly, the SSE technology allows the server to actively push data to the client without frequent requests from the client, which is suitable for real-time progress feedback in the diagnosis and treatment process.

[0033] The workflow event sequence follows a standardized format to ensure that the client can clearly track the processing progress.

[0034] ​SSE connection stability evaluation calculation is performed:

[0035]

[0036] Wherein, S is the connection stability index; T up is the connection duration; T down is the connection interruption duration.

[0037] In one of the embodiments, the knowledge graph retrieval and reasoning module calls the integrated medical knowledge graph through a professional tool set, retrieves relevant medical knowledge and performs reasoning, and the specific method is as follows:

[0038] First, the medical knowledge graph is stored in a graph database with entities as nodes and relationships as edges, and a graph neural network is used for knowledge reasoning to mine the potential association between entities;

[0039] In addition, the entity association strength is calculated as follows:

[0040]

[0041] Wherein, R(e1,e2) is the association strength between entities e1 and e2; w direct is the direct relationship weight; P is the set of intermediate entities connecting e1 and e2; w p is the weight of intermediate entity p; R(e1,p) and R(p,e2) are the association strengths between e1 and p, and p and e2, respectively;

[0042] The knowledge retrieval accuracy is evaluated, and the calculation is as follows:

[0043]

[0044] Wherein, A is the retrieval accuracy; N relevant is the number of relevant knowledge in the retrieval result; N retrieved is the total number of retrieval results.

[0045] In one of the embodiments, the diagnosis and treatment logic reasoning and scheme generation module performs diagnosis and treatment logic reasoning and generates preliminary auxiliary diagnosis and treatment suggestions based on the retrieved knowledge and clinical rules, and the specific method is as follows:

[0046] First, a hybrid reasoning mechanism combining rule reasoning and machine learning is used, rule reasoning ensures compliance with clinical guidelines, and machine learning models handle personalized recommendations for complex clinical scenarios.

[0047] Wherein, the diagnosis and treatment scheme confidence is calculated as follows:

[0048] Conf = λ × C rule + (1-λ) × C model ;

[0049] Formula effect: Integrate the results of rule-based reasoning and machine learning models to evaluate the reliability of treatment plans.

[0050] Where Conf is the confidence of the treatment plan; λ is the weight of rule-based reasoning; C rule is the consistency score of rule-based reasoning; C model is the prediction probability of the machine learning model.

[0051] The plan individualization index is calculated as follows:

[0052]

[0053] Where P I is the individualization index (value range 0-1, the closer to 1, the more individualized); n is the number of patient feature dimensions; x i is the actual value of the i-th feature of the patient; is the average value of the i-th feature of the disease population; σ i is the standard deviation of the i-th feature of the disease population.

[0054] In one embodiment, the streaming result output and node state update module pushes the generated treatment results to the user in blocks through the SSE mechanism, and updates the node processing state in real time. The specific method is as follows:

[0055] The treatment results are divided into blocks according to logical units, and the text_chunk event is used for streaming output, while the node_started and node_finished events are used to mark the state of each processing node, improving user experience.

[0056] The streaming output efficiency evaluation calculation method is as follows:

[0057]

[0058] Where E is the output efficiency (characters / second); L total is the total number of characters in the output result; T trans is the total transmission time.

[0059] In one embodiment, the dynamic knowledge update triggering and verification module triggers the dynamic knowledge update mechanism according to new information or external medical progress during the treatment process, and verifies the validity of the updated content. The specific method is as follows:

[0060] First, based on the Graphiti dynamic knowledge graph technology, new knowledge is automatically identified by monitoring clinical data, academic literature and guideline updates, and the knowledge base is updated after verification to ensure the timeliness of the system knowledge;

[0061] The knowledge update necessity assessment calculation is performed:

[0062] U = δ × D time + (1 - δ) × R impact ;

[0063] Wherein, U is the update necessity index (value range 0-1); δ is the time weight (δ ∈ [0.3, 0.5]); D time is the time attenuation coefficient after the last knowledge update (linearly increases with time); R impact is the clinical impact factor of new knowledge (based on reference quantity, evidence level assessment).

[0064] The update knowledge credibility assessment calculation is performed:

[0065] K cred = w1 × Q source + w2 × L evidence + w3 × C consensus ;

[0066] Wherein, K cred is the knowledge credibility (value range 0-1); w1, w2, w3 are weights (w1 + w2 + w3 = 1); Q source is the knowledge source quality (such as top journals, guideline committees); L evidence is the evidence level (such as randomized controlled trial > cohort study); C consensus is the expert consensus degree.

[0067] In one embodiment, the workflow completion and result summary module pushes the workflow completion event after all processing nodes are completed, and the specific method for generating an auxiliary diagnosis report is as follows:

[0068] After all the text blocks are output, the system integrates the contents of each part to generate a structured diagnosis and treatment report (such as SOAP medical record, follow-up plan), and notifies the user through the workflow_finished event to ensure the completeness and availability of the results. - Report integrity assessment

[0069]

[0070] Wherein, I is the integrity index (value range 0-1); N sections is the number of content modules actually contained in the report; N required is the number of necessary modules for this type of report.

[0071] The beneficial effects of the present application are: compared with the prior art, in the present application, at the system integration level, the efficient integration of multi-specialty medical resources is realized based on the role as a service architecture, and through the role service registration effectiveness evaluation mechanism, it is ensured that the multi-specialty knowledge modules such as internal medicine, surgery and oncology are completely covered, at the user interaction level, the high-level natural language processing technology breaks the structured input limit, and doctors can complete information submission through natural language description of the disease, combined with the high accuracy guarantee of intent recognition and entity extraction, especially for emergency department and primary hospital scene, the information collection efficiency is greatly improved, and the diagnosis delay caused by complicated input is avoided. The intelligent routing mechanism realizes the accurate matching of diagnosis and treatment demand and professional tool set, the optimal tool set is automatically selected through multi-dimensional matching degree calculation, the SSE flow response mechanism constructs a real-time dynamic feedback mode, through stable connection and standardized event sequence, the multi-hop retrieval and correlation reasoning ability of the knowledge graph, the potential relationship between medical entities is deeply mined, compared with traditional database query, a complete knowledge network can be constructed, the flow result output decomposes the complex report into logical modules, and the information acquisition efficiency is improved. Let the doctor can preferentially view the key content, reduce the omission caused by information overload, the dynamic knowledge update mechanism realizes real-time synchronization of medical knowledge, shortens the update cycle from months to days, ensures that the system timely incorporates the latest guidelines and research results, the standardization report generated by the work flow aggregation is verified through integrity, and the content such as medical record and follow-up plan is ensured to be complete and complete. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a system framework diagram of the present application; DETAILED DESCRIPTION

[0073] In the description of the present application, it should be explained that the terms "front", "upper", "lower", "left", "right", "vertical", "horizontal" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore cannot be understood as a limitation on the present application.

[0074] In order to solve the problems in the background art, the present application proposes the following technical scheme: an intelligent doctor assistant diagnosis and treatment system, comprising:

[0075] The system initialization and role service registration module (complete basic environment configuration and role service registration when the system starts, provide underlying support for subsequent diagnosis and treatment process), the system initialization and role service registration module based on role-as-a-service (Role-as-a-Service) architecture, different medical specialties (such as internal medicine, surgery, oncology, etc.) are abstracted as independent role services, and dynamic calling is realized through service registration mechanism, each role service corresponds to a set of professional tool set, including the knowledge graph, diagnosis and treatment rules and decision model of the field;

[0076] In the system initialization and role service registration module, the role service registration effectiveness evaluation formula R valid As follows:

[0077]

[0078] The formula is used to evaluate the completeness of the role service registration, and to ensure that the registered role service covers the core knowledge and functional requirements of the target medical specialty.

[0079] Wherein, R valid is the role service registration effectiveness index (the value range is 0-1, the closer to 1 indicates that the registration is more complete); n is the number of core knowledge modules of the specialty; w i is the weight of the i-th knowledge module (according to the clinical importance, ∑w i =1); S ki is the i-th knowledge module content set registered by the system; S ri is the standard content set of the i-th knowledge module of the specialty.

[0080] Implementation process: the system reads the pre-defined medical specialty list, loads the metadata of the corresponding role service one by one (including professional name, core knowledge module, tool set version, etc.), calculates the registration effectiveness index of each role service, and when R valid ≥0.9, it is determined that the registration is successful, otherwise the automatic completion mechanism is triggered to supplement the missing knowledge module from the dynamic knowledge base.

[0081] In the above technical solution, the integration of multi-specialty medical resources is achieved through the role-as-a-service architecture, completely changing the limitations of traditional medical auxiliary systems running in a single field. For example, when a medical institution deploys the system, it does not need to configure independent systems for internal medicine, surgery, and oncology, but through a one-time registration to complete the loading of multi-specialty role services. The effectiveness evaluation mechanism ensures that each role service covers core knowledge modules, such as tumor staging, chemotherapy regimens, and side effect management. The registration effectiveness index ≥ 0.9 standard guarantees service quality. For comprehensive hospitals, this mechanism can reduce system deployment costs and reduce system switching operations when collaborating across departments. Taking comprehensive internal medicine as an example, the system can simultaneously call cardiovascular and endocrine sub-specialty role services. When a doctor faces a patient with "diabetes combined with coronary heart disease," they do not need to switch multiple systems to obtain comprehensive diagnosis and treatment support, significantly improving work efficiency. The automatic completion mechanism avoids the one-sidedness of diagnosis and treatment recommendations caused by missing knowledge modules, ensuring that the system can provide comprehensive professional support after initialization.

[0082] The user input analysis and intent recognition module receives diagnosis and treatment related information submitted by the user in natural language form, analyzes the input content through natural language processing technology, and identifies the user's intent.

[0083] The user input analysis and intent recognition module receives diagnosis and treatment related information submitted by the user in natural language form, analyzes the input content through natural language processing technology, and identifies the user's intent. The specific method is as follows:

[0084] A pre-training language model based on Transformer (such as BERT) is used in combination with domain fine-tuning to analyze medical natural language. Through an intent recognition model, user demand types are distinguished, including diagnosis suggestion generation, medical record writing assistance, and evidence-based medicine queries.

[0085] User intent confidence calculation C j :

[0086]

[0087] Formula function: Calculate the probability of user input belonging to the jth intent type, used to determine the user's core needs.

[0088] Where C j is the confidence of the jth intent; S j is the output score of the jth intent recognition model; and m is the total number of preset intent types.

[0089] Entity extraction accuracy evaluation P:

[0090]

[0091] Formula effect: Evaluate the accuracy of extracting key medical entities (such as symptoms, diseases, examination items) from user input.

[0092] Letter meaning: P is the accuracy of entity extraction; T p is the number of correctly extracted entities; F p is the number of incorrectly extracted entities.

[0093] Implementation process: The user submits a natural language description (such as "patient 55-year-old male, with a history of smoking, cough with chest pain for 1 week") in a single input field, and the system inputs the fine-tuned BERT model after preprocessing (word segmentation, stop word removal), outputs the intent classification result (by calculating the confidence, taking the intent type corresponding to the maximum value) and the entity extraction result (by evaluating the accuracy, when P≥0.85, enter the next step, otherwise prompt the user to supplement information). j

[0094] In the above technical solution, the dependence on structured input of traditional systems is broken through advanced natural language processing technology, greatly reducing the operation threshold of doctors. For example, the doctor inputs "55-year-old male, 20 years of smoking, cough with blood in sputum for nearly 1 week", the system accurately judges the "diagnosis suggestion generation" intent (confidence ≥0.9) through the intent recognition model, and at the same time accurately extracts key entities such as "age 55 years old", "male", "20 years of smoking history", "cough", "blood in sputum" through entity extraction. Compared with the operation mode of traditional systems that need to manually check symptoms and fill in parameters, the input time is shortened. In actual application, facing ambiguous expressions such as "the patient recently feels tired all the time and loses weight quickly", the system can combine the context to supplement the recognition of potential symptom associations, avoiding the omission of key information. For emergency department scenarios, doctors can quickly input the condition description, and the system can instantly analyze the intent and entity, saving valuable time for emergency treatment decisions, reducing diagnosis delays caused by cumbersome information input, especially for elderly patients and patients with unclear expressions, which can make up for the deficiency of medical history collection through semantic understanding.

[0095] The intelligent routing and professional tool set selection module automatically selects the most matched professional tool set for processing according to the user intent and entity information.

[0096] The intelligent routing and professional tool set selection module automatically selects the most matched professional tool set for processing according to the user intent and entity information. The specific method is as follows:

[0097] Based on the entity association analysis and rule reasoning of the knowledge graph, the accurate mapping from user demand to professional tool set is realized, and the intelligent routing mechanism considers the disease type, symptom characteristics, and patient population factors.

[0098] ​Toolset matching degree calculation M t ;

[0099] M t = a x R et + b x R it + g x R dt ;

[0100] Formula function: quantize the matching degree of the tth professional toolset with user demand, used for selecting the optimal toolset.

[0101] Letter meaning: M t is the matching degree of the tth toolset; a, b, g are weight coefficients (a + b + g = 1); R et is the entity matching rate (the overlapping proportion of user input entity and toolset coverage entity); R it is the intent matching rate (the overlapping proportion of user intent and toolset supported function); R dt is the disease field matching rate (the overlapping proportion of user input involving disease and toolset professional field).

[0102] Implementation process: the system extracts entity set E, intent type I and disease field D from the analysis result, traverses all registered professional toolsets, calculates the matching degree M t of each toolset, selects the toolset corresponding to the maximum M t (value, when M t ≥ 0.7, directly select, otherwise start multi-toolset collaboration mechanism), and sends a processing request to the toolset.

[0103] In the above technical solution, the intelligent routing mechanism solves the pain point of "manual selection of professional modules" in traditional systems, realizing accurate matching of diagnosis and treatment demand and professional tools. For example, when the user inputs "gastric cancer postoperative chemotherapy review suggestion", the system calculates the matching degree of each toolset: the entity matching rate (gastric cancer, chemotherapy related entity coverage), the intent matching rate (review suggestion function support), and the disease field matching rate of the oncology toolset are significantly higher than those of other toolsets, and the matching degree is ≥ 0.7, so as to automatically route to the oncology toolset. This process does not require the doctor to manually switch modules. For cross-disciplinary cases, such as "lung cancer bone metastasis patient with hypercalcemia", the system will start multi-toolset collaboration due to insufficient single toolset matching degree, and simultaneously call the oncology and endocrinology toolsets to generate a comprehensive treatment suggestion. In primary hospitals, non-specialist doctors face complex cases, and the system can automatically select the most suitable specialist toolset to provide professional guidance, make up for the short board of insufficient medical resources in primary hospitals, and reduce the risk of misdiagnosis. For example, when a township doctor encounters a "fever with rash of unknown cause" case, the system can route to the infectious disease toolset to assist in determining the diagnosis direction.

[0104] The SSE stream response initialization and workflow start module starts the Server-Sent Events stream response mechanism, sends a workflow start event to the user end, and establishes a real-time communication channel.

[0105] The SSE stream response initialization and workflow start module starts the Server-Sent Events stream response mechanism, sends a workflow start event to the user end, and establishes a real-time communication channel. The specific method is as follows:

[0106] First, the SSE technology allows the server to actively push data to the client without frequent requests from the client, which is suitable for real-time progress feedback in the diagnosis and treatment process.

[0107] The sequence of workflow events follows a standardized format, ensuring that the client can clearly track the processing progress.

[0108] SSE connection stability evaluation calculation:

[0109]

[0110] Formula function: Evaluate the stability of the SSE connection to ensure the continuity of real-time feedback.

[0111] Letter meaning: S is the connection stability index (value range 0-1); T up is the normal connection time; T down is the connection interruption time.

[0112] Implementation process: After selecting the professional tool set, the system initializes the SSE server, generates a unique session ID and establishes a connection with the user end. By monitoring the connection stability in real time (set threshold S≥0.95), when the connection is stable, the workflow_started event (including session ID, tool set name, and expected processing steps) is pushed to the client. After the client receives it, it displays the "diagnosis and treatment processing has started" progress prompt.

[0113] In the above technical solution, the SSE streaming response mechanism completely changes the traditional "request-wait-response" static interaction mode, providing real-time dynamic feedback of the diagnosis and treatment process for doctors. For example, when a doctor submits a "postoperative follow-up plan generation for breast cancer" request, the system immediately establishes a connection (connection stability index ≥ 0.95) through SSE and pushes the workflow_started event to inform that "the oncology follow-up module has been started, and 3 processing nodes are expected." The doctor does not need to wait for the result generation throughout the process, and can perform other work while receiving real-time feedback, greatly improving time utilization. When dealing with complex diagnosis and treatment tasks, such as "multidrug-resistant bacteria infection treatment plan development", the system will push the node status in stages, and the doctor can understand the progress of "drug sensitivity test data retrieval" and "treatment guideline matching" in real time, avoiding anxiety caused by excessive waiting time. For teaching scenarios, students can intuitively learn the formation process of diagnosis and treatment ideas through streaming events, understand the role of each processing node, and improve teaching effectiveness compared to traditional static result output. Connection stability ensures that the teaching process is not interrupted.

[0114] The knowledge graph retrieval and reasoning module retrieves and reasons related medical knowledge by calling the integrated medical knowledge graph through a professional tool set.

[0115] The knowledge graph retrieval and reasoning module retrieves and reasons related medical knowledge by calling the integrated medical knowledge graph through a professional tool set, with the following specific methods:

[0116] First, the medical knowledge graph is stored in a graph database with entities (such as diseases, symptoms, and drugs) as nodes and relationships (such as causal relationships and treatment relationships) as edges, and graph neural networks are used for knowledge reasoning to mine potential associations between entities.

[0117] In addition, the entity association strength is calculated as follows:

[0118]

[0119] Formula function: Calculate the association strength between two medical entities e1 and e2 for knowledge reasoning and diagnosis and treatment suggestion generation.

[0120] Where R(e1, e2) is the association strength between entities e1 and e2; w direct is the direct relationship weight; P is the set of intermediate entities connecting e1 and e2; w p is the weight of intermediate entity p; R(e1, p) and R(p, e2) are the association strengths between e1 and p, and p and e2, respectively.

[0121] The knowledge retrieval accuracy is evaluated as follows:

[0122]

[0123] Formula effect: Evaluate the accuracy of knowledge graph search results to ensure that the acquired knowledge is related to user needs.

[0124] Where A is the search accuracy; N relevant is the number of relevant knowledge in the search results; N retrieved is the total number of search results.

[0125] Implementation process: The tool set performs multi-hop search in the knowledge graph based on the entity input by the user (such as "gastric cancer T3N2M0 stage" "3rd cycle of chemotherapy"), filters out the association relationship of R(e1, e2) ≥ 0.6 by calculating the association strength between entities. At the same time, the search accuracy is evaluated (A ≥ 0.8 is required), and if it does not meet the standard, the search range is expanded or the weight parameter is adjusted until enough relevant knowledge is obtained.

[0126] In the above technical solution, the multi-hop search and association reasoning capability of the knowledge graph enable the system to deeply mine the potential relationship between medical entities, surpassing the simple query function of traditional databases. For example, facing the problem of "gastric cancer T3N2M0 stage chemotherapy regimen selection", the system finds that the association strength of "T3N2M0 stage" with "adjuvant chemotherapy" and "oxaliplatin + capecitabine regimen" is ≥ 0.6, and also retrieves the potential association between "patient age 65 years old" and "drug dose adjustment". Avoid irrelevant information interference. Compared with traditional systems that can only return isolated knowledge points, it can construct a complete knowledge network, such as when analyzing "diabetic patients using statins", the system will also retrieve multi-dimensional associated knowledge such as blood sugar influence and drug interactions. In the diagnosis and treatment of rare diseases, such as "tuberous sclerosis combined with renal angiomyolipoma", the knowledge graph can connect scattered knowledge such as rare disease characteristics and complication management through multi-hop reasoning, providing comprehensive reference for doctors and significantly reducing the risk of missed diagnosis caused by knowledge fragmentation.

[0127] Diagnosis and treatment logic reasoning and scheme generation module, the diagnosis and treatment logic reasoning and scheme generation module is based on the knowledge and clinical rules retrieved, carries out diagnosis and treatment logic reasoning and generates the preliminary auxiliary diagnosis and treatment suggestion.

[0128] The diagnosis and treatment logic reasoning and scheme generation module is based on the knowledge and clinical rules retrieved, carries out diagnosis and treatment logic reasoning and generates the preliminary auxiliary diagnosis and treatment suggestion, and the specific method is as follows:

[0129] Firstly, a hybrid reasoning mechanism combining rule reasoning and machine learning is adopted, rule reasoning ensures compliance with clinical guidelines, and machine learning models (such as decision tree, random forest) handle personalized recommendations for complex clinical scenarios.

[0130] Wherein, the diagnosis and treatment scheme confidence is calculated as follows:

[0131] Conf = λ × C + (1 - λ) × C rule model ;

[0132] Formula function: Integrate the results of rule-based reasoning and machine learning models to evaluate the reliability of the treatment plan.

[0133] where Conf is the confidence of the treatment plan (value range 0-1); λ is the weight of rule-based reasoning (λ ∈ [0.5, 0.8]); C rule is the consistency score of rule-based reasoning (the degree to which the plan meets clinical guidelines); C model is the prediction probability of the machine learning model.

[0135] The plan individualization index is calculated as follows:

[0136]

[0137] Function: Evaluate the individualization degree of the treatment plan to ensure that the plan fits the individual characteristics of the patient.

[0138] where P I is the individualization index (value range 0-1, the closer to 1 indicates more individualization); n is the number of patient feature dimensions; x i is the actual value of the patient's i-th feature; μ is the mean of the i-th feature of the disease population; σ i is the standard deviation of the i-th feature of the disease population.

[0139] Implementation process: The tool set first applies the clinical rule base (such as NCCN guidelines, clinical pathways) for rule-based reasoning to obtain C rule ; at the same time, the patient's characteristics are input into the pre-trained machine learning model to obtain C model ; by calculating the plan confidence (require Conf ≥ 0.75), and by evaluating the individualization index (require P I ≥ 0.6). If the conditions are not met, adjust the reasoning parameters or supplement the patient's characteristics, and regenerate the plan.

[0140] ​In the above technical solution, the mixed inference mechanism combines the advantages of rule-based reasoning and machine learning, ensuring the standardization of diagnosis and treatment plans and achieving individualized adaptation. For example, when developing a targeted treatment plan for "HER2-positive breast cancer patients", it is ensured that the treatment standard of trastuzumab recommended by the NCCN guideline is met, and the machine learning model combines patient age, liver and kidney function, etc. to give dose adjustment suggestions, and the final plan confidence is ≥0.75. The individualization index ≥0.6 ensures that the plan adapts to individual differences, such as automatically reducing the starting dose of drugs for elderly patients. Compared with the "one-size-fits-all" plan recommendation of traditional systems, the generated treatment plan is more in line with the actual situation of patients. In chronic disease management, such as "medication plan for type 2 diabetes patients with kidney disease", the system can balance blood glucose control and kidney function protection, avoiding the risk of worsening kidney function that may be caused by traditional plans. For complex cases, such as "multiple myeloma relapse patients", mixed reasoning can integrate the latest clinical research and the patient's past treatment response to generate an optimized salvage treatment plan, improving treatment effectiveness.

[0141] The streaming result output and node state updating module pushes the generated diagnosis and treatment results to the user in blocks through the SSE mechanism and updates the node processing state in real time.

[0142] The streaming result output and node state updating module pushes the generated diagnosis and treatment results to the user in blocks through the SSE mechanism and updates the node processing state in real time. The specific method is as follows:

[0143] The diagnosis and treatment results are divided into blocks according to logical units (such as symptom analysis, examination suggestions, treatment plans), and output in text_chunk event stream, while pushing node_started and node_finished events to mark the status of each processing node, improving user experience.

[0144] The efficiency evaluation calculation method of streaming output is as follows:

[0145]

[0146] Formula function: Evaluate the efficiency of streaming output to ensure the timeliness of result pushing.

[0147] Letter meaning: E is the output efficiency (characters / second); L total is the total number of characters of the output result; T trans is the total transmission time.

[0148] Implementation process: The system splits the generated diagnosis and treatment plan into multiple text blocks (such as "symptom analysis", "examination suggestion", "treatment plan"), each text block corresponds to a processing node. First push the node_started event (mark the current processing node), then push the text_chunk event (contains the text block content), and then push the node_finished event. By monitoring the output efficiency in real time (require E≥500 characters / second), ensure that users can quickly obtain the step-by-step generated results.

[0149] In the above technical solution, the streaming result output mechanism decomposes complex diagnosis and treatment results into logically clear modules, and updates the node state to allow doctors to gradually understand the diagnosis and treatment ideas rather than passively receiving the final results. For example, when generating a "PCI postoperative management plan for patients with coronary heart disease", the system first pushes the "node_started: risk stratification assessment" event, then outputs the risk stratification analysis text block, and then pushes the "node_finished" event, and then sequentially outputs the medication suggestion, follow-up arrangement and other modules. Ensure that the output efficiency is ≥500 characters / second to ensure that doctors do not need to wait for a long time. In teaching ward rounds, teachers can explain to students the clinical logic of "why risk stratification is performed first before developing a medication plan" based on the node sequence of the streaming output, significantly enhancing the interactivity of teaching. For lengthy diagnosis and treatment reports such as "comprehensive treatment plan for advanced lung cancer", block output allows doctors to prioritize viewing key modules (such as treatment goals and first-line solutions), and if necessary, delve into details, improving information acquisition efficiency and reducing the omission of important information due to information overload.

[0150] The dynamic knowledge update trigger and verification module triggers the dynamic knowledge update mechanism according to new information or external medical progress during the diagnosis and treatment process, and verifies the effectiveness of the updated content.

[0151] The dynamic knowledge update trigger and verification module triggers the dynamic knowledge update mechanism according to new information or external medical progress during the diagnosis and treatment process, and verifies the effectiveness of the updated content, the specific method is as follows:

[0152] First, based on the Graphiti dynamic knowledge graph technology, new knowledge is automatically identified by monitoring clinical data, academic literature and guideline updates, and the knowledge base is updated after verification to ensure the timeliness of the system knowledge;

[0153] Knowledge update necessity assessment calculation:

[0154] U = δ × D time + (1-δ) × R impact ;

[0155] Formula function: assess the necessity of knowledge update and decide whether to start the update process.

[0156] Wherein, U is the update necessity index (value range 0-1); δ is the time weight (δ∈[0.3, 0.5]); D time is the time attenuation coefficient of the last knowledge update (linearly increasing with time); R impact is the clinical impact factor of new knowledge (based on the amount of reference, evidence level assessment).

[0157] Update knowledge credibility assessment calculation:

[0158] K cred = w1×Q source + w2×L evidence + w3×C consensus ;

[0159] Formula function: assess the credibility of the knowledge to be updated, and ensure that the knowledge included in the knowledge base is reliable.

[0160] Wherein, K cred is the knowledge credibility (value range 0-1); w1, w2, w3 are weights (w1 + w2 + w3 = 1); Q source is the quality of knowledge sources (such as top journals, guideline committees); L evidence is the evidence level (such as randomized controlled trial > cohort study); C consensus is the expert consensus degree.

[0161] Implementation process: the system regularly monitors academic databases (such as PubMed) and guideline platforms, when new knowledge is detected, through the calculation of U (when U≥0.6 trigger update). Through the assessment of the credibility of new knowledge (require K cred ≥0.7), through expert review, the entity and relationship of knowledge graph are updated by using Graphiti technology, and the update log is recorded.

[0162] In the above technical solution, the dynamic knowledge updating mechanism solves the problem of knowledge lag in traditional medical systems, ensuring that the system is always synchronized with the latest medical progress. For example, when international guidelines update the indications for targeted therapy for a certain type of lung cancer, the system triggers an update assessment through the monitoring mechanism, calculates the time decay coefficient and clinical impact factor, and obtains an update necessity index ≥ 0.6, starting the update process. The guideline content is evaluated for credibility (source quality, evidence level, and expert consensus level meet the standards), and after review, the knowledge graph is updated through the Graphiti technology. This process shortens the knowledge update cycle from months to days in the traditional sense, and doctors can immediately receive the latest medication guidance when faced with "newly approved targeted drugs." In the event of a public health emergency such as an epidemic, the system can quickly incorporate newly discovered clinical features and treatment plans to provide timely support to frontline doctors. For primary care institutions, this mechanism ensures that they can synchronize knowledge updates from top medical centers, narrowing the gap in diagnosis and treatment levels between different levels of hospitals.

[0163] The workflow completion and result summary module pushes a workflow completion event and summarizes an auxiliary diagnosis report after all processing nodes are completed.

[0164] The specific method of the workflow completion and result summary module pushing a workflow completion event and summarizing an auxiliary diagnosis report after all processing nodes are completed is as follows:

[0165] After all text block outputs are completed, the system integrates the contents of each part to generate a structured diagnosis and treatment report (such as a SOAP medical record or a follow-up plan), and notifies the user through a workflow_finished event to ensure the completeness and usability of the results.

[0166] The report integrity assessment is calculated as follows:

[0167]

[0168] Formula function: Evaluate the completeness of the diagnosis and treatment report to ensure that all necessary content modules are included.

[0169] Letter meaning: I is the integrity index (value range 0-1); N sections is the actual number of content modules included in the report; N required is the number of necessary modules for this type of report.

[0170] Implementation process: After the system detects that all processing nodes have pushed the node_finished event, it integrates all text_chunk content to generate a standardized report (such as a SOAP format medical record containing subjective, objective, assessment, and plan sections). By checking the integrity of the report (I = 1), it ensures that there are no missing necessary module blocks, and then pushes the workflow_finished event (containing the complete report link), allowing users to view or download the complete results.

[0171] In the above technical solution, workflow completion and result aggregation are achieved through standardized report generation and integrity checking, ensuring the standardization and usability of diagnosis and treatment results. For example, when generating a SOAP case, the system integrates all text blocks to automatically generate a complete report containing subjective (S), objective (O), assessment (A), and plan (P) sections, and the integrity index = 1 ensures that there are no missing modules. Compared to the tedious process of repeatedly checking the format when doctors manually write medical records, the completion time of medical records is reduced by more than 40%, and the format is uniform, making it easy to review and analyze data in the future. In multi-center clinical research, standardized reports can reduce data entry errors and improve research data quality. For follow-up management, the summarized follow-up plan clearly lists examination items, time nodes, and precautions, allowing doctors to directly reference and execute, avoiding follow-up omissions due to memory bias. When patients are transferred, a complete and standardized diagnosis and treatment report can help receiving doctors quickly grasp the patient's condition, reduce repeated examinations, and improve medical continuity, especially for complex cases across hospitals, the effect is more significant.

[0172] To sum up, in the present application, at the system integration level, the efficient integration of multi-specialty medical resources is realized based on the role as a service architecture, and through the role service registration effectiveness evaluation mechanism, it is ensured that the multi-specialty knowledge modules such as internal medicine, surgery and oncology are completely covered, and the disadvantages of traditional systems needing to deploy multiple independent modules are completely changed, so that when the doctors of the comprehensive department face cross-disciplinary cases such as "diabetes combined with coronary heart disease", they can obtain comprehensive support without switching systems. At the user interaction level, the high-level natural language processing technology breaks through the limitation of structured input, and the doctor can complete information submission through natural language description of the disease, combined with high-accuracy guarantee of intent recognition and entity extraction, especially for emergency department and primary hospital scenes, the information collection efficiency is greatly improved, and the diagnosis delay caused by complicated input is avoided. The intelligent routing mechanism realizes the accurate matching of diagnosis and treatment demand and professional tool set, automatically selects the optimal tool set through multi-dimensional matching degree calculation, the SSE streaming response mechanism constructs a real-time dynamic feedback mode, through stable connection and standardized event sequence, the multi-hop retrieval and association reasoning ability of the knowledge graph, the potential relationship between medical entities is deeply mined, compared with traditional database query, a complete knowledge network can be constructed, scattered knowledge points can be connected in rare disease diagnosis and treatment, the mixed reasoning mechanism combines the advantages of rule reasoning and machine learning, which not only ensures that the scheme conforms to the normativity of clinical guidelines, but also guarantees the adaptation to the individual characteristics of patients through personalized indexes, the streaming result output decomposes the complex report into logical modules, and cooperates with node state update to improve information acquisition efficiency, so that the doctor can preferentially view the key content, reduce the omission caused by information overload, and the dynamic knowledge update mechanism realizes medical knowledge real-time synchronization through the Graphiti technology, shortens the update cycle from months to days, ensures that the system timely incorporates the latest guidelines and research results, and the standardized report generated by the workflow summary is verified for integrity, ensuring that the medical records, follow-up plans and other contents are complete and complete.

[0173] Although embodiments of the present application have been shown and described, the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent doctor-assisted diagnosis and treatment system, characterized in that, include: The system initialization and role service registration module is based on the role-as-a-service architecture. It abstracts different medical specialties into independent role services and realizes dynamic invocation through the service registration mechanism. Each role service corresponds to a set of professional tools, which includes the knowledge graph, diagnosis and treatment rules and decision models of the field. The user input parsing and intent recognition module receives diagnosis and treatment-related information submitted by the user in natural language form, and parses the input content and recognizes the user's intent through natural language processing technology. The intelligent routing and professional toolset selection module automatically selects the most suitable professional toolset for processing based on user intent and entity information. The SSE streaming response initialization and workflow startup module starts the Server-SentEvents streaming response mechanism, sends a workflow start event to the user terminal, and establishes a real-time communication channel. The knowledge graph retrieval and reasoning module uses a professional toolset to call an integrated medical knowledge graph to retrieve relevant medical knowledge and perform reasoning. The diagnostic logic reasoning and plan generation module performs diagnostic logic reasoning and generates preliminary auxiliary diagnostic and treatment suggestions based on retrieved knowledge and clinical rules. The streaming results output and node status update module pushes the generated diagnosis and treatment results to the user in blocks through the SSE mechanism and updates the node processing status in real time. The dynamic knowledge update triggering and verification module triggers the dynamic knowledge update mechanism based on new information during the diagnosis and treatment process or external medical advancements, and verifies the validity of the updated content. The workflow completion and result summary module pushes a workflow completion event after all processing nodes are completed, and summarizes and generates an auxiliary diagnostic report.

2. The intelligent doctor-assisted diagnosis and treatment system according to claim 1, characterized in that, In the system initialization and role service registration module, the formula for evaluating the validity of role service registration is R. valid as follows: Among them, R valid The registration validity index for role services; n is the number of core knowledge modules for this major; w i S represents the weight of the i-th knowledge module; ki S is the set of content for the i-th knowledge module registered in the system. ri This is the set of standard content for the i-th knowledge module of this major.

3. The intelligent doctor-assisted diagnosis and treatment system according to claim 2, characterized in that, The user input parsing and intent recognition module receives diagnosis-related information submitted by the user in natural language form, and parses the input content and recognizes the user's intent using natural language processing technology. The specific method is as follows: We employ a Transformer-based pre-trained language model combined with domain fine-tuning to parse medical natural language. We use an intent recognition model to distinguish user need types, including diagnosis suggestion generation, medical record writing assistance, and evidence-based medicine query. Perform user intent confidence calculation C j : Among them, C j S represents the confidence level of the j-th intention; j is the output score of the intent recognition model for the j-th intent; m is the preset total number of intent types; Entity extraction accuracy assessment P: Where P is the entity extraction accuracy; T p F represents the number of entities correctly extracted. p This represents the number of entities that were incorrectly extracted.

4. The intelligent doctor-assisted diagnosis and treatment system according to claim 3, characterized in that, The intelligent routing and professional toolset selection module automatically selects the most suitable professional toolset for processing based on user intent and entity information. The specific method is as follows: Based on knowledge graph entity association analysis and rule reasoning, it achieves accurate mapping of user needs to professional toolsets, and the intelligent routing mechanism takes into account disease type, symptom characteristics and patient population factors. Calculate the matching degree of toolsets M t ; M t =α×R et +β×R it +γ×R dt ; Among them, M t R represents the matching degree of the t-th toolset; α, β, γ are weight coefficients; et R represents the entity matching rate. it R represents the intent matching rate. dt Matching rate for disease domains.

5. The intelligent doctor-assisted diagnosis and treatment system according to claim 4, characterized in that, The SSE streaming response initialization and workflow startup module initiates the Server-SentEvents streaming response mechanism, sends a workflow start event to the user terminal, and establishes a real-time communication channel. The specific method is as follows: First, SSE technology allows the server to proactively push data to the client without requiring frequent client requests, making it suitable for real-time progress feedback during the diagnosis and treatment process; Workflow event sequences follow a standardized format to ensure that clients can clearly track processing progress; Perform SSE connection stability assessment calculations: Where S is the connection stability index; T up For normal connection duration; T down This represents the duration of the connection interruption.

6. The intelligent doctor-assisted diagnosis and treatment system according to claim 5, characterized in that, The knowledge graph retrieval and reasoning module uses a professional toolset to call upon an integrated medical knowledge graph to retrieve relevant medical knowledge and perform reasoning. The specific method is as follows: First, the medical knowledge graph is stored in a graph database with entities as nodes and relationships as edges. Then, a graph neural network is used for knowledge reasoning to uncover potential relationships between entities. In addition, the entity association strength is calculated as follows: Where R(e1,e2) represents the association strength between entities e1 and e2; w direct The weight represents the direct relationship; P is the set of intermediate entities connecting e1 and e2; w p Let R(e1,p) be the weight of the intermediate entity p; R(e1,p) and R(p,e2) are the association strengths between e1 and p, and between p and e2, respectively. The accuracy of knowledge retrieval was evaluated and calculated as follows: Where A is the retrieval accuracy; N relevant N represents the number of relevant pieces of information in the search results. retrieved This represents the total number of search results.

7. The intelligent doctor-assisted diagnosis and treatment system according to claim 6, characterized in that, The diagnostic logic reasoning and treatment plan generation module performs diagnostic logic reasoning based on retrieved knowledge and clinical rules, and generates preliminary auxiliary diagnostic and treatment suggestions. The specific method is as follows: First, a hybrid reasoning mechanism combining rule-based reasoning and machine learning is adopted. Rule-based reasoning ensures compliance with clinical guidelines, while machine learning models handle personalized recommendations for complex clinical scenarios. The confidence level of the treatment plan is calculated as follows: Conf=λ×C rule +(1-λ)×C model ; Formula purpose: To evaluate the reliability of treatment plans by combining the results of rule-based reasoning and machine learning models; Where Conf represents the confidence level of the treatment plan; λ represents the rule inference weight; C rule For the consistency score of rule-based reasoning; C model The predicted probability of the machine learning model; The personalization index of the plan is calculated as follows: Among them, P I The personalization index (ranging from 0 to 1, with values ​​closer to 1 indicating greater personalization); n represents the number of patient characteristic dimensions; x i This represents the actual value of the i-th feature of the patient. σ is the average value of the i-th characteristic in this disease population. i Let be the standard deviation of the i-th characteristic of the disease population.

8. The intelligent doctor-assisted diagnosis and treatment system according to claim 7, characterized in that, The streaming results output and node status update module pushes the generated diagnosis and treatment results to the user in blocks through the SSE mechanism, and updates the node processing status in real time. The specific method is as follows: The diagnosis results are divided into logical blocks and output in a streaming manner through the text_chunk event, while the node_started and node_finished events are pushed to mark the status of each processing node; The calculation method for evaluating streaming output efficiency is as follows: Where E is the output efficiency; L total T represents the total number of characters in the output; trans This represents the total transmission time.

9. The intelligent doctor-assisted diagnosis and treatment system according to claim 8, characterized in that, The dynamic knowledge update triggering and verification module triggers the dynamic knowledge update mechanism based on new information during the diagnosis and treatment process or external medical advancements, and verifies the validity of the updated content. The specific method is as follows: First, based on Graphiti dynamic knowledge graph technology, new knowledge is automatically identified by monitoring clinical data, academic literature and guideline updates, and the knowledge base is updated after verification to ensure the timeliness of the system's knowledge; Perform a knowledge update necessity assessment calculation: U=δ×D time +(1-δ)×R impact ; Where U is the update necessity index; δ is the time weight; D time R is the time decay coefficient of knowledge since the last update; impact Clinical impact factors for new knowledge; Perform updated knowledge credibility assessment calculations: K cred =w1×Q source +w2×L evidence +w3×C consensus ; Among them, K cred Knowledge credibility; w1, w2, w3 are weights; Q source For the quality of knowledge sources (such as top journals, guide committees); L evidence Level of evidence: C consensus This represents the consensus among experts.

10. The intelligent doctor-assisted diagnosis and treatment system according to claim 9, characterized in that, The workflow completion and result summary module pushes a workflow completion event and summarizes and generates an auxiliary diagnostic report after all processing nodes have completed. The specific method is as follows: After all text blocks are output, the system integrates the content of each part, generates a structured diagnosis and treatment report (such as SOAP medical records and follow-up plans), and notifies the user through the workflow_finished event to ensure the integrity and availability of the results; The report completeness assessment is calculated as follows: Where I is the integrity index; N sections The report contains the actual number of content modules; N required The number of modules required for this type of report.

Images