A multi-disciplinary joint diagnosis and treatment method and system

Abstract

The application discloses a kind of multidisciplinary joint diagnosis and treatment method and system, it is related to the technical field of intelligent medical system, it solves the problems that traditional artificial consultation efficiency is low, is greatly interfered by human factor, simple system recommendation is not accurate, it is difficult to meet the needs of patient diagnosis and treatment, its method includes: generating consultation scheme based on the matching degree standard of pre-set, illness condition and diagnosis and treatment standard, trigger payment process after system compliance verification, lock consultation resource after payment success;According to the emergency degree of patient illness, expert scheduling information and hospital resource occupation condition, consultation time arrangement is generated by intelligent dispatching, and consultation invitation and detailed information are sent to relevant experts;After consultation, the system sorts out the written record, fuses expert opinion, generates structured consultation report and pushes to patient end.The application has the following effects: realize multidisciplinary joint diagnosis and treatment intelligent management, optimize medical resource allocation, improve diagnosis and treatment efficiency and accuracy, improve patient medical experience.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent medical systems, and in particular to a multidisciplinary collaborative diagnosis and treatment method and system. Background Technology

[0002] With the continuous advancement of medical technology and the increasing demands of patients for higher quality medical services, the importance of multidisciplinary collaborative diagnosis and treatment methods in the modern medical system is becoming increasingly prominent. By integrating the resources and expertise of specialists from different disciplines, it provides patients with comprehensive and precise treatment plans, demonstrating unique value in addressing complex diseases and improving medical outcomes, thus contributing to enhancing the hospital's medical service level and overall competitiveness.

[0003] Currently, multidisciplinary collaborative diagnosis and treatment methods in practice mainly rely on manual analysis of patient medical data and the organization of consultations according to fixed procedures. Manual analysis is easily influenced by subjective factors, resulting in low efficiency and insufficient accuracy. While some systems can process data in a simple manner, they can only perform basic analysis and recommendations, lacking in-depth intelligent decision-making. For example, some hospitals have attempted to match consultation departments based on medical record data, but the accuracy and personalization of recommendations are low.

[0004] Traditional manual consultation organization relies on medical staff coordinating expert time and resources, which is easily affected by factors such as staff experience and workload, leading to delays in consultation arrangements and unreasonable resource allocation. Furthermore, the manual consultation process is cumbersome, time-consuming, and inefficient; it is difficult to standardize the consultation process, potentially resulting in insufficient depth of discussion or uneven participation from different disciplines, leading to inconsistent consultation quality. Simple systems, when recommending consultation types and assembling expert teams, do not fully consider key factors such as the complexity of the patient's condition and the experts' collaborative history, resulting in recommendations that are not highly targeted and fail to meet the individualized treatment needs of patients, thus affecting the effectiveness and advantages of multidisciplinary collaborative treatment. Summary of the Invention

[0005] In order to automate multidisciplinary joint diagnosis and treatment and improve the accuracy, efficiency and personalization of diagnosis and treatment, this application provides a multidisciplinary joint diagnosis and treatment method and system.

[0006] Firstly, this application provides a multidisciplinary collaborative diagnosis and treatment method, employing the following technical solution:

[0007] A multidisciplinary collaborative diagnosis and treatment method includes:

[0008] The system receives patient medical records submitted through the joint outpatient portal, performs preliminary analysis using natural language processing algorithms, extracts key information, and automatically recommends joint outpatient types by matching the hospital's disease knowledge base with knowledge graph technology.

[0009] Based on the recommended joint outpatient clinic type, and according to the complexity of the patient's condition, historical consultation data, and the area of ​​expertise of the experts, the system intelligently recommends the range of consultation fees and expert team combinations and pushes them to the patient's end.

[0010] Based on preset matching standards, patient conditions, and treatment guidelines, a consultation plan is generated. After the system verifies compliance, the payment process is triggered, and the consultation resources are locked once the payment is successful.

[0011] Based on the urgency of the patient's condition, the expert's scheduling information, and the hospital's resource utilization, the system intelligently generates a consultation schedule and sends consultation invitations and detailed information to relevant experts.

[0012] At the start of the consultation, speech recognition algorithms are used to assist experts in real-time annotation and analysis. At the same time, the system automatically records the speech discussions during the consultation and converts them into text records.

[0013] After the consultation, the system organizes the written records, integrates expert opinions, generates a structured consultation report, and pushes it to the patient.

[0014] By adopting the above-mentioned technical solutions, this multidisciplinary collaborative diagnosis and treatment method, with the help of intelligent technology, can accurately recommend outpatient types and expert teams, intelligently schedule consultations, and automatically generate reports to push to patients, thereby improving the efficiency of diagnosis and treatment, optimizing the allocation of medical resources, enhancing the patient's medical experience and the accuracy of diagnosis and treatment, and ensuring timely handling of emergency conditions.

[0015] Optionally, this also includes steps prior to receiving patient medical records submitted through the joint outpatient portal, as follows:

[0016] The system collects multimodal data from patients and uses a multimodal fusion algorithm to extract and align features from the collected data, generating a disease feature vector with a unified dimension. The multimodal data from patients includes language, images, and text.

[0017] Based on the constructed medical knowledge graph, the graph query algorithm is used to retrieve nodes related to patient input and generate guidance questions. A reinforcement learning-knowledge graph joint model is adopted to dynamically adjust the priority of guidance questions based on historical guidance effect data.

[0018] The system utilizes a Conditional Random Field (CRF) model to perform integrity checks on the input text, automatically annotating missing entities and providing examples. A bidirectional LSTM classifier is used for consistency checks, detecting and probing contradictory information. For ambiguous descriptions, an entity linking algorithm is applied to match standard terms, providing options for patients to choose from.

[0019] Background feature data is extracted from patient registration information to construct patient profile vectors. Input examples of similar cases are retrieved based on collaborative filtering algorithm to guide patients to supplement information.

[0020] When the patient's input information passes the integrity and consistency checks, and standard terms are selected for fuzzy descriptions, the input is considered complete.

[0021] By adopting the above technical solutions, this multidisciplinary joint diagnosis and treatment method first collects multimodal data to generate disease feature vectors, and then, guided by knowledge graphs, validates multiple models and constructs patient profiles, it ensures that the patient's input information is complete and accurate, unifies standards, improves the quality and efficiency of diagnosis and treatment information, lays a solid foundation for subsequent precision diagnosis and treatment, and promotes the intelligent and precise development of medical services.

[0022] Optionally, based on the constructed medical knowledge graph, a graph query algorithm is used to retrieve nodes related to patient input, generate guidance questions, and a reinforcement learning-knowledge graph joint model is adopted to dynamically adjust the priority of guidance questions based on historical guidance effect data, including:

[0023] We use a Transformer-based model to encode the nodes and edges of a medical knowledge graph to capture deep semantic information.

[0024] Model the context of patient input, and combine LSTM or Transformer models to understand its meaning in different contexts and accurately match relevant knowledge points;

[0025] Calculate the semantic similarity between patient input and knowledge graph nodes to determine the most relevant knowledge points;

[0026] By combining semantic similarity and contextual analysis results, a suitable guidance question template is matched from a predefined guidance question template library;

[0027] Based on the patient's profile information, the matched guidance question templates are personalized.

[0028] By using a reinforcement learning-knowledge graph joint model, the generated guidance questions are prioritized based on historical guidance effect data.

[0029] By adopting the above technical solutions, this method comprehensively utilizes a variety of cutting-edge technologies to deeply mine and accurately match medical knowledge graphs. It can not only generate appropriate guiding questions based on patient input, but also dynamically adjust priorities based on historical data through reinforcement learning, thereby comprehensively improving the accuracy and personalization of guiding questions, making the patient input information more complete and accurate, and fully serving subsequent diagnosis and treatment.

[0030] Optional, automatically recommended joint outpatient clinic types include:

[0031] Receive multimodal medical records, including voice, text, and images, submitted by patients through the joint outpatient portal;

[0032] Using a multimodal fusion algorithm based on the Transformer architecture, feature extraction and cross-modal alignment are performed on data from different modalities to generate disease feature vectors of a unified dimension;

[0033] By inputting the disease feature vector into the medical knowledge graph, the disease knowledge base is retrieved through the graph query algorithm, and the disease nodes and related departments most relevant to the patient's condition are matched to automatically recommend joint outpatient clinic types.

[0034] By adopting the above technical solution, this method receives multimodal disease data, generates a unified disease feature vector through a Transformer architecture multimodal fusion algorithm, and then uses a knowledge graph to match disease nodes and related departments, automatically recommends joint outpatient clinic types, efficiently integrates multi-source information, improves the efficiency and accuracy of diagnosis and treatment information integration, assists in precision medicine, provides patients with personalized diagnosis and treatment path suggestions, and improves the quality and efficiency of diagnosis and treatment.

[0035] Optionally, after inputting the disease feature vector into the medical knowledge graph, searching the disease knowledge base using a graph query algorithm, matching the disease nodes and related departments most relevant to the patient's condition, and automatically recommending joint outpatient clinic types, the following steps are also performed:

[0036] The recommendation information is visualized using natural language generation technology and pushed to the patient in the form of interactive cards, with the basis for the recommendation annotated simultaneously, allowing the patient to confirm with one click or initiate a type change request.

[0037] By adopting the above technical solutions, natural language generation technology is used to visualize recommendation information, which is pushed out in interactive cards with annotations, allowing patients to confirm or change their information with one click. This improves information readability and patient participation, optimizes the human-computer interaction experience, and ensures the flexibility and accuracy of the recommendation process.

[0038] Optionally, based on the recommended joint outpatient clinic type, and considering the complexity of the patient's condition, historical consultation data, and the areas of expertise of the specialists, the system intelligently recommends consultation fee ranges and expert team combinations and pushes them to the patient's end, including:

[0039] Construct a medical knowledge graph that includes expert expertise, collaboration history, and successful cases;

[0040] Using a graph attention network algorithm, the matching degree between experts and patients' conditions is calculated based on the complexity and type of the patient's condition. Experts with high matching degree are selected, and the collaborative relationship between experts is analyzed to form a complementary collaborative team.

[0041] By combining historical consultation data, a cost prediction model was trained using a multi-task learning framework.

[0042] Using a trained cost prediction model, the basic cost range for a consultation is predicted based on the patient's condition and the selected expert team combination.

[0043] The recommended expert team combinations and consultation fee ranges are integrated into structured information and pushed to the patient's end, allowing the patient to view the basis for the recommendations and the details of the fees.

[0044] By adopting the above technical solutions, and by constructing knowledge graphs and graph attention algorithms, we can achieve accurate matching of experts and the formation of complementary teams; by combining multi-task learning to predict cost ranges, we can push structured information containing recommendation criteria, thereby improving the accuracy of expert matching, the accuracy of cost prediction, and the transparency of patient information.

[0045] Optionally, the intelligent scheduling system can generate consultation time arrangements including:

[0046] Collect and integrate patient pre-assessment levels, expert schedules, hospital resource status, and medical standard constraints;

[0047] A fuzzy logic algorithm is introduced to convert the urgency of the illness into an urgency index of 0-1;

[0048] By combining backtracking search algorithm with heuristic function optimization, the heuristic function is based on the Pareto optimality principle and simultaneously optimizes conflict resolution efficiency, priority of emergency patients, and expert workload balance.

[0049] Based on the optimized scheduling results, a consultation schedule is directly generated.

[0050] By adopting the above technical solution, this method collects information from multiple sources, introduces fuzzy logic algorithms to quantify the urgency of the illness, and combines backtracking search algorithms and heuristic functions to optimize scheduling results. Based on the Pareto optimality principle, it balances multiple factors and ultimately directly generates a consultation schedule. This significantly improves the efficiency and accuracy of medical resource scheduling, making consultation arrangements more scientific, reasonable, and efficient. It effectively improves the timeliness of treatment for emergency patients, while ensuring the rational use of expert resources and optimizing the overall operational efficiency of the hospital.

[0051] Optionally, this also includes a step following the direct generation of the consultation schedule based on the optimized scheduling results, as follows:

[0052] Establish a real-time monitoring mechanism to continuously monitor the sign-in status of experts and patients before the consultation begins;

[0053] If it is detected that an expert has temporarily changed the consultation status, a dynamic rearrangement mechanism based on a genetic algorithm will be immediately triggered to quickly generate alternative consultation plans and push them to the patient.

[0054] By adopting the above technical solutions, real-time monitoring of check-in status and dynamic rearrangement using genetic algorithms can be achieved, enabling real-time monitoring and rapid reallocation of consultation resources, improving the anti-interference capability of consultation arrangements, and ensuring the continuity of the treatment process and the efficiency of resource utilization.

[0055] Secondly, this application provides a multidisciplinary joint diagnosis and treatment system, which adopts the following technical solution:

[0056] A multidisciplinary collaborative diagnosis and treatment system includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements the multidisciplinary collaborative diagnosis and treatment method as described in the first aspect. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall process of a multidisciplinary joint diagnosis and treatment method according to an embodiment of this application.

[0058] Figure 2 This is a flowchart illustrating another embodiment of the present application, representing the steps preceding the receipt of patient medical information submitted through the joint outpatient portal. Detailed Implementation

[0059] The present application will be further described in detail below with reference to the accompanying drawings.

[0060] Reference Figure 1 This application discloses a multidisciplinary collaborative diagnosis and treatment method, comprising:

[0061] Step S100: Receive the patient's medical information submitted through the joint outpatient portal, perform preliminary analysis using natural language processing algorithms, extract key information, and combine knowledge graph technology with the hospital's disease knowledge base to automatically recommend the type of joint outpatient clinic.

[0062] Natural Language Processing (NLP) algorithms are used by computers to understand and process human language, transforming unstructured text into structured data to extract information (such as identifying key information like symptoms and time in a patient's description). Knowledge Graph technology stores entities and their relationships in a structured semantic network format, facilitating rapid retrieval and matching of knowledge (such as associating diseases with corresponding symptoms and departments).

[0063] Hospital Disease Knowledge Base: A database that centrally stores medical knowledge such as disease symptoms, diagnostic criteria, and treatment plans.

[0064] The following methods can be used to obtain medical information: Patients can submit their medical information through the hospital's APP, mini-program and other joint outpatient portals in the form of text (such as "recurrent syncope for more than 3 years, 1-2 times a year, all of which are manifested as dizziness and headache induced by exertion, followed by loss of consciousness"), voice, and imaging examination reports (this is an example to illustrate the submission method).

[0065] The specific process is as follows: The system uses the BERT named entity recognition algorithm to analyze the medical condition text. For example, "fainting, dizziness, headache, loss of consciousness" are considered as symptoms, "more than three years" as the time frame, and "1-2 times per year" as the frequency. Through rule-based knowledge graph matching, a recommendation is triggered when the extracted symptom entities match the symptom attributes of disease nodes in the hospital's disease knowledge base with a match rate ≥80%. If the above symptoms match the dizziness knowledge node with a match rate of 86%, a multidisciplinary joint clinic for dizziness is recommended.

[0066] Step S200: Based on the recommended joint outpatient clinic type, and according to the complexity of the patient's condition, historical consultation data, and the areas of expertise of the experts, intelligently recommend the consultation fee range and expert team combination and push it to the patient's end.

[0067] The complexity of the condition is quantified using a weighted scoring method, with 1 point for each symptom, 3 points for each disease, and 5 points for each complication. The higher the total score, the more complex the condition. For example, with 3 symptoms, 1 disease, and 1 complication, the complexity of the condition is 1×3+3×1+5×1=11 points.

[0068] Expert Specialization Areas: The expert database details the diseases each expert specializes in. For example, Neurology Expert A is noted to specialize in migraines and epilepsy, and has treated over 500 migraine cases with a cure rate of 82%. Intelligent Recommendation: Utilizing algorithms and models, predictions are made based on historical data. By analyzing the correlation between various factors and outcomes in historical data, recommended solutions are generated.

[0069] The methods for obtaining the above content are as follows:

[0070] Disease complexity: The total score is automatically calculated based on the information extracted from step S100, including symptoms (1 point per item), diseases (3 points per type), and complications (5 points per type). For example, if a patient has two symptoms (cough and fever), one disease (pneumonia), and no complications, the complexity score is 1×2+3×1=5 points.

[0071] Expertise area information: retrieved from the hospital's expert information database, including data such as the types of diseases the expert is good at, the number of cases treated, the cure rate, and the available time slots in the current week's schedule.

[0072] For details of the process, please refer to steps S210 to S250, which will not be elaborated here.

[0073] In step S300, a consultation plan is generated based on preset matching standards, patient condition, and treatment guidelines. After the system's compliance verification, the payment process is triggered, and the consultation resources are locked once the payment is successful.

[0074] Among them, the preset matching degree standard is an indicator that measures the degree of fit between the recommended consultation plan and the patient's condition and hospital standards. For example, it requires that the expert's expertise match ≥90% and that the consultation process comply with the treatment guidelines (such as the nasopharyngeal carcinoma MDT need to include experts from the radiotherapy department and the pathology department).

[0075] Consultation resources include expert scheduling (e.g., Professor Wang from the Department of Otolaryngology-Head and Neck Surgery is available at 15:00 on February 19, 2025), consultation rooms (e.g., the demonstration room on the 5th floor of Building 5 in a certain hospital campus), and their status is synchronized to the hospital resource management system in real time.

[0076] Data acquisition methods are as follows: Matching criteria: Stored in the hospital's MDT management system rule base, such as the "Head and Neck Tumor MDT Consultation Guidelines," which requires the participation of at least 3 experts from relevant departments. Disease status: Retrieved symptoms from step S100 (e.g., headache, nasopharyngeal carcinoma diagnosis) and complexity scores calculated by step S200 (e.g., a total score of 11). Hospital standards: Retrieved from a standardized document library, such as the "Requirements for MDT Consultation Expert Team Members," which specifies that nasopharyngeal carcinoma MDTs must include experts from radiology and oncology. Resource status: Real-time capture from the hospital's scheduling system (e.g., experts have ≥2 consecutive hours of free time within 3 days).

[0077] The consultation plan was automatically confirmed as follows:

[0078] Algorithm rules: A matching algorithm based on a rule engine is adopted. The triggering condition is: the expert's expertise matching degree is ≥90% (e.g., Professor Zhang has treated a total of 500 cases of nasopharyngeal carcinoma with a cure rate of 82% and a matching degree of 92%).

[0079] The consultation process complies with standards (e.g., including experts from the radiotherapy and pathology departments); the complexity of the case is rated 8 points, and the estimated consultation time is 4 hours (meeting the rule of ≤2 points / hour). Example: The patient is diagnosed with nasopharyngeal carcinoma, and the system recommends Dr. Li from the radiotherapy department (95% match rate) and Dr. Wang from the pathology department (93% match rate), and automatically confirms the treatment plan (e.g., the patient confirms the consultation request form after the secretary issues it).

[0080] Payment terms and resource lockout are as follows:

[0081] Payment process: The system calls the hospital's HIS system payment interface, supporting WeChat / Alipay payments. Upon successful payment, the system receives a "200" status code from the payment platform, triggering resource locking.

[0082] Resource locking rules:

[0083] Expert Resources: In the scheduling system, mark the time slot of Director Li and Director Wang from 15:00 to 17:00 on February 19, 2025 as "MDT Occupied";

[0084] Consultation Room: Automatically book a classroom on the 5th floor of Building 5 in a specific hospital area, and synchronize it with the hospital's space management system.

[0085] In step S400, based on the urgency of the patient's condition, the expert's scheduling information, and the hospital's resource occupancy, an intelligent scheduling system generates a consultation time arrangement and sends consultation invitations and detailed information to relevant experts.

[0086] The urgency level of the illness is categorized based on the risk of life-threatening symptoms. For example, chest pain, difficulty breathing, critical values, and patients with high-risk hospitalization warnings are marked as "urgent," while common chronic diseases are marked as "routine." Expert scheduling information shows the available time slots for experts, such as an expert's availability from 14:00 to 16:00 on February 19, 2025. Hospital resource occupancy indicates the usage status of consultation rooms, such as MDT-03 room being available after 15:00.

[0087] The data acquisition and rules are as follows:

[0088] Urgency level: Identified from symptoms extracted in step S100 (e.g., "persistent chest pain" is marked as urgent). Scheduling data: Synchronized with the hospital's scheduling system in real time, filtering for consecutive available time slots (≥1 hour) for experts within the next 7 days. Resource status: Retrieved from the consultation room management system (e.g., occupancy status of the MDT room on the 3rd floor of the outpatient building).

[0089] The specific process can be found in steps S410 to S470, which will not be elaborated here.

[0090] In step S500, at the start of the consultation, a speech recognition algorithm is used to assist the expert in real-time annotation and analysis. At the same time, the system automatically records the speech discussion during the consultation and converts it into text records.

[0091] Speech recognition algorithms refer to Automatic Speech Recognition (ASR) technology based on deep learning. By training a medical-specific model, it converts expert speech signals into text and supports accurate recognition of medical terms (such as "atrial fibrillation" and "paclitaxel"). Key features include: Real-time performance: Speech input and text output latency ≤500ms, meeting the needs of real-time discussions; Professionalism: Built-in database of 30,000+ medical terms (including alias mappings, such as "cerebral infarction" corresponding to "ischemic stroke"), with a professional term recognition accuracy of ≥98%; Multi-source differentiation: Identifying the speech of different experts through voiceprint features, achieving a "who speaks, who is labeled" association.

[0092] Real-time annotation refers to the operation of experts marking or supplementing the patient's medical records, images and other materials through voice commands, and the annotation results are synchronized to all participating terminals in real time.

[0093] The process of converting voice discussions into text records refers to the system recording the entire consultation process's voice dialogue, converting it into structured text with timestamps through a voice recognition algorithm, and automatically classifying key information (such as diagnostic conclusions and treatment recommendations) to form a traceable electronic file.

[0094] The consultation process is recorded as follows:

[0095] Automatic transcription: The speech recognition algorithm converts the discussion content into text with an accuracy of ≥95% (such as "It is recommended to improve the PET-CT examination" is automatically recorded), and the text is segmented according to the expert role to form a structured draft. After being revised and confirmed by the experts, a complete consultation document is formed.

[0096] In addition, the multidisciplinary collaborative diagnosis and treatment method also includes processing steps that run parallel to step S500, which can be specifically designed as follows:

[0097] For scenarios such as preoperative multidisciplinary discussions, difficult case discussions, general consultations, and single-department consultations for hospitalized patients, online consultation mode is prioritized. Through the integration of cloud video conferencing platform with the hospital information system (HIS), the patient's electronic medical records and imaging data are automatically retrieved to the consultation interface.

[0098] Blockchain technology is used to hash and store the audio and video recordings and expert annotations of online consultations, generating an immutable timestamp.

[0099] Using pre-set standardized consultation process templates (such as surgical risk assessment templates and difficult case diagnosis templates), experts are guided to complete the consultation according to the process through visual methods such as interface pop-ups and progress bars.

[0100] The system monitors the time spent in each stage of the consultation in real time and automatically triggers an alert for stages that exceed the preset time (such as experts who have not signed in 5 minutes beyond the consultation time), and pushes the alert to the medical management terminal to facilitate medical management personnel to supervise the quality of consultations.

[0101] After the consultation, the encrypted video recording and structured text record are automatically linked to the patient's medical record in the HIS, allowing medical management departments to access and review them through digital signature verification.

[0102] In step S600, after the consultation, the system organizes the written records, integrates expert opinions, generates a structured consultation report, and pushes it to the patient.

[0103] Natural language generation technology automatically converts structured data (such as expert opinions and examination results) into medical report text, supporting templated output. Structured consultation reports are documents containing standardized content such as clinical diagnoses, treatment recommendations, and expert signatures, with a unified format for easy archiving and retrieval.

[0104] The report generation process is as follows:

[0105] Data Integration and Processing: Input Sources: Speech-to-text transcripts automatically recorded in step S500 (e.g., expert discussion "consider nasopharyngeal carcinoma with lung metastasis"); real-time image analysis results annotated during consultations (e.g., "lung nodule diameter 2.3cm, biopsy recommended"); expert electronic signatures (via handwriting pad or system-preset signature templates). Processing Rules: Logically organize the text records according to "chief complaint → present illness → auxiliary examinations → diagnostic opinion → treatment plan opinion"; extract key data and fill them into the report template.

[0106] The automated generation and review process is as follows: Technical implementation: Template matching + rule engine technology is used. For example, if the disease is breast cancer MDT, the "Breast Cancer MDT Multidisciplinary Discussion Case Record" is automatically called and standardized suggestions such as "requires consultation with the radiotherapy department" are filled in; the consultation opinions of experts from various specialties are displayed in detail in the report, and the comprehensive summary of consultation opinions is filled into the report template.

[0107] The report push and archiving are as follows: Push mechanism: The system automatically sends the PDF format report to the patient's end (such as the "My Consultation" page in the APP) and attaches an SMS reminder (such as "Your consultation report has been generated, click to view"); it is also synchronously archived to the hospital's electronic medical record system and linked to the patient ID for subsequent diagnosis and treatment reference (similar to the process of the patient viewing the consultation conclusion).

[0108] Furthermore, the system has built-in record quality verification rules to detect missing items in important sections of the report, such as "disease summary, expert opinions, and treatment plan" (using a rule engine to match preset templates). If key information is missing, a forced interception is triggered (prompting "Please supplement XX content before submitting"). Missing item data is automatically calculated (e.g., "treatment plan missing item rate = number of missing reports / total number of reports") and stored as quality control indicators to support subsequent report generation.

[0109] Reference Figure 2 A multidisciplinary collaborative diagnosis and treatment approach also includes steps prior to receiving patient medical records submitted through the collaborative outpatient portal, as follows:

[0110] Step Sa00 involves collecting multimodal data from the patient and using a multimodal fusion algorithm to extract and align features from the collected data, generating a disease feature vector with a unified dimension. The patient's multimodal data includes language, images, and text.

[0111] Multimodal data includes patient information provided in various forms, such as: language: voice descriptions (e.g., recordings of “recurrent syncope for more than 3 years, 1-2 times a year, all of which are induced by fatigue, dizziness and headache, followed by loss of consciousness”); images: medical images (e.g., CT, MRI films) or examination report photos; and text: medical record text and content entered in the symptom input box (e.g., “syncope”).

[0112] Multimodal fusion algorithm: A technique that transforms data in different forms into unified features. For example, it extracts image features through convolutional neural networks (CNN) and language / text features through recurrent neural networks (RNN), and then generates a comprehensive vector by concatenating the features. Disease feature vector: Structured data that represents the disease in numerical form, such as [symptoms: syncope, dizziness, headache, loss of consciousness; duration: more than three years; frequency: 1-2 times per year].

[0113] Language data: Patients can record their symptoms via voice input through the app (such as recording an audio message like "coughing with chest pain"), which is then automatically transcribed into text by the system.

[0114] Image data: Patients upload images of imaging reports (such as chest CT film photos) or examination reports in PDF format, supporting JPEG / PNG / PDF formats (up to 9 attachments, similar to "Upload Examination Data" in PowerPoint).

[0115] Text data: Patients fill in a description of their symptoms in the input box (such as "fatigue and decreased appetite in the past week"), or synchronize the diagnosis record in the electronic medical record (such as a history of "hypertension").

[0116] The specific process is as follows:

[0117] Multimodal feature extraction is as follows:

[0118] Image Features: ResNet convolutional neural network is used to analyze CT images to identify abnormal areas such as pulmonary nodules and pleural effusion, and output feature vectors [nodule location: left upper lobe, diameter: 2.3cm] (e.g., extracting consolidation features from CT images of pneumonia patients). Language / Text Features: Text is analyzed using the BERT natural language processing model to extract entities such as symptoms ("cough", "fever"), time ("2 weeks"), and disease ("pneumonia"), generating vectors [cough: 1, fever: 1, duration: 14 days] (e.g., extracting time and symptoms from the text "recurrent fever for 14 days").

[0119] Feature alignment and fusion are as follows: Alignment rules: Use time and symptoms as anchors to associate data from different modalities. For example: The text description "headache started 3 days ago" and the voice mention "headache lasted for 3 days" are aligned as a unified time feature [duration: 3 days]; The image "brain MRI shows vascular abnormalities" and the text "past history of stroke" are aligned as a disease feature [stroke: 1].

[0120] Fusion Algorithm: An early fusion strategy is adopted to concatenate the image feature vector (CNN output) and the language / text feature vector (BERT output) into a unified dimension vector [cough: 1, fever: 1, duration: 14 days, nodule location: left upper lobe, diameter: 2.3cm] for subsequent analysis.

[0121] Step Sb00: Based on the constructed medical knowledge graph, a graph query algorithm is used to retrieve nodes related to the patient's input, generate guidance questions, and a reinforcement learning-knowledge graph joint model is adopted to dynamically adjust the priority of guidance questions based on historical guidance effect data.

[0122] Among them, the medical knowledge graph is a semantic network with diseases as core nodes, linking entities such as symptoms, examinations, departments, and experts (e.g., the node "migraine" connects to "headache," "nausea and vomiting," and "neurology"). Graph query algorithms are methods for retrieving associated nodes from the knowledge graph, such as the SPARQL query language. The reinforcement learning-knowledge graph joint model is an algorithmic framework that dynamically adjusts question priorities based on historical patient feedback data (e.g., response rate, information effectiveness rate).

[0123] The specific process is as follows:

[0124] Generating guiding questions: Based on the disease feature vector (e.g., "headache + dizziness") from step Sa00, the medical knowledge graph is retrieved using the SPARQL query algorithm to generate relevant questions (e.g., "Is the headache accompanied by blurred vision?" "Is there tinnitus during a dizziness attack?"). For example, if the feature vector contains "pulmonary nodules", the relevant nodes for "differentiation between benign and malignant nodules" in the graph are queried to generate questions such as "Are there any tumor marker test results?" and "Are the nodule edges smooth?".

[0125] Dynamic priority adjustment: The Q-learning reinforcement learning algorithm is used to adjust the ranking based on the "effective information contribution" of the questions in the historical data (e.g., a question improves the diagnostic accuracy by 15%).

[0126] Example rule: For patients with "headache", if the effective response rate for the "headache frequency" question in historical data reaches 80%, it will be displayed in priority over the "headache location" question.

[0127] The output is as follows: A list of guiding questions sorted by priority (such as 3-5 key questions) is generated and pushed to the patient's interactive interface to guide supplementary information (similar to the logic of "the system automatically asking for details during consultation").

[0128] Step Sc00 uses a Conditional Random Field (CRF) model to perform integrity checks on the input text, automatically annotating missing entities and providing examples. A bidirectional LSTM classifier is used for consistency checks, detecting and questioning contradictory information. For ambiguous descriptions, an entity linking algorithm is invoked to match standard terms, providing options for the patient to choose from.

[0129] Among them, the Conditional Random Field (CRF) model is a machine learning model used for sequence labeling to detect missing entities (such as symptoms, time, and test results) in text. The Bidirectional LSTM classifier is an algorithm based on recurrent neural networks to identify contradictory information in text (such as the coexistence of "no history of allergies" and "penicillin allergy"). The entity linking algorithm matches ambiguous descriptions to standard medical terms (such as mapping "shaking" to "chills").

[0130] The specific process is as follows:

[0131] 1. Completeness Validation: The input text is analyzed using a CRF model, and missing entities (such as "duration of symptoms" not mentioned) are annotated. A prompt is automatically generated: "Please complete the duration of cough (example: 3 days / 1 week)." For example: If a patient describes "fever and cough," and CRF detects the missing "degree of fever," the prompt will be: "Please specify the highest temperature (example: 38.5℃)."

[0132] 2. Consistency Check: A bidirectional LSTM classifier is used to compare the text before and after to detect contradictory information. For example, if a patient first fills in "no history of smoking" and then mentions "smoking 10 cigarettes a day", the system will mark it and ask, "Please confirm whether the description of the smoking history is consistent."

[0133] 3. Standardization of fuzzy descriptions: For vague expressions such as "chest pain", the entity linking algorithm is used to match standard terms (such as "chest pain") and provide options "① chest pain ② palpitations ③ chest tightness" for patients to choose from.

[0134] Step Sd00: Extract background feature data from patient registration information, construct patient profile vector, retrieve input examples of similar cases based on collaborative filtering algorithm, and guide patients to supplement information.

[0135] The background feature data includes structured data such as age, gender, past medical history, and allergy history from patient registration information (e.g., "50 years old / male / 10-year history of hypertension"). The patient profile vector converts the background features into a numerical vector (e.g., [age:50, gender:1, hypertension:1, allergy history:0]) for case similarity calculation. The collaborative filtering algorithm recommends reference examples based on the similarity between the patient profile and historical cases (e.g., retrieving symptom description templates for patients of similar age and medical history).

[0136] The specific process is as follows:

[0137] 1. Construct patient profiles: Extract background features from registration information and generate vectors using one-hot encoding. For example, age "50 years old" is encoded as [0,0,1,0] (assuming age range: <30 / 30-40 / 50-60 / >60 years old); past medical history "hypertension" is encoded as [hypertension:1, diabetes:0, coronary heart disease:0].

[0138] 2. Similar Case Retrieval: Using the cosine similarity algorithm, cases with a similarity of ≥70% to the patient profile vector are retrieved from the historical case database. For example, for a 50-year-old male patient with hypertension, a historical case with a similar profile, "headache accompanied by nausea and vomiting, diagnosed as hypertensive encephalopathy," is retrieved. The symptom description "sudden onset of severe headache, frequent vomiting, and blurred vision" is extracted as a guiding example.

[0139] 3. Information guidance: Show patients input examples of similar cases (such as "Do you have similar symptoms to 'sudden headache'?") to help supplement missing information (such as details such as "blurred vision").

[0140] Step Se00: When the patient's input information passes the integrity check and consistency check, and the standard terminology selection is completed for the fuzzy description, the input is considered complete.

[0141] The process includes the following steps: Integrity verification: A Conditional Random Field (CRF) model is used to check whether the patient's medical information contains necessary entities (such as symptoms, time, and examination results). Consistency verification: A bidirectional LSTM classifier is used to identify whether there are contradictory descriptions in the text (such as "no history of allergies" versus "penicillin allergy"). Standard terminology selection: For ambiguous descriptions, an entity linking algorithm is used to match standard medical terms (such as "shivering" → "chills") and the options are confirmed.

[0142] The judgment rules are as follows:

[0143] 1. Completeness requirement: All required entities (such as symptom duration and major medical history) are labeled and there are no missing items (refer to the CRF model output results). Example: The patient's input includes "cough for 3 days, fever 38.5℃, chest CT showed pneumonia", which meets the completeness requirement of "symptoms + time + imaging examination".

[0144] 2. Consistency Passing Criteria: The bidirectional LSTM classifier did not detect contradictory information (e.g., consistent descriptions of "smoking history" throughout). Example: The patient confirmed "no smoking history" and did not mention smoking-related content throughout the entire process; therefore, consistency was deemed passed.

[0145] 3. Fuzzy description processing completed: All fuzzy expressions are matched with standard terms through entity linking algorithm and confirmed by the patient (e.g., “heartburn” → select “chest pain”).

[0146] Based on the constructed medical knowledge graph, a graph query algorithm is used to retrieve nodes related to patient input, generate guidance questions, and a reinforcement learning-knowledge graph joint model is adopted to dynamically adjust the priority of guidance questions based on historical guidance effect data, including:

[0147] Step Sb10: Encode the nodes and edges of the medical knowledge graph using a Transformer-based model to capture deep semantic information.

[0148] Among these, Transformer-based models employ self-attention mechanisms in deep learning, such as BioBERT (a pre-trained model optimized for the biomedical field). Node and edge encoding transforms entities (e.g., "diabetes") and relationships (e.g., "complications") in the knowledge graph into low-dimensional vector representations. Deep semantic information uncovers implicit connections between medical concepts (e.g., the risk relationship between "hypertension" and "cardiovascular disease").

[0149] The specific process is as follows:

[0150] 1. Knowledge Graph Vectorization: Use BioBERT to text-encode nodes (such as "migraine") in the medical knowledge graph to generate 768-dimensional vectors (such as [0.12,0.34,...,0.78]).

[0151] For edges (such as “symptom-headache”), the TransE algorithm is used to map the relationship to a vector offset (such as headache vector = migraine vector + symptom relationship vector).

[0152] 2. Semantic Information Capture: Through a multi-head self-attention mechanism, the association weights between nodes are calculated (e.g., the co-occurrence probability of "migraine" and "nausea and vomiting"). Example: When the input is "headache", the model identifies the association degree of the "migraine" node as 0.85 through attention weights, which is higher than the 0.62 of "tension headache".

[0153] Step Sb20 involves modeling the context of the patient's input, combining LSTM or Transformer models to understand its meaning in different contexts, and accurately matching relevant knowledge points.

[0154] Contextual modeling: Analyze the contextual relationships of words in the patient's input text (e.g., the different contexts of "cough" in "smoking causes cough" and "cough caused by a cold").

[0155] LSTM model: Long Short-Term Memory network, which captures long-distance dependencies in text sequences (such as identifying the temporal correlation in "diagnosed with lung cancer last year, cough has worsened recently").

[0156] Transformer model: An architecture based on self-attention mechanism that processes semantic associations in text in parallel (such as simultaneously analyzing the mutual influence of "chest pain, shortness of breath, and history of hypertension").

[0157] The specific process is as follows:

[0158] Text sequence encoding: Convert patient input (e.g., “coughing for the past week, shortness of breath after activity”) into word embedding vectors (e.g., using Word2Vec to generate vector representations of “coughing” and “shortness of breath”).

[0159] Contextual Analysis: LSTM Path: Captures the temporal relationship between "the past week" and "cough" through hidden layer state propagation, outputting a context vector containing temporal information. Transformer Path: Utilizes multi-head self-attention to calculate the association weights between "cough" and "after the activity" (e.g., a weight of 0.78 indicates a strong correlation), generating parallel semantic encoding.

[0160] Knowledge point matching: Compare the encoded context vector with the knowledge graph vector generated in step Sb10 (e.g., calculate the cosine similarity between the "shortness of breath" vector and the "cardiopulmonary disease" node) to accurately locate relevant knowledge points (e.g., "heart failure" and "chronic obstructive pulmonary disease").

[0161] Step Sb30: Calculate the semantic similarity between the patient's input and the knowledge graph nodes to determine the most relevant knowledge points.

[0162] Semantic similarity: Measures the degree of semantic association between the patient's input text and knowledge graph nodes, with a value range of [0,1] (higher values ​​indicate greater relevance). Cosine similarity algorithm: Evaluates semantic similarity by calculating the cosine of the angle between vectors.

[0163] The specific process is as follows:

[0164] 1. Vector Preparation: Patient Input Vector: Context modeling results from step Sb20 (e.g., the encoded vector for "chest pain with palpitations"). Knowledge Graph Node Vector: Transformer encoding results from step Sb10 (e.g., vectors for nodes such as "coronary artery disease" and "arrhythmia").

[0165] 2. Similarity Calculation: For each knowledge graph node, the cosine similarity algorithm is used to calculate the degree of relevance with the patient's input vector. Example: The cosine value of the patient's input vector "chest pain after activity" is 0.89 with the "coronary heart disease" node vector, and 0.45 with the "intercostal neuralgia" node vector. The former is judged to be more relevant.

[0166] 3. Knowledge point selection: Sort by similarity value in descending order, and retain the top 5 nodes (such as "coronary heart disease", "angina pectoris", "myocardial ischemia", etc.) as candidate knowledge points for generating subsequent guiding questions.

[0167] Step Sb40: Combining semantic similarity and context analysis results, a suitable guidance question template is matched from a predefined guidance question template library.

[0168] The results include: semantic similarity score: the correlation score between the patient input and the knowledge graph node calculated in step Sb30 (e.g., the similarity score for the "pneumonia" node is 0.92). Contextual analysis results: the contextual information of the patient input extracted in step Sb20 (e.g., the correlation between time and symptoms in "fever with cough for 1 week"). Guiding question template library: a predefined set of standardized questions, such as templates for the "pneumonia" node: "Have you experienced any worsening of {symptoms}?" and "Have you recently been in contact with {source of infection}?".

[0169] The specific process is as follows:

[0170] 1. Template Matching Rules: Based on the high similarity nodes in step Sb30 (such as the top 3 "pneumonia", "bacterial pneumonia", and "viral pneumonia"), retrieve the corresponding disease categories in the template library. Example: After matching the "pneumonia" node, retrieve the template set under this category: "Is there a high fever (body temperature ≥38.5℃) or chills?" "Is the cough accompanied by yellow-green sputum or bloody sputum?"

[0171] 2. Contextual adaptation: Combine the contextual information from step Sb20 (such as the patient mentioning "worsening cough"), filter out irrelevant templates (such as excluding tuberculosis-related questions such as "whether there is night sweats"), and retain templates that match the current context (such as "is the worsening cough related to the activity?").

[0172] Step Sb50: Based on the patient's profile information, the matched guidance question template is personalized.

[0173] The patient profile information includes structured data such as age, gender, past medical history, and allergy history (e.g., "65 years old / male / 10-year history of hypertension"). The guiding question template consists of predefined standardized questions (e.g., "Does this symptom recur?"), which need to be dynamically filled in or adjusted based on the patient's characteristics.

[0174] The specific process is as follows:

[0175] Feature extraction and matching: Extract key features from the patient profile (such as age "65 years old" and medical history "hypertension"). Example: If the patient is a child (age <12 years old), adjust "smoking history" in the template to "passive smoking history"; if there is "penicillin allergy", filter out questions containing "antibiotic use".

[0176] Template personalization rules: Age adaptation: For elderly patients, add "Do you have any underlying diseases that are not well controlled?", and for children, add "Were you vaccinated before the onset of the disease?". Medical history correlation: If there is a history of "diabetes", insert relevant questions such as "How has your blood sugar been controlled recently?" into the template.

[0177] Output: Generate personalized questions for the current patient (e.g., a guiding question for a 65-year-old male with hypertension: "Does chest pain occur with elevated blood pressure?"), improving the relevance and specificity of the questions.

[0178] Step Sb60: Using a reinforcement learning-knowledge graph joint model, the generated guidance questions are prioritized based on historical guidance effect data.

[0179] The reinforcement learning-knowledge graph joint model combines the dynamic optimization capabilities of reinforcement learning (RL) with the semantic associations of knowledge graphs (KG) to prioritize guidance questions. Historical guidance effect data records metrics such as the frequency of question responses and the improvement in diagnostic accuracy after a response (e.g., a certain question reduces diagnosis time by 20%).

[0180] The specific process is as follows:

[0181] Reward function definition: Set reward indicators: Information effectiveness rate: The proportion of newly added effective diagnostic information after answering a question (e.g., if a question increases the acquisition rate of key diagnostic indicators by 30%, the reward is +0.5).

[0182] Answering time: The average time it takes for a patient to answer a question (e.g., <2 minutes, reward +0.3; >5 minutes, reward -0.2).

[0183] Priority ranking rule: The order of questions is iteratively optimized using the Q-learning algorithm, prioritizing questions with higher historical reward values ​​(e.g., the "nature of chest pain" question is ranked first due to its high effectiveness). Example: For patients with "headache", the reward value of the "headache location" question (0.8) in historical data is higher than that of the "headache trigger" question (0.6), so the former is displayed first.

[0184] Dynamic adjustment mechanism: The reward value is updated in real time based on the feedback from new patients (e.g., if the efficiency of a certain question decreases in the last 100 cases, the priority is automatically reduced).

[0185] The output results are as follows: A priority list of guiding questions is generated (e.g., Top 3 questions: "Is the headache accompanied by nausea and vomiting?", "How long does the pain last?", "Is there a family history of migraines?"), ensuring that key information is obtained first.

[0186] A multidisciplinary collaborative diagnosis and treatment approach also includes a step that runs parallel to receiving patient medical data submitted through the joint outpatient portal, as follows:

[0187] When a doctor initiates a consultation, the system retrieves medical records from the patient's information file and adds supplementary information.

[0188] Among them, the HIS system integrates a dedicated module that provides multi-dimensional patient retrieval (supporting medical record number / name / visit time), consultation type selection (preset tumor MDT, difficult disease consultation, etc.), urgency level classification (normal / urgent / special urgent), and key concern question input box (supporting voice and structured terminology association input). Through the above dedicated module, consultations initiated by doctors can be received.

[0189] The system automatically parses the structured fields annotated by doctors and combines them with the electronic medical records retrieved from the HIS to generate standardized disease feature vectors. The structured fields include consultation type and key concerns.

[0190] Based on the consultation type and urgency level marked by the doctor, the corresponding specialty MDT template is retrieved from the knowledge graph, the joint outpatient type is automatically matched and recommended, and the process jumps to step S200.

[0191] Additionally, the hospital information system (HIS) and the internet hospital platform will exchange personnel data and manage access permissions, specifically including:

[0192] The system monitors personnel data change events in the HIS in real time (such as the entry of information for newly hired doctors and job adjustments). Based on preset mapping rules, it automatically generates corresponding accounts on the Internet hospital platform. The account information includes fields such as employee number, name, department, and title, avoiding manual duplicate registration.

[0193] Construct an RBAC (Role-Based Access Control) model, setting consultation initiation permissions and internet hospital patient access permissions as independent permission groups: Resident physicians and attending physicians are assigned consultation initiation permissions by default, allowing them to access HIS medical records and submit consultation requests; only doctors who have passed the internet medical qualification review and have the title of attending physician or above are granted internet hospital patient access permissions, supporting operations such as text and image consultations and video consultations.

[0194] When personnel in the HIS are promoted, transferred, or have their qualifications expire, the system automatically triggers a permission verification process, synchronously updates the permission configuration of the Internet Hospital platform, and pushes permission change notifications to relevant personnel through a message queue.

[0195] All permission-related operations (such as permission allocation, modification, and revocation) are stored on the blockchain to generate an immutable log containing the operation time, operator, and target, which can be used by medical management departments for retrospective review.

[0196] Furthermore, it is also possible to consider interoperability with the hospital's HIS system to achieve resource sharing of consultation personnel and layout. Therefore, there can be a parallel step to receiving patient medical information submitted through the joint outpatient portal, as follows:

[0197] Through API interfaces or message queue mechanisms, data exchange between the Hospital Information System (HIS) and the multidisciplinary consultation platform can be achieved, and medical staff scheduling information (including available time slots and on-duty status), expert specialty tags, and departmental resource occupancy can be synchronized in real time.

[0198] Upon receiving a consultation request initiated by a doctor on the consultation platform, the system automatically retrieves the patient's electronic medical record (including historical diagnoses, examination reports, and medication records) from the HIS and provides a structured form for doctors to supplement and annotate fields such as consultation type and key concerns.

[0199] The Conditional Random Field (CRF) model is used to parse the structured fields labeled by doctors. Combined with the electronic medical record data retrieved from HIS, multimodal feature fusion is performed through the Transformer architecture to generate a standardized disease feature vector containing disease severity, treatment history, and urgency level.

[0200] Based on real-time scheduling data synchronized with HIS and expert specialty tags, an expert-patient condition matching knowledge graph is constructed. Collaborative filtering algorithms are used to calculate the matching degree between experts and patients' conditions. Combined with the edge weights of historical collaboration relationships in the knowledge graph, expert team combinations are automatically recommended. At the same time, based on the expert scheduling conflict detection results, more than 3 sets of optional consultation time slots and resource occupancy status reports are generated.

[0201] Based on the consultation type (such as preoperative MDT at level 4, difficult case discussion, general consultation, etc.) and urgency level marked by the doctor, the corresponding specialty MDT template is retrieved from the knowledge graph, the semantic similarity between the template and the disease feature vector is calculated, the joint outpatient type with a matching degree greater than the preset ratio is automatically recommended, and the process jumps to step S200.

[0202] The types of joint outpatient clinics automatically recommended include:

[0203] Step S110: Receive multimodal medical data, including voice, text, and images, submitted by the patient through the joint outpatient portal.

[0204] Among them, the Joint Clinic entrance is an online channel for patients to initiate MDT applications (such as the "Joint Clinic" module in the hospital's APP / mini-program), which supports the submission of multimodal data.

[0205] Multimodal medical data: Voice: Audio description of symptoms recorded through a microphone (e.g., "cough with fever for 3 days"); Text: Manually entered symptom text (e.g., "recurrent headache for half a year") or electronic medical records; Images: Uploaded medical images (CT / MRI films), photos of examination reports, or screenshots of laboratory reports.

[0206] The data reception and format requirements are as follows:

[0207] Input method: After logging into the system, patients can select "Quick Application" or "Online Consultation" on the Joint Outpatient Clinic page to enter the data submission interface. Multiple modalities of data can be uploaded simultaneously (such as text descriptions + CT images + voice supplements), and up to 9 attachments (image / document formats) can be added.

[0208] Formatting specifications: Audio: mp3 / wav format, single recording ≤ 5 minutes; Images: JPEG / PNG format, resolution ≥ 300dpi (supports stitching multiple images); Text: supports plain text input or PDF / Word document upload (automatic parsing of key information).

[0209] Step S120: Using a multimodal fusion algorithm based on the Transformer architecture, feature extraction and cross-modal alignment are performed on data from different modalities to generate a disease feature vector with a unified dimension.

[0210] Feature extraction involves extracting key medical information from various modal data, such as symptoms in speech ("chest pain"), lesions in images ("lung nodules"), and medical history in text ("hypertension for 5 years"). Cross-modal alignment maps features from different modalities to a vector space of a unified dimension, bridging the semantic gap (e.g., the speech description, text record, and image representation of "cough" correspond to the same feature vector).

[0211] The specific process is as follows:

[0212] 1. Modal feature extraction: Speech: Use the Wav2Vec2.0 model to extract acoustic features, identify symptom keywords (such as "fever" and "shortness of breath"), and generate speech feature vectors (256 dimensions).

[0213] Text: Natural language processing is performed using the BERT model to extract entities (such as "diabetes" and "abnormal electrocardiogram") and generate text feature vectors (768 dimensions). Images: Medical images are analyzed using a ResNet50 convolutional neural network to identify the location / type of lesions (such as "ground-glass opacity in the upper lobe of the left lung") and generate image feature vectors (1024 dimensions).

[0214] 2. Cross-modal alignment and fusion: The association weights of different modal features are calculated through a multi-head self-attention mechanism (e.g., the association degree between "chest pain" in speech and "enlarged heart" in image is 0.91), and a unified dimension vector (e.g., 256+768+1024=2048 dimensions) is generated by feature concatenation method.

[0215] 3. Output results: Generate disease feature vectors containing multimodal semantic information (e.g., [chest pain: 0.89, enlarged heart: 0.91, diabetes: 0.75]) for disease node matching in step S130.

[0216] Step S130: Input the disease feature vector into the medical knowledge graph, retrieve the disease knowledge base through the graph query algorithm, match the disease nodes and related departments most relevant to the patient's condition, and automatically recommend joint outpatient clinic types.

[0217] The medical knowledge graph is a semantic network with diseases as core nodes, linking entities such as symptoms, examinations, departments, and specialists (e.g., the "lung cancer" node connects to "cough," "CT scan," and "thoracic surgery"). Graph query algorithms are techniques for retrieving associated nodes from the knowledge graph, such as the SPARQL protocol or graph traversal algorithms. Joint outpatient clinic types are pre-defined multidisciplinary treatment combinations based on disease complexity and involved departments (e.g., "oncology MDT," "neurological disease MDT," and "cardiovascular joint outpatient clinic").

[0218] The specific process is as follows:

[0219] 1. Knowledge Graph Retrieval: Input the disease feature vector generated in step S120 (e.g., [lung nodules: 0.95, chest pain: 0.88]) into the medical knowledge graph, and match relevant disease nodes (e.g., "lung cancer", "lung nodules", "pneumonia") using the SPARQL query algorithm. Example: If the feature vector contains "lung ground-glass opacity", retrieve the disease node in the graph with the highest correlation to this image (e.g., "early lung cancer" with a similarity of 0.92).

[0220] 2. Department Association Matching: Based on the "Department" edge relationship of the disease node (e.g., "Lung Cancer" → "Thoracic Surgery", "Medical Oncology", "Radiation Oncology"), extract a list of associated departments. Rules: If a single disease is associated with ≥3 departments, automatically match "Oncology MDT"; if it involves cross-system diseases (e.g., "Diabetes + Retinopathy"), match "Chronic Disease Joint Clinic".

[0221] 3. Recommended Joint Outpatient Clinic Types: Results are generated by sorting the number of related departments and the severity of the disease. For example: Lung cancer (thoracic surgery + medical oncology + radiotherapy) → Recommended "Oncology MDT Clinic"; Headache with cerebrovascular abnormalities (neurology + neurosurgery + radiology) → Recommended "Neurological MDT Clinic".

[0222] Output results: Show patients 1-3 matching joint outpatient clinic types (e.g., "Recommendation: Multidisciplinary Joint Outpatient Clinic for Oncology"), with descriptions of the department combinations (e.g., "Thoracic Surgery + Medical Oncology + Pathology"), for patients to confirm or adjust.

[0223] After inputting the disease condition feature vector into the medical knowledge graph, the graph query algorithm retrieves the disease knowledge base, matches the disease nodes and related departments most relevant to the patient's condition, and automatically recommends joint outpatient clinic types, the following steps are still required:

[0224] Step S140: Visualize the recommendation information using natural language generation technology and push it to the patient's device in the form of an interactive card, simultaneously annotating the basis for the recommendation, and supporting the patient to confirm with one click or initiate a type change request.

[0225] Among them, Natural Language Generation Technology: An algorithm that transforms structured data (such as outpatient type, priority, and reasons for matching) into human-readable text, supporting templated and dynamically populated content. Interactive Cards: Mobile interface components that display recommendation information in a combination of text and images, supporting interactive operations such as clicking and swiping. Recommendation Basis: Explanatory content generated based on patient data and algorithmic logic (e.g., "Your location is closer to the XX Hospital Oncology MDT Center, therefore, we give you priority recommendation").

[0226] The specific process is as follows:

[0227] Structured information transformation: Using the T5 text generation model, the recommendation results of step S140 (such as "tumor MDT clinic") are transformed into natural language descriptions, such as "We recommend tumor multidisciplinary joint clinics to you".

[0228] Visual presentation: Embed the text content into a card-style UI template, including outpatient type icons (such as tumor icons), priority labels (red "priority" markers), recommendation paragraphs, and a thumbnail of the hospital's location map (example image).

[0229] Interactive functionality: The card has "One-click Confirmation" and "Change Request" buttons at the bottom. Clicking "One-click Confirmation" will automatically redirect you to the appointment process; clicking "Change Request" will bring up a reason selection box (such as "too far away" or "too high cost"), and after submission, it will trigger manual review.

[0230] Output: A visual recommendation card is pushed to the patient's terminal (APP / Mini Program), clearly showing the type of joint outpatient clinic, recommendation priority and matching basis, simplifying the decision-making process and supporting flexible adjustments.

[0231] Based on the recommended type of joint outpatient visit, and considering the complexity of the patient's condition, historical consultation data, and the areas of expertise of the specialists, the system intelligently recommends consultation fee ranges and expert team combinations, and pushes these recommendations to the patient's end, including:

[0232] Step S210: Construct a medical knowledge graph that includes expert expertise, collaboration history, and successful cases.

[0233] The specific construction process is as follows:

[0234] Data collection: Extract basic information and professional field tags of experts from the hospital's expert database; obtain historical data on expert collaboration from the consultation record system; screen successful cases from the case database and associate them with the responsible experts.

[0235] Graph construction: Abstract experts, professional fields, diseases, etc. into nodes; connect nodes through relationships such as "expertise", "collaboration", and "treatment". For example, expert A is connected to the "lung cancer targeted therapy" node by an edge of "expertise", and expert A and expert B are connected by an edge of "collaboration"; add attributes to edges and nodes, such as labeling the number of collaborations on the "collaboration" edge and labeling the years of practice on the "expert" node.

[0236] Step S220: Using a graph attention network algorithm, the matching degree between experts and patients' conditions is calculated based on the complexity and type of the patient's condition. Experts with high matching degree are selected, and the collaborative relationship between experts is analyzed to form a complementary collaborative team.

[0237] Among them, the graph attention network algorithm is a graph neural network algorithm based on the attention mechanism, which can automatically learn the association weights between nodes and uncover hidden relationships. Disease complexity is a numerical value quantified by indicators such as the number of symptoms and the severity of the disease. Matching degree is a numerical value measuring the degree of fit between the expert's professional ability and the patient's condition, ranging from [0,1].

[0238] The specific process is as follows:

[0239] Calculating expert matching degree: The disease feature vector generated in step S120 and the medical knowledge graph constructed in step S210 are input into the graph attention network. The algorithm automatically calculates the association weights between each expert node and disease nodes related to the patient's condition, thus obtaining the matching degree for each expert. For example, for lung cancer patients, expert C, who specializes in lung cancer diagnosis and treatment, has a matching degree of 0.85, which is higher than other experts.

[0240] Filter highly matched experts: Set a matching threshold (e.g., ≥0.7) to filter out a list of suitable experts.

[0241] Building complementary teams: Based on the historical collaboration relationships among experts in the knowledge graph, analyze the level of teamwork and prioritize combinations with frequent collaborations and high success rates. Simultaneously consider professional complementarity; for example, assign experts in oncology, surgery, and radiology to teams treating patients with complex tumors, forming a collaborative team with complementary diagnostic and treatment capabilities.

[0242] Output: Generates a list of teams composed of highly matched and complementary experts.

[0243] Step S230: Combine historical consultation data and train the cost prediction model using a multi-task learning framework.

[0244] Among them, the multi-task learning framework simultaneously optimizes the machine learning architecture of the main task (cost prediction) and auxiliary tasks (such as duration prediction and outpatient type classification) to improve the model's generalization ability.

[0245] Historical consultation data: sourced from HIS (Hospital Information System) and EMR (Electronic Medical Record Database), including: disease characteristics (complexity, disease type, etc., taken from the processing results of steps S100-S220); expert characteristics (title, number, etc., extracted from the expert database); and consultation attributes (duration, whether expedited, etc., obtained from consultation records).

[0246] The key processes are as follows:

[0247] Data preprocessing: Cleaning historical data (filling missing values, correcting abnormal costs), and converting information such as disease conditions and experts into feature vectors that the model can recognize (such as numerical standardization and category coding).

[0248] Model Construction: a) Shared Layer: Extracts common features for each task (e.g., the correlation pattern between disease complexity and cost) through a neural network; b) Task Layer: The main task outputs the cost prediction value, and auxiliary tasks (duration, outpatient type) provide supplementary constraints. The model is optimized through a weighted loss function (integrating the errors of each task); c) Training and Optimization: Iteratively trains using divided training and validation sets. The prediction error is reduced by adjusting network parameters (e.g., learning rate, number of iterations), ultimately enabling the model to achieve a preset accuracy on the test set (e.g., cost prediction error ≤ 15%).

[0249] Training and optimization: Iterative training is performed using divided training and validation sets. Prediction error is reduced by adjusting network parameters (such as learning rate and number of iterations), and finally the model achieves the preset accuracy on the test set (such as cost prediction error ≤15%).

[0250] Step S240: Using the trained cost prediction model, predict the basic cost range for the consultation based on the patient's condition and the selected expert team combination.

[0251] Among them, the cost prediction model refers to the multi-task learning-based model trained in step S230, which is used to estimate the consultation cost.

[0252] Patient condition: including condition complexity (symptom / disease / complication weighted score), disease type (ICD-10 code), etc., generated through natural language processing and knowledge graph matching in step S100.

[0253] Expert team composition: A list of highly matched and complementary experts (including features such as professional title, department, and collaboration history) selected through step S220 graph attention network screening.

[0254] The key processes are as follows:

[0255] 1. Input Feature Integration: Disease Characteristics: Extract structured data generated in step S100 (complexity score, disease type vector, etc.). Expert Characteristics: Expert titles (e.g., chief physician / associate chief physician), number, departmental distribution, and collaboration synergy in the coding team (based on historical collaboration records from the knowledge graph).

[0256] 2. Cost Prediction and Interval Generation: a. Point Estimation: The model outputs a basic cost prediction (e.g., 1500 yuan), based on a multi-task learning fusion model of the association between the patient's condition, expert, and consultation attributes. b. Confidence Interval: The standard deviation is calculated based on the model's historical prediction residuals to generate a 95% confidence interval (e.g., ±196 yuan, forming a range of 1304-1696 yuan), reflecting the predicted fluctuation range. c. Dynamic Adjustment: The cost is increased by 10%-20% based on the urgency level (urgent / express), plus additional fees for special examinations.

[0257] 3. Structured Output: Generates cost information including the following: 3.1. Basic Cost: Core prediction value of the model; 3.2. Cost Range: Fluctuation range combined with confidence level; 3.3. Component Details: Itemized explanations of expert consultation fees, disease premiums, duration fees, etc.; 3.4. Prediction Quality: Error indicators (such as MAE) and model version, improving transparency.

[0258] Step S250: The recommended expert team combination and consultation fee range are integrated into structured information and pushed to the patient's end, allowing the patient to view the basis for the recommendation and the fee details.

[0259] The structured information includes data on joint outpatient clinic types, consultation cost ranges, and expert team combinations, all integrated in a standardized format (such as JSON or XML). Recommendation criteria include explanatory content generated based on patient condition, expert matching degree, and cost prediction model output. Cost details include a breakdown of basic costs, medical insurance reimbursement amounts, and additional costs.

[0260] Push notifications and interactions are as follows:

[0261] The integrated information will be pushed to patients through hospital apps, text messages, and other channels.

[0262] The patient-side interface allows users to click to view the recommendation criteria (such as "Expert A has a matching degree of 0.85 and is skilled in lung cancer diagnosis and treatment") and cost details, and also allows users to submit questions or requests for adjustments.

[0263] Furthermore, it is advisable to add cross-institutional expert matching logic within the expert's area of ​​expertise. In this case, the intelligent recommendation of expert team combinations would include:

[0264] A cross-hospital knowledge graph covering multiple institutions within a regional medical alliance is constructed. Each institution uploads encrypted expert feature vectors (including professional fields, cross-hospital consultation history, treatment success rate, etc.) to the central server through federated learning technology. The feature vectors are protected by homomorphic encryption algorithms (such as the Paillier algorithm) to ensure privacy and security.

[0265] The graph attention network (GAT) algorithm is used to calculate the matching degree scores between in-hospital experts and patients' conditions by combining patient condition characteristics with cross-hospital knowledge graphs. At the same time, institutional collaboration weight factors (such as the number of successful cross-hospital consultations and the effectiveness of joint treatment) are introduced to optimize the matching results and generate complementary team combinations that include in-hospital and out-of-hospital experts.

[0266] The institutional collaboration weighting factor automatically compiles historical cross-institutional consultation data through smart contracts. The automatic compilation of this data via smart contracts includes the following steps: 1. Data Triggering Input: After each cross-institutional consultation, the participating institutions' systems automatically submit core data (including the consultation institution pair, consultation results (success / failure), combined treatment case IDs, and follow-up effects) to the smart contract, triggering data recording; 2. Data Verification and Storage: The contract first verifies the authenticity of the data (comparing the consultation records of both institutions to ensure consistency), and then stores it on the blockchain (immutable) to prevent data tampering or omission; 3. Categorized Statistical Storage: Statistics are categorized by "institution pair + disease type," such as A and B hospitals, or the lung cancer field. The system automatically accumulates the total number of consultations, the number of successful consultations, and the number of effective combined treatment cases for this combination. The data is updated and searchable in real time.

[0267] The steps for calculating the institutional collaboration weight factor are as follows: 1. Basic weight calculation: Retrieve the consultation success rate (number of successful consultations ÷ total number of consultations) of the "institution pair" from the contract. The higher the success rate, the higher the basic weight (e.g., if the success rate of Hospital AB is 90%, the basic weight is higher than that of Hospital AC (70%)). 2. Effectiveness weight bonus: Retrieve the combined treatment effectiveness rate (number of effective cases ÷ total number of combined treatment cases). For every 10% increase in effectiveness rate, an additional weight score is added (e.g., an effectiveness rate of 90% adds more points than 80%). 3. Comprehensive weight ranking: The comprehensive weight is obtained by adding the basic weight and the effectiveness bonus. The institutions are ranked from highest to lowest score, and experts from high-scoring institutional pairs are prioritized for matching.

[0268] Intelligent scheduling and generation of consultation schedules include:

[0269] Step S410: Collect and integrate patient pre-assessment level, expert schedule, hospital resource status, and medical standard constraints.

[0270] The pre-assessment level is a priority classification based on indicators such as the severity of the patient's condition and the urgency of symptoms (e.g., Level I is critical, Level III is routine). Expert schedule: The time range during which experts can participate in consultations (e.g., "Dr. Zhang is available Wednesday morning from 9:00-11:30") and the capacity within that time period. Hospital resource status: The occupancy status of physical resources such as consultation rooms (e.g., "MDT consultation room 2 is available today from 14:00-16:00"). Medical standard constraints: Treatment process rules (e.g., "Critical patients must complete consultations within 4 hours"), restrictions on the scope of expertise of specialists, etc.

[0271] The data integration content is as follows:

[0272] 1. Patient dimension: pre-assessment level (from the initial screening results of the condition), available time window for participation in consultation (e.g., "Patient A can only attend on Thursday afternoon").

[0273] 2. Expert Dimension: Each expert's schedule for the next 7 days (accurate to 30-minute intervals) and the number of consultations already scheduled (to avoid overloading).

[0274] 3. Resource dimension: Consultation room uses calendar (marking occupied time slots).

[0275] 4. Rule dimensions: The urgency level corresponds to the time requirement (e.g., Level I patients must be matched with an expert available within 48 hours), and the rule of having multidisciplinary experts present at the same time.

[0276] Output: Generates a comprehensive dataset that includes patient needs, expert resources, and rule constraints.

[0277] Step S420: Introduce a fuzzy logic algorithm to convert the urgency of the illness into an urgency index of 0-1.

[0278] Among them, fuzzy logic algorithm: a mathematical tool for processing imprecise and fuzzy information, converting natural language descriptions (such as "critical condition") into numerical values ​​through membership functions. Severity of condition: a critical state assessed based on a comprehensive evaluation of the patient's symptoms, examination results, etc. (such as "chest pain accompanied by difficulty breathing" being an emergency). Emergency index: quantifies the urgency of the condition into a numerical value between 0 and 1, with higher values ​​indicating greater urgency (0 for non-urgent, 1 for critical).

[0279] The specific process is as follows:

[0280] Key indicator extraction: Extract emergency-related indicators such as symptoms (e.g., "hemoptysis" or "coma"), vital signs (heart rate > 120 beats / min), and examination results (CT scan showing cerebral hemorrhage) from the patient's pre-assessment information.

[0281] Fuzzy rule setting:

[0282] Establish a fuzzy rule base, for example: if "symptom = chest pain and abnormal vital signs", then the urgency level is "high"; if "symptom = common cough", then the urgency level is "low".

[0283] Numerical transformation: Fuzzy rules are mapped to emergency indices through membership functions, such as 0.9 for "critical condition" and 0.1 for "stable condition".

[0284] Step S430 combines the backtracking search algorithm with the heuristic function optimization. The heuristic function is based on the Pareto optimality principle and simultaneously optimizes conflict resolution efficiency, priority of emergency patients, and expert workload balance.

[0285] The algorithm includes the following components: A backtracking search algorithm: This algorithm uses a depth-first strategy to traverse the solution space tree. In consultation scheduling, each node represents a patient's consultation arrangement. When the current arrangement leads to expert conflict or insufficient resources, it immediately backtracks to the previous node to reselect. A heuristic function: This is a "guided scoring tool" that scores intermediate solutions during the search process, prioritizing higher-scoring paths to reduce blind spots and improve search efficiency. The Pareto optimality principle: This is a multi-objective optimization concept that refers to the optimal state where "one objective cannot be improved without compromising another." Here, it is used to balance the three objectives of "conflict resolution, urgency priority, and load balancing," avoiding sacrificing one for another.

[0286] Conflict resolution efficiency: This measures the effectiveness of controlling expert time and resource conflicts; fewer conflicts mean higher efficiency. It is calculated as "Conflict rate = (Number of conflicts ÷ Total number of scheduling attempts) × 100%". Data comes from real-time scheduling records (such as event logs of overlapping expert times and consultation room occupancy conflicts) and the expert schedule (HIS system scheduling data) and resource status table (consultation room occupancy records) integrated in step S410. For example, if there are 2 conflicts out of 12 scheduling attempts, the conflict rate is approximately 16.7%.

[0287] Emergency Patient Priority: This measures whether patients with a high urgency index (generated in step S420, 0-1) are given priority in scheduling. It is calculated as "Priority Matching Degree = (Number of patients in the top N% of urgency and ranked in the top N% of time slots ÷ Total number of patients in the top N% of urgency) × 100%". Data comes from the urgency index file (JSON format) output in step S420 and the timestamp list generated by the scheduling system (the actual patient scheduling order). For example, if N=20%, one of two high-urgency patients is prioritized, resulting in a matching degree of 50%.

[0288] Expert workload balance: measures the fairness of the allocation of consultation workload among experts, avoiding some experts being overworked while others are idle. Expert workload balance data: based on the expert schedules (available consultation time slots) and the number of scheduled consultations integrated in step S410, it is calculated by statistically analyzing the actual number of consultations each expert receives (e.g., using the Gini coefficient to assess allocation differences, data from expert schedules and real-time scheduling records).

[0289] The overall processing steps are as follows:

[0290] 1. Construct the solution space tree:

[0291] Centered on "patient consultation arrangements," and combining the patient availability window (from patient confirmation records and inpatient / outpatient schedule system), expert available time slots (from hospital scheduling system and expert personal schedule), consultation room occupancy status (from resource management system), and patient urgency index generated in step S410, a tree structure is constructed:

[0292] Hierarchical sorting: The tree is divided into levels according to the "patient urgency index from high to low" (e.g., patients with an urgency index of 0.9 are treated first, followed by patients with an urgency index of 0.6), reducing subsequent adjustment costs;

[0293] Node content: Each node represents a possible arrangement for a patient (including the matched expert, specific time, and consultation room used), and only nodes that comply with medical standards are generated (e.g., the node time for Level I patients must be within 4 hours), excluding invalid options such as "expert overload" and "resources already occupied";

[0294] Hierarchical relationship: The upper-level node is "patients who have been scheduled", and the lower-level node is "patients who are to be scheduled". The hierarchy expands to include all possible combinations that meet the constraints.

[0295] 2. Backtracking search traversal:

[0296] The solution space tree is traversed using a depth-first strategy, with real-time verification of the constraints from step S410 during the process.

[0297] Initialization traversal: Starting with the first patient with the highest urgency index, based on S410 data (e.g., specialist A is available Wednesday 9:00-11:00, MDT1 room is available Wednesday morning), generate all feasible scheduling options for the patient and add them to the traversal queue;

[0298] Expanding layer by layer: Select one feasible node to enter the next layer. When generating a new node for the next patient, it is necessary to exclude "expert time slots and consultation rooms already occupied by previous patients" (e.g., if a previous patient used expert A's Wednesday 9:00-10:00, the new node will not select that time slot).

[0299] Backtracking trigger: If a conflict is found in the current arrangement at a certain level (such as the selected expert's time slot being occupied by a previous patient, or the consultation room being occupied), or if medical standards are violated (such as the arrangement of a Level I patient exceeding the 4-hour time limit), then immediately "backtrack" to the previous level, abandon the current selection, and reselect for the previous patient from the remaining feasible options.

[0300] 3. Heuristic function-based scoring guidance:

[0301] For each patient's appointment completed (i.e., reaching a node in the solution space tree), the heuristic function, based on the Pareto optimality principle, comprehensively scores the current plan by combining data from previous steps:

[0302] Conflict resolution efficiency score (weight 40%): The data comes from real-time scheduling records (number of conflicts) and S410 data. The lower the proportion of the number of conflicts to the total number of attempts, the higher the score (e.g., if there is only 1 conflict in 10 attempts, the score is close to 1).

[0303] Emergency Patient Priority Score (Weight 40%): Data is derived from the patient urgency index and current scheduling order in S420. Patients with higher urgency indices are scheduled earlier and have higher scores (e.g., a patient with an urgency index of 0.9 is scheduled first and has a score close to 1).

[0304] Expert workload balance score (weight 20%): Data comes from the expert schedule and real-time arrangement records of S410. The difference in the number of consultations by experts is assessed by the Gini coefficient. The smaller the difference (the lower the Gini coefficient), the higher the score (if all experts have similar number of consultations, the score is close to 1).

[0305] The overall score is calculated as follows: "Conflict resolution score × 0.4 + Emergency priority score × 0.4 + Load balancing score × 0.2".

[0306] 4. Optimize search direction:

[0307] Based on a comprehensive score using heuristic functions, the search path is dynamically adjusted to ensure efficient focus on high-quality solutions.

[0308] Pathway ranking: Pathways to be explored are ranked from high to low according to their comprehensive scores. High-scoring paths (such as those with a comprehensive score of 0.8 or above) that have "few conflicts, priority for emergency patients, and balanced expert workload" are given priority for further exploration.

[0309] Path pruning: If the score of a certain path continues to decrease (e.g., drops below 0.5), or triggers backtracking conditions (e.g., an irreconcilable conflict occurs), the path is terminated early and other better paths are switched to.

[0310] Solution generation: Continuously iterate through all patients to complete the arrangements, generate a complete solution, and then verify again whether it meets all constraints of S410 (such as the scope of expertise of experts and the time requirements of patients) to ensure that the solution is feasible.

[0311] Step S440: Based on the optimized scheduling results, directly generate the consultation time schedule.

[0312] The specific process is as follows:

[0313] Organize the scheduling plan: Based on the optimized scheduling plan in step S430, organize the matching relationships between patients, specialists, and time into a clear consultation schedule list according to chronological order. For example, schedule patients with high urgency levels first, then schedule other patients in sequence to avoid scheduling chaos.

[0314] Clarify consultation information: For each consultation arrangement in the list, specify the patient's specific consultation time (accurate to the minute, such as 10:30-11:00 on June 5, 2025), the list of experts participating in the consultation (marking the attending expert and collaborating experts), and the consultation resources used (such as the designated MDT consultation room, etc.).

[0315] Inspection and Correction: Conduct a comprehensive review of the generated consultation schedule, focusing on whether there are any issues such as experts being scheduled for multiple consultations at the same time, or resource conflicts such as consultation room occupation. Once a conflict is found, immediately adjust the relevant arrangements, re-matching time, experts, and resources to ensure that all consultation arrangements are reasonable and feasible.

[0316] The output is as follows:

[0317] Patient notification: Send patients a notification containing detailed information such as consultation time, location, and names and titles of participating experts via SMS, APP push, etc. For example: "Dear patient, your multidisciplinary consultation will be held at 10:30 on June 5, 2025 in MDT Consultation Room 2 on the 5th floor of Building 3 of the hospital. The attending expert is Chief Physician Wang from the Department of Oncology, and the participating experts are Associate Chief Physician Li from the Department of Respiratory Medicine and Attending Physician Zhao from the Department of Radiology. Please arrive 10 minutes in advance."

[0318] Expert schedule updates: Automatically updates experts' electronic schedules, marking patient information and consultation types for corresponding time periods, allowing experts to understand consultation details in advance and prepare accordingly.

[0319] A multidisciplinary collaborative diagnosis and treatment method also includes a step following the direct generation of consultation schedules based on the optimized scheduling results, as follows:

[0320] Step S450: Establish a real-time monitoring mechanism to continuously monitor the check-in status of experts and patients before the consultation begins.

[0321] The system includes a real-time monitoring mechanism: an intelligent monitoring system built on IoT, big data, and cloud computing technologies. Through real-time collection and analysis of multi-source data, it dynamically tracks consultation-related information, ensuring the smooth progress of the consultation process. The system also tracks attendance: accurately recording the time points of experts and patients, covering various statuses such as on-time attendance, late arrival, and no attendance, and presenting this information intuitively through color-coding (e.g., green for attended, yellow for late, and red for no attendance).

[0322] Furthermore, after the system collects expert / patient sign-in data in real time, it automatically generates a "Dynamic Sign-in Report," which includes a "List of Latecomers (Name + Department + Duration of Lateness)" and "Details of Absentees (Remarks on Reasons for Not Signing In, such as Temporary Cessation of Service / Non-Response)," and synchronizes it to the quality control platform dashboard. The sign-in rate (actual number of attendees / expected number of attendees) is included in the data source for calculating indicators such as "surgeon participation rate in preoperative discussions" (e.g., preoperative discussion sign-in data is directly linked to surgeon participation rate statistics).

[0323] The specific process is as follows:

[0324] Data collection:

[0325] For experts: Experts can complete the check-in process by scanning a QR code through the check-in module in the hospital's OA system. If an expert fails to check in within 10 minutes of the scheduled time, they will be automatically marked as "late".

[0326] For patients: There is a dedicated "Consultation Check-in" entry on the hospital's APP, which can be clicked to complete the check-in; patients at the hospital can complete the check-in through the self-service check-in machine in the outpatient hall or consultation area.

[0327] Status Tracking: Based on the consultation start time, the system automatically refreshes the check-in status of experts and patients every 3 minutes. If an expert has not checked in 5 minutes after the consultation starts, the system will call them via a specific mobile app to remind them; if a patient has not checked in 30 minutes before the consultation starts, the app will send a reminder notification.

[0328] For patient data, personalized thresholds are set based on different disease types and individual patient conditions. For example, for heart disease patients, in addition to routine heart rate and blood pressure thresholds, monitoring thresholds are also set for indicators such as ST segment changes on electrocardiograms. Once the data triggers a threshold, the system immediately pops up a red warning box on the monitoring interface, marking the abnormal data item and the degree of exceedance, while simultaneously generating a record of the change in the patient's condition.

[0329] The output results are as follows: The system continuously generates a structured real-time information table, which includes information such as expert name, title, check-in status, and actual check-in time; patient name, ID number, and check-in status.

[0330] Furthermore, the participation rate of the aforementioned surgeons in the preoperative discussion will be calculated in real time based on the check-in data of step S450 + the preoperative discussion record (associated with the surgeon ID) according to the formula "number of surgeons participating in the preoperative discussion / total number of surgeons who should participate", and updated to the quality control platform daily.

[0331] In step S460, if it is detected that the expert temporarily adjusts the consultation status, the dynamic rearrangement mechanism based on the genetic algorithm is immediately triggered to quickly generate alternative consultation plans.

[0332] Among them, genetic algorithm: a heuristic search algorithm that simulates the biological evolution process, and iteratively optimizes the solution through selection, crossover and mutation operations.

[0333] Dynamic rescheduling mechanism: An emergency response mechanism that reschedules consultation times and personnel arrangements in the event of unforeseen circumstances such as expert cancellations.

[0334] Alternative consultation plan: A new consultation arrangement generated based on the original plan, with adjustments to factors such as experts and time, to ensure the smooth conduct of the consultation.

[0335] The specific process is as follows:

[0336] Abnormal Trigger: After step S450 detects that the expert has stopped seeing patients, the system immediately starts the dynamic rescheduling mechanism and extracts key data such as patient information, consultation time, and resource requirements from the original consultation plan.

[0337] Algorithm Execution: Population Initialization: Generate multiple random consultation times and expert combinations as initial plans. Fitness Assessment: Calculate the fitness score for each plan based on indicators such as patient urgency, remaining expert matching degree, and resource availability. Evolutionary Operations: Generate new plans through selection (retaining high-scoring plans), crossover (integrating the advantages of some plans), and mutation (randomly adjusting local arrangements). Iterative Optimization: Repeat the above operations until the plan that satisfies the constraints and has the highest fitness is found. Plan Generation: After determining the optimal plan, generate alternative consultation plans that include new consultation times, participating experts, and resources used.

[0338] Output: Alternative consultation plans are pushed to the consultation coordinator and the patient. The coordinator can manually confirm or make minor adjustments. After receiving the notification, the patient understands the latest consultation arrangements, ensuring that the consultation process can quickly return to normal.

[0339] In step S470, if the patient's vital signs and other quantitative indicators show that the condition is deteriorating, the patient's consultation priority is increased in real time. For patients with increased priority, the system automatically puts them at the front of the consultation queue and notifies the relevant experts in a timely manner.

[0340] Among them, quantitative indicators refer to measurable parameters of the patient's condition, such as vital signs (e.g., heart rate, blood pressure, blood oxygen saturation) and laboratory data (e.g., complete blood count, biochemical values), obtained through medical equipment monitoring or testing. Consultation priority is the order of consultations set according to the urgency of the patient's condition; higher priority patients are scheduled for consultation first. A queue-jumping mechanism automatically moves the patient to a more appropriate position in the consultation queue when their condition worsens and their priority is increased.

[0341] The specific process is as follows:

[0342] Determination of deterioration of condition: The system continuously monitors the patient's quantitative indicators collected in step S450. If the data exceeds the preset threshold (such as heart rate > 130 beats / min, blood oxygen saturation < 90%), or if the test results deteriorate sharply (such as serum creatinine increasing by 50% within 24 hours), the patient's condition is determined to have deteriorated.

[0343] Priority upgrade: Immediately raise the patient's consultation priority to the highest level and recalculate its ranking position in the consultation queue.

[0344] Queue Adjustment: Automatically moves other patients with lower priority in the consultation queue to the back and inserts patients with worsening conditions to the front of the queue.

[0345] Notify experts: Immediately send notifications of changes in patient priority and earlier consultations to the originally scheduled experts and relevant department experts via SMS, hospital internal communication systems, etc., including the patient's latest condition data and emergency treatment suggestions.

[0346] Furthermore, a new "death case marking" function has been added to the system (doctors can mark cases when initiating consultations). If a case is marked as a "death case", the system will automatically trigger a 5-day countdown monitoring. Starting from the date the case is confirmed as dead, the system will count the percentage of cases that have been discussed within 5 days. If the discussion is not completed within the time limit, a red warning will be issued on the quality control platform.

[0347] In addition, regarding the completion rate of preoperative multidisciplinary discussion for Level IV surgery, when receiving patient data in step S100, the "Level IV surgery" tag is matched through the knowledge graph to automatically associate the preoperative discussion requirements and calculate "the number of Level IV surgery cases that completed MDT discussion / the total number of Level IV surgery cases". The data comes from the expert team combination record (including preoperative discussion identifier) ​​in step S200.

[0348] To further facilitate data management and control, a periodic quality control analysis step has been added to the end of the existing process: daily / weekly automatic data aggregation to generate three types of reports: "Difficult / Death Case Discussion List Report": This report integrates consultation records marked as "Difficult / Death" (including discussion time, participating experts, and conclusions), and links them to indicators such as "Department Head Chair Rate"; "General Consultation Report": This report statistically analyzes general consultation volume, on-time completion rate, cancellation rate, etc., with data sourced from the scheduling records of step S400 and the dynamic rescheduling records of S460; "MDT Consultation Report": This report categorizes and statistically analyzes consultations by "joint outpatient type" (e.g., oncology MDT / vertigo MDT), including average consultation duration, expert matching degree, etc., with data taken from the records of steps S200-S300. All reports are displayed through a visual interface (e.g., line charts / pie charts), supporting export by the quality control department or online viewing.

[0349] In addition, the system accumulates the number of consultations for each disease in real time (based on the disease type extracted in step S100), and generates a "Heat Map of Consultation Disease Distribution" (e.g., "Lung Cancer MDT accounts for 25%, Stroke MDT accounts for 18%)" through clustering algorithm, which serves as a reference for the allocation of expert resources (e.g., prioritizing expert scheduling for high-frequency diseases). At the same time, this distribution data is regularly synchronized to the quality control platform to support "Disease Structure Rationality Analysis".

[0350] Based on the same inventive concept, embodiments of the present invention provide a multidisciplinary collaborative diagnosis and treatment system, including a memory and a processor, wherein the memory stores information that can run on the processor to implement, as described above. Figures 1 to 2 The procedure for any method.

[0351] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A multidisciplinary collaborative diagnosis and treatment method, characterized in that, include: Multimodal data of patients is collected, and feature extraction and alignment of the collected data are performed through a multimodal fusion algorithm to generate a disease feature vector with a unified dimension. Based on the constructed medical knowledge graph, the graph query algorithm is used to retrieve nodes related to patient input and generate guidance questions. A reinforcement learning-knowledge graph joint model is adopted to dynamically adjust the priority of guidance questions based on historical guidance effect data. The system uses a conditional random field model to perform integrity checks on the input text, automatically labels missing entities and provides examples, uses a bidirectional LSTM classifier for consistency checks, detects and asks follow-up questions on contradictory information, calls an entity linking algorithm to match standard terms for ambiguous descriptions, and provides options for patients to choose from. Background feature data is extracted from patient registration information to construct patient profile vectors. Input examples of similar cases are retrieved based on collaborative filtering algorithm to guide patients to supplement information. When the patient's input information passes the integrity and consistency checks, and standard terms are selected for fuzzy descriptions, the input is considered complete. The system receives patient medical records submitted through the joint outpatient portal, performs preliminary analysis using natural language processing algorithms, extracts key information, and automatically recommends joint outpatient types by matching the hospital's disease knowledge base with knowledge graph technology. Based on the recommended joint outpatient clinic type, and according to the complexity of the patient's condition, historical consultation data, and the area of ​​expertise of the experts, the system intelligently recommends the range of consultation fees and expert team combinations and pushes them to the patient's end. Based on preset matching standards, patient conditions, and treatment guidelines, a consultation plan is generated. After the system verifies compliance, the payment process is triggered, and the consultation resources are locked once the payment is successful. Based on the urgency of the patient's condition, the expert's scheduling information, and the hospital's resource utilization, the system intelligently generates a consultation schedule and sends consultation invitations and detailed information to relevant experts. At the start of the consultation, speech recognition algorithms are used to assist experts in real-time annotation and analysis. At the same time, the system automatically records the speech discussions during the consultation and converts them into text records. After the consultation, the system organizes the written records, integrates expert opinions, generates a structured consultation report, and pushes it to the patient.

2. The multidisciplinary joint diagnosis and treatment method according to claim 1, characterized in that, Based on the constructed medical knowledge graph, a graph query algorithm is used to retrieve nodes related to patient input, generate guidance questions, and a reinforcement learning-knowledge graph joint model is adopted to dynamically adjust the priority of guidance questions based on historical guidance effect data, including: We use a Transformer-based model to encode the nodes and edges of a medical knowledge graph to capture deep semantic information. Model the context of patient input, and combine LSTM or Transformer models to understand its meaning in different contexts and accurately match relevant knowledge points; Calculate the semantic similarity between patient input and knowledge graph nodes to determine the most relevant knowledge points; By combining semantic similarity and contextual analysis results, a suitable guidance question template is matched from a predefined guidance question template library; Based on the patient's profile information, the matched guidance question templates are personalized. By using a reinforcement learning-knowledge graph joint model, the generated guidance questions are prioritized based on historical guidance effect data.

3. The multidisciplinary joint diagnosis and treatment method according to claim 1, characterized in that, This also includes a parallel step to receiving patient medical records submitted through the joint outpatient portal, as follows: When a doctor initiates a consultation, the system retrieves medical records from the patient's HIS and supplements the information with additional data. The system automatically parses the structured fields annotated by doctors and combines them with the electronic medical records retrieved from HIS to generate standardized disease feature vectors. The structured fields include consultation type and key concerns. Based on the consultation type and urgency level marked by the doctor, the corresponding specialty MDT template is retrieved from the knowledge graph, the type of joint outpatient clinic is automatically matched and recommended, and the user is redirected to the step of intelligently recommending the consultation fee range and expert team combination, and then the subsequent steps are executed.

4. The multidisciplinary joint diagnosis and treatment method according to claim 1, characterized in that, The types of joint outpatient clinics automatically recommended include: Receive multimodal medical records, including voice, text, and images, submitted by patients through the joint outpatient portal; Using a multimodal fusion algorithm based on the Transformer architecture, feature extraction and cross-modal alignment are performed on data from different modalities to generate disease feature vectors of a unified dimension; By inputting the disease feature vector into the medical knowledge graph, the disease knowledge base is retrieved through the graph query algorithm, and the disease nodes and related departments most relevant to the patient's condition are matched to automatically recommend joint outpatient clinic types.

5. The multidisciplinary joint diagnosis and treatment method according to claim 4, characterized in that, After inputting the disease condition feature vector into the medical knowledge graph, the graph query algorithm retrieves the disease knowledge base, matches the disease nodes and related departments most relevant to the patient's condition, and automatically recommends joint outpatient clinic types, the following steps are still required: The recommendation information is visualized using natural language generation technology and pushed to the patient in the form of interactive cards, with the basis for the recommendation annotated simultaneously, allowing the patient to confirm with one click or initiate a type change request.

6. A multidisciplinary collaborative diagnosis and treatment method according to any one of claims 1 to 5, characterized in that, Based on the recommended type of joint outpatient visit, and considering the complexity of the patient's condition, historical consultation data, and the areas of expertise of the specialists, the system intelligently recommends consultation fee ranges and expert team combinations, and pushes these recommendations to the patient's end, including: Construct a medical knowledge graph that includes expert expertise, collaboration history, and successful cases; Using a graph attention network algorithm, the matching degree between experts and patients' conditions is calculated based on the complexity and type of the patient's condition. Experts with high matching degree are selected, and the collaborative relationship between experts is analyzed to form a complementary collaborative team. By combining historical consultation data, a cost prediction model was trained using a multi-task learning framework. Using a trained cost prediction model, the basic cost range for a consultation is predicted based on the patient's condition and the selected expert team combination. The recommended expert team combinations and consultation fee ranges are integrated into structured information and pushed to the patient's end, allowing the patient to view the basis for the recommendations and the details of the fees.

7. The multidisciplinary joint diagnosis and treatment method according to claim 6, characterized in that, The intelligent scheduling system generates a consultation timetable, including: Collect and integrate patient pre-assessment levels, expert schedules, hospital resource status, and medical standard constraints; A fuzzy logic algorithm is introduced to convert the urgency of the illness into an urgency index of 0-1; By combining backtracking search algorithm with heuristic function optimization, the heuristic function is based on the Pareto optimality principle and simultaneously optimizes conflict resolution efficiency, priority of emergency patients, and expert workload balance. Based on the optimized scheduling results, a consultation schedule is directly generated.

8. The multidisciplinary joint diagnosis and treatment method according to claim 7, characterized in that: It also includes the steps following the direct generation of the consultation schedule based on the optimized scheduling results, as detailed below: Establish a real-time monitoring mechanism to continuously monitor the sign-in status of experts and patients before the consultation begins; If it is detected that an expert has temporarily changed the consultation status, a dynamic rearrangement mechanism based on a genetic algorithm will be immediately triggered to quickly generate alternative consultation plans and push them to the patient.

9. A multidisciplinary collaborative diagnosis and treatment system, characterized in that, It includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements a multidisciplinary joint diagnosis and treatment method as described in any one of claims 1 to 8.

Images