Regional collaborative multi-disciplinary integrated rehabilitation management method and system for tumor patients

Abstract

The application discloses a regional collaborative multidisciplinary fusion rehabilitation management method and system for tumor patients, and relates to the technical field of medical rehabilitation management. The method comprises the following steps: constructing a tumor rehabilitation management knowledge graph based on a clinical path and a diagnosis and treatment specification combed by a multidisciplinary MDT expert team; constructing a patient portrait vector comprising a disease dimension, a treatment dimension, a constitution dimension and a risk dimension; generating a patient personalized intervention path by using a large language model according to the patient portrait vector and the tumor rehabilitation management knowledge graph and verifying and confirming the patient personalized intervention path; automatically triggering a node process at a specified time according to a node in the intervention path, wherein the node process comprises multi-dimensional rehabilitation information pushing, diagnosis and treatment process reminding and follow-up interaction; continuously collecting patient data and automatically starting intervention path adjustment when a trigger condition is detected; and jointly performing regional collaborative management with grassroots and home. The application constructs a new mode of whole-course closed-loop management and realizes a leap from general management to personalized service.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation management technology, and in particular to a regional collaborative multidisciplinary rehabilitation management method and system for cancer patients. Background Technology

[0002] Malignant tumors are chronic diseases with complex symptoms, a high risk of recurrence and metastasis, often requiring multidisciplinary comprehensive treatment with multiple courses and long treatment cycles. Healthcare needs are not confined to hospital settings. Outside of hospitals, cancer patients lack effective channels for communication with doctors. After discharge, patients frequently encounter problems such as post-operative recovery issues, adverse drug reactions after chemotherapy, and questions about how to consult a doctor about sudden changes in their condition. They also lack follow-up care and timely, detailed interpretation of disease monitoring data. These issues are difficult to address promptly and conveniently, easily causing patient anxiety and hindering early detection and intervention of changes in the patient's condition.

[0003] Cancer patients face a long recovery period, requiring long-term medication, a balanced diet, regular follow-up visits, and exercise rehabilitation. Due to a lack of systematic understanding of the disease and professional guidance, or the influence of unhealthy lifestyle habits, patients often exhibit low compliance and self-management levels after discharge. This can lead to poor treatment outcomes, relapses, and complications. Hospitals often lack effective management measures, resulting in the inability to monitor and track patients' health status in a timely manner, hindering post-treatment recovery and preventing prompt intervention in risky situations. Furthermore, a lack of effective doctor-patient communication channels makes it difficult to understand patients' health needs and changes in their condition, potentially leading to decreased patient satisfaction and selection to transfer to another hospital or discontinue treatment.

[0004] Current cancer treatment emphasizes multidisciplinary team (MDT) approaches, but management suffers from disconnects. Patients receive isolated treatment and rehabilitation services across different departments and hospitals (especially between different levels of hospitals), resulting in significant information barriers and a break in the "treatment-rehabilitation-follow-up" chain. This not only hinders the effective implementation and continuation of advanced MDT treatment plans outside the hospital but also leads to a waste of medical resources and an increased burden on patients. Standardized and personalized treatment plans developed by hospitals cannot be effectively synchronized to primary care settings or dynamically adjusted based on real-time data from patients at home, significantly diminishing the effectiveness of multidisciplinary treatment in the outpatient management phase. Furthermore, there are issues such as uneven distribution of medical resources, hospitals operating independently, weak primary care capabilities, a lack of regional collaboration on standardized treatment plans, and patients' inability to access consistent healthcare services.

[0005] Therefore, strengthening post-diagnosis management of cancer patients is of great significance for improving the quality of medical services, enhancing patient compliance with post-diagnosis treatment, and reducing patient attrition. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a regional collaborative multidisciplinary rehabilitation management method and system for cancer patients. By constructing a multidisciplinary knowledge engine and a regional collaborative management mechanism, the plan formulated by the MDT is transformed into a personalized, executable, traceable and adjustable full-process management path, thus constructing a new model of closed-loop management throughout the entire disease course and realizing a leap from general management to personalized services.

[0007] In a first aspect, the present invention provides a regional collaborative multidisciplinary integrated rehabilitation management method for cancer patients, comprising: The process of constructing a knowledge graph for tumor rehabilitation management: Based on the clinical pathways and treatment guidelines sorted out by a multidisciplinary team of experts (MDT), a knowledge graph for tumor rehabilitation management is constructed. The knowledge graph for tumor rehabilitation management adopts a three-layer architecture: schema layer, instance layer and rule layer. The schema layer is used to define entity types, the instance layer includes specific medical knowledge instances, and the rule layer contains clinical decision-making rules. The process of constructing a multidimensional patient profile involves extracting data from the patient's electronic medical record system and constructing a patient profile vector that includes disease, treatment, physical condition, and risk dimensions. The process of generating personalized intervention paths for patients is as follows: Patient profile vectors are used as query conditions, and subgraph matching and reasoning are performed in the knowledge graph of tumor rehabilitation management to obtain a set of candidate intervention measures; Patient profile vectors and the set of candidate intervention measures are used as input to a large language model to generate personalized intervention paths for patients, which are then submitted to doctors for review and confirmation after being verified by multidisciplinary constraints. Rehabilitation management process: Based on the nodes in the intervention path, the node process is automatically triggered at specified times, including multi-dimensional rehabilitation information push, treatment process reminders, and follow-up interaction; when performing follow-up interaction, intelligent interaction is carried out by analyzing the user's voice information through speech recognition and semantic understanding, extracting and automatically summarizing key slot information, and generating a structured follow-up report by combining intent classification results; based on a large language model, the risk of the patient's questions is judged, low-risk questions are automatically answered by the patient, and high-risk questions are automatically transferred to the health management team or doctor for reply; The dynamic adjustment process of the intervention pathway: Patient data is continuously collected, and the intervention pathway adjustment is automatically initiated when a trigger condition is detected. The intervention pathway adjustment adopts a triple mechanism of large language model generation + multidisciplinary constraint verification + doctor review. Regional collaborative management process: The multidisciplinary team of experts from tertiary hospitals is responsible for developing standardized management plan templates, training and updating maps and models, and handling high-risk warnings; primary hospitals / communities receive the management plan and conduct daily data collection and low-risk consultation; the full-dimensional data generated by patients at the primary level / at home is transmitted back to the tertiary hospital platform in real time to form a data closed loop.

[0008] Furthermore, the construction of the tumor rehabilitation management knowledge graph adopts a human-computer coupling approach. First, a large language model is used to automatically extract triples from clinical guidelines. Then, a multidisciplinary team of experts (MDT) reviews and supplements the data. Finally, knowledge reasoning is used to complete the missing relationships.

[0009] Furthermore, the large language model is fine-tuned in stages: In the first stage, based on the already pre-trained general medical industry model, further pre-training is conducted using corpora including desensitized medical record texts, clinical guidelines, and expert consensus to adapt the model to the professional context of cancer patient management; In the second stage, instruction data with three complexity levels is designed for different patient management tasks, including low-complexity instructions of single-turn question-and-answer type, medium-complexity instructions of multi-step reasoning type, and high-complexity instructions of multi-turn dialogue type. The high-complexity instructions embed "follow-up question triggering conditions," and the model actively asks follow-up questions when the patient's answer contains specific key information; In the third stage, based on reinforcement learning alignment with expert feedback, a multi-evaluation dimension expert reward model is constructed, and the algorithm gradually aligns the model's decisions with expert knowledge.

[0010] Furthermore, the personalized patient intervention pathway undergoes multidisciplinary constraint verification. After verification, it is submitted to the physician for review and confirmation. The multidisciplinary constraint verification includes: Knowledge graph consistency check: Check whether the intervention measures conflict with the contraindication rules of each specialty; Timing rationality check: Check whether the time arrangement conforms to the timing constraints of interdisciplinary clinical treatment; Integrity check: Checks whether all management dimensions are covered.

[0011] Furthermore, in the rehabilitation management process, follow-up interactions are performed based on intent recognition and knowledge graph constraints, including: Step 1, Speech Recognition and Preprocessing: An end-to-end speech recognition model is used to convert the patient's speech into text, and domain adaptation is performed using hospital follow-up recording data based on a general ASR model; Step 2, Intent Recognition and Key Slot Extraction: Design an intent-slot system for cancer patient follow-up scenarios; Intent category definition: Symptom feedback: Report new symptoms, describe changes in symptoms, inquire about the causes of symptoms; Medication-related categories: Inquiring about medication administration methods, reporting adverse reactions, and consulting about drug interactions; Follow-up related matters: Inquire about the follow-up time and report the examination results; Life guidance services: dietary advice, exercise advice, and psychological support. Emergency situations: Reporting acute symptoms or requesting urgent medical attention; Key slot definition: Symptom slot: Symptom name, onset time, duration, severity, accompanying symptoms; Drug compartment: Drug name, dosage, frequency, and description of adverse reactions; Time slot: time of symptom onset, duration, and time of follow-up examination; Severity gauge: pain score, degree of impact on daily life; Key slot extraction employs an end-to-end approach based on a large language model: taking patient statements and predefined slot schemas as input, the large language model directly outputs structured slot filling results; Step 3: Based on knowledge graph constraints, design a hybrid dialogue management strategy that combines "knowledge graph constraints + large model generation": Dialogue status tracking: Maintains filled slots, unconfirmed messages, and dialogue history summaries.

[0012] Follow-up decision mechanism: Based on the symptom-disease association in the knowledge graph, when a patient reports a certain symptom, the system automatically retrieves the information required for differential diagnosis associated with that symptom and determines whether the collected information is sufficient for risk assessment; if insufficient, it generates targeted follow-up questions. Risk classification and response strategy: Assess the risk of the problem and implement the corresponding response strategy; Low-risk issues: Large language models directly generate responses, combined with personalized expressions based on patient profiles; Medium-risk issues: Generate responses and mark them to remind doctors to pay attention. High-risk questions: Do not answer directly; generate a transfer script and notify the doctor immediately. Step 4: Follow-up report is automatically generated. Key slot information is automatically summarized and combined with intent classification results to generate a structured follow-up report.

[0013] Furthermore, the large language model is combined with a hierarchical, structured multidisciplinary knowledge base and retrieval-enhanced generation; the first layer of the multidisciplinary knowledge base is a clinical experience knowledge base, with data sources including clinical guidelines, expert consensus, MDT discussion records, and typical cases from various disciplines; the second layer is a multidisciplinary medical knowledge graph, with data sources including ICD-10 codes, drug instructions, clinical pathway documents, and expert annotations; the third layer is a clinical knowledge base, with data sources including examination report templates, imaging description standards, and pathology report standards.

[0014] Secondly, the present invention provides a regional collaborative multidisciplinary integrated rehabilitation management system for cancer patients, comprising: The tumor rehabilitation management knowledge graph construction module is used to construct a tumor rehabilitation management knowledge graph based on clinical pathways and treatment guidelines compiled by a multidisciplinary MDT expert team. The tumor rehabilitation management knowledge graph adopts a three-layer architecture: schema layer, instance layer and rule layer. The schema layer is used to define entity types, the instance layer includes specific medical knowledge instances, and the rule layer contains clinical decision-making rules. The multidimensional patient profile building module is used to extract data from the patient's electronic medical record system and build a patient profile vector that includes disease dimension, treatment dimension, physical condition dimension and risk dimension; The patient-personalized intervention path generation module uses patient profile vectors as query conditions to perform subgraph matching and reasoning in the tumor rehabilitation management knowledge graph to obtain a set of candidate intervention measures; it uses patient profile vectors and the set of candidate intervention measures as input to a large language model to generate a patient-personalized intervention path, which is then submitted to the doctor for review and confirmation after being verified by multidisciplinary constraints. The rehabilitation management module is used to automatically trigger node processes at specified times based on nodes in the intervention path, including multi-dimensional rehabilitation information push, treatment process reminders, and follow-up interactions. When performing follow-up interactions, it uses speech recognition and semantic understanding to analyze the user's voice information for intelligent interaction, extracts and automatically summarizes key slot information, and generates a structured follow-up report based on intent classification results. It judges the risk of patient questions based on a large language model, automatically answers patients with low-risk questions, and automatically forwards high-risk questions to the health management team or doctor for reply. The intervention pathway dynamic adjustment module is used to continuously collect patient data. When a trigger condition is detected, the intervention pathway adjustment is automatically initiated. The intervention pathway adjustment adopts a triple mechanism of large language model generation + multidisciplinary constraint verification + doctor review. The regional collaborative management module is used to develop standardized management plan templates, train and update maps and models, and handle high-risk warnings through multidisciplinary MDT expert teams from tertiary hospitals; primary hospitals / communities receive the management plan and conduct daily data collection and low-risk consultation processing; and the full-dimensional data generated by patients at the primary level / at home is transmitted back to the tertiary hospital platform in real time to form a data closed loop.

[0015] Furthermore, the large language model is fine-tuned in stages: In the first stage, based on the already pre-trained general medical industry model, further pre-training is conducted using corpora including desensitized medical record texts, clinical guidelines, and expert consensus to adapt the model to the professional context of cancer patient management; In the second stage, instruction data with three complexity levels is designed for different patient management tasks, including low-complexity instructions of single-turn question-and-answer type, medium-complexity instructions of multi-step reasoning type, and high-complexity instructions of multi-turn dialogue type. The high-complexity instructions embed "follow-up question triggering conditions," and the model actively asks follow-up questions when the patient's answer contains specific key information; In the third stage, based on reinforcement learning alignment with expert feedback, a multi-evaluation dimension expert reward model is constructed, and the algorithm gradually aligns the model's decisions with expert knowledge.

[0016] Furthermore, the personalized patient intervention pathway undergoes multidisciplinary constraint verification. After verification, it is submitted to the physician for review and confirmation. The multidisciplinary constraint verification includes: Knowledge graph consistency check: Check whether the intervention measures conflict with the contraindication rules of each specialty; Timing rationality check: Check whether the time arrangement conforms to the timing constraints of interdisciplinary clinical treatment; Integrity check: Checks whether all management dimensions are covered.

[0017] Furthermore, the large language model is combined with a hierarchical, structured multidisciplinary knowledge base and retrieval-enhanced generation; the first layer of the multidisciplinary knowledge base is a clinical experience knowledge base, with data sources including clinical guidelines, expert consensus, MDT discussion records, and typical cases from various disciplines; the second layer is a multidisciplinary medical knowledge graph, with data sources including ICD-10 codes, drug instructions, clinical pathway documents, and expert annotations; the third layer is a clinical knowledge base, with data sources including examination report templates, imaging description standards, and pathology report standards.

[0018] The technical solutions provided in the embodiments of the present invention have at least the following technical effects: 1. Extending and implementing the concept of multidisciplinary integrated diagnosis and treatment outside the hospital: This invention breaks the status quo that multidisciplinary diagnosis and treatment is limited to in-hospital decision-making. By constructing a knowledge engine that integrates multiple disciplines and a regional collaborative management mechanism, it transforms the personalized plans formulated by MDT into an executable, traceable, and adjustable full-process management path, ensuring that patients can continue to recover under the guidance of a unified and collaborative plan after leaving the hospital, and truly realizing the full coverage of the value of MDT.

[0019] 2. Empowering medical staff and building an intelligent multidisciplinary collaborative workflow: The system automatically processes a large number of routine consultation and follow-up tasks through an AI assistant, and makes preliminary judgments and triages based on a multidisciplinary knowledge system, freeing medical staff (especially experts from tertiary hospitals) from repetitive labor, so that they can focus more on high-risk, interdisciplinary and complex decision-making.

[0020] 3. Facilitating the construction of a patient-centered, integrated service ecosystem: By covering the entire process from outpatient to inpatient care to home care and follow-up visits, and injecting multidisciplinary collaborative intelligent management into each stage, the system ensures patient recovery and significantly improves patient compliance and satisfaction. Patients, regardless of their location, can experience continuous, consistent, and professional hospital services, enhancing their medical experience, strengthening patient loyalty, and promoting the effective implementation of hierarchical medical services.

[0021] 4. It provides a unified collaborative work platform for medical staff at different levels and in different disciplines, based on the same multidisciplinary knowledge foundation, thereby improving management efficiency and scientific research collaboration capabilities.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart illustrating the overall process of the method in Embodiment 1 of the present invention. Figure 3 This is a schematic diagram of the rehabilitation management business process in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the personalized intervention path for patients in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation

[0025] This application provides a regional collaborative multidisciplinary rehabilitation management method and system for cancer patients. By constructing a multidisciplinary knowledge engine and a regional collaborative management mechanism, the plan formulated by the MDT is transformed into a personalized, executable, traceable, and adjustable full-process management path, thus constructing a new model of closed-loop management throughout the entire disease course and realizing a leap from general management to personalized services.

[0026] The overall concept of the technical solutions in the embodiments of the present invention is as follows: By applying artificial intelligence (AI) technology, a cancer patient management platform is built to provide patients with personalized rehabilitation plans. From outpatient to inpatient to home care and follow-up visits, AI technology precisely reaches patients, improving their medical experience, ensuring their treatment and recovery, and enabling patients to track their progress at every stage. This primarily includes the following technological innovations: 1. Construction of a knowledge graph and multidimensional patient profile for tumor rehabilitation management Based on clinical pathways and treatment guidelines compiled by a multidisciplinary team of experts (MDT), a knowledge graph for tumor rehabilitation management is constructed. This knowledge graph adopts a three-layer architecture: a schema layer, an instance layer, and a rule layer. The schema layer defines entity types, the instance layer includes specific medical knowledge instances, and the rule layer contains clinical decision-making rules. Patient profile vectors are constructed, encompassing disease, treatment, physical condition, and risk dimensions.

[0027] 2. Based on the technology for generating large-scale models of refined patient management By leveraging artificial intelligence (AI) technology, combined with a knowledge graph for tumor rehabilitation management and multi-dimensional patient profiles, personalized patient management plans are generated, constructing a refined patient management platform that offers new possibilities for individualized medical services. Utilizing the latest AI technologies such as large language models, personalized intervention pathways are generated based on patients' medical records and physical examination information. These pathways include important medical reminders, medication guidance, dietary recommendations, exercise rehabilitation programs, and rehabilitation guidance. After physician review and confirmation, these plans are used to encourage patients to follow them more effectively via intelligent voice and mini-programs. During home-based rehabilitation, patients' physiological parameters can be monitored, and their health status can be assessed in real time based on the collected data. Potential risks are identified, and management plans are dynamically adjusted to better meet patients' needs. This series of measures aims to help patients better manage their health, improve management effectiveness, and thus enhance their quality of life and overall health.

[0028] 3. Development technology for instruction sets of large-scale personalized patient follow-up models By leveraging a pre-trained large-scale medical industry model, fine-tuning is performed in patient management scenarios to enhance its capabilities in such situations. Instruction annotation data adapted to each task is constructed. This instruction data requires designing instructions of varying complexity based on task characteristics and existing data to improve the accuracy of the large-scale medical industry model on the target task.

[0029] 4. Intelligent personalized health human-computer interaction technology based on large models Based on expert feedback learning and controllable generation technology, this research studies personalized health interaction technology, enabling a large model to dynamically interact with user profiles and proactively engage with them using medical professional pathways. By integrating expert experience and reinforcement learning, controllable response technology is implemented to improve doctors' work efficiency while meeting patients' personalized needs. An intelligent personalized health human-computer interaction model simulates doctor-resident interaction, achieving human-computer dialogue. First, speech information is analyzed through speech recognition and semantic understanding, and key slot information is extracted. Then, response content is generated based on the dialogue scenario and business knowledge, and communication with residents is conducted through speech synthesis. During follow-up, basic analysis and feedback are performed based on residents' symptoms and other information, and the records are synchronized with the doctor. After the follow-up, reports are generated through statistical analysis of patient intent and key slot information, providing doctors with more comprehensive data support.

[0030] Before introducing specific embodiments, the system framework corresponding to the method in the embodiments of this application will be introduced first, such as... Figure 1 As shown, the system roughly consists of a three-layer architecture: an application layer, a service layer, a capability layer, and a data layer. The application layer includes management, healthcare, and patient terminals; the service layer includes basic services, business services, and statistical analysis; the capability layer includes AI capabilities and platform channel capabilities; and the data layer includes business data and unified model data. Example 1

[0031] This embodiment provides a regional collaborative multidisciplinary integrated rehabilitation management method for cancer patients, such as... Figure 2 and Figure 3 As shown, it includes the following steps: S1. Construction Process of Cancer Rehabilitation Management Knowledge Graph: Based on the clinical pathways and treatment guidelines compiled by a multidisciplinary MDT expert team (covering oncology, surgery, radiation oncology, nutrition, psychology, rehabilitation, etc.), a cancer rehabilitation management knowledge graph is constructed. The cancer rehabilitation management knowledge graph adopts a three-layer architecture: Schema layer: Defines entity types (diseases, symptoms, drugs, examinations, interventions, etc.) and relationship types (indications, contraindications, temporal relationships, causal relationships, etc.); Instance layer: Contains specific medical knowledge examples, such as "post-breast cancer surgery → requires → endocrine therapy → continues → 5 years"; Rule layer: Contains clinical decision-making rules, such as "white blood cell count <3.0×10⁻⁶ after chemotherapy". 9 / L→Trigger→Whitening Treatment Reminder.

[0032] The system consists of a schema layer, an instance layer, and a rule layer. The schema layer defines entity types, the instance layer contains specific medical knowledge instances, and the rule layer contains clinical decision-making rules.

[0033] The knowledge graph for tumor rehabilitation management is constructed using a human-computer coupling approach. First, a large language model is used to automatically extract triples from clinical guidelines. Then, a multidisciplinary team of experts (MDT) reviews and supplements the data. Finally, knowledge reasoning is used to complete the missing relationships.

[0034] S2. Multidimensional Patient Profile Construction Process: Extract data from the patient's electronic medical record system to construct a patient profile vector that includes disease dimension, treatment dimension, physical condition dimension, and risk dimension.

[0035] Specifically, structured data (diagnosis codes, surgical procedure codes, pathological staging, gene testing results, comorbidities, etc.) and unstructured data (discharge summaries, outpatient medical record texts, examination reports, etc.) can be extracted from the patient's electronic medical record system. For the unstructured text, a medical named entity recognition model is used to extract key medical entities. The extracted entities are then fused with the structured data to construct a patient profile vector containing the following dimensions: Disease dimensions: tumor type, TNM stage; Treatment dimensions: surgical methods, chemotherapy regimens, radiotherapy regimens, targeted / immunotherapy regimens; Physical fitness dimensions: age, BMI, comorbidities, organ function scores; Risk dimensions: recurrence risk score, complication risk score, and psychological risk score.

[0036] The electronic medical record system with a full-process data fusion mechanism has established a unified patient index and data standard, realizing the structured integration and long-term archiving of outpatient medical records, inpatient records, home monitoring data and follow-up visit information, thus solving the problem of data fragmentation.

[0037] S3. Personalized Intervention Path Generation Process: Using patient profile vectors as query conditions, subgraph matching and reasoning are performed within the tumor rehabilitation management knowledge graph to obtain a set of candidate intervention measures. The patient profile vectors and the candidate intervention measure set are then used as input to a large language model to generate a personalized intervention path. After multidisciplinary constraint verification (such as conflict detection with contraindication rules of various specialties and verification of the rationality of interdisciplinary treatment sequence), the path is submitted to physicians for review and confirmation to ensure the safety and synergy of the plan. The personalized intervention path, tailored to different diseases and disease stages (acute phase, recovery phase, stable phase), constructs a dynamic intervention strategy that includes follow-up plans, health education, and critical value warnings, achieving a leap from general management to personalized services.

[0038] For example, input the Prompt template: "Patient Basic Information: [Portrait Summary]; Diagnosis: [Diagnostic Information]; Treatment Plan: [Treatment Performed]; Current Status: [Latest Examination Results]; Candidate Interventions: [Knowledge Graph Matching Results]. Please generate a personalized rehabilitation management path for this patient in the future [time period], including time nodes, intervention content, implementation methods, and expected goals." The generated path undergoes triple constraint verification: Knowledge graph consistency check: Check whether the intervention measures conflict with the contraindication rules; Timing rationality check: Check whether the time arrangement conforms to the timing constraints of the clinical pathway; Completeness check: Check whether all management dimensions (medication, follow-up visits, rehabilitation, education, etc.) are covered; After verification, submit it to the doctor for review and confirmation.

[0039] S4. Rehabilitation Management Process: Based on the nodes in the intervention pathway, the node process is automatically triggered at specified times, including multi-dimensional rehabilitation information push, treatment process reminders, and follow-up interactions (such as...). Figure 4 As shown, the node process types include: rehabilitation goals and expectations, key points of rehabilitation, pharmaceutical services, rehabilitation guidance, lifestyle intervention, follow-up, re-examination reminders, health knowledge, rehabilitation instructions, and key concerns. When performing follow-up interactions, intelligent interaction is achieved by analyzing the user's voice information through speech recognition and semantic understanding, extracting and automatically summarizing key slot information, and generating a structured follow-up report based on intent classification results. Based on a large language model, the risk of the patient's questions is judged, with low-risk questions automatically answered by the patient and high-risk questions automatically transferred to the health management team or doctor for reply.

[0040] Based on a patient's diagnosis, surgical procedure, complications, and laboratory test results, a personalized intervention path is generated. This allows for the delivery of discharge instructions and post-discharge precautions to patients nearing discharge via mini-programs and official accounts, reducing the workload of medical staff. For patients at home, intelligent push notifications of medication reminders, rehabilitation guidance, and health education support recovery. Smart hardware enables real-time transmission of patient vital signs data to doctors, facilitating remote intervention. Based on discharge orders, intelligent reminders for follow-up visits are generated, guiding patients to their outpatient appointments. From outpatient to inpatient to home (or primary care hospital) to follow-up visits, artificial intelligence technology precisely reaches patients, improving their healthcare experience, ensuring their recovery, and providing a traceable path for each stage of their journey.

[0041] During the rehabilitation management process, follow-up interactions are performed based on intent recognition and knowledge graph constraints, including: Step 1, Speech Recognition and Preprocessing: An end-to-end speech recognition model is used to convert the patient's speech into text, and domain adaptation is performed using hospital follow-up recording data based on a general ASR model; Step 2, Intent Recognition and Key Slot Extraction: Design an intent-slot system for cancer patient follow-up scenarios; Intent category definition: Symptom feedback: Report new symptoms, describe changes in symptoms, inquire about the causes of symptoms; Medication-related categories: Inquiring about medication administration methods, reporting adverse reactions, and consulting about drug interactions; Follow-up related matters: Inquire about the follow-up time and report the examination results; Life guidance services: dietary advice, exercise advice, and psychological support. Emergency situations: Reporting acute symptoms or requesting urgent medical attention; Key slot definition: Symptom slot: Symptom name, onset time, duration, severity, accompanying symptoms; Drug compartment: Drug name, dosage, frequency, and description of adverse reactions; Time slot: time of symptom onset, duration, and time of follow-up examination; Severity scale: Pain Rating Scale (NRS), degree of impact on life; Key slot extraction employs an end-to-end approach based on a large language model: taking patient statements and predefined slot schemas as input, the large language model directly outputs structured slot filling results.

[0042] Compared to traditional sequence labeling methods, this method can handle implicit information extraction (such as extracting "shortness of breath after activity" from "I've been out of breath lately") and slot inheritance in multi-turn dialogues.

[0043] Step 3: Based on knowledge graph constraints, design a hybrid dialogue management strategy that combines "knowledge graph constraints + large model generation": (1) Dialogue status tracking: Maintain filled slots, unconfirmed information, and dialogue history summary.

[0044] (2) Follow-up decision mechanism: Based on the symptom-disease association in the knowledge graph, when a patient reports a certain symptom, the system automatically retrieves the information required for differential diagnosis associated with that symptom and determines whether the collected information is sufficient for risk assessment; if insufficient, it generates targeted follow-up questions.

[0045] Example: Patient reports "I've been feeling dizzy lately" → Knowledge graph association: Dizziness may be related to anemia (after chemotherapy), brain metastases, orthostatic hypotension, etc. → System follow-up questions: "Under what circumstances does the dizziness occur? Is it when you stand up or is it constant? Is it accompanied by nausea or blurred vision?" (3) Risk classification and response strategy: Assess the risk of the problem and implement the corresponding response strategy; Low-risk issues: Large language models directly generate responses, combined with personalized expressions based on patient profiles; Medium-risk issues: Generate responses and mark them to remind doctors to pay attention. High-risk questions: Do not answer directly; generate a transfer script and notify the doctor immediately. Step 4: Follow-up report is automatically generated, which automatically summarizes key slot information and combines it with intent classification results to generate a structured follow-up report: summary of the patient's current symptoms, medication adherence assessment, risk warning items, and suggestions for doctors to pay attention to.

[0046] Intelligent follow-up is achieved through smart outbound calling and human-machine coupling, enabling real-time monitoring of patient recovery, fostering close relationships with patients, and precisely reaching them. This expands the service radius without increasing doctors' workload, improves doctor-patient relationships, enhances management capabilities, and helps departmental medical staff reduce patient consultation workload by 60%. Abnormalities detected during follow-up are promptly alerted. The comprehensive patient management platform establishes a bridge for doctor-patient communication, providing dynamic health guidance, resolving patients' questions about home rehabilitation in real time, and offering precise services. This effectively improves patients' adherence to medical advice, disease awareness, and self-health management awareness, enabling early detection, intervention, and treatment of post-diagnosis risks, thereby improving patients' overall health.

[0047] Through intelligent follow-up, follow-up plans are automatically created and executed, and manual intervention and adjustments can also be made. During the follow-up process, AI automatically records, transcribes and structures the follow-up content, and generates follow-up reports, greatly reducing data processing work. Intelligent vital sign monitoring allows real-time monitoring of patients' vital sign data, detection of risks, timely reminders, and avoidance of risks to patients' conditions.

[0048] S5. Dynamic Adjustment Process of Intervention Pathway: Patient data is continuously collected, and the intervention path adjustment is automatically initiated when a trigger condition is detected, such as: Abnormal vital signs: such as body temperature >38.5°C for 24 hours or blood pressure exceeding the safe range; Symptom changes: Patients report new symptoms or worsening of existing symptoms; Abnormal test results: The retested indicators exceeded the expected range.

[0049] Cross-scenario business triggering logic: An automatic state transition mechanism based on a rule engine was designed. For example, discharge events automatically trigger home follow-up tasks, and abnormal home data automatically triggers follow-up visit reminders, realizing "event-driven" closed-loop management.

[0050] The intervention pathway adjustment employs a triple mechanism: large language model generation, multidisciplinary constraint verification, and physician review. When patient data triggers pathway adjustment, the system can assess the impact of changes in a single indicator on the overall treatment plan based on multidisciplinary knowledge, generating a comprehensive adjustment suggestion for review by the multidisciplinary team, thereby maintaining multidisciplinary collaboration in outpatient management.

[0051] To make the large language model more suitable for patient management scenarios, the large language model was fine-tuned in stages: The first stage is domain-adaptive pre-training: Based on the large medical industry model that has already completed general pre-training, further pre-training is carried out using corpora including desensitized medical record texts, clinical guidelines, and expert consensus to adapt the model to the professional context of cancer patient management. The second phase involves fine-tuning multi-task instructions: For different tasks in patient management, instruction data at three complexity levels is designed, including: Low-complexity instructions (single-round question-and-answer type): Task types: medication reminder generation, dietary suggestion generation, exercise program generation; Command format: "Generate [specific task] based on the following patient information [X]"; Data source: Based on existing patient management records, with correct output annotated by clinical experts.

[0052] Medium complexity instructions (multi-step reasoning type): Task types: risk assessment, anomaly warning, path adjustment suggestions; Instruction format: "Based on patient information [X] and latest data [Y], analyze the [problem] and provide [suggestions]"; Data source: Based on real clinical cases, with expert annotations of reasoning processes and conclusions.

[0053] High-complexity instructions (multi-turn dialogue type): Task types: Intelligent follow-up dialogue, patient consultation response, symptom assessment Command format: Multi-turn dialogue history + current patient input → model response Special design: Embedded "follow-up question trigger conditions" so that the model actively asks follow-up questions when the patient's answer contains specific key information.

[0054] Instead of simply applying general instructions with minor adjustments, we design hierarchical instruction construction strategies. Based on the characteristics of the task and the existing data, we design instructions of different complexities to improve the accuracy of large medical models on target tasks.

[0055] The third stage involves reinforcement learning alignment based on expert feedback: constructing a multi-dimensional expert reward model, where evaluation dimensions may include: Medical accuracy: Does the answer conform to clinical guidelines and expert consensus? Personalization level: This refers to whether patient profile information is fully utilized; Safety: Whether high-risk issues are correctly identified and referred to the doctor; Naturalness of interaction: Whether the answers are easy to understand and have a warm feel.

[0056] The PPO algorithm enables the model's decisions to be gradually aligned with expert knowledge.

[0057] Ideally, to address the knowledge illusion problem in large language models, a method combining a hierarchical, structured multidisciplinary knowledge system with retrieval-enhanced generation (RAG) is proposed: A. Construction of a multi-level knowledge base First layer: Clinical experience knowledge base (unstructured) Data sources: clinical guidelines, expert consensus, MDT discussion records, and typical cases from various disciplines; Construction method: Use a large language model to semantically segment the document, maintain the semantic integrity of each chunk, and generate vector embeddings to be stored in a vector database; Update mechanism: Incremental updates are made when new guidelines are released or experts provide supplementary information.

[0058] Second layer: Multidisciplinary medical knowledge graph (structured) Data sources: ICD-10 codes, drug instructions, clinical pathway documents, expert annotations; Construction method: (a) Automatic extraction of triples by large language model; (b) Expert review, error correction and supplementation; (c) GNN link prediction to complete missing relations; (d) Conflict resolution according to the priority of "latest guidelines > expert consensus > textbooks".

[0059] Third layer: Clinical knowledge base (semi-structured) Data sources: Examination report templates, imaging description guidelines, pathology report standards; Construction method: Based on image and text parsing technology, documents containing tables, images, and flowcharts are parsed into a semi-structured format.

[0060] B. Search Enhancement Generation (RAG) Process (1) Intent analysis: Analyze the query intent and determine the type of knowledge to be retrieved. (2) Multi-path recall: vector retrieval + graph retrieval + keyword retrieval (3) Results fusion and sorting: deduplication, fusion, and sorting by relevance. (4) Knowledge injection generation: Top-K results are injected as context into the Prompt. (5) Factual verification: Verify the consistency between the generated results and the retrieved knowledge, and mark potentially illusory content.

[0061] By using RAG technology, the system ensures that the suggestions provided in any interaction (such as answering patient inquiries or generating follow-up questions) are anchored to authoritative, up-to-date, and multidisciplinary validated structured knowledge, thus guaranteeing the consistency and security of interdisciplinary suggestions.

[0062] Knowledge enhancement based on a multi-level knowledge base and retrieval enhancement effectively reduces knowledge illusion.

[0063] S6. Regional Collaborative Management Process: (1) Hierarchical management model: Tertiary hospitals: Multidisciplinary team (MDT) experts are responsible for developing standardized management program templates, training and updating atlases and models, and handling high-risk warnings; Primary care hospitals / communities: Receive and implement management plans developed by tertiary hospitals that incorporate multidisciplinary expertise, and conduct daily data collection and low-risk consultations to ensure that patients at the primary care level can receive the same multidisciplinary collaborative management services as those at tertiary hospitals; Patient side: Receive management instructions, provide symptom data, and manage themselves.

[0064] (2) Knowledge synchronization and data closed loop: The latest knowledge generated by multidisciplinary experts in tertiary hospitals (updated through maps and models) is dynamically synchronized to primary care facilities. Simultaneously, comprehensive data generated by patients at the primary care level or at home is transmitted back to the tertiary hospital platform in real time, forming a closed loop of "multidisciplinary treatment plan development → regional collaborative execution → comprehensive data transmission → dynamic multidisciplinary optimization." This extends multidisciplinary comprehensive treatment from a one-off hospital meeting to a dynamic management process that covers the entire patient lifecycle, spans multiple regions, and is continuously optimized.

[0065] Based on the same inventive concept, this application also provides a system corresponding to the method in Embodiment 1, as detailed in Embodiment 2. Example 2

[0066] This embodiment provides a regional collaborative multidisciplinary integrated rehabilitation management system for cancer patients, such as... Figure 5 As shown, it includes: The tumor rehabilitation management knowledge graph construction module is used to construct a tumor rehabilitation management knowledge graph based on clinical pathways and treatment guidelines compiled by a multidisciplinary MDT expert team. The tumor rehabilitation management knowledge graph adopts a three-layer architecture: schema layer, instance layer and rule layer. The schema layer is used to define entity types, the instance layer includes specific medical knowledge instances, and the rule layer contains clinical decision-making rules. The multidimensional patient profile building module is used to extract data from the patient's electronic medical record system and build a patient profile vector that includes disease dimension, treatment dimension, physical condition dimension and risk dimension; The patient-personalized intervention path generation module uses patient profile vectors as query conditions to perform subgraph matching and reasoning in the tumor rehabilitation management knowledge graph to obtain a set of candidate intervention measures; it uses patient profile vectors and the set of candidate intervention measures as input to a large language model to generate a patient-personalized intervention path, which is then submitted to the doctor for review and confirmation after being verified by multidisciplinary constraints. The rehabilitation management module is used to automatically trigger node processes at specified times based on nodes in the intervention path, including multi-dimensional rehabilitation information push, treatment process reminders, and follow-up interactions. When performing follow-up interactions, it uses speech recognition and semantic understanding to analyze the user's voice information for intelligent interaction, extracts and automatically summarizes key slot information, and generates a structured follow-up report based on intent classification results. It judges the risk of patient questions based on a large language model, automatically answers patients with low-risk questions, and automatically forwards high-risk questions to the health management team or doctor for reply. The intervention pathway dynamic adjustment module is used to continuously collect patient data. When a trigger condition is detected, the intervention pathway adjustment is automatically initiated. The intervention pathway adjustment adopts a triple mechanism of large language model generation + multidisciplinary constraint verification + doctor review. The regional collaborative management module is used to develop standardized management plan templates, train and update maps and models, and handle high-risk warnings through multidisciplinary MDT expert teams from tertiary hospitals; primary hospitals / communities receive the management plan and conduct daily data collection and low-risk consultation processing; and the full-dimensional data generated by patients at the primary level / at home is transmitted back to the tertiary hospital platform in real time to form a data closed loop.

[0067] Preferably, the large language model is fine-tuned in stages: In the first stage, based on the already pre-trained general medical industry model, further pre-training is performed using corpora including desensitized medical record texts, clinical guidelines, and expert consensus, so that the model can adapt to the professional context of cancer patient management; In the second stage, instruction data with three complexity levels is designed for different patient management tasks, including low-complexity instructions of single-turn question-and-answer type, medium-complexity instructions of multi-step reasoning type, and high-complexity instructions of multi-turn dialogue type. The high-complexity instructions embed "follow-up question triggering conditions", and when the patient's answer contains specific key information, the model actively asks follow-up questions; In the third stage, based on reinforcement learning alignment with expert feedback, a multi-evaluation dimension expert reward model is constructed, and the algorithm enables the model's decisions to gradually align with expert knowledge.

[0068] Preferably, the patient-specific intervention pathway undergoes multidisciplinary constraint verification. After verification, it is submitted to the physician for review and confirmation. The multidisciplinary constraint verification includes: Knowledge graph consistency check: Check whether the intervention measures conflict with the contraindication rules of each specialty; Timing rationality check: Check whether the time arrangement conforms to the timing constraints of interdisciplinary clinical treatment; Integrity check: Checks whether all management dimensions are covered.

[0069] Preferably, the large language model is combined with a hierarchical, structured multidisciplinary knowledge base and retrieval-enhanced generation; the first layer of the multidisciplinary knowledge base is a clinical experience knowledge base, with data sources including clinical guidelines, expert consensus, MDT discussion records, and typical cases from various disciplines; the second layer is a multidisciplinary medical knowledge graph, with data sources including ICD-10 codes, drug instructions, clinical pathway documents, and expert annotations; the third layer is a clinical knowledge base, with data sources including examination report templates, imaging description standards, and pathology report standards.

[0070] Since the system described in Embodiment 2 of the present invention is an apparatus used to implement the method of Embodiment 1 of the present invention, those skilled in the art can understand the specific structure and variations of the system based on the method described in Embodiment 1 of the present invention, and therefore will not be described again here. All systems used in the method of Embodiment 1 of the present invention fall within the scope of protection of this invention.

[0071] The technical solutions provided in the embodiments of the present invention have at least the following technical effects: 1. Extending and implementing the concept of multidisciplinary integrated diagnosis and treatment outside the hospital: This invention breaks the status quo that multidisciplinary diagnosis and treatment is limited to in-hospital decision-making. By constructing a knowledge engine that integrates multiple disciplines and a regional collaborative management mechanism, it transforms the personalized plans formulated by MDT into an executable, traceable, and adjustable full-process management path, ensuring that patients can continue to recover under the guidance of a unified and collaborative plan after leaving the hospital, and truly realizing the full coverage of the value of MDT.

[0072] 2. Empowering medical staff and building an intelligent multidisciplinary collaborative workflow: The system automatically processes a large number of routine consultation and follow-up tasks through an AI assistant, and makes preliminary judgments and triages based on a multidisciplinary knowledge system, freeing medical staff (especially experts from tertiary hospitals) from repetitive labor, so that they can focus more on high-risk, interdisciplinary and complex decision-making.

[0073] 3. Facilitating the construction of a patient-centered, integrated service ecosystem: By covering the entire process from outpatient to inpatient care to home care and follow-up visits, and injecting multidisciplinary collaborative intelligent management into each stage, the system ensures patient recovery and significantly improves patient compliance and satisfaction. Patients, regardless of their location, can experience continuous, consistent, and professional hospital services, enhancing their medical experience, strengthening patient loyalty, and promoting the effective implementation of hierarchical medical services.

[0074] 4. It provides a unified collaborative work platform for medical staff at different levels and in different disciplines, based on the same multidisciplinary knowledge foundation, thereby improving management efficiency and scientific research collaboration capabilities.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A regional collaborative multidisciplinary integrated rehabilitation management method for cancer patients, characterized in that, include: The process of constructing a knowledge graph for tumor rehabilitation management: Based on the clinical pathways and treatment guidelines sorted out by a multidisciplinary team of experts (MDT), a knowledge graph for tumor rehabilitation management is constructed. The knowledge graph for tumor rehabilitation management adopts a three-layer architecture: schema layer, instance layer and rule layer. The schema layer is used to define entity types, the instance layer includes specific medical knowledge instances, and the rule layer contains clinical decision-making rules. The process of constructing a multidimensional patient profile involves extracting data from the patient's electronic medical record system and constructing a patient profile vector that includes disease, treatment, physical condition, and risk dimensions. The process of generating personalized intervention pathways for patients involves using patient profile vectors as query conditions, performing subgraph matching and reasoning within the tumor rehabilitation management knowledge graph, and obtaining a set of candidate intervention measures. Using patient profile vectors and candidate intervention sets as input to a large language model, a personalized intervention path for the patient is generated. After being validated by multidisciplinary constraints, it is submitted to the doctor for review and confirmation. Rehabilitation management process: Based on the nodes in the intervention path, the node process is automatically triggered at the specified time, including multi-dimensional rehabilitation information push, diagnosis and treatment process reminders, and follow-up interaction; when performing follow-up interaction, the user's voice information is analyzed through speech recognition and semantic understanding to carry out intelligent interaction, extract and automatically summarize key slot information, and generate a structured follow-up report by combining the intent classification results. Based on a large language model, the risk of a patient's question is judged. Low-risk questions are automatically answered by the patient, while high-risk questions are automatically forwarded to the health management team or doctor for reply. The dynamic adjustment process of the intervention pathway: Patient data is continuously collected, and the intervention pathway adjustment is automatically initiated when a trigger condition is detected. The intervention pathway adjustment adopts a triple mechanism of large language model generation + multidisciplinary constraint verification + doctor review. Regional collaborative management process: A multidisciplinary team of experts from tertiary hospitals is responsible for developing standardized management plan templates, training and updating maps and models, and handling high-risk warnings; A management plan for receiving patients from primary care hospitals / communities, including daily data collection and handling of low-risk consultations; The comprehensive data generated by patients at the primary care level or at home is transmitted back to the tertiary hospital platform in real time, forming a data closed loop.

2. The method according to claim 1, characterized in that: The knowledge graph for tumor rehabilitation management is constructed using a human-computer coupling approach. First, a large language model is used to automatically extract triples from clinical guidelines. Then, a multidisciplinary team of experts (MDT) reviews and supplements the data. Finally, knowledge reasoning is used to complete the missing relationships.

3. The method according to claim 1, characterized in that, The large language model is fine-tuned in stages: In the first stage, based on the large medical industry model that has been pre-trained, the model is further pre-trained using corpora including desensitized medical record texts, clinical guidelines, and expert consensus, so that the model can be adapted to the professional context of cancer patient management. In the second stage, three levels of instruction data with varying complexity were designed for different patient management tasks: low-complexity instructions (single-turn question-and-answer type), medium-complexity instructions (multi-step reasoning type), and high-complexity instructions (multi-turn dialogue type). The high-complexity instructions embed "follow-up question triggering conditions," whereby the model actively asks follow-up questions when the patient's answer contains specific key information. In the third stage, a multi-dimensional expert reward model was constructed based on reinforcement learning alignment with expert feedback. The algorithm gradually aligns the model's decisions with expert knowledge.

4. The method according to claim 1, characterized in that: The patient-specific intervention pathway undergoes multidisciplinary constraint verification. After verification, it is submitted to the physician for review and confirmation. The multidisciplinary constraint verification includes: Knowledge graph consistency check: Check whether the intervention measures conflict with the contraindication rules of each specialty; Timing rationality check: Check whether the time arrangement conforms to the timing constraints of interdisciplinary clinical treatment; Integrity check: Checks whether all management dimensions are covered.

5. The method according to claim 1, characterized in that: During the rehabilitation management process, follow-up interactions are performed based on intent recognition and knowledge graph constraints, including: Step 1, Speech Recognition and Preprocessing: An end-to-end speech recognition model is used to convert the patient's speech into text, and domain adaptation is performed using hospital follow-up recording data based on a general ASR model; Step 2, Intent Recognition and Key Slot Extraction: Design an intent-slot system for cancer patient follow-up scenarios; Intent category definition: Symptom feedback: Report new symptoms, describe changes in symptoms, inquire about the causes of symptoms; Medication-related categories: Inquiring about medication administration methods, reporting adverse reactions, and consulting about drug interactions; Follow-up related matters: Inquire about the follow-up time and report the examination results; Life guidance services: dietary advice, exercise advice, and psychological support. Emergency situations: Reporting acute symptoms or requesting urgent medical attention; Key slot definition: Symptom slot: Symptom name, onset time, duration, severity, accompanying symptoms; Drug compartment: Drug name, dosage, frequency, and description of adverse reactions; Time slot: time of symptom onset, duration, and time of follow-up examination; Severity gauge: pain score, degree of impact on daily life; Key slot extraction employs an end-to-end approach based on a large language model: taking patient statements and predefined slot schemas as input, the large language model directly outputs structured slot filling results; Step 3: Based on knowledge graph constraints, design a hybrid dialogue management strategy that combines "knowledge graph constraints + large model generation": Dialogue status tracking: Maintains filled slots, unconfirmed messages, and dialogue history summaries; Follow-up decision mechanism: Based on the symptom-disease association in the knowledge graph, when a patient reports a certain symptom, the system automatically retrieves the information required for differential diagnosis associated with that symptom and determines whether the collected information is sufficient for risk assessment; if insufficient, it generates targeted follow-up questions. Risk classification and response strategy: Assess the risk of the problem and implement the corresponding response strategy; Low-risk issues: Large language models directly generate responses, combined with personalized expressions based on patient profiles; Medium-risk issues: Generate responses and mark them to remind doctors to pay attention. High-risk questions: Do not answer directly; generate a transfer script and notify the doctor immediately. Step 4: Follow-up report is automatically generated. Key slot information is automatically summarized and combined with intent classification results to generate a structured follow-up report.

6. The method according to claim 1, characterized in that: The large language model is combined with a hierarchical, structured multidisciplinary knowledge base and enhanced retrieval generation. The first layer of the multidisciplinary knowledge base is a clinical experience knowledge base, with data sources including clinical guidelines, expert consensus, MDT discussion records, and typical cases from various disciplines. The second layer is a multidisciplinary medical knowledge graph, with data sources including ICD-10 codes, drug instructions, clinical pathway documents, and expert annotations. The third layer is a clinical knowledge base, with data sources including examination report templates, imaging description standards, and pathology report standards.

7. A regional collaborative multidisciplinary integrated rehabilitation management system for cancer patients, characterized in that, include: The tumor rehabilitation management knowledge graph construction module is used to construct a tumor rehabilitation management knowledge graph based on clinical pathways and treatment guidelines compiled by a multidisciplinary MDT expert team. The tumor rehabilitation management knowledge graph adopts a three-layer architecture: schema layer, instance layer and rule layer. The schema layer is used to define entity types, the instance layer includes specific medical knowledge instances, and the rule layer contains clinical decision-making rules. The multidimensional patient profile building module is used to extract data from the patient's electronic medical record system and build a patient profile vector that includes disease dimension, treatment dimension, physical condition dimension and risk dimension; The patient personalized intervention path generation module is used to use patient profile vectors as query conditions to perform subgraph matching and reasoning in the tumor rehabilitation management knowledge graph to obtain a set of candidate intervention measures; Using patient profile vectors and candidate intervention sets as input to a large language model, a personalized intervention path for the patient is generated. After being validated by multidisciplinary constraints, it is submitted to the doctor for review and confirmation. The rehabilitation management module is used to automatically trigger node processes at specified times based on nodes in the intervention path, including multi-dimensional rehabilitation information push, diagnosis and treatment process reminders, and follow-up interaction; when performing follow-up interaction, it uses speech recognition and semantic understanding to analyze the user's voice information for intelligent interaction, extracts and automatically summarizes key slot information, and generates a structured follow-up report by combining intent classification results; Based on a large language model, the risk of a patient's question is judged. Low-risk questions are automatically answered by the patient, while high-risk questions are automatically forwarded to the health management team or doctor for reply. The intervention pathway dynamic adjustment module is used to continuously collect patient data. When a trigger condition is detected, the intervention pathway adjustment is automatically initiated. The intervention pathway adjustment adopts a triple mechanism of large language model generation + multidisciplinary constraint verification + doctor review. The regional collaborative management module is used to enable multidisciplinary team (MDT) experts from tertiary hospitals to develop standardized management plan templates, train and update maps and models, and handle high-risk warnings. A management plan for receiving patients from primary care hospitals / communities, including daily data collection and handling of low-risk consultations; The comprehensive data generated by patients at the primary care level or at home is transmitted back to the tertiary hospital platform in real time, forming a data closed loop.

8. The system according to claim 6, characterized in that, The large language model is fine-tuned in stages: In the first stage, based on the large medical industry model that has been pre-trained, the model is further pre-trained using corpora including desensitized medical record texts, clinical guidelines, and expert consensus, so that the model can be adapted to the professional context of cancer patient management. In the second stage, three levels of instruction data with varying complexity were designed for different patient management tasks: low-complexity instructions (single-turn question-and-answer type), medium-complexity instructions (multi-step reasoning type), and high-complexity instructions (multi-turn dialogue type). The high-complexity instructions embed "follow-up question triggering conditions," whereby the model actively asks follow-up questions when the patient's answer contains specific key information. In the third stage, a multi-dimensional expert reward model was constructed based on reinforcement learning alignment with expert feedback. The algorithm gradually aligns the model's decisions with expert knowledge.

9. The system according to claim 6, characterized in that: The patient-specific intervention pathway undergoes multidisciplinary constraint verification. After verification, it is submitted to the physician for review and confirmation. The multidisciplinary constraint verification includes: Knowledge graph consistency check: Check whether the intervention measures conflict with the contraindication rules of each specialty; Timing rationality check: Check whether the time arrangement conforms to the timing constraints of interdisciplinary clinical treatment; Integrity check: Checks whether all management dimensions are covered.

10. The system according to claim 6, characterized in that: The large language model is combined with a hierarchical, structured multidisciplinary knowledge base and enhanced retrieval generation. The first layer of the multidisciplinary knowledge base is a clinical experience knowledge base, with data sources including clinical guidelines, expert consensus, MDT discussion records, and typical cases from various disciplines. The second layer is a multidisciplinary medical knowledge graph, with data sources including ICD-10 codes, drug instructions, clinical pathway documents, and expert annotations. The third layer is a clinical knowledge base, with data sources including examination report templates, imaging description standards, and pathology report standards.

Images