Full-cycle path chronic disease management system and method based on artificial intelligence

Abstract

The invention discloses a full-cycle path chronic disease management system and method based on artificial intelligence, and belongs to the technical field of intelligent medical treatment, and the method comprises the steps: generating a model result based on an artificial intelligence chronic disease risk prediction model and a prescription through an active diagnosis and treatment module, and carrying out the active recognition, intervention and management of a target group; the personalized diagnosis and treatment module is used for providing refined follow-up visit, prescription and screening services according to health states and risk characteristics of different individuals; and semantic search, index generality identification and multi-source data integration capabilities of clinical data are provided through an intelligent data governance and decision support module. According to the invention, an intelligent medical new mode of active, personalized and intelligent chronic disease management is constructed, clinical decision is assisted, the management efficiency is improved, the chronic disease management level is improved, reasonable flow of medical resources is promoted, and chronic disease prevention, management and referral are promoted to develop towards the full-life-cycle management direction.

Description

Technical Field

[0001] The present invention belongs to the field of smart medical technology, and specifically relates to a full-cycle path-based chronic disease management system and method based on artificial intelligence. Background Art

[0002] Currently, the treatment and management of chronic diseases has become a key focus of healthcare systems worldwide. With socioeconomic development and advancements in medical information technology, the chronic disease management model is gradually shifting from a treatment-centric approach to a health-centric approach, emphasizing health maintenance throughout the life cycle. However, current chronic disease management practices still face the following three prominent issues, which severely hinder the full-lifecycle management of chronic diseases.

[0003] 1. Reliance on individual physician capabilities, with insufficient standardized and personalized support capabilities: In the existing medical model, the disease diagnosis and treatment decision-making process is highly dependent on the individual abilities of doctors. Although evidence-based medicine clinical guidelines provide a scientific basis for diagnosis and treatment practices, the actual application effect of the guidelines is mainly subject to the professional quality and execution ability of individual doctors. Specifically, there are significant differences in the depth of doctors' understanding of the guidelines, their willingness to apply them, and their ability to integrate multi-source health information (such as patient historical data and current disease information). This individual difference in execution, especially in the management of chronic diseases, directly leads to uneven management results, making it difficult to simultaneously take into account both standardized and personalized treatment needs. The main problems with the existing model can be summarized as follows: Over-reliance on individual physician capabilities: The core aspects of chronic disease management (screening, diagnosis, and treatment) are severely limited by the personal experience and judgment of the attending physician. Different physicians vary significantly in their understanding of the same guidelines, their compliance, and their attention to individual patient situations (such as adequacy of communication and interpretation of past medical history). Under high workloads, physicians find it difficult to systematically review a patient's complete medical history and accurately assess potential risks, significantly increasing the risk of misdiagnosis and missed diagnosis. Low level of homogeneity in diagnosis and treatment: Due to the uneven allocation of medical resources across regions, some doctors in primary care institutions find it difficult to master and effectively apply the latest clinical guidelines. This results in a lack of standardized and systematic support for the management of chronic disease patients, and a significant disparity in diagnosis and treatment quality between medical institutions at different levels and in different regions. Inefficient allocation of medical resources: Inconsistent or biased individual physicians' judgments can lead to erroneous assessments of the severity of patients' conditions. Consequently, mildly ill patients may receive unnecessary overdiagnosis and treatment, while intervention for severely ill patients may be delayed. This situation not only impacts patient outcomes but also results in suboptimal allocation of precious medical resources.

[0004] Second, passive response is the main focus, with a lack of proactive management mechanisms: The current chronic disease management system is still characterized by "post-event processing," passively responding to the disease's "onset-progression-serious adverse events" and lacking proactive and forward-looking management interventions. The main manifestations are: Delayed medical treatment: Patients usually seek medical attention for the first time only after symptoms become apparent or complications occur, thus missing the window for early intervention; Referral delays: Due to limited experience and understanding of the disease, primary care physicians often recommend referrals only after treatment is ineffective or the condition worsens, resulting in delayed diagnosis and treatment. Patients then need to receive more complex and costly treatment at higher-level hospitals, exacerbating the burden on medical resources. Delayed prescription / treatment response: The development and adjustment of prescriptions and treatment plans often rely on periodic review and evaluation, making it difficult to respond promptly to changes in the condition. This delayed response may lead to inadequate intervention, increase the risk of complications, and affect patient outcomes.

[0005] 3. Weak intelligent collaboration and data closed-loop capabilities: Current medical information systems still primarily serve the purpose of recording information in chronic disease management, lacking support for proactive and personalized services throughout the entire process: Severe information fragmentation: There is a lack of effective data interoperability between clinical systems, laboratory systems, and public health systems. Data coding standards are not standardized across different institutions, making it difficult to create a complete patient health record. This not only leads to duplicate examinations and waste of resources, but also limits comprehensive assessment of disease progression. Duplicate reporting and redundant processes: Because data between different platforms is not interoperable, doctors need to repeatedly fill in follow-up, referral, prescription, and other data on multiple platforms, increasing their workload and consuming clinical energy. Single system support: The existing medical information system focuses on data collection and storage, making it difficult to provide dynamic guidance and services tailored to patients' specific conditions. It also lacks proactive functions such as clinical decision support, risk warning, and follow-up plan formulation.

[0006] To sum up, the current chronic disease management system has obvious shortcomings in terms of "proactiveness", "personalization" and "intelligence", and needs further optimization and improvement. Summary of the Invention

[0007] In view of the above, the present invention aims to provide a full-cycle path-based chronic disease management system and method based on artificial intelligence. By integrating artificial intelligence technology with regional multi-source medical data, a full-cycle path-based management system covering pre-disease prevention for high-risk groups and post-disease progression control for chronic disease groups is constructed, promoting the three major transformations in chronic disease management: (1) Transformation from passive response to disease to closed-loop management with active early warning and intervention; (2) Transformation from individual ability-dependent decision-making to dynamic collaborative personalized path driven by intelligence; (3) Transformation from rule-based logical judgment to intelligent support system with deep understanding of clinical logic.

[0008] Through the above transformation, the core objectives of the present invention are: to enhance management initiative: establish a forward-looking intervention mechanism based on risk prediction to reduce health losses caused by delayed medical treatment / referral; strengthen standardization and personalized collaboration: improve the level of diagnosis and treatment homogeneity through path-based navigation, and dynamically adjust the plan based on individual health trajectory; break through the bottleneck of intelligent support: build an AI-assisted decision-making system with a deep understanding of clinical logic, reduce the cognitive load of doctors, and optimize resource allocation efficiency; and ultimately achieve a systematic improvement in the effectiveness of chronic disease management, significantly reduce the incidence, disability and mortality of chronic diseases, and improve the treatment compliance rate and health management coverage.

[0009] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides an artificial intelligence-based full-cycle path-based chronic disease management system, comprising: Active diagnosis and treatment module: The active follow-up unit is used to proactively push medical reminders, personalized management paths, and follow-up plans based on the chronic disease risk prediction model; the active health prescription generation unit is used to automatically generate structured health prescriptions by combining the chronic disease risk prediction model and the health prescription generation model; the active referral unit is used to automatically generate referral recommendations based on the chronic disease risk prediction model and track referral status; Personalized diagnosis and treatment module: The personalized follow-up unit is used to automatically generate and present customized follow-up plans and content based on patient data; the personalized health prescription generation unit is used to automatically generate customized health prescriptions through the health prescription generation model; the personalized disease screening unit is used to conduct intelligent screening based on individual characteristics for residents who have not yet been diagnosed with specific chronic diseases; Intelligent data governance and decision support module: The indicator intelligent correlation identification unit is used to automatically identify potential correlations between clinical indicators through artificial intelligence models and realize automatic classification and synchronization; the natural language semantic retrieval unit is used to provide natural language query services with semantic understanding capabilities; the data closed-loop integration management unit is used to access and integrate multi-source heterogeneous medical systems in the region.

[0010] Preferably, the active medical follow-up unit is used to actively push medical reminders, personalized management paths and follow-up plans based on the chronic disease risk prediction model, including: The system backend periodically runs a chronic disease risk prediction model to assess the risk of onset and progression of chronic diseases in high-risk groups. It automatically screens out individuals with risk above the thresholds, generating a high-risk list of onset and a high-risk list of progression, respectively. Based on the lists, the system pushes medical reminders to contracted chronic disease doctors. When a patient arrives at a medical institution for treatment, the system automatically matches the patient with a corresponding management pathway and follow-up plan based on their risk tags, providing real-time guidance on the doctor's interface. The management pathway is displayed in a list, recording the diagnosis and treatment links and task execution status throughout the follow-up process. The follow-up plan is based on the patient's individual characteristics and disease stage, setting the frequency of visits, examination items, and the next follow-up date, and is dynamically updated. After the consultation is completed, the system automatically synchronizes the patient's consultation and follow-up records, and supports one-click synchronization to the public health platform for superior supervision and quality evaluation, and also serves as training samples and reference data for the next cycle of chronic disease risk prediction model operation.

[0011] Preferably, the active health prescription generation unit is used to automatically generate a structured health prescription by combining the chronic disease risk prediction model and the health prescription generation model, including: The chronic disease risk prediction model adopts a modular structure to support parallel prediction of multiple diseases. The Transformer-based multivariate time series prediction submodule processes dense features, the large language model submodule processes sparse features, the filter submodule processes missing patterns and time interval matrices, and the static feature encoding submodule processes static features. The feature vectors output by each submodule are fused by the classification head module to output the incidence and progression risks of multiple chronic diseases. These are then compared with the set risk threshold to form a risk label. The health prescription generation model generates a physical examination summary by summarizing physical examination abnormalities and chronic disease label information and retrieves the patient's historical doctor's advice. It is constructed into a prompt word that includes the physical examination summary, risk label, and historical advice. It calls the large language model to generate a structured health prescription that includes disease risk assessment results, disease progression prediction results, and personalized lifestyle recommendations. The prescription text is output after deduplication and compliance review.

[0012] Preferably, the active referral unit is used to automatically generate referral recommendations and track referral status based on the chronic disease risk prediction model, including: Based on the chronic disease incidence risk score and chronic disease progression risk score generated by the chronic disease prediction model and the referral rules set by the system, a decision recommendation on whether to refer a patient is automatically generated, including key referral information such as the reason for referral, recommended transfer institution, and recommended transfer department. Among them, the reason for referral is automatically generated by the AI ​​referral engine based on the large language model. The model is trained by introducing chronic disease-related referral decision samples and combines the thinking chain reasoning mechanism to achieve step-by-step decision-making capabilities. At the same time, the retrieval enhancement generation technology is used to perform similarity retrieval on the patient assessment data text and the referral rule vector database at the model input stage, and the matching referral rules are dynamically injected into the model input as external knowledge context to improve the rationality and standardization of the generated results. When the doctor accepts the recommendation, the system will automatically embed the referral suggestion into the health prescription, displaying information such as the reason for the referral, the recommended institution, and the department. When the doctor rejects the recommendation, he or she must provide a reason for the rejection. The reason for the rejection will be sent back to the fine-tuning database as an important training sample for iterative optimization of the model to further enhance the model's judgment accuracy. In referral tracking, the system automatically tracks whether the patient has completed the recommended referral operation and assigns a referral label to form a closed-loop management of the referral results. At the same time, the system connects to the regional health data center based on the patient's ID number, monitors the changes in the patient's medical records in medical institutions at all levels in real time, and then judges their referral behavior. In addition, the referral label will be fed back to the system for subsequent referral effect evaluation, medical behavior analysis, and supplementation of model training data.

[0013] Preferably, the personalized medical follow-up unit is used to automatically generate and present a customized follow-up plan and content based on patient data, including: The system first automatically identifies or allows doctors to manually select one or more chronic diseases that the patient currently suffers from based on the target chronic diseases managed by the system. For each identified or selected chronic disease, the system automatically matches and displays its related core symptoms, common complications, and comorbidity follow-up content items based on the chronic disease complication database. Based on the patient's individualized onset / progression risk prediction results, previous health data, and medical records, a personalized test checklist and follow-up plan adjustment recommendations are generated, including: (1) Based on the patient's current chronic disease type and the predicted risk of onset / progression, the follow-up frequency and follow-up items are dynamically adjusted on the basis of a standardized follow-up pathway generated based on clinical guidelines and local public health service requirements to form a personalized follow-up plan; (2) Integrate health data center information to provide patients with a historical list query function for previous test and examination items across medical institutions, support hierarchical viewing by time range and specific items to avoid repeated examinations; (3) Based on the personalized follow-up plan and the patient's recent test and examination completion status, a list of personalized test and examination items recommended for completion in the future cycle is dynamically generated and prompted; for items that have been completed recently, the system automatically identifies and marks the status, and only prompts the items to be completed.

[0014] Preferably, the personalized health prescription generating unit is used to automatically generate a customized health prescription through a health prescription generating model, including: Customized health prescriptions include visualization of key chronic disease indicators, chronic disease onset / progression risk prediction results, condition summary, and lifestyle diagnosis and treatment recommendations, including: (1) Visualization of key indicators of chronic diseases: Based on the cross-institutional health data of patients, identify chronic diseases that require special attention, and dynamically extract and visualize their key physiological indicator trends; key indicators are determined based on the actual illness of the patient and the risk prediction results, and abnormal values ​​are highlighted; (2) Chronic disease onset / progression risk prediction section: Integrate the results of the chronic disease risk prediction model to generate a graded onset / progression risk assessment report; the risk level determination threshold is configurable; (3) Condition summary: Input the patient's medical information and key indicator data, and output a structured condition summary, covering the test frequency of important / abnormal indicators, historical abnormalities, current status and future monitoring priorities; (4) Lifestyle diagnosis and treatment recommendations: Input multidimensional patient data to generate personalized non-drug intervention plans, including lifestyle recommendations, key indicator observation points, and precautions; prescription weights and details are dynamically adjusted according to the condition and risk level; At the same time, the personalized health prescription generation unit supports the review and confirmation mechanism of chronic disease doctors for generated prescriptions. If the content is reasonable, it will be signed and approved. If some suggestions are rejected, clear reasons for rejection must be provided. The system will automatically record the review behavior and incorporate the reasons for rejection into the model optimization closed loop for subsequent model training and reasoning logic correction, so as to continuously improve the clinical adaptability of the model-generated content.

[0015] Preferably, after the personalized disease screening unit identifies potential high-risk groups through intelligent screening, it also includes: For high-risk groups that are newly identified during medical consultations but have not yet signed up for a contract, we push signing suggestions to doctors, recommending them to sign up for corresponding chronic disease management services, and authorizing primary care chronic disease doctors to dynamically manage them, thus establishing a closed-loop service path from screening to follow-up. For high-risk groups who have signed contracts, customized disease screening plans will be automatically generated, including specific diseases to be screened, screening items, recommended screening institutions, and recommended screening frequencies. All screening results will be automatically archived in the residents' electronic health records after completion and serve as the data source for subsequent model iterative training.

[0016] Preferably, the indicator intelligent association identification unit is used to automatically identify potential associations of clinical indicators through an artificial intelligence model and achieve automatic classification and synchronization, including: Construct a medical knowledge graph, covering seven types of nodes including diseases, symptoms, drugs, foods, examination items, and departments, and the multiple relationships between them, including department-department relationship, disease-avoided food relationship, disease-recommended food relationship, disease-general drug relationship, disease-popular drug relationship, disease-examination relationship, manufacturer-drug relationship, disease-symptom relationship, disease-concurrency relationship, and disease-department relationship. The system writes the patient's historical structured data into the graph in the form of event nodes or attributes, and combines it with structured health data including physiological indicators, laboratory test results, and diagnosis results. Through rule reasoning and path query algorithms, it realizes dynamic judgment of potential concurrent relationships and causal connections between diseases, and then drives the automatic checking logic of the corresponding indicator items in the doctor's end interface, returns the relevant disease / indicator / complication information that can be checked, and realizes intelligent rendering as checked state.

[0017] Preferably, the natural language semantic retrieval unit is used to provide a natural language query service with semantic understanding capabilities, including: Based on the fusion of large language models and artificial intelligence agent technology, a natural language semantic retrieval unit that supports multi-round conversational query modes is constructed. The semantic retrieval process is completed autonomously by the large model, including call path judgment, resource scheduling, parameter organization and semantic generation. At the same time, it supports the fusion retrieval of structured and unstructured information. Doctors can set limited dimensions in the query, including specified time periods, medical institutions or indicator types. The system will retrieve information efficiently and in a targeted manner within the limited range.

[0018] In a second aspect, an embodiment of the present invention further provides a full-cycle path-based chronic disease management method based on artificial intelligence, which is implemented using the above-mentioned full-cycle path-based chronic disease management system based on artificial intelligence, and includes the following steps: Based on the active diagnosis and treatment module, the active follow-up unit is used to proactively push out visit reminders, personalized management paths, and follow-up plans based on the chronic disease risk prediction model. The active health prescription generation unit is used to automatically generate structured health prescriptions by combining the chronic disease risk prediction model and the health prescription generation model. The active referral unit is used to automatically generate referral recommendations based on the chronic disease risk prediction model and track referral status. Based on the personalized diagnosis and treatment module, the personalized follow-up unit is used to automatically generate and present customized follow-up plans and content based on patient data; the personalized health prescription generation unit is used to automatically generate customized health prescriptions through the health prescription generation model; and the personalized disease screening unit is used to conduct intelligent screening based on individual characteristics for residents who have not yet been diagnosed with specific chronic diseases. Based on the intelligent data governance and decision support module, the indicator intelligent association identification unit is used to automatically identify potential associations between clinical indicators through artificial intelligence models and realize automatic classification and synchronization; the natural language semantic retrieval unit is used to provide natural language query services with semantic understanding capabilities; and the data closed-loop integration management unit is used to access and integrate multi-source heterogeneous medical systems in the region.

[0019] Compared with the prior art, the present invention has the following beneficial effects: (1) Breakthrough in full-cycle active management capabilities: Establish a closed-loop management system covering high-risk groups and chronic disease groups, realize full-process active management of consultation follow-up, prescription generation, and referral coordination, and support the dual-layer prevention and control goals of "preventing illness before it occurs and preventing progression after illness occurs"; (2) Collaborative optimization of individualization and standardization: providing system support for personalized pathway collaboration for chronic disease management, achieving precise guidance in the stages of follow-up visits, prescription generation, and referrals, with doctors making the final decision, thus improving the efficiency of diagnosis and treatment services; (3) Intelligent decision-making and data governance upgrade: realize regional medical data interconnection and interoperability, realize intelligent association and natural language query of clinical follow-up indicators, and reduce clinical workload; (4) Dual improvement of clinical resource optimization and medical insurance cost control: Through intelligent analysis and accurate prediction, optimize clinical resource allocation and ensure efficient use of medical resources; at the same time, guide the diagnosis and treatment process, promote the rational flow of medical insurance funds, avoid resource waste, effectively control overall medical expenses, and achieve dual optimization management of medical resources and medical insurance costs; (5) Significant results have been achieved in the homogenization of primary medical care and the prevention and control of chronic diseases: By combining standardized active management processes with personalized service plans, the homogenization level of primary medical services has been improved, the incidence and mortality of chronic diseases have been reduced, the quality of life of patients has been significantly improved, and the average life expectancy has been extended; (6) Regionalized systematic promotion and strengthening of public health capabilities: Support systematic and modular promotion and deployment on a regional basis, improve the overall health level of the region by strengthening public health management capabilities and primary medical service capabilities, and provide strong guarantees for building a healthy and harmonious social environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1This is a structural diagram of an artificial intelligence-based full-cycle path-based chronic disease management system provided in an embodiment; Figure 2 is a schematic overview of the active diagnosis and treatment module provided in the embodiment; Figure 3 This is a system interface diagram of an automated medical follow-up unit provided in an embodiment; Figure 4 This is a system interface diagram of an automated / personalized health prescription generation unit provided in an embodiment; Figure 5 This is a system interface diagram of an active referral unit provided in an embodiment; Figure 6 is a schematic diagram of an overview of the personalized diagnosis and treatment module provided in the embodiment; Figure 7 This is a diagram of a follow-up content system interface in a personalized medical follow-up unit provided in an embodiment; Figure 8 This is a diagram of an interface of a test checklist system in a personalized medical follow-up unit provided in an embodiment; Figure 9 This is a system interface diagram of a personalized disease screening unit provided in an embodiment; Figure 10 Schematic diagram of an overview of the intelligent data governance and decision support module provided in the embodiment; Figure 11 This is a system interface diagram of the indicator intelligent association identification unit provided in the embodiment; Figure 12 This is a system interface diagram of the natural language semantic retrieval unit provided in the embodiment. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0023] The inventive concept of the present invention is: in response to the key problems of the existing chronic disease management model, such as insufficient degree of diagnosis and treatment homogeneity, dominance of doctors' personal abilities, weak personalized support, passive responsive diagnosis and treatment, and weak system intelligent support, the embodiment provides a full-cycle path-based chronic disease management system and method based on artificial intelligence, and the risk prediction and prescription generation technology based on artificial intelligence to achieve effective management of chronic disease onset in high-risk groups and chronic disease progression in chronic disease groups. While improving the level of clinical diagnosis and treatment homogeneity, it has three core features: forward-looking active management, personalized path collaborative management adapted to individual health trajectories, and human-computer collaborative intelligent support with a deep understanding of clinical logic. It can give full play to the role of primary medical institutions under the premise of reducing clinical burden, build a full-cycle chronic disease active management system covering the health-high-risk-onset-progression stages, and realize full-cycle closed-loop intelligent chronic disease management from risk prediction, intervention suggestion generation, diagnosis and treatment behavior recommendation to data closed-loop governance.

[0024] Furthermore, the system provided by embodiments of the present invention is applicable to the continuous, dynamic, and precise integrated management of various chronic diseases at the resident level, including but not limited to major chronic diseases such as hypertension, diabetes, coronary heart disease, stroke, chronic obstructive pulmonary disease (COPD), and chronic kidney disease (CKD). It is primarily deployed in primary healthcare institutions, such as community health service centers, community health service stations, and village clinics (hereinafter collectively referred to as the "doctor-side"). The service targets all residents in the region who are at high risk for chronic diseases or already have chronic diseases and who require screening, initial diagnosis, follow-up visits, and long-term follow-up management. The system also supports management services for chronic disease control statistics and performance evaluation for medical institutions and government departments, as well as mobile services for patients, implementing a multi-dimensional chronic disease management architecture that integrates "doctors, managers, and patients." The system provided by embodiments of the present invention is suitable for managing the onset of chronic diseases in high-risk populations and the progression of chronic diseases in chronic disease populations.

[0025] Specifically, high-risk individuals are those who haven't been diagnosed with a chronic disease in the electronic medical system, but whose risk score for a particular chronic disease exceeds a specific threshold (default value 0.8) as assessed by the system's embedded AI chronic disease risk prediction model. This threshold can be flexibly adjusted by local health departments based on regional public health management capabilities. Although undiagnosed, these individuals are at significant risk of developing the disease and should be included in this invention to receive proactive contracting, follow-up intervention, and health education, enabling early identification, intervention, and management.

[0026] The onset of a chronic disease means that residents meet the system diagnostic criteria or the chronic disease diagnosis is included in the electronic medical system. The system diagnostic criteria are based on the clinical guidelines for the chronic disease and the needs of public health control, adapted to the application of the present invention, and evaluated by a team of clinical experts as feasible diagnostic criteria. For example, the system diagnostic criteria for diabetes are that the typical symptoms of diabetes appear and the patient's fasting blood glucose ≥11.1 mmol / l or random blood glucose ≥11.1 mmol / l or blood glucose ≥11.1 mmol / l 2 hours after an oral glucose tolerance test.

[0027] Chronic disease population refers to individual residents who have met the onset criteria of a certain type of chronic disease defined by the system from a certain moment. This onset standard takes into account national clinical guidelines, public health control requirements and actual grassroots implementation capabilities, and is formulated by a team of clinical experts in combination with actual application scenarios. All those who meet the diagnostic conditions will be included in the chronic disease population of the present invention for full-cycle management. The system further supports a hierarchical management mechanism for chronic diseases based on the definition of clinical guidelines, and can adjust its management level according to changes in the patient's condition. At the same time, the system automatically records and identifies chronic disease progression events, including concomitant other chronic diseases (such as hypertension with diabetes), the disease entering a higher level stage, or the occurrence of clinical adverse events such as new hospitalizations and deaths, and promptly triggers high-risk warnings and referral recommendations to support closed-loop dynamic management.

[0028] Chronic disease progression refers to the development of other chronic diseases in patients with a chronic disease, or the progression of a chronic disease to the next system clinical grade, or the occurrence of adverse events such as new hospitalization or death. For example, a patient with grade 1 hypertension develops diabetes, or progresses to grade 2 diabetes, or is newly hospitalized or dies. The system clinical grade is based on the clinical guidelines for the chronic disease and the needs of public health management and control, adapted to the application of the present invention, and evaluated by a team of clinical experts for feasible clinical grades. For example, the system clinical grade of chronic kidney disease is shown in Table 1.

[0029] Table 1 Systematic clinical classification of chronic kidney disease

[0030] The population managed by the system of the present invention (including high-risk population and chronic disease population) signs a one-to-one contract with a primary care chronic disease management physician (hereinafter referred to as a chronic disease physician). The chronic disease physician is the primary physician responsible for the patient and uses the system of the present invention to manage the patient.

[0031] like Figure 1 As shown, the embodiment provides an AI-based, full-cycle, path-based chronic disease management system, including a proactive diagnosis and treatment module, a personalized diagnosis and treatment module, and an intelligent data governance and decision support module. The system integrates multi-source regional health data and, through AI modeling and modular services, forms a complete system for proactive identification, risk prediction, follow-up intervention, prescription recommendations, intelligent referrals, and closed-loop data governance.

[0032] 1. Activated diagnosis and treatment module.

[0033] like Figure 2 As shown, this module is used to proactively identify, intervene, and manage the target population based on the results of the artificial intelligence chronic disease risk prediction model and prescription generation model, and to achieve proactive follow-up visits, health prescription generation, and referrals. Specifically, it includes: (1) Active medical follow-up unit, which is used to conduct real-time scanning and dynamic risk prediction of high-risk groups and chronic disease groups that have signed contracts based on the AI ​​chronic disease risk prediction model, automatically identify high-risk individuals with the risk of disease onset or disease progression, and push medical reminders, personalized management paths and follow-up plans to their corresponding chronic disease contracted doctors, to realize an intelligent, closed-loop follow-up management mechanism.

[0034] Specifically, the system periodically runs (every 10 minutes) an AI-powered chronic disease risk prediction model in the background to assess the risk of onset and progression (adverse events such as graded progression, complications, hospitalization, or death) of high-risk individuals. The system automatically screens individuals above these thresholds by setting onset and progression risk thresholds (both set to a default value of 0.8), generating a "high-risk list for onset" and a "high-risk list for progression," respectively. These thresholds can be flexibly adjusted by local health authorities based on local medical capacity. Based on these high-risk lists for onset and progression, the system sends follow-up reminders to contracted chronic disease physicians through various channels, including creating a tag under the patient's name in the system's patient list and via SMS push notifications. For example, a tag indicating "high risk for onset of hypertension" or "high risk for progression of hypertension" will appear under the patient's name in the patient list. Following the system's prompts, chronic disease physicians can contact the patient by phone, SMS, or other means to arrange for a follow-up visit at a medical institution.

[0035] When a patient arrives at a medical institution for treatment, the system will automatically match the corresponding management path and follow-up plan based on their risk tags, and provide real-time guidance on the doctor's interface. Figure 3 As shown, the management path is displayed in a list format, recording the diagnosis and treatment links and task execution status throughout the follow-up process. The system will mark the completion status of the completed steps and clarify the content to be executed. For example, the system doctor side shows that the current patient has completed the "physical sign collection", "condition inquiry", "lifestyle habits" before the diagnosis, "pre-diagnosis information", "medication status", "health assessment" during the diagnosis, and "confirmation of submission to social services" after the diagnosis, but has not yet completed the "health prescription, annual assessment" step during the diagnosis. The follow-up plan is based on the individual characteristics of the patient and the stage of the disease, sets the frequency of visits, examination items and the next follow-up date, and updates it dynamically. This process ensures that doctors have clear execution guidance and path visualization support during the clinical process, which improves the standardization and operational efficiency of chronic disease management.

[0036] After the consultation is completed, the system will automatically synchronize the patient's consultation and follow-up records to the data center, and support one-click synchronization to the public health platform for government supervision and quality evaluation, and serve as training samples and reference data for the next cycle of AI model operation.

[0037] (2) An active health prescription generation unit is used to automatically generate structured health prescriptions including disease risk assessment, disease progression prediction and personalized lifestyle recommendations based on the patient's historical health data and current medical data (including changes in disease symptoms, changes in disease complications and comorbidities, changes in prescriptions, etc.), combined with an artificial intelligence chronic disease risk prediction model and a health prescription generation model, for review and implementation by chronic disease doctors, thereby enhancing patients' forward-looking health awareness and active participation.

[0038] Specifically, if Figure 4 As shown, a health prescription includes a chronic disease risk prediction section (future chronic disease onset / progression risk grading, risk warnings, and health guidance) and a lifestyle prescription generation section (personalized non-drug intervention plans, including lifestyle recommendations such as exercise, diet, and weight, key indicator observation points, and precautions). For example, the prescription would read, "Based on your current lifestyle and medication use, your future diabetes risk is estimated to be 95% (high). You can reduce your risk factors by living a healthy lifestyle and following your doctor's advice." In the health prescription, the risk of onset and progression is categorized as "high" and "low." The system defaults to high for patients with a risk greater than 0.8, and this value can be adjusted by local health authorities based on regional medical capabilities.

[0039] To predict the risk of onset and progression of disease, a chronic disease risk prediction model consisting of multiple submodules was designed for six chronic disease types in this embodiment. This model automatically assesses the risk of all chronic diseases for a user based on input. To address the issue of uneven time series clinical data, which can be caused by varying follow-up intervals between patients and the different test items tested each time, a new AI-powered chronic disease risk prediction model approach was developed.

[0040] Specifically, the input of the chronic disease risk prediction model includes four types of key feature data: (a) Dense features: The top 50 dense feature data with the least missing data among the test items screened by missing rate, as well as several important feature data proposed by the clinical expert team based on medical prior knowledge; (b) Sparse features: other sparse feature data in the inspection items except the dense feature data screened out; (c) Missing pattern and time interval matrix: The missing pattern of the user's dense input data and the user's time interval matrix (recording the time interval between each examination and the last examination with data) are used to characterize the temporal frequency and observation missing characteristics of the patient's physical examination data; (d) Static characteristics: such as the user’s gender, age, and other demographic and sociological information; The chronic disease risk prediction model consists of four sub-modules: a multivariate time series prediction sub-module based on the Transformer structure, which is responsible for processing dense input; a large language model sub-module, which processes sparse input and unstructured data; a filter sub-module, which is used to extract missing patterns and time-frequency information; and a static feature encoding sub-module, which processes auxiliary information.

[0041] When making predictions, the model first obtains all user input data: 1. Basic user information, including name, gender, date of birth, and ID number (unique identifier); 2. Physical examination data, including examination date, examination item name, test result value, and unit of measurement; 3. Historical medical information, namely, historical medical labels generated from multiple sources such as outpatient, hospitalization, and medical records; 4. Other features, including age and examination interval (the time interval between two consecutive physical examinations). Data governance is then performed, including standardization and unification of examination item names and units. All features are organized into a data format suitable for prediction model input, divided into four parts: dense input data, sparse input data, missing patterns, and other static features, and then loaded into the prediction model.

[0042] During the model inference phase, the organized data is input into the corresponding submodule. The dense input data is used as a vector for multivariate time series interpolation, and then input into a prediction model based on the Transformer architecture suitable for time series analysis tasks. After passing through multiple layers of Transformer decoder layers, a feature vector of the dense data is obtained. The sparse input data is input as jsonb format data into the large language model submodule for processing to obtain a feature vector of the sparse data. The large language model used here is also the large model fine-tuned according to its own data set in the embodiment of the present invention.

[0043] In addition, the user's missing pattern and time interval matrix are input as two sets of important information into a filter submodule used to process time-frequency information to obtain the feature vector of the missing pattern. In view of the widespread missing and irregular sampling phenomena in clinical time series data, the present invention regards the missing pattern matrix and the time interval matrix as inputs with important time dynamic properties, and designs a separate time-frequency filter extraction structure. A dedicated filter group is introduced to perform multi-scale and multi-band time-frequency feature extraction. This operation can significantly improve the model's sensitivity to complex time-frequency patterns such as periodicity, persistence, and mutation of missing events; for time interval heterogeneity, its actual impact on observed dynamics and causal structures can be modeled. The features pre-processed by the filter, whether periodic missing, sparse distribution or time-varying nature of the interval, are efficiently encoded, so that when fused with the main data features, they can better reflect the structural uncertainty and potential physiological rhythm of the sequence, providing more discriminative supplementary information for the downstream prediction layer.

[0044] In addition to the aforementioned inputs, the model also uses other important static features of the user as auxiliary reference information, performing one-hot encoding to produce a static feature vector. Finally, a classification head submodule integrates the four feature vectors to perform classification prediction. Given the four feature vectors, the classification head outputs the user's risk of all chronic diseases (between 0 and 1). These probabilities are compared with a system-defined threshold (default 0.8). If the predicted probability of a particular chronic disease exceeds the threshold, the risk label for that chronic disease is set to "high," and vice versa. Multiple labels can be output simultaneously for multiple diseases. After the prediction is complete, the system generates a health summary of the chronic disease prediction (for example, "You have no abnormal indicators. Please maintain this."). Finally, the disease probabilities, risk labels, and prediction summaries derived by the chronic disease prediction algorithm are stored in a database.

[0045] In summary, the embodiments introduce a time series analysis model based on the Transformer architecture, perform certain optimizations on time series analysis tasks, and combine it with learnable positional encoding to better adapt to multivariate time series analysis tasks. The prediction model based on the Transformer architecture consists of multiple Transformer Decoder layers. In particular, the prediction model design of the present invention adopts a hierarchical feature extraction and top-level dynamic fusion structure, in which the dense observation part of the data, missing patterns, and time interval information are processed separately through independent feature extraction channels. Dedicated filters are introduced for the latter two to achieve multi-scale time-frequency feature modeling. The high-order features extracted from the auxiliary information are not directly merged with the main data at the input end, but are integrated within the dynamic fusion unit at the top level of the prediction network, which is mainly based on attention and gating mechanisms. After theoretical derivation and experimental demonstration, this method is significantly superior to the traditional input layer hybrid method, and can more accurately capture the missing mechanism, sampling heterogeneity, and potential time dependence in time series data, greatly improving the model's prediction effect and generalization ability for complex, multivariate clinical time series data. This method has been deduced and tested and found to effectively improve the prediction effect.

[0046] Furthermore, dense input data requires multivariate time series interpolation when entering the model for analysis. This invention utilizes the missing pattern matrix as a feature for interpolation of dense input data, abandoning the traditional "interpolation first, modeling later" approach. Instead, it adopts a joint modeling and interpolation approach, optimizing the interpolation process alongside the target task, using the missing values ​​themselves as an information source to achieve end-to-end learning. In practical applications, this invention adds two auxiliary training objectives to the main model training objective (prediction loss function): mask interpolation loss and observation reconstruction loss. By establishing these two objectives, the model learns how to better fill missing values ​​during training, demonstrating significant advantages for time series interpolation.

[0047] After the prediction is complete, the system generates a health risk summary (e.g., "Your current risk is high; dietary adjustments are recommended") and stores the prediction results, labels, and summary in a database for use by the health prescription generation model. This model is used to generate dietary, medication, symptom, and clinical indicator considerations, as well as other precautions, based on future risk of developing and progressing chronic diseases. This helps strengthen patients' proactive awareness and proactive participation in health management.

[0048] Health Prescription Generation Model Process: The model input data includes: 1. Basic user information, including age and gender; 2. Physical examination results, including examination items, results, and abnormalities; 3. Chronic disease risk tags, output by the chronic disease risk prediction algorithm; and 4. Doctor's recommendations (optional), derived from the user's historical treatment recommendations and notes in the medical database. When implementing a health guidance plan, a brief summary is generated based on the physical examination results. The chronic disease tags are then integrated. If any historical doctor's recommendations exist, the user's historical doctor's recommendations are retrieved and relevant content from the current chronic disease is extracted as a reference for generating health recommendations. A prompt is constructed, combining information such as the user's age, physical examination summary, chronic disease tags, and historical recommendations to form input text. The fine-tuned large language model of the present invention is then invoked to generate 3-5 multi-faceted, structured health recommendations in a doctor's voice, covering aspects such as lifestyle, diet, and exercise, avoiding specific prescriptions and medication recommendations. Finally, the generated recommendations are deduplicated, content that does not match the prompt is removed, and compliance processing is performed to ensure that the recommendations are compliant, concise, and unambiguous. The final output is personalized health recommendations in HTML format.

[0049] Health Prescription Generation Model Training: First, based on the existing patient prescription data in the system, a structured, high-quality prescription dataset was constructed, combining patient basic information, diagnosis results, test indicators, disease type, allergy history, complications, previous treatment plans, medication usage, and subsequent efficacy feedback. To protect privacy and ensure model generalization, the dataset underwent rigorous desensitization, noise filtering, and annotation consistency verification. Furthermore, data normalization, feature annotation, and label refinement were performed using domain knowledge to ensure that the data used for model training was comprehensive, accurate, and medically relevant. On this basis, Deepseek-r1, a currently popular open-source large language pre-training model, was selected as the base model. The base model was incrementally trained twice using the efficient parameter fine-tuning algorithm "LoRA (Low-Rank Adaptation)." LoRA utilizes low-rank matrices to fine-tune key parameters, enabling the model to more quickly converge to the specialized knowledge distribution for health management and medical advice generation while maintaining its original knowledge base. This improves the input-output ratio and algorithm maintainability, thereby enhancing the reliability and professionalism of the large language model-generated health advice solutions, making it more suitable for the target population of this invention. After deployment, the model can be continuously fine-tuned based on doctor feedback and patient follow-up data to dynamically optimize its accuracy and professional intelligence.

[0050] Finally, the system submits the automatically generated health prescription to the patient's contracted chronic disease physician for review. If the physician agrees with the prescription, they sign and approve it. If they reject a particular item in the prescription, they are required to provide a reason for the rejection. The health prescription is provided to the patient via printed paper or mobile push notification. The system automatically records the physician's approval or rejection of the health prescription, and the rejection reason is evaluated and used for model optimization.

[0051] (3) Active referral unit, which is used to automatically generate referral recommendations based on the model's assessment of the patient's condition (taking into account factors such as disease severity, complication risk, and the handling capacity of lower-level medical institutions) and push them to the doctor's end; the doctor can refer to the risk prompts provided by the system to sign the decision; the referral results can be fed back to the system (for iterative optimization of model parameters) to achieve referral tracking.

[0052] Specifically, if Figure 5 As shown in the figure, referral recommendations are automatically generated based on the chronic disease incidence and progression risk scores generated by the chronic disease prediction model and the system-defined referral rules. These recommendations include key referral information, such as the reason for referral, the recommended institution, and the recommended department. Referral reasons are automatically generated by an AI referral engine based on a large language model. The recommended institution and department are dynamically determined based on a standardized referral pathway provided by local health authorities. The referral pathway, based on regional differences, supports a three-tiered structure: "community health service station / village clinic → community health service center → district hospital," with clear upstream and downstream connecting institutions and departments at each level. In some regions, due to the high level of integration between community health service stations / village clinics and community health service centers, the referral pathway consists of only a two-tiered structure: "grassroots community health service institution → district hospital."

[0053] The AI ​​referral engine is built on a fine-tuned large language model. This model is trained using samples of referral decisions related to chronic diseases and incorporates a Chain of Thought (CoT) reasoning mechanism to enable step-by-step decision-making. Furthermore, to ensure referral decisions adhere to standardized referral pathways and rules, the system employs Retrieval-Augmented Generation (RAG) technology. During the model input phase, similarity searches are performed between patient assessment data text and a database of referral rule vectors. Matching referral rules are dynamically injected into the model input as external knowledge context, enhancing the rationality and standardization of generated results.

[0054] During execution, the system integrates the patient's basic information, chronic disease onset and progression risk scores, and examination data into standardized structured input, which is then fed into a large language model. Based on contextual knowledge and prompt words, the large language model generates output including a referral decision and the reason. After standardization, the output is saved as part of the referral recommendation report and simultaneously pushed to the patient's contracted physician for review.

[0055] For example, if a patient's chronic disease progression risk score is "high," and based on the patient's hospital examination results and medical history, the large model will gradually execute the above steps according to the referral rules to determine that the patient needs to be referred. Based on the score and other input data, it will provide what it believes to be a scientific reason for the referral. The standardized referral judgment and referral reason will then be added to the database as part of the referral recommendation and written into the referral recommendation report.

[0056] In referral tracking, the system backend automatically searches the regional health data center (including historical medical records at all levels of medical institutions) for the patient's ID number. By automatically tracking changes in the medical institution visited, it automatically tracks whether the patient has completed the recommended referral and assigns a referral label, forming a closed-loop management of referral results. The system connects to the regional health data center based on the patient's ID number, monitoring changes in the patient's medical records at all levels of medical institutions in real time to determine their referral behavior.

[0057] Specifically, the system will determine the referral results based on changes in medical institutions and assign the following three referral labels: (1) “Transferred out”: means the patient has been transferred to a designated institution according to the referral recommendation; (2) “Not referred as recommended”: means that the patient was referred but not to the recommended institution or department; (3) “No”: indicates that the patient was not referred.

[0058] The referral label results will be fed back to the system for subsequent referral effect evaluation, medical behavior analysis, and supplementation of AI model training data.

[0059] 2. Personalized diagnosis and treatment module.

[0060] like Figure 6 As shown, it is used to provide personalized medical follow-up, prescription and screening services based on the health status and risk characteristics of different individuals, including: (1) Personalized medical follow-up unit, which is used to automatically generate and present customized follow-up plans and contents to the doctor based on the individual's disease risk label (generated by the onset risk prediction model or the progression risk prediction model), historical health data and current diagnosis. The follow-up plan includes the follow-up time, frequency, specific follow-up items and recommended test and inspection list to improve the efficiency and accuracy of follow-up and reduce the follow-up workload of doctors.

[0061] Specifically, if Figure 7 As shown, the system first automatically identifies or allows physicians to manually select one or more chronic diseases currently affecting the patient based on the target chronic diseases managed by the system. This provides personalized follow-up content on the physician side, precisely matching the patient's actual condition, reducing irrelevant information, and alleviating the follow-up workload. For each identified or selected chronic disease, the system automatically matches and displays relevant core symptoms, common complications, and comorbidity follow-up content based on the chronic disease complication database. For example, for a patient with both hypertension and diabetes, the system will display hypertension and diabetes follow-up content on the interface, such as common hypertension complications such as "coronary atherosclerotic heart disease," "hypertrophic cardiomyopathy," "hypertensive retinopathy," "hypertensive nephropathy," and "hypertensive encephalopathy," and common diabetes complications such as "diabetic nephropathy," "diabetic foot," "diabetic retinopathy," "diabetic peripheral neuropathy," and "coronary atherosclerotic heart disease." Common complications of a particular chronic disease are assessed and provided by a team of clinical experts. Physicians can also use fuzzy searches on disease names to match diseases in the chronic disease complication database, which is constructed based on the International Classification of Diseases (ICD) medical insurance code and incorporating consensus from the clinical expert team.

[0062] Furthermore, if Figure 8 As shown in the personalized follow-up unit, based on the patient's individualized onset / progression risk prediction results, previous health data, and medical records, a personalized inspection checklist and follow-up plan adjustment suggestions are generated, including: (a) Based on the patient's current chronic disease type and the predicted risk of onset / progression, the frequency and items of follow-up visits are dynamically adjusted to form a personalized follow-up plan, based on a standardized follow-up pathway generated based on clinical guidelines and local public health service requirements; (b) Integrate health data center information to provide patients with a historical list query function for previous test items across medical institutions, supporting hierarchical viewing by time range (e.g., within the past 90 days) and specific items (e.g., blood routine - hemoglobin) to avoid duplicate examinations; (c) Based on the personalized follow-up plan and the patient's recent test and examination completion status, a list of personalized test and examination items recommended for completion in the future cycle (e.g., within the next 90 days) is dynamically generated and prompted; for items that have been completed recently, the system automatically identifies and marks the status (e.g., "Completed"), and only prompts items that are yet to be completed.

[0063] For example, based on the information in the health data center, a search can be performed based on the patient's ID number to provide a list of previous test items. If the patient has undergone multiple routine blood tests in different medical institutions in the past 90 days, the doctor can click on "Test items in the past 90 days - Routine blood tests - Hemoglobin" layer by layer to view the cross-institutional hemoglobin data in the data center. Specifically, based on the results of the onset / progression risk prediction and the patient follow-up plan, a personalized list of recommended future tests is provided. For example, for patients with both hypertension and diabetes, according to the follow-up plan, they should complete "blood lipids", "fasting blood sugar", "dynamic blood pressure" and other items in the next 90 days. Since the "dynamic blood pressure" check has been completed recently, the system automatically prompts that the "dynamic blood pressure" check has been completed, and it is recommended to complete "blood lipids", "fasting blood sugar" and other checks in the next 90 days.

[0064] (2) Personalized health prescription generation unit, which is used to integrate the patient's multi-dimensional health data (including current and past health data, medical records, and chronic disease incidence / progression risk prediction results) after the screening or follow-up medical process is completed, and automatically generate a structured customized health prescription through an artificial intelligence prescription generation model. The health prescription at least includes visualization of key indicators of chronic diseases, chronic disease incidence / progression risk prediction results, condition summary and life-oriented diagnosis and treatment suggestions; this unit also supports chronic disease doctors to review, modify and electronically sign the generated prescription content, and use the review feedback information for model optimization.

[0065] Specifically, after the screening and follow-up visits using this system are completed, the AI ​​chronic disease risk prediction model and the AI ​​prescription generation model generate a health prescription, which includes the following four parts: (a) Visualization of key indicators of chronic diseases: Based on the cross-institutional health data of patients, identify chronic diseases that need to be focused on, dynamically extract and visualize the trends of their key physiological indicators; key indicators are determined based on the actual condition of the patient and the risk prediction results, and abnormal values ​​are highlighted. For example, it is presented in the form of a line graph to show the trend of a certain indicator across medical institutions in the region. Specifically, the system displays "triglycerides", "total cholesterol", "fasting blood sugar", "postprandial blood sugar", "systolic blood pressure", and "diastolic blood pressure" by default. According to the patient's actual condition, the risk of chronic disease onset / chronic disease progression, the chronic disease that the patient needs to focus on is generated, and the key indicators of the chronic disease are generated based on the chronic disease that needs to be focused on, and the key indicators and abnormal indicators are output. For example, a patient is a diabetic patient and has abnormal blood lipids, and "fasting blood sugar", "postprandial blood sugar", "glycated hemoglobin", "triglycerides", and "total cholesterol" are displayed; (b) Chronic Disease Onset / Progression Risk Prediction: This section integrates the results of the chronic disease risk prediction model to generate a graded (high / low) onset / progression risk assessment report; the risk level determination threshold is configurable. For example, the prescription will display "Based on your current lifestyle and medication use, your estimated future diabetes risk is 95% (high). You can reduce your risk factors by living a healthy lifestyle and following your doctor's advice." In the health prescription, the risk of onset and progression is divided into two levels: "high" and "low." The system defaults to high for patients with a risk greater than 0.8, and this value can be adjusted by local health departments based on regional medical capabilities. (c) Condition summary: Input the patient's medical information and key indicator data, and output a structured condition summary through the AI ​​health prescription generation model, covering the inspection frequency of important / abnormal indicators, historical abnormalities, current status and future monitoring priorities; (d) Lifestyle diagnosis and treatment recommendations: Input the patient's multi-dimensional data and generate a personalized non-drug intervention plan through the AI ​​health prescription generation model, including lifestyle recommendations (exercise, diet, weight, etc.), key indicator observation points and precautions; the prescription weight and details are dynamically adjusted according to the condition and risk level.

[0066] Furthermore, within the personalized health prescription generation unit, a mechanism supports the review and confirmation of generated prescriptions by chronic disease physicians. Specifically, chronic disease management physicians can review the content of generated prescriptions and sign off if the content is reasonable. If they reject any suggestions, they must provide clear reasons for the rejection. The system will automatically record this review and incorporate the rejection reasons into the model optimization loop for subsequent model training and inference logic correction, continuously improving the clinical adaptability of AI-generated content.

[0067] The final health prescription can be delivered to patients in various forms such as paper printing, SMS / WeChat push, App push, etc., to achieve the continuity and accessibility of health management after follow-up.

[0068] (3) Personalized disease screening unit, which is used to conduct intelligent screening based on individual characteristics for residents who have not yet been diagnosed with specific chronic diseases, so as to identify potential high-risk groups and achieve the chronic disease management goals of early screening, early prevention, and early intervention.

[0069] Specifically, the personalized disease screening unit can assess a resident's risk of developing a specific chronic disease at any primary healthcare institution where the system is deployed, based on their historical health data (including high-risk factors, lifestyle, previous medical history, family history, and test results), combined with an AI-powered chronic disease risk prediction model (morbidity risk prediction model). If a resident's chronic disease risk exceeds a preset threshold, the system automatically includes them in high-risk population management and initiates a targeted screening process.

[0070] For high-risk groups that are newly discovered during medical treatment but have not signed a contract, push contract suggestions to the doctor, recommend signing the corresponding chronic disease management service, and authorize grassroots chronic disease doctors to dynamically manage them. Figure 9 Conduct medical inquiries based on the content shown and build a closed-loop service path of screening and follow-up.

[0071] For high-risk individuals who have signed up, the personalized disease screening unit automatically generates customized disease screening plans, including the specific diseases to be screened, screening items, recommended screening institutions, and recommended screening frequency. For example, individuals at high risk for chronic kidney disease are required to complete annual urinalysis, urine albumin-to-creatinine ratio (UACR), and serum creatinine tests at a medical institution with a chronic disease doctor, and receive CKD prevention and treatment education.

[0072] All screening results will be automatically archived in the residents' electronic health records after completion, and can be used as a data source for subsequent model iterative training to continuously improve the accuracy of the risk prediction model and the adaptability of individualized screening recommendations.

[0073] Through the deployment of this unit, the system can realize intelligent identification and forward intervention of high-risk individuals with chronic diseases among residents in the region, improve the disease prevention capabilities at the grassroots level, and help move the chronic disease checkpoints forward and accurately prevent and control them.

[0074] 3. Intelligent data governance and decision support module.

[0075] like Figure 10 As shown, it is used to provide semantic search of clinical data, identification of common indicators and multi-source data integration capabilities, including: (1) Indicator intelligent association identification unit, which is used to automatically identify the potential association between different clinical indicators (such as key indicators shared by multiple cardiovascular diseases) through an artificial intelligence model (based on the knowledge graph reasoning mechanism), and realize the automatic classification and synchronous operation (such as checking) of clinical indicators (symptoms, complications, comorbidities, smoking and drinking and other lifestyles) in the medical follow-up interface and screening interface, so as to reduce the repetitive work of doctors in actual follow-up and data entry.

[0076] Specifically, the indicator intelligent association identification unit is based on a self-built medical knowledge graph, combined with structured health data (such as physiological indicators, laboratory test results, diagnostic results, etc.), and through rule reasoning and path query algorithms, it realizes dynamic judgment of potential concurrent relationships and causal connections between diseases, and then drives the automatic checking logic of corresponding indicator items in the doctor-side interface, effectively reducing redundant entry operations.

[0077] For example, if a patient is identified by the system as having "diabetes", "hypertension" and "coronary heart disease", "coronary artery atherosclerotic heart disease" will be automatically checked in the diabetes complication type and hypertension complication type (e.g. Figure 11 shown).

[0078] For example, before seeing a doctor, a male patient measures his waist circumference at the all-in-one machine at the nurse station and finds it is 100 cm, which meets the male obesity standard (waist circumference > 90 cm) and satisfies the logical inference of "obesity". The chronic disease doctor imports the all-in-one machine data during the follow-up visit. In the doctor-side interface of the present invention, the option about "obesity" will be automatically checked for screening of related diseases.

[0079] For example, in a patient's historical test items, the low-density cholesterol in the blood biochemistry test is 4.0mmol / L, which is inferred to meet the clinical criteria for hyperlipidemia. In the doctor's interface, the option about "hyperlipidemia" will be automatically checked to remind screening for coronary heart disease and stroke.

[0080] To support the aforementioned reasoning mechanism, an embodiment of the present invention builds a dedicated knowledge graph database based on Neo4j, combining public open-source medical knowledge with supplementary data from a team of clinical experts. When constructing the knowledge graph database, the present invention defines entity nodes, relationship modeling, and attribute definitions. It then accesses structured data (basic patient data) and simultaneously writes or synchronizes it to a data table structure available within the knowledge graph. Rules are defined within the knowledge graph, for example: if a patient has both "diabetes" and "hypertension" → if they also have "coronary heart disease" → automatically check "coronary heart disease" under "diabetes complication type" or "hypertension complication type." Neo4j supports the Cypher query language, enabling complex path searches and conditional judgments. It uses nodes such as diseases and test results and their relationships to detect the status of relevant indicators and output selected results. Data is automatically synchronized in the system backend, and patient test data is written to Neo4j as attributes or event nodes. A series of pre-written Cypher statements are provided for each indicator type to be automatically selected. When the doctor opens the patient interface, the system backend queries Neo4j in real time, retrieves all the indicators / diseases / complications that should be checked for the patient, and returns them to the frontend through the interface. The frontend automatically renders them in the "checked" state. The database contains 7 types of nodes: drugs, foods, examinations, departments, major drug categories, diseases, and symptoms. Each disease has disease information. It includes node entity relationships: department-department relationship, disease-avoided food relationship, disease-recommended food relationship, disease-common drug relationship, disease-popular drug relationship, disease-examination relationship, manufacturer-drug relationship, disease-symptom relationship, disease concurrency relationship, and disease-department relationship. Compared with traditional if-else rules, knowledge graphs can realize reasoning and identify the logical relationship between disease indicators through nodes and the relationship between nodes, realize intelligent association, and relationship adjustment, update, and visual reasoning are very intuitive and dynamic.

[0081] (2) Natural language semantic retrieval units, such as Figure 12 As shown, it is used to provide natural language query services with semantic understanding capabilities based on large language models and artificial intelligence agent (AI Agent) technology, realize structured extraction of medical information and semantic hierarchical scheduling query, assist doctors in performing multiple rounds of interactive retrieval through natural language, and greatly improve the accessibility of medical record data and the efficiency of clinical decision-making support.

[0082] This unit supports doctors to input natural language questions (including but not limited to "Patient Zhang San's blood sugar control status last year", "Trends in renal function changes in the past six months", etc.). The system uses semantic analysis and knowledge-driven matching mechanism to output relevant historical medical records, test data, diagnostic conclusions, indicator trends and other information, and presents them in a structured manner.

[0083] Specifically, the natural language semantic retrieval unit is built based on the integration of a large language model and the ModelContext Protocol (MCP) technology. Its internal implementation process is as follows: (1) The MCP server pre-sets query tools, resources, and execution templates and listens for connections; (2) When a doctor initiates a natural language query request on the client interface, the system starts a subprocess to connect to the server and obtain available query components; (3) The model determines whether to call a tool based on the input content. If so, the system automatically determines which tool to call, which resources to use, and constructs the input parameters; (4) After the tool is executed, the model can determine whether to further call other tools based on the returned results until the final response is obtained; (5) The large language model merges all intermediate results with the generated text to generate a comprehensive query result output that the doctor can read. The above semantic retrieval process is completed autonomously by the large model, including the judgment of the call path, resource scheduling, parameter organization, and semantic generation, and has high adaptability and context memory capabilities. Furthermore, the unit supports the integrated retrieval of structured and unstructured information. Doctors can set limited dimensions in the query, such as a specified time period, medical institution or indicator type, and the system will retrieve information within the limited range to improve the targetedness and efficiency of the retrieval.

[0084] The system also supports multi-round conversational querying. Once the doctor enters the natural language search interface, the system automatically records the current conversation state and memorizes the context for each query. As the doctor continues to ask questions, the system maintains conversational coherence and optimizes the current output based on historical interactions. When the doctor closes the query window or enters the "Exit" command, the system automatically ends the session and clears the context.

[0085] (3) A data closed-loop integration management unit, which is used to access and integrate multi-source heterogeneous medical systems in the region (based on the regional health data middle-end architecture of the present invention), including electronic medical record systems, public health platforms, medical insurance settlement systems and other data sources. Through identity unification, data standardization processing and interface integration, it can achieve structured governance and cross-system closed-loop sharing of residents' health information, improve data interoperability efficiency and the coordination of full-process clinical services.

[0086] Specifically, the system database uses the patient's ID number as a unique identifier to unify historical health data from different institutions, ensuring long-term continuous management and vertical alignment of various medical records, follow-up information, test results, etc. The data mapping engine supports automatic cleaning and standardization of test data in various heterogeneous systems, including the following dimensions: variable name mapping, variable unit conversion, reference value standardization, and error value filtering. The variable name unifies the names of certain test items in different institutions to standardized variable names. For example, "creatinine", "creatinine (CREA)", "creatinine (enzymatic method)", and "creatinine determination" are standardized to "blood creatinine".

[0087] The cleaned and standardized data supports multi-institutional data visualization and interactive query. Doctors can view the historical records of a certain indicator in multiple institutions in the form of lists, trend line charts, etc. on the platform to assist in judging the continuity of long-term chronic disease management and cross-hospital medical treatment.

[0088] Furthermore, the unit supports regional information sharing and assessment integration. After a doctor completes a follow-up visit, prescribes a prescription, or sets a referral plan, the system supports a one-click "sync to regional platform" operation. For example, clicking the "Sync to Social Service" button pushes the follow-up information to the regional public health service platform in real time, automatically completing the submission of performance assessment data, greatly simplifying the doctor's operation process and reducing the burden of manual repetitive data entry.

[0089] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A full-cycle path-based chronic disease management system based on artificial intelligence, characterized by: include: Active diagnosis and treatment module: The active diagnosis and follow-up unit is used to proactively push diagnosis reminders, personalized management pathways, and follow-up plans based on the chronic disease risk prediction model; The active health prescription generation unit is used to automatically generate structured health prescriptions by combining the chronic disease risk prediction model and the health prescription generation model; The proactive referral unit is used to automatically generate referral recommendations and track referral status based on the chronic disease risk prediction model; Personalized diagnosis and treatment module: The personalized follow-up unit is used to automatically generate and present customized follow-up plans and content based on patient data; The personalized health prescription generation unit is used to automatically generate customized health prescriptions through the health prescription generation model; the personalized disease screening unit is used to conduct intelligent screening based on individual characteristics for residents who have not yet been diagnosed with specific chronic diseases; Intelligent data governance and decision support module: The indicator intelligent correlation identification unit is used to automatically identify potential correlations between clinical indicators through artificial intelligence models and realize automatic classification and synchronization; the natural language semantic retrieval unit is used to provide natural language query services with semantic understanding capabilities; the data closed-loop integration management unit is used to access and integrate multi-source heterogeneous medical systems in the region.

2. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The proactive medical follow-up unit is used to proactively push medical reminders, personalized management pathways, and follow-up plans based on the chronic disease risk prediction model, including: The system backend periodically runs a chronic disease risk prediction model to assess the risk of onset and progression of chronic diseases in high-risk groups. It automatically screens out individuals with risk above the thresholds, generating a high-risk list of onset and a high-risk list of progression, respectively. Based on the lists, the system pushes medical reminders to contracted chronic disease doctors. When a patient arrives at a medical institution for treatment, the system automatically matches the patient with a corresponding management pathway and follow-up plan based on their risk tags, providing real-time guidance on the doctor's interface. The management pathway is displayed in a list, recording the diagnosis and treatment links and task execution status throughout the follow-up process. The follow-up plan is based on the patient's individual characteristics and disease stage, setting the frequency of visits, examination items, and the next follow-up date, and is dynamically updated. After the consultation is completed, the system automatically synchronizes the patient's consultation and follow-up records, and supports one-click synchronization to the public health platform for superior supervision and quality evaluation, and also serves as training samples and reference data for the next cycle of chronic disease risk prediction model operation.

3. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The active health prescription generation unit is used to automatically generate a structured health prescription by combining the chronic disease risk prediction model and the health prescription generation model, including: The chronic disease risk prediction model adopts a modular structure to support parallel prediction of multiple diseases. The Transformer-based multivariate time series prediction submodule processes dense features, the large language model submodule processes sparse features, the filter submodule processes missing patterns and time interval matrices, and the static feature encoding submodule processes static features. The feature vectors output by each submodule are fused by the classification head module to output the incidence and progression risks of multiple chronic diseases. These are then compared with the set risk threshold to form a risk label. The health prescription generation model generates a physical examination summary by summarizing physical examination abnormalities and chronic disease label information and retrieves the patient's historical doctor's advice. It is constructed into a prompt word that includes the physical examination summary, risk label, and historical advice. It calls the large language model to generate a structured health prescription that includes disease risk assessment results, disease progression prediction results, and personalized lifestyle recommendations. The prescription text is output after deduplication and compliance review.

4. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The active referral unit is used to automatically generate referral recommendations and track referral status based on the chronic disease risk prediction model, including: Based on the chronic disease incidence risk score and chronic disease progression risk score generated by the chronic disease prediction model and the referral rules set by the system, a decision recommendation on whether to refer a patient is automatically generated, including key referral information such as the reason for referral, recommended transfer institution, and recommended transfer department. Among them, the reason for referral is automatically generated by the AI ​​referral engine based on the large language model. The model is trained by introducing chronic disease-related referral decision samples and combines the thinking chain reasoning mechanism to achieve step-by-step decision-making capabilities. At the same time, the retrieval enhancement generation technology is used to perform similarity retrieval on the patient assessment data text and the referral rule vector database at the model input stage, and the matching referral rules are dynamically injected into the model input as external knowledge context to improve the rationality and standardization of the generated results. When the doctor accepts the recommendation, the system will automatically embed the referral suggestion into the health prescription, displaying information such as the reason for the referral, the recommended institution, and the department. When the doctor rejects the recommendation, he or she must provide a reason for the rejection. The reason for the rejection will be sent back to the fine-tuning database as an important training sample for iterative optimization of the model to further enhance the model's judgment accuracy. In referral tracking, the system automatically tracks whether the patient has completed the recommended referral operation and assigns a referral label to form a closed-loop management of the referral results. At the same time, the system connects to the regional health data center based on the patient's ID number, monitors the changes in the patient's medical records in medical institutions at all levels in real time, and then judges their referral behavior. In addition, the referral label will be fed back to the system for subsequent referral effect evaluation, medical behavior analysis, and supplementation of model training data.

5. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The personalized follow-up unit is used to automatically generate and present customized follow-up plans and content based on patient data, including: The system first automatically identifies or allows doctors to manually select one or more chronic diseases that the patient currently suffers from based on the target chronic diseases managed by the system. For each identified or selected chronic disease, the system automatically matches and displays its related core symptoms, common complications, and comorbidity follow-up content items based on the chronic disease complication database. Based on the patient's individualized onset / progression risk prediction results, previous health data, and medical records, a personalized test checklist and follow-up plan adjustment recommendations are generated, including: (1) Based on the patient's current chronic disease type and the predicted risk of onset / progression, the follow-up frequency and follow-up items are dynamically adjusted on the basis of a standardized follow-up pathway generated based on clinical guidelines and local public health service requirements to form a personalized follow-up plan; (2) Integrate health data center information to provide patients with a historical list query function for previous test and examination items across medical institutions, support hierarchical viewing by time range and specific items to avoid repeated examinations; (3) Based on the personalized follow-up plan and the patient's recent test and examination completion status, a list of personalized test and examination items recommended for completion in the future cycle is dynamically generated and prompted; for items that have been completed recently, the system automatically identifies and marks the status, and only prompts the items to be completed.

6. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The personalized health prescription generation unit is used to automatically generate a customized health prescription through a health prescription generation model, including: Customized health prescriptions include visualization of key chronic disease indicators, chronic disease onset / progression risk prediction results, condition summary, and lifestyle diagnosis and treatment recommendations, including: (1) Visualization of key indicators of chronic diseases: Based on the cross-institutional health data of patients, identify chronic diseases that require special attention, and dynamically extract and visualize their key physiological indicator trends; key indicators are determined based on the actual illness of the patient and the risk prediction results, and abnormal values ​​are highlighted; (2) Chronic disease onset / progression risk prediction section: Integrate the results of the chronic disease risk prediction model to generate a graded onset / progression risk assessment report; the risk level determination threshold is configurable; (3) Condition summary: Input the patient's medical information and key indicator data, and output a structured condition summary, covering the test frequency of important / abnormal indicators, historical abnormalities, current status and future monitoring priorities; (4) Lifestyle diagnosis and treatment recommendations: Input multidimensional patient data to generate personalized non-drug intervention plans, including lifestyle recommendations, key indicator observation points, and precautions; prescription weights and details are dynamically adjusted according to the condition and risk level; At the same time, the personalized health prescription generation unit supports the review and confirmation mechanism of chronic disease doctors for generated prescriptions. If the content is reasonable, it will be signed and approved. If some suggestions are rejected, clear reasons for rejection must be provided. The system will automatically record the review behavior and incorporate the reasons for rejection into the model optimization closed loop for subsequent model training and reasoning logic correction, so as to continuously improve the clinical adaptability of the model-generated content.

7. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: After identifying potential high-risk groups through intelligent screening, the personalized disease screening unit also includes: For high-risk groups that are newly identified during medical consultations but have not yet signed up for a contract, we push signing suggestions to doctors, recommending them to sign up for corresponding chronic disease management services, and authorizing primary care chronic disease doctors to dynamically manage them, thus establishing a closed-loop service path from screening to follow-up. For high-risk groups who have signed contracts, customized disease screening plans will be automatically generated, including specific diseases to be screened, screening items, recommended screening institutions, and recommended screening frequencies. All screening results will be automatically archived in the residents' electronic health records after completion and serve as the data source for subsequent model iterative training.

8. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The indicator intelligent association identification unit is used to automatically identify potential associations of clinical indicators through an artificial intelligence model and achieve automatic classification and synchronization, including: Construct a medical knowledge graph, covering seven types of nodes including diseases, symptoms, drugs, foods, examination items, and departments, and the multiple relationships between them, including department-department relationship, disease-avoided food relationship, disease-recommended food relationship, disease-general drug relationship, disease-popular drug relationship, disease-examination relationship, manufacturer-drug relationship, disease-symptom relationship, disease-concurrency relationship, and disease-department relationship. The system writes the patient's historical structured data into the graph in the form of event nodes or attributes, and combines it with structured health data including physiological indicators, laboratory test results, and diagnosis results. Through rule reasoning and path query algorithms, it realizes dynamic judgment of potential concurrent relationships and causal connections between diseases, and then drives the automatic checking logic of the corresponding indicator items in the doctor's end interface, returns the relevant disease / indicator / complication information that can be checked, and realizes intelligent rendering as checked state.

9. The full-cycle path-based chronic disease management system based on artificial intelligence according to claim 1 is characterized in that: The natural language semantic retrieval unit is used to provide a natural language query service with semantic understanding capabilities, including: Based on the fusion of large language models and artificial intelligence agent technology, a natural language semantic retrieval unit that supports multi-round conversational query modes is constructed. The semantic retrieval process is completed autonomously by the large model, including call path judgment, resource scheduling, parameter organization and semantic generation. At the same time, it supports the fusion retrieval of structured and unstructured information. Doctors can set limited dimensions in the query, including specified time periods, medical institutions or indicator types. The system will retrieve information efficiently and in a targeted manner within the limited range.

10. A full-cycle path-based chronic disease management method based on artificial intelligence, implemented using the full-cycle path-based chronic disease management system based on artificial intelligence according to any one of claims 1 to 9, characterized in that: The following steps are involved: Based on the proactive diagnosis and treatment module, the proactive follow-up unit proactively pushes out visit reminders, personalized management pathways, and follow-up plans based on the chronic disease risk prediction model; Utilize the active health prescription generation unit combined with the chronic disease risk prediction model and the health prescription generation model to automatically generate structured health prescriptions; Utilize an active referral unit to automatically generate referral recommendations and track referral status based on a chronic disease risk prediction model; Based on the personalized diagnosis and treatment module, the personalized follow-up unit is used to automatically generate and present customized follow-up plans and content based on patient data; Utilize the personalized health prescription generation unit to automatically generate customized health prescriptions through the health prescription generation model; Utilize personalized disease screening units to conduct intelligent screening based on individual characteristics for residents who have not yet been diagnosed with specific chronic diseases; Based on the intelligent data governance and decision support module, the indicator intelligent correlation identification unit uses artificial intelligence models to automatically identify potential correlations between clinical indicators and achieve automatic classification and synchronization; Use the natural language semantic retrieval unit to provide natural language query services with semantic understanding capabilities; use the data closed-loop integration management unit to access and integrate multi-source heterogeneous medical systems in the region.

Images