A primary-level chronic disease management system based on dynamic multi-model and knowledge enhancement

Abstract

This disclosure presents an embodiment of a primary-level chronic disease management system based on dynamic multi-model and knowledge enhancement. One specific implementation of the system includes: a data integration server, a knowledge graph server, a task scheduling server, a decision generation server, and a user interaction terminal. The data integration server is used to standardize primary-level chronic disease patient data, generating structured feature vectors and dynamic indicator datasets. The knowledge graph server is used to construct a medical knowledge graph. The task scheduling server is used to perform task scheduling and generate primary-level chronic disease management task assignment instructions. The decision generation server is used to drive the collaboration between a first model and a second model to generate patient treatment information. The user interaction terminal is used to convert patient treatment information into a preset format and display it on the terminal. This implementation ensures that diagnostic results are consistent with reality and reduces the waste of computing resources.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to a primary-level chronic disease management system based on dynamic multi-model and knowledge enhancement. Background Technology

[0002] Chronic non-communicable diseases (hereinafter referred to as "chronic diseases") have become a global public health challenge. As the first line of defense in chronic disease prevention and control, grassroots communities bear the heavy responsibility of long-term, dynamic, and highly personalized management. Currently, when using artificial intelligence technology to assist in chronic disease management, the common approaches are: utilizing Large Language Models (LLM) for tasks such as medical knowledge question answering and medical record summary generation; dynamically scheduling different functional agents (diagnosis, report generation, etc.) to collaborate based on task complexity; and improving decision-making transparency by simulating the step-by-step reasoning process of doctors—"symptoms, analysis, hypothesis, verification."

[0003] However, when using the above methods to assist in the management of chronic diseases, the following technical problems often arise:

[0004] The general-purpose large language model lacks verified deep medical knowledge, which can easily lead to "fact illusions" that contradict the facts. Furthermore, existing knowledge graph augmentation methods ignore temporal information, resulting in diagnostic results that do not match the facts and wasting computing resources.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion that follows. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure propose a primary-level chronic disease management system based on dynamic multi-model and knowledge enhancement to address one or more of the technical problems mentioned in the background section above.

[0008] Some embodiments of this disclosure provide a primary-level chronic disease management system based on dynamic multi-model and knowledge enhancement. The system includes: a data integration server, a knowledge graph server, a task scheduling server, a decision generation server, and a user interaction terminal. The data integration server is used to standardize primary-level chronic disease patient data to generate structured feature vectors and dynamic indicator datasets. The primary-level chronic disease patient data includes: electronic health records, self-reported texts from primary-level chronic disease patients, and time-series monitoring data. The knowledge graph server is used to construct a medical knowledge graph based on the structured feature vectors and the dynamic indicator datasets. The knowledge graph includes a structured knowledge network for chronic diseases at the grassroots level and a dynamic weight parameter set based on the aforementioned time-series monitoring data; the task scheduling server is used to schedule tasks based on the aforementioned structured feature vectors to generate instructions for the division of labor in the treatment of chronic diseases at the grassroots level; the decision generation server is used to drive the collaboration between the first model and the second model based on the aforementioned instructions for the division of labor in the treatment of chronic diseases at the grassroots level, the aforementioned dynamic weight parameter set, and the aforementioned medical knowledge graph to generate patient treatment information, wherein the aforementioned patient treatment information includes target treatment templates for chronic diseases at the grassroots level, intervention measures, and time-series warnings; the aforementioned user interaction terminal is used to convert the aforementioned patient treatment information into a preset format and display it on the terminal.

[0009] The various embodiments of this disclosure have the following beneficial effects: The grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement, as described in some embodiments of this disclosure, utilizes crucial temporal information in chronic disease management to provide a medically validated medical knowledge graph and diagnostic process for a large language model, ensuring that diagnostic results align with reality and reducing the waste of computational resources. Specifically, the reason for discrepancies between diagnostic results and reality, leading to wasted computational resources, lies in the fact that general-purpose large language models lack validated deep medical knowledge, easily generating "factual illusions" that contradict reality, and existing knowledge graph enhancement methods ignore temporal information. Based on this, the grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement, as described in some embodiments of this disclosure, includes a data integration server, a knowledge graph server, a task scheduling server, a decision generation server, and a user interaction terminal. First, the data integration server is used to standardize grassroots chronic disease patient data to generate structured feature vectors and dynamic indicator datasets. This completes the data cleaning and time alignment of heterogeneous data. Secondly, the aforementioned knowledge graph server is used to construct a medical knowledge graph based on the structured feature vectors and the dynamic indicator dataset. This integrates entity relationships such as diseases, drugs, and treatment plans to form a reasonable knowledge base. Then, the aforementioned task scheduling server is used to schedule tasks based on the structured feature vectors to generate instructions for the division of labor in primary care chronic disease management. This allows for dynamic scheduling of computing resources and dynamic allocation of models based on patient risk levels, reducing waste of computing resources. Next, the aforementioned decision generation server is used to drive the collaboration between the first and second models based on the instructions for the division of labor in primary care chronic disease management, the dynamic weight parameter set, and the medical knowledge graph to generate patient treatment information. This generates diagnostic hypotheses and treatment plans that are consistent with reality. Finally, the aforementioned user interaction terminal is used to convert the patient treatment information into a preset format and display it on the terminal. This adapts the treatment information to various terminals. This implementation utilizes the crucial temporal information in chronic disease management to provide a medically validated medical knowledge graph and diagnostic process for the large language model, ensuring that diagnostic results are consistent with reality and reducing waste of computing resources. Attached Figure Description

[0010] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0011] Figure 1 These are schematic diagrams illustrating the structure of some embodiments of a primary chronic disease management system based on dynamic multi-model and knowledge enhancement, as disclosed herein. Detailed Implementation

[0012] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0013] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0014] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0015] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0016] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0017] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Please see Figure 1 , Figure 1 This is a schematic diagram 100 of some embodiments of the grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement disclosed herein. The grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement includes: a data integration server 101, a knowledge graph server 102, a task scheduling server 103, a decision generation server 104, a user interaction terminal 105, and a conflict resolution server 106.

[0019] Data integration server 101 is a server used to standardize data of patients with chronic diseases at the grassroots level to generate structured feature vectors and dynamic indicator datasets.

[0020] In some embodiments, the above data integration server 101 may perform standardization processing on the grass-roots chronic disease patient data to generate a structured feature vector and a dynamic index data set. Among them, the above data integration server 101 may include, but is not limited to, at least one of the following: a processor, a storage, a database, and a data integration dedicated module (ETL engine, API gateway, etc.). The above grass-roots chronic disease patient data includes: electronic health records, self-reported texts of grass-roots chronic disease patients, and time-series monitoring data. The above self-reported texts of grass-roots chronic disease patients are subjective texts actively described by patients about symptom feelings, medical history details, and life impacts, and may include dialects and non-universal terms. The above time-series monitoring data may be a time-ordered physiological index sequence continuously collected by a device, such as the recorded value per minute of a blood glucose meter and an electrocardiogram waveform. It includes a timestamp, an index type (such as heart rate), and a measured value, and is used for real-time monitoring of disease trends and early warning of abnormalities.

[0021] In some optional implementation manners of some embodiments, the above data integration server 101 may be configured to perform the following steps:

[0022] In the first step, perform entity recognition and timestamp alignment operations on the above self-reported texts of grass-roots chronic disease patients to generate structured self-reported texts, and perform the following sub-steps:

[0023] Sub-step one, perform multi-level semantic segmentation on the above self-reported texts of grass-roots chronic disease patients to generate a set of segmented phrases. In practice, the above data integration server 101 may perform multi-level semantic segmentation on the above self-reported texts of grass-roots chronic disease patients through punctuation and conjunction segmentation and dependency syntax correction. The above punctuation and conjunction segmentation may be through terminal punctuation marks (such as:.,?,!, ;), or through conjunctions (such as: but, because, so). Among them, the above dependency grammar correction may include subject-predicate separation and attributive-headword merging. For example, split gastrointestinal diarrhea into gastrointestinal and diarrhea, and merge fasting and blood glucose into fasting blood glucose.

[0024] Sub-step two, based on a preset trigger word library, perform keyword recognition on the above set of segmented phrases to obtain an initial keyword set. Among them, the above preset trigger word library may be a basic medical word library and a grass-roots dialect word library. In practice, the above data integration server 101 may perform keyword recognition on the above set of segmented phrases by replacing the dialects in the above set of segmented phrases in the above grass-roots dialect word library to obtain a set of replaced segmented phrases. Then, search for the same keywords as those in the above set of replaced segmented phrases in the above basic medical word library to obtain an initial keyword set.

[0025] Sub-step three involves performing initial entity recognition on the initial keyword set based on a preset recognition model to generate initial entities. In practice, the data integration server 101 can perform initial entity recognition on the initial keyword set through the following steps: First, the data integration server 101 converts the initial keyword set into a word vector set using BERT word vectors finely tuned for the medical field, and adds positional label features to the word vector set to represent parts of speech (e.g., nouns, quantifiers), obtaining a keyword feature matrix to input into the preset recognition model. Second, the data integration server 101 defines a medical-specific label set for the preset recognition model. For example, B-SYMP (symptom onset), I-SYMP (symptom continuation), B-DRUG (drug onset), MED_DOSE (dosage unit), and NEG (negation modifier). For example, a dosage must follow a drug, and a unit is prohibited after a symptom. The preset recognition model is a Natural Language Processing (NLP) model used to assign a label to each element (such as a word, character, etc.) in text. These tasks include, but are not limited to, Named Entity Recognition (NER), Part-of-Speech Tagging (POS), syntactic parsing, and chunking. For example, the aforementioned preset recognition model could be a BiLSTM-CRF model.

[0026] Sub-step four involves post-processing the initial entities to generate structured self-narrative text. In practice, the post-processing operations performed by the data integration server 101 on the initial entities may include: adding attributes (e.g., adding "intensity" as a degree word), reorganizing the initial entities (e.g., adding a degree word after the symptom, such as difficulty breathing), filtering negative words (e.g., "none" or "not"), symptom splitting rules (e.g., changing palpitations and shortness of breath to palpitations and difficulty breathing), and probabilistic classification (e.g., if the modifier "somewhat" is present, the confidence of the corresponding symptom is reduced, such as symptom confidence = 0.6).

[0027] The second step involves preprocessing the electronic health records and the structured self-report text to generate structured feature vectors. In practice, the data integration server 101 can preprocess the electronic health records and the structured self-report text using the following steps: First, the data integration server 101 can align the fields of the electronic health records and the structured self-report text using the UMLS semantic network mapping method to resolve terminology differences between heterogeneous data sources (e.g., the patient says "diabetes," while the electronic health record uses "T2DM"); next, the data integration server 101 can perform time window calibration on the electronic health records and the structured self-report text (e.g., converting "last 3 months" in the structured self-report text to "2025Q2" based on time estimation). Then, in response to conflicts between the electronic health records and the structured self-report text, the data integration server 101 can prioritize adopting the new data with time attributes (e.g., the electronic health records record "no diabetes," while the structured self-report text records "diabetes for 10 years," adopting the self-report data). Finally, the data integration server 101 can encode the electronic health record and the structured self-report text into feature vectors to generate structured feature vectors (e.g., combining gender, age, number of complications, symptoms, and symptom duration into a vector).

[0028] The third step is to determine the fluctuation coefficient of the aforementioned time-series monitoring data to generate a dynamic indicator dataset. In practice, the data integration server 101 can determine the fluctuation coefficient of the aforementioned time-series monitoring data through the following steps: First, the aforementioned time-series monitoring data (e.g., blood glucose and blood pressure) are sampled at a uniform frequency (e.g., once every 15 minutes). The fluctuation coefficient can be determined using the following formula:

[0029]

[0030] Where σ represents the standard deviation, μ represents the mean, α represents the trend weight (0.3 for blood glucose and 0.2 for blood pressure), and trend represents the trend factor.

[0031] Then, the names, fluctuation coefficients, standard deviations, and means of the aforementioned time-series monitoring data are used as a dynamic indicator dataset.

[0032] The knowledge graph server 102 is a server used to construct a medical knowledge graph based on the structured feature vectors and dynamic indicator datasets received from the aforementioned data integration server 101.

[0033] In some embodiments, the knowledge graph server 102 can construct a medical knowledge graph based on the structured feature vectors and the dynamic indicator dataset. The knowledge graph server 102 may include, but is not limited to, at least one of the following: a graph database, a knowledge reasoning engine, a knowledge extraction tool, and a vector database. The medical knowledge graph includes a structured knowledge network for primary care chronic diseases and a dynamic weight parameter set based on the time-series monitoring data.

[0034] In some alternative implementations of certain embodiments, the knowledge graph server 102 described above can be configured to perform the following steps:

[0035] The first step is to perform feature semantic mapping on the above structured feature vectors to generate an initial set of medical concepts. In practice, the knowledge graph server 102 can achieve feature semantic mapping by mapping numerical ranges to medical concepts to obtain the initial set of medical concepts (for example, concepts of age greater than or equal to 60 years old can be mapped to elderly patients according to the WHO elderly standard, concepts of the number of complications greater than or equal to 3 can be mapped to high-risk complication states according to the ADA diabetes diagnosis and treatment standards, and concepts of blood glucose fluctuation coefficient greater than 30% can be mapped to unstable blood glucose control according to the Chinese type 2 diabetes prevention and treatment standards).

[0036] The second step involves classifying and binding the initial medical concept set to its attributes to generate medical entities. In practice, the knowledge graph server 102 can classify and bind the initial medical concept set to its attributes using the following steps: First, based on the entity classification rule table, each initial medical concept in the initial medical concept set is divided into disease entities, symptom entities, intervention entities, and management entities. Then, attribute binding is performed on the disease entities, symptom entities, intervention entities, and management entities. The medical entities can include: disease entities, symptom entities, intervention entities, and management entities. The entity classification rule table can include: initial medical concepts that conform to the ICD-11 symptom terminology definition are classified as symptom entities; initial medical concepts with matching entries in the National Drug Catalog are classified as intervention entities; initial medical concepts that conform to the ICD-11 disease ontology are classified as disease entities; and initial medical concepts belonging to clinical management behaviors (not diseases or symptoms) are classified as management entities. The attributes of the aforementioned disease entities may include staging criteria (e.g., based on the ADA Diabetes Care Standards, "Stage": "Stage III"), risk (e.g., based on the number of complications and biochemical indicators, "Risk": "High Risk"), and diagnostic criteria (e.g., linked to the ICD-11 diagnostic code, "Diagnostic Code": "5A11.2"). The attributes of the aforementioned symptom entities may include quantification thresholds (e.g., based on the abnormal range defined by the ADA Diabetes Care Standards, "Threshold": "Fasting Blood Glucose > 7.0 mmol / L"), intensity grading (e.g., based on the aforementioned structured self-written text and the ADA Diabetes Care Standards, "Intensity": "Moderate"), and timeliness (e.g., based on the symptom duration in the aforementioned structured feature vector, "Duration": "2025Q2"). The attributes of the aforementioned intervention entities may include: drug name, dosage range, contraindications (e.g., contraindicated in renal insufficiency), and route of administration (e.g., oral).

[0037] The third step involves parsing the pre-defined clinical guidelines to generate clinical guideline rules. In practice, this parsing can be performed using a data integration server 101. The pre-defined clinical guidelines can be the ADA diabetes treatment standards. These clinical guideline rules can include conditional statements, logical connectors, and action statements. Action statements are statements that include one or more actions.

[0038] The fourth step involves defining the relational types of the aforementioned clinical guideline rules to generate entity relations, and then using these entity relations and medical entities as a structured knowledge network. In practice, the knowledge graph server 102 can define the relational types of the aforementioned clinical guideline rules based on production rules (IF-THEN rules) to generate entity relations. For example, in a treatment path: a conditional statement triggers the clinical characteristics of the treatment path (e.g., HbA1c > 7.0% and no renal insufficiency (glomerular filtration rate ≥ 45 mL / min / 1.73 mcg)). 2 Conditional or action statements explicitly define the corresponding treatment or management measures (e.g., metformin treatment, initial dose 500 mg / day). Logical connectors use natural language to connect conditional and action statements (e.g., when a patient meets the diagnostic criteria for hypertension and has diabetes, ACE inhibitors are preferred). Contraindications: Conditional statements trigger the clinical characteristics of contraindications, action statements explicitly define the corresponding constraints, and logical connectors use natural language to connect conditional and action statements (e.g., metformin is contraindicated in patients with a glomerular filtration rate <45%. Or, when glibenclamide is used in pregnant patients, the risks need to be carefully assessed). Temporal relationships: Conditional statements trigger the clinical characteristics of temporal relationships, action statements explicitly define the corresponding time requirements, and logical connectors use natural language to connect conditional and action statements (e.g., for post-operative management of breast cancer, tumor markers need to be monitored every 90 days). The above entity relationships include: treatment pathways, contraindications, and temporal relationships.

[0039] The fifth step involves performing dynamic indicator knowledge path matching based on the aforementioned dynamic indicator dataset and entity relationships to generate a dynamic indicator knowledge path mapping. In practice, the knowledge graph server 102 can create an association table between the aforementioned dynamic indicator dataset and entity relationships based on threshold judgment and time-series windows (e.g., if the blood glucose fluctuation coefficient is >25% within 30 days, the path in the entity relationship is triggered: unstable blood glucose control, medication adjustment; if the blood pressure target achievement rate is <70% within 90 days, the path in the entity relationship is triggered: blood pressure not achieved, intensified medication), to generate a dynamic indicator knowledge path mapping.

[0040] Step 6: Based on the aforementioned dynamic indicator knowledge path mapping, construct the path weight function. In practice, the knowledge graph server 102 can construct the aforementioned path weight function:

[0041]

[0042] γ=0.4×S 证据 +0.3×R 风险 +0.3×A 年龄 .

[0043]

[0044] Where p represents a diagnostic path in the knowledge graph (e.g., "blood sugar fluctuations require medication adjustment"). t (p) represents the dynamic weighting factor for path p. β represents the baseline static weight (based on the evidence level from the aforementioned pre-defined clinical guidelines, Class I = 0.9, Class II = 0.7). e -λΔt The time-dependent decay function is represented by λ = 0.03 / day. Δt represents the difference between the current time and the data time. γ represents the gain coefficient (controlling the strength of the data matching effect). S 证据 The weights represent the levels of evidence (based on the pre-defined clinical guidelines above: Class I = 1.0, Class IIa = 0.7, Class IIb = 0.4). R 风险 This represents the risk score for complications (no complications = 0.3, high risk = 1.0). A 年龄 Indicates age weight.

[0045] Step seven involves injecting time-sensitivity constraints into the aforementioned path weight function to generate a dynamic weight parameter set. The structured knowledge network and the dynamic weight parameter set are then defined as a medical knowledge graph. In practice, the knowledge graph server 102 can adjust the time-sensitivity decay function e. -λΔt The following constraints are applied to bind the validity period rules, resulting in a new path weight function, which serves as the dynamic weight parameter set:

[0046]

[0047] Among them, e -λΔt This represents the time decay function (λ = 0.03 / day). Δt represents the difference between the current time and the data time. T max Indicates the validity period of the indicator (e.g., blood glucose fluctuation T). max =90 days). δ represents the overdue penalty factor (which can be 0.7).

[0048] The aforementioned steps one through seven and related content, as an inventive point of this disclosure, solve the technical problem that "general-purpose large language models lack verified deep medical knowledge, easily generating factual illusions that contradict reality, leading to inconsistent diagnostic results and wasting computing resources." Factors leading to inconsistent diagnostic results and wasted computing resources often include: general-purpose large language models lacking verified deep medical knowledge, easily generating factual illusions that contradict reality. Solving these factors can achieve the effect of consistent diagnostic results and reduced waste of computing resources. To achieve this effect, firstly, feature semantic mapping is performed on the aforementioned structured feature vectors to generate an initial set of medical concepts. This transforms the structured feature vectors into standardized medical concepts. Secondly, entity classification and attribute binding are performed on the initial set of medical concepts to generate medical entities. This assigns types (disease / symptom / intervention / management) to medical concepts and binds clinical attributes (such as drug dosage, contraindications), forming computable medical entities. Then, rule parsing is performed on preset clinical guidelines to generate clinical guideline rules. Therefore, the natural language guideline text is parsed to generate machine-executable logical rules. Then, the clinical guideline rules are defined using relational typology to generate entity relationships, which, along with the medical entities, form a structured knowledge network. This transforms the rules into three types of entity relationships: treatment pathways, contraindications, and temporal associations, constructing the knowledge network framework. Next, based on the dynamic indicator dataset and the entity relationships, dynamic indicator knowledge path matching is performed to generate dynamic indicator knowledge path mappings. This associates real-time patient data (e.g., blood glucose fluctuations) with knowledge paths, generating personalized mappings (e.g., matching high fluctuation values ​​with insulin adjustment paths). Furthermore, based on these dynamic indicator knowledge path mappings, path weight functions are constructed. This quantifies the dynamic relevance of paths, supporting priority decisions. Finally, time-sensitivity constraints are injected into the path weight functions to generate a dynamic weight parameter set, and the structured knowledge network and dynamic weight parameter set are defined as a medical knowledge graph. This adds a time decay parameter to the weight functions, generating a dynamic weight parameter set and forming a spatiotemporally aware knowledge graph that supports real-time reasoning. Ultimately, the goal is to provide validated deep medical knowledge for large language models, ensuring that diagnostic results match reality and reducing the waste of computing resources.

[0049] The task scheduling server 103 is a server used to perform task scheduling based on the structured feature vector received from the aforementioned data integration server 101, in order to generate instructions for the division of labor in the treatment of chronic diseases at the grassroots level.

[0050] In some embodiments, the task scheduling server 103 can perform task scheduling based on the structured feature vector to generate instructions for the division of labor in the treatment of chronic diseases at the grassroots level. The task scheduling server 103 may include, but is not limited to, at least one of the following: a multi-core processor, a scheduling center core (such as Quartz, XXL-JOB), an executor cluster, and a metadata database.

[0051] Furthermore, in the process of adopting technical solutions to address the technical problems mentioned in the background, the following technical issues often arise: traditional scheduling systems struggle to respond in real-time to resource fluctuations and changes in task requirements, leading to competition for or idle computing resources. The conventional solution to these technical problems is generally to utilize traditional resource scheduling systems for resource scheduling. However, the above conventional solutions still have the following problems: they do not consider the important temporal characteristics of chronic diseases at the grassroots level, and they do not perform risk classification for specific symptoms. Considering the shortcomings of the above conventional solutions, and combining the advantages / current state of multimodal data deduplication technology possessed by the applicant's research institute partners in this field, we have decided to adopt the following solution:

[0052] Optionally, the task scheduling server 103 described above can be configured to perform the following steps:

[0053] The first step involves extracting key features from the structured feature vectors to generate static and dynamic temporal features based on primary care chronic diseases. The static features can be obtained using a pre-defined static index. The dynamic temporal features can be obtained using a pre-defined dynamic feature index. The pre-defined static index may include, but is not limited to, age, gender, and number of comorbidities. The pre-defined dynamic feature index may include, but is not limited to, blood glucose fluctuation coefficient and blood pressure target achievement rate.

[0054] The second step involves determining the risk values ​​of the aforementioned static and dynamic time-series features based on a pre-defined classifier, thereby identifying the risk values ​​for chronic diseases at the grassroots level. The pre-defined classifier is used to determine the risk values ​​of the static and dynamic time-series features according to their respective risk values. For example, it could be a random forest classifier. The number of decision trees in the random forest classifier could be 100. The feature subset sampling ratio of the random forest classifier could be... (n is the total number of features of the above static features and the above dynamic time-series features).

[0055] In practice, before classification, it is necessary to concatenate the aforementioned static features and dynamic time-series features to dynamically adjust the weights and resolve feature conflicts (e.g., low risk in static features but high risk in dynamic features).

[0056] Fusion characteristics = η·V 静态 +(1-η)V动态 .

[0057] Where η represents the static feature weight (the default value can be 0.6, but it can be reduced to 0.4 for critically ill patients). V 静态 This represents the aforementioned static characteristics. V 动态 This represents the aforementioned dynamic temporal characteristics.

[0058] The task scheduling server 103 can input the aforementioned fused features into the aforementioned random forest classifier and determine the risk value of chronic diseases at the grassroots level according to the following formula:

[0059]

[0060] Where K represents the number of decision trees, and k represents the decision tree index. T k This represents the classification result of the k-th decision tree (which can be low risk or high risk). This indicates an indicator function (outputs 1 if the condition is met, otherwise 0).

[0061] The third step is to classify the aforementioned chronic disease risk values ​​at the grassroots level into risk levels to generate grassroots chronic disease risk labels. These labels can include: low risk (e.g., 0 ≤ grassroots chronic disease risk value < S1); medium risk (e.g., S1 ≤ grassroots chronic disease risk value < S2); and high risk (e.g., S2 ≤ grassroots chronic disease risk value ≤ 1).

[0062] S1 = 0.3 + 0.1 × C complication .

[0063] S² = 0.7 - 0.05 × C age .

[0064] Where S1 represents the low-risk and medium-risk classification value. S2 represents the medium-risk and high-risk classification value. C complication This represents the risk weight for the number of complications (e.g., no complications = 0, ≥3 complications = 1). C age This represents age-related risk weights (e.g., for ages < 60, C). age =0. Age ≥ 80 years, C age =1).

[0065] The fourth step involves constructing a risk resource mapping table based on the aforementioned risk level labels for chronic diseases at the grassroots level, and determining the allocation of task leadership according to this table. In practice, this allocation can be achieved through attention distribution. For low-risk cases, the attention distribution can be 95% for the second model and 5% for the first model, representing the second model executing automated tasks while the first model monitors for anomalies in the background. For medium-risk cases, the attention distribution can be 30% for the second model and 70% for the first model, representing the first model leading core decisions (diagnosis or medication adjustment) while the second model synchronously generates execution plans (education or monitoring). For high-risk cases, the attention distribution can be 100% for both the second and first models, representing the first model exclusively utilizing diagnostic resources while the second model executes support tasks in parallel. The aforementioned grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement simultaneously calls upon all functional modules of both the first and second models to participate in the work. The second model can be a large language model in a multi-agent system (MAS) responsible for generating tasks such as routine follow-up reminders and health education (e.g., the nurse agent in Agent Hospital). The first model mentioned above can be a large language model responsible for diagnostic analysis and treatment plan adjustment in a medical multi-agent system (MAS) (e.g., the physician agent in an Agent Hospital). The attention mentioned above can be used to characterize the scheduling strategy of computing resources and decision weights in the aforementioned primary chronic disease management system based on dynamic multi-model and knowledge enhancement.

[0066] The fifth step involves generating resource allocation coefficients based on the aforementioned risk level labels for chronic diseases at the grassroots level, to allocate computing power. Here, computing power represents the relative quota of computing resources allocated to a single patient's graphics processing unit (GPU). For example, the proportion of computing power resources allocated to low-risk patients could be 15%. The proportion of computing power resources allocated to medium-risk patients could be 30%. The proportion of computing power resources allocated to high-risk patients could be 80%.

[0067] Step 6: Based on the aforementioned task leadership allocation and the aforementioned grassroots chronic disease risk level labels, a dual-channel task flow is constructed to generate grassroots chronic disease management task assignment instructions. These instructions include a second model-led instruction, a first model-led instruction, and a collaborative-led instruction. These instructions can be used to instruct the first and second models in the decision generation server 104 to collaborate, and for allocating computing resources (GPUs). The second model-led instruction corresponds to the aforementioned grassroots chronic disease risk level label and is used to characterize low risk. The first model-led instruction corresponds to the aforementioned grassroots chronic disease risk level label and is used to characterize medium risk. The collaborative-led instruction corresponds to the aforementioned grassroots chronic disease risk level label and is used to characterize high risk.

[0068] The first to sixth steps and related content described above, as an inventive point of this disclosure, solve the technical problem that "traditional scheduling systems struggle to respond in real-time to resource fluctuations and changes in task requirements, leading to competition for or idle computing resources." Factors causing competition for or idle computing resources often include the inability of traditional scheduling systems to respond in real-time to resource fluctuations and changes in task requirements. Solving these factors can achieve the effect of rational utilization of computing resources. To achieve this effect, firstly, key features are extracted from the structured feature vector to generate static and dynamic temporal features. This allows for the extraction of the features of interest. Secondly, based on a preset classifier, risk values ​​are determined for the static and dynamic temporal features to identify the risk values ​​of chronic diseases at the grassroots level. This allows for the effective fusion and quantification of the static and dynamic temporal features. Then, the risk values ​​of chronic diseases at the grassroots level are classified into levels to generate risk level labels. This enables task risk level classification. Finally, based on the risk level labels, a risk resource mapping table is constructed, and the task leadership for handling chronic diseases at the grassroots level is determined according to the risk resource mapping table. Therefore, attention can be dynamically allocated based on different task types. Then, based on the aforementioned grassroots chronic disease risk level labels, resource allocation coefficients are generated to allocate computing power. This allows for real-time adjustment of resource supply according to dynamic changes in task risk. Finally, based on the aforementioned task leadership attribution and grassroots chronic disease risk level labels, a dual-channel task flow is constructed to generate grassroots chronic disease treatment division instructions. This allocates independent resource pools to different channels, avoiding resource contention. Ultimately, this achieves the effect of rational utilization of computing resources and reduced computational redundancy.

[0069] The decision generation server 104 is a server used to drive the collaboration between the first model and the second model to generate patient treatment information based on the division of labor instructions for primary chronic disease treatment, dynamic weight parameter set and medical knowledge graph received from the knowledge graph server 102 and the task scheduling server 103.

[0070] In some embodiments, the decision generation server 104 can drive the collaboration between the first model and the second model to generate patient treatment information based on the aforementioned grassroots chronic disease management task allocation instructions, the aforementioned dynamic weight parameter set, and the aforementioned medical knowledge graph. The decision generation server 104 may include, but is not limited to, at least one of the following: a rule engine (e.g., Drools), a distributed database, a distributed task scheduler (e.g., Airflow / XXL-JOB), and a dynamic resource allocator. The patient treatment information includes grassroots chronic disease target treatment templates, intervention measures, and time-series early warnings.

[0071] Furthermore, in the process of adopting technical solutions to address the technical problems mentioned in the background, the following technical issues often arise: existing multi-agent frameworks are mostly of general design, and their role allocation and collaboration mechanisms are not optimized for the relatively rigid and highly standardized workflows in grassroots chronic disease management. This results in the system being unable to allocate model resources for specific tasks, wasting computing resources. The conventional solution to these technical problems is generally to perform diagnostic tasks based on a general-designed multi-agent framework. However, the conventional solution still suffers from the following problem: its role allocation and collaboration mechanisms are not optimized for the relatively rigid and highly standardized workflows in grassroots chronic disease management, resulting in the system being unable to allocate model resources for specific tasks, wasting computing resources. Considering the shortcomings of the conventional solution and combining the advantages / current state of multimodal data deduplication technology possessed by the applicant's research institute partners in this field, we have decided to adopt the following solution:

[0072] Optionally, the decision generation server 104 described above can be configured to perform the following steps:

[0073] In response to the detection of the aforementioned primary care chronic disease management division of labor instruction characterizing a collaborative leadership instruction, the following sub-steps are executed to generate patient treatment information:

[0074] The first sub-step involves generating a candidate diagnosis set and diagnosis probabilities based on the first model and the aforementioned structured feature vectors. In practice, the decision generation server 104 can input the aforementioned structured feature vectors into the aforementioned first model, which will then perform chained diagnostic reasoning and output the candidate diagnosis set and diagnosis probabilities.

[0075] D={d i}

[0076] P LLM (d i |E).

[0077] Where D represents the candidate diagnostic set. i Let P represent the i-th candidate diagnosis (medical entity). Let i represent the candidate diagnosis number. Let E represent the structured feature vector. LLM (d i |E) represents the generation of candidate diagnoses d under the condition of structured feature vector E. i The probability of diagnosis.

[0078] The second sub-step involves determining the total path-weighted score of the candidate diagnostic set based on the aforementioned dynamic weight parameter set. This total path-weighted score can be calculated using the following formula:

[0079]

[0080] Among them, W(d) i ) represents the diagnostic hypothesis d i The total path weighted score. φ(p) represents the basic static score of the path (based on the pre-set medical evidence strength), ranging from [0,1]. p represents the knowledge graph path (entity relationship). P(d i ) indicates the relationship with d i The set of all related knowledge graph paths. t (p) represents the dynamic weight parameter of path p.

[0081] The third sub-step involves determining the final confidence level based on the total path-weighted score and the diagnostic probability described above. This final confidence level can be achieved using the following formula:

[0082]

[0083] Among them, P final (d i |E, μ, ρ) represent the diagnostic hypothesis d i The final confidence level ranges from [0, 1]. `MinMaxScaler()` represents normalization. `μ` represents the probability gating coefficient (controlling the contribution strength of the original LLM probability), ranging from [0.5, 1.0]. `ρ` represents the weighted scaling parameter (controlling the weighted scaling parameter of W(d)). i The impact of the confidence level is [0.1, 0.3].

[0084] The fourth sub-step involves determining the primary care chronic disease target treatment template associated with the aforementioned candidate diagnosis set, based on the final confidence level and the aforementioned medical knowledge graph. In practice, the first model selects a structured knowledge network (which may include medical entities and entity relationships) from the aforementioned medical knowledge graph that is associated with the diagnostic hypothesis with the highest final confidence level in the candidate diagnosis set, as the primary care chronic disease target treatment template.

[0085] The fifth sub-step involves generating intervention measures based on the second model, the taboo rule filtering mechanism, and the temporal compatibility check mechanism, using the aforementioned primary care chronic disease target treatment template. In practice, firstly, the decision generation server 104 inputs the primary care chronic disease target treatment template into the second model and filters it using the taboo rule filtering mechanism to obtain the first part of the intervention measures (e.g., if the patient's glomerular filtration rate is <45%, replace insulin with an SGLT2 inhibitor). Secondly, the temporal compatibility check mechanism checks the primary care chronic disease target treatment template to set its validity period, resulting in the second part of the intervention measures. Finally, the first and second parts of the intervention measures are used as the intervention measures. The taboo rule filtering mechanism can be a mechanism that dynamically intercepts high-risk treatment plans through a preset taboo rule library to ensure that the intervention measures comply with the patient's individual taboos. The preset taboo rule library can be a rule library composed of the ADA diabetes treatment standards and the national medical insurance drug catalog, used to avoid conflicting treatment plans. The aforementioned time-compatibility check mechanism is a mechanism that verifies the temporal logical feasibility of intervention measures based on pharmacokinetics and physiological rhythms.

[0086] The sixth sub-step involves using the time-series prediction model and the aforementioned intervention measures to perform time-series prediction on the time-series monitoring data, thereby generating a time-series warning. In practice, firstly, the decision generation server can convert the aforementioned intervention measures (e.g., drug dosage and exercise intensity) into feature vectors and concatenate them with the aforementioned time-series monitoring data to obtain a two-dimensional vector. Secondly, the input layer of the aforementioned time-series prediction model is dimensionally expanded to conform to the aforementioned two-dimensional vector. Thirdly, the aforementioned two-dimensional vector is input into the input layer of the aforementioned time-series prediction model. The hidden layer of the aforementioned time-series prediction model can extract time-series features and retain long-term dependencies. Then, through a fully connected layer, the abstract features can be transformed into specific task results, outputting predicted values. Finally, threshold judgment is performed on the aforementioned predicted values ​​to generate a time-series warning (e.g., predicting blood glucose < 4.0 mmol / L triggers a hypoglycemia warning; predicting blood glucose > 13.9 mmol / L for 2 hours triggers a hyperglycemia warning). The aforementioned time-series detection model can be a model that predicts the future physiological indicator trends of patients and identifies potential risks in advance by analyzing historical and real-time time-series monitoring data of patients and combining them with current intervention measures. For example, the aforementioned time series prediction model could be a Long Short-Term Memory (LSTM) network.

[0087] The seventh sub-step involves generating patient treatment information based on the aforementioned primary care chronic disease target treatment template, the aforementioned intervention measures, and the aforementioned time-series early warning. In practice, the aforementioned decision generation server 104 can adjust the corresponding medical entities and entity relationships in the aforementioned primary care chronic disease target treatment template according to the aforementioned intervention measures and the aforementioned time-series early warning to generate patient treatment information. The aforementioned decision generation server 104 can send the aforementioned patient treatment information to the aforementioned user interaction terminal 105 for display.

[0088] The first to seventh sub-steps and related content described above, as an inventive point of this disclosure, solve the technical problem that "existing multi-agent frameworks are mostly general designs, and their role allocation and collaboration mechanisms are not optimized for the relatively fixed and highly standardized workflows in grassroots chronic disease management, resulting in the system's inability to allocate model resources for specific tasks and wasting computing resources." The factors that cause the system to be unable to allocate model resources for specific tasks and waste computing resources are often as follows: existing multi-agent frameworks are mostly general designs, and their role allocation and collaboration mechanisms are not optimized for the relatively fixed and highly standardized workflows in grassroots chronic disease management. If these factors are resolved, the system can allocate model resources for specific tasks, reducing the waste of computing resources. To achieve this effect, firstly, based on the first model and the aforementioned structured feature vectors, a candidate diagnosis set and diagnosis probability are generated. This avoids the limitations of a single diagnosis and covers the possibilities of differential diagnosis. Secondly, based on the aforementioned dynamic weight parameter set, the total path weighted score of the aforementioned candidate diagnosis set is determined. This quantifies the priority of multidisciplinary collaborative paths and improves the reliability of complex disease diagnosis. Then, based on the total path weighted score and the diagnostic probability, the final confidence level is determined. Therefore, combining the diagnostic probability and the interdisciplinary path score, a comprehensive confidence index is generated to screen for highly reliable diagnoses. Next, based on the final confidence level and the medical knowledge graph, a primary care chronic disease target treatment template associated with the candidate diagnosis set is determined. This achieves a combination of standardized treatment and personalized needs. Then, based on the second model, the contraindication rule filtering mechanism, and the temporal compatibility check mechanism, intervention measures are generated for the primary care chronic disease target treatment template. This mitigates treatment risks and ensures the safety of the treatment plan. Subsequently, based on the temporal prediction model and the intervention measures, temporal prediction is performed on the temporal monitoring data to generate temporal early warnings. This proactively prevents complications and optimizes the treatment pace. Finally, based on the primary care chronic disease target treatment template, the intervention measures, and the temporal early warnings, patient treatment information is generated. The decision generation server 104 can send the patient treatment information to the user interaction terminal 105 for display. This provides an operable end-to-end management path for primary care chronic disease treatment. Ultimately, the system can allocate resources according to specific tasks, reducing the waste of computing resources.

[0089] User interaction terminal 105 is a terminal used to convert patient treatment information received from the decision generation server 104 into a preset format and display it on the terminal.

[0090] In some embodiments, the user interaction terminal 105 can convert the patient treatment information into a preset format and display it on the terminal. The preset format may include, but is not limited to, text and audio.

[0091] The conflict resolution server 106 is a server used to resolve conflicts in patient treatment information received from the decision generation server 104, based on a weighted voting method, to generate a conflict resolution result.

[0092] In some embodiments, the conflict resolution server 106 can, in response to conflicts existing in the patient treatment information, perform conflict resolution based on a weighted voting method to generate a conflict resolution result. In practice, the conflict resolution server 106 can perform conflict detection on the patient treatment information based on a preset conflict rule base to generate a conflict resolution result, and then send the conflict resolution result to the user interaction terminal 105 for display. The conflict resolution server 106 may include, but is not limited to, at least one of the following: a stream processing engine (such as Kafka / Pulsar), a coordination service (such as ZooKeeper / etcd), and a database (such as PostgreSQL / Redis). The preset conflict rule base can be composed of industry standards (such as the FDA Adverse Event Reporting System, Micromedex Decision Support Library), clinical guidelines (such as the NCCN Clinical Practice Guidelines for Malignant Tumors), and a historical conflict case database, serving as a rule base to avoid conflicting treatment plans. The weighted voting method is as follows:

[0093]

[0094] Where u represents the candidate protocol number (e.g., Protocol 1: Increase insulin; Protocol 2: Adjust diet). v represents the model role number (v=1: First model; v=2: Second model). n represents the total number of model roles. G uv P represents the confidence level of model v in solution u. final (u|E,μ,ρ) represents the candidate diagnosis d corresponding to scheme u. i The final confidence level. μ represents the probability gating coefficient, ranging from [0.5, 1.0]. ρ represents the weighted scaling parameter, ranging from [0.1, 0.3]. H v The weights of model role v are represented (H1 = 0.7 for the first model, H2 = 0.3 for the second model). The final solution represents the conflict resolution result.

[0095] The various embodiments of this disclosure have the following beneficial effects: The grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement, as described in some embodiments of this disclosure, utilizes crucial temporal information in chronic disease management to provide a medically validated medical knowledge graph and diagnostic process for a large language model, ensuring that diagnostic results align with reality and reducing the waste of computational resources. Specifically, the reason for discrepancies between diagnostic results and reality, leading to wasted computational resources, lies in the fact that general-purpose large language models lack validated deep medical knowledge, easily generating "factual illusions" that contradict reality, and existing knowledge graph enhancement methods ignore temporal information. Based on this, the grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement, as described in some embodiments of this disclosure, includes a data integration server, a knowledge graph server, a task scheduling server, a decision generation server, and a user interaction terminal. First, the data integration server is used to standardize grassroots chronic disease patient data to generate structured feature vectors and dynamic indicator datasets. This completes the data cleaning and time alignment of heterogeneous data. Secondly, the aforementioned knowledge graph server is used to construct a medical knowledge graph based on the structured feature vectors and the dynamic indicator dataset. This integrates entity relationships such as diseases, drugs, and treatment plans to form a reasonable knowledge base. Then, the aforementioned task scheduling server is used to schedule tasks based on the structured feature vectors to generate instructions for the division of labor in primary care chronic disease management. This allows for dynamic scheduling of computing resources and dynamic allocation of models based on patient risk levels, reducing waste of computing resources. Next, the aforementioned decision generation server is used to drive the collaboration between the first and second models based on the instructions for the division of labor in primary care chronic disease management, the dynamic weight parameter set, and the medical knowledge graph to generate patient treatment information. This generates diagnostic hypotheses and treatment plans that are consistent with reality. Finally, the aforementioned user interaction terminal is used to convert the patient treatment information into a preset format and display it on the terminal. This adapts the treatment information to various terminals. This implementation utilizes the crucial temporal information in chronic disease management to provide a medically validated medical knowledge graph and diagnostic process for the large language model, ensuring that diagnostic results are consistent with reality and reducing waste of computing resources.

[0096] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A primary-level chronic disease management system based on dynamic multi-model and knowledge enhancement, comprising a data integration server, a knowledge graph server, a task scheduling server, a decision generation server, and a user interaction terminal, wherein: The data integration server is used to standardize the data of patients with chronic diseases at the grassroots level to generate structured feature vectors and dynamic indicator datasets. The data of patients with chronic diseases at the grassroots level includes: electronic health records, self-reported texts of patients with chronic diseases at the grassroots level, and time-series monitoring data. The data integration server is configured to: Entity recognition and timestamp alignment are performed on the self-reported text of patients with chronic diseases at the grassroots level to generate structured self-reported text; The electronic health record and the structured self-described text are preprocessed to generate a structured feature vector; Determine the fluctuation coefficient of the time-series monitoring data to generate a dynamic indicator dataset, including: The time-series monitoring data are sampled at a uniform frequency, and the fluctuation coefficient is determined using the following formula: ； ； in, Indicates standard deviation, This represents the mean. Indicates trend weight. Indicates trend factor; The knowledge graph server is used to construct a medical knowledge graph based on the structured feature vectors and the dynamic indicator dataset received from the data integration server. The medical knowledge graph includes a structured knowledge network for chronic diseases at the grassroots level and a dynamic weight parameter set based on the time-series monitoring data. The knowledge graph server is configured to: The structured feature vectors are subjected to feature semantic mapping to generate an initial set of medical concepts; The initial set of medical concepts is classified and its attributes are bound to generate medical entities, which include: disease entities, symptom entities, intervention entities, and management entities. The pre-defined clinical guidelines are parsed to generate clinical guideline rules; The clinical guideline rules are defined by relational typology to generate entity relations. The entity relations and the medical entities are used as a structured knowledge network. The entity relations include: treatment pathways, contraindications, and temporal associations. Based on the dynamic indicator dataset and the entity relationship, dynamic indicator knowledge path matching is performed to generate dynamic indicator knowledge path mapping. Based on the dynamic indicator knowledge path mapping, a path weight function is constructed; Time-sensitivity constraints are injected into the path weight function to generate a dynamic weight parameter set, and the structured knowledge network and the dynamic weight parameter set are determined as a medical knowledge graph. The task scheduling server is used to perform task scheduling based on the structured feature vector received from the data integration server, so as to generate instructions for the division of labor in the treatment of chronic diseases at the grassroots level. The decision generation server is used to drive the collaboration between the first model and the second model based on the division of labor instructions for primary chronic disease treatment, the dynamic weight parameter set, and the medical knowledge graph received from the knowledge graph server and the task scheduling server, to generate patient treatment information, wherein the patient treatment information includes primary chronic disease target treatment templates, intervention measures, and time-series early warnings; The user interaction terminal is used to receive the patient treatment information from the decision generation server, convert it into a preset format, and display it on the terminal.

2. The system according to claim 1, wherein, The grassroots chronic disease management system based on dynamic multi-model and knowledge enhancement also includes: a conflict resolution server; and The conflict resolution server is used to respond to conflicts in the patient treatment information by performing conflict resolution based on a weighted voting method to generate a conflict resolution result.

3. The system according to claim 1, wherein, The data integration server is further configured to: Multi-level semantic segmentation was performed on the self-reported texts of patients with chronic diseases at the grassroots level to generate a set of segmented phrases; Based on a preset trigger word library, keyword recognition is performed on the segmented phrase set to obtain an initial keyword set; Based on a preset recognition model, the initial keyword set is used to perform initial entity recognition in order to generate initial entities; Post-processing operations are performed on the initial entity to generate a structured readme text.

Images