Chronic disease intervention safety detection system and method based on large model multi-agent cooperation

Abstract

The invention discloses a chronic disease intervention safety detection system and method based on large model multi-agent cooperation, and relates to the technical field of natural language processing and multi-agent cooperation in the medical health field. A standardized interface and an adaptive cleaning algorithm are adopted to realize data desensitization, missing value filling and format unification, and a dynamically updated patient health portrait is constructed; complex tasks are decomposed based on a chain reasoning technology, special tools such as a drug interaction detection tool and a nutrition gap calculation engine are developed, and a clinical knowledge base is integrated for parallel analysis; suggestion conflicts are eliminated through a multi-agent debate mechanism, the priority is calculated in combination with a weight rule, manual auditing is triggered to process low-confidence disputes, and finally a structured health report containing medication adjustment, diet optimization and behavior intervention is generated. The limitation of a traditional method in data integration, cross-domain reasoning and conflict resolution is solved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and multi-agent collaboration technology in the medical and health field, and more specifically to a chronic disease intervention safety detection system and method based on large-model multi-agent collaboration. Background Art

[0002] With the continuous advancement of medical informatization, the variety and volume of medical data are exploding. Multi-source data, such as electronic health records (EHRs), wearable devices, and user-uploaded records of diet and lifestyle habits, provides a rich foundation for personalized health management. However, this data often exists in heterogeneous forms and lacks a unified structure and semantics, posing significant challenges to data integration and analysis.

[0003] Existing technologies have the following main problems when processing multi-source health data:

[0004] Data integration is complex: Different data sources provide information in different formats, such as structured data in electronic medical records, dietary logs in the form of pictures uploaded by users, and sensor data from wearable devices, lacking effective standardized processing methods.

[0005] Limited reasoning ability: Traditional methods usually rely on a single model or fixed rules for reasoning, and it is difficult to integrate knowledge from multiple fields (such as drug interactions, nutritional standards, the impact of lifestyle habits, etc.) for in-depth analysis.

[0006] Lack of collaborative mechanisms: Agents in existing systems typically work independently, lack effective collaboration and debate mechanisms, and are unable to dynamically adjust and optimize conclusions during the reasoning process.

[0007] Insufficient tool development and integration: Existing methods have limitations in the development of dedicated tools and integration of external knowledge bases, and cannot efficiently support the generation of health recommendations in multiple fields.

[0008] Therefore, how to provide a chronic disease intervention safety detection system and method based on large-scale model multi-agent collaboration is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0009] In view of this, the present invention provides a chronic disease intervention safety detection system and method based on large-scale model multi-agent collaboration. Through multi-agent collaboration and dynamic knowledge enhancement, it solves the limitations of traditional methods in data integration, cross-domain reasoning and conflict resolution, and significantly improves the scientificity, safety and operability of chronic disease management programs.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] A chronic disease intervention safety detection system based on large-scale model multi-agent collaboration, including:

[0012] Data collection and standardization processing module: used to obtain and clean multi-source health data and generate patient health portraits;

[0013] Task decomposition and chain reasoning module: Based on the patient's health profile, data integrity is verified. Through chain reasoning, complex tasks are decomposed into subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification. The task logic is verified using a medical knowledge base, and conflicts are resolved through a multi-agent negotiation mechanism to generate a standardized task list.

[0014] Dedicated tool construction and knowledge base integration module: used to build dedicated tools for drugs, diet, and habits based on the standardized task list, integrate drug interaction maps and nutrition databases, encapsulate them into containerized microservices, and deploy them to the cloud platform;

[0015] Data analysis and recommendation generation module: This module concurrently calls the dedicated tools to perform drug safety assessment, nutritional gap calculation, and habit risk identification, and generates preliminary recommendations based on clinical rules;

[0016] Multi-agent debate and conclusion revision module: used to identify contradictions in the preliminary suggestions, resolve conflicts through multi-agent debate, and generate structured consensus conclusions;

[0017] Report generation and semantic description module: Based on the consensus conclusions, multi-field recommendations are merged and prioritized, and standardized medical documents containing data traceability and reasoning paths are generated through semantic modeling, and output as machine-readable PDF reports.

[0018] Preferably, the data collection and standardization processing module has the following specific processing steps:

[0019] Real-time access to medication records through standardized interfaces;

[0020] Use optical character recognition technology to parse user-uploaded food diary images and combine them with a nutrition database to quantify nutritional content;

[0021] Synchronize exercise and sleep data from wearable devices through the application program interface and integrate user-entered lifestyle records;

[0022] Perform data cleaning, missing value filling, and outlier detection on the collected medication records, nutritional composition, and lifestyle records;

[0023] After pre-processing, a hash algorithm is used to desensitize the patient's identity information and construct a semantic document containing labels for medication rules, dietary patterns, and lifestyle habits;

[0024] After desensitization, the information is aggregated and stored in a hybrid database according to patient dimensions, and a dynamically updated patient health profile is generated.

[0025] Preferably, conflicts are resolved through a multi-agent negotiation mechanism, specifically including:

[0026] Conflict detection: The rules engine flags contradictory instructions;

[0027] Pre-debate: Based on the conflict detection results, agents in each field submit evidence and vote;

[0028] Task reconstruction: Generate a new priority-weighted task list based on the pre-debate results.

[0029] Preferably, the special tool includes:

[0030] Drug analysis tools: realize drug contraindication detection, dosage safety range verification and medication timing optimization;

[0031] Dietary assessment tools: Calculate the difference between nutritional intake and disease management goals, screen for allergens, and generate personalized recipes;

[0032] Habit monitoring tools: assessing the compatibility of exercise intensity with disease states, analyzing sleep quality patterns, and identifying unhealthy behaviors.

[0033] Preferably, the data analysis and suggestion generation module is further used to dynamically fill in the standardized suggestion template and cache intermediate data.

[0034] Preferably, the multi-agent debate includes: organizing multi-domain executors to conduct three rounds of debate and evidence update, combining weight algorithm to calculate priority, and triggering manual review mechanism.

[0035] Preferably, organize multi-domain executors to conduct three rounds of debate and evidence update, calculate priorities using a weighting algorithm, and trigger a manual review mechanism, including:

[0036] First round of presentation: Each executor submits their claims and supporting evidence;

[0037] Cross-examination: executors question and refute each other's claims;

[0038] Evidence update: Re-verify dispute points using specialized tools in the field and dynamically update evidence content;

[0039] Among them, during the debate process, the medical semantic network is connected in real time to obtain the latest clinical evidence to supplement or correct the original evidence;

[0040] The priority of each suggestion is evaluated using the evidence collected above. When the difference in scores between multiple suggestions is less than the preset threshold, the automatic decision is frozen and a manual review is triggered.

[0041] Preferably, the semantic modeling includes:

[0042] Use JSON-LD to define the patient-recommendation relationship and cross-recommendation association logic to generate structured data;

[0043] Design a clinical document template that includes an executive summary, detailed protocol, and appendices;

[0044] Convert structured data into clinical documents based on clinical document templates, and support automatic matching of multi-language versions and machine-readable parsing.

[0045] A method for detecting the safety of chronic disease interventions based on large-scale multi-agent collaboration, including:

[0046] Acquire and clean multi-source health data to generate patient health profiles;

[0047] Based on the patient's health profile, data integrity is verified. Chain reasoning is used to decompose complex tasks into subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification. The task logic is verified using a medical knowledge base, and conflicts are resolved through a multi-agent negotiation mechanism to generate a standardized task list.

[0048] Build dedicated tools for drugs, diet, and habits based on the standardized task list, integrate drug interaction maps and nutrition databases, encapsulate them as containerized microservices, and deploy them to the cloud platform;

[0049] The dedicated tools are used in parallel to perform drug safety assessment, nutritional gap calculation, and habit risk identification, and preliminary recommendations are generated in combination with clinical rules;

[0050] Identify contradictions in the initial proposals, resolve conflicts through multi-agent debate, and generate structured consensus conclusions;

[0051] Based on the consensus conclusions, multi-field recommendations are merged and prioritized, and standardized medical documents containing data traceability and reasoning paths are generated through semantic modeling and output as machine-readable PDF reports.

[0052] It can be seen from the above technical solutions that compared with the existing technology, the present invention discloses a chronic disease intervention safety detection system and method based on large-scale model multi-agent collaboration, which realizes comprehensive analysis of multi-source health data and generation of personalized health recommendations through multi-agent collaboration. The system has shown significant advantages in data collection, task decomposition, tool construction, data analysis, multi-agent debate and report generation. Through standardized processing and chain reasoning, the system ensures the accuracy of the data and the depth of analysis. The multi-agent debate mechanism effectively eliminates conflicts and improves the reliability of recommendations. The structured health report finally generated provides users with clear and practical health management guidance, significantly improving the scientificity and practicality of health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] 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 merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0054] Figure 1 This is a schematic diagram of the structure of the chronic disease intervention safety detection system based on large-model multi-agent collaboration provided by the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] The embodiment of the present invention discloses a chronic disease intervention safety detection system based on large model multi-agent collaboration, such as Figure 1 Shown, including:

[0057] Data collection and standardization processing module: used to obtain and clean multi-source health data and generate patient health portraits;

[0058] Task decomposition and chain reasoning module: Verify data integrity based on patient health profiles. Use chain reasoning to decompose complex tasks into subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification. Use the medical knowledge base to verify task logic and resolve conflicts through a multi-agent negotiation mechanism to generate a standardized task list.

[0059] Dedicated tool building and knowledge base integration module: This module is used to build specialized tools for drugs, diets, and habits based on standardized task lists, integrating drug interaction maps and nutritional databases. These include the pharmaceutical analysis module, nutritional calculation engine, and behavioral pattern analysis module mentioned below, packaged as containerized microservices, and deployed on the cloud platform.

[0060] Data analysis and recommendation generation module: This module uses dedicated tools in parallel to perform drug safety assessments, calculate nutritional gaps, and identify habitual risks, and generates preliminary recommendations based on clinical rules.

[0061] Multi-agent debate and conclusion revision module: used to identify contradictions in preliminary suggestions, resolve conflicts through multi-agent debate, and generate structured consensus conclusions;

[0062] Report generation and semantic description module: Based on consensus conclusions, it merges multi-field recommendations and prioritizes them. Through semantic modeling, it generates standardized medical documents containing data traceability and reasoning paths, and outputs them as machine-readable PDF reports.

[0063] The following describes the specific implementation process of each module:

[0064] (1) Data collection and standardization processing module

[0065] Connect to the hospital system in real time through a standardized interface to obtain medication records, including key information such as drug name, dosage, and medication time;

[0066] Use optical character recognition technology to analyze user-uploaded diet log images and combine with authoritative nutrition databases to quantitatively analyze nutrients such as calories, protein, and sodium content;

[0067] Synchronize exercise and sleep data from wearable devices through the application program interface and integrate user-entered lifestyle records;

[0068] The collected medication records, nutritional composition and living habits records are cleaned, missing values ​​are filled, and outlier detection is performed; it can also include: standardization of the collected heterogeneous data: filling missing values ​​based on historical behavior patterns, and predicting rationality parameters through machine learning models; defining unified data structure specifications to ensure the standardization of field formats such as time and numbers; using automated processes to convert data formats, and combining outlier detection algorithms to identify abnormal data, and correcting them after manual review.

[0069] A hash algorithm is used to desensitize patient identity information, and encrypted association keys are retained to achieve cross-source data matching; the cleaned data is aggregated according to patient dimensions to construct semantic documents containing medication rules, dietary patterns and lifestyle habit tags, supporting intelligent retrieval based on knowledge graphs; a hybrid storage architecture is used to manage data, with unstructured text stored in a highly scalable document library and structured data stored in a relational database, and distributed indexing technology is used to achieve fast multi-condition joint queries, ultimately generating a dynamically updated patient health portrait.

[0070] (2) Task decomposition and chain reasoning module

[0071] Data integrity is verified based on health profiles, and chain reasoning is used to decompose subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification. Task logic is verified using SPARQL queries against the clinical knowledge base, and task conflicts are resolved through a multi-agent negotiation mechanism. Ultimately, a standardized task list is generated and pushed to downstream modules via a message queue.

[0072] Specifically: (201) Input data:

[0073] Patient health profile (including medication records, dietary quantitative indicators, and lifestyle habit data).

[0074] (202) Data integrity verification phase, including:

[0075] 1) Define structured validation rules: clarify required fields (e.g., drug name, dosage unit, dietary calorie value) and data format standards (e.g., timestamp format, value range);

[0076] 2) Verify the integrity of basic fields layer by layer, including:

[0077] Hierarchical validation: Checks whether basic fields are missing;

[0078] Logical verification: Verify the rationality of the association between data.

[0079] (203) Task triggering and decomposition phase

[0080] Task template triggering: Based on health profile features, automatically match predefined task templates and generate initial task instruction sets according to preset clinical rules:

[0081] Task 1: Compare current medication combinations with disease contraindications and identify potential conflicts;

[0082] Task 2: Analyze the nutritional composition of the diet and assess the degree of deviation from the target disease management standards;

[0083] Task 3: Detect the negative impact of lifestyle habits on treatment outcomes;

[0084] Knowledge base collaborative verification: Search the medical knowledge base through semantic query language to verify the feasibility of the task;

[0085] (204) Task conflict detection and negotiation mechanism

[0086] If the subtask logic conflicts, the multi-agent negotiation mechanism is triggered:

[0087] Conflict detection: The rules engine flags contradictory instructions;

[0088] Pre-debate: Agents from each field present evidence and vote;

[0089] Task Refactoring: Generate a new, prioritized, and standardized list of tasks.

[0090] Also includes:

[0091] (205)Task structured encapsulation and task scheduling

[0092] Task structured definition: Encapsulate subtasks into semantic data units, including: task type; detection target; associated data fields; judgment threshold; output weighted task list and execution dependency graph.

[0093] Input a structured task list and prioritize the tasks: calculate the execution order based on clinical urgency and data confidence;

[0094] Task channel allocation: push medication tasks to the pharmaceutical analysis module; push diet tasks to the nutrition calculation engine; push habit tasks to the behavior pattern analysis module;

[0095] Execution status monitoring: records the task distribution time, receiving module, and expected completion time; if the task times out without feedback, triggers a retry or manual intervention process.

[0096] This step ensures input data quality through rule-driven validation, leverages knowledge-enhanced reasoning to decompose executable subtasks, and ensures logical consistency through conflict resolution. Task definitions clearly define detection objectives, data dependencies, and judgment criteria, forming an end-to-end traceable decision chain and providing standardized input for subsequent tool invocation and recommendation generation.

[0097] (3) Special tool construction and knowledge base integration module

[0098] We built specialized tools for medication, diet, and habits based on subtask requirements. We integrated knowledge sources like drug interaction maps and nutritional databases, packaged them into containerized microservices, and deployed them on a cloud platform. We provided functionality like dosage verification, allergen matching, and sleep pattern analysis through a REST API, automatically generating interactive documentation to support invocation instructions.

[0099] Specifically, (301) input data

[0100] Standardized task list (medication analysis, diet assessment, habit monitoring).

[0101] Tool Capability Mapping:

[0102] (302) Build tools for drugs, diet, and habits

[0103] Drug analysis tools: realize drug contraindication detection, dosage safety range verification, and medication timing optimization; among them, the drug contraindication detection function identifies conflicting drug combinations based on the drug interaction map; the dosage safety range verification function dynamically adjusts the medication threshold based on the patient's physiological indicators; the medication timing optimization function adapts to the rules of taking before / after meals.

[0104] Dietary assessment tools: Calculate the difference between nutritional intake and disease management goals, screen for allergens, and generate personalized recipes;

[0105] Habit monitoring tools: assessing the compatibility of exercise intensity with disease states, analyzing sleep quality patterns, and identifying adverse behaviors;

[0106] (303) Interface specification design:

[0107] Define standardized input and output formats (JSON structure). All tools use the JSON format for data exchange. For example, the input of the drug tool includes the drug list and patient physiological indicators, and the output includes risk level and adjustment suggestions.

[0108] Label field constraints to ensure data compatibility across tools.

[0109] (304) Knowledge Base Integration and Modeling

[0110] Build a pharmaceutical analysis module: connect to the authoritative drug interaction database (Drugs.com InteractionChecker) to obtain drug conflict rules and dosage threshold data; build a local drug interaction map to store drug-disease contraindications, metabolic pathway associations, etc.; develop a conflict detection algorithm to support real-time risk scanning of multi-drug combinations and provide alternative medication recommendations and dosage adjustment plans; integrate pharmacogenomics data to achieve high-risk drug sensitivity warning based on patient genotype.

[0111] Develop a nutrition calculation engine: connect to a standardized nutrition database, establish a disease-nutrition association model, and support personalized nutrition rule configuration for chronic diseases such as diabetes and kidney disease; design a difference calculation module to compare the patient's daily intake with the recommended value based on the patient's age, weight, and metabolic status, and generate a nutrition deviation report (such as calories, trace elements, and dietary fiber); implement an allergen matching engine to perform risk tagging based on the ingredient list of the food and the patient's allergy history; expand the cooking method influencing factor model to evaluate the loss or conversion of nutrients during food processing.

[0112] Even better, a behavioral pattern analysis module can be developed: access the wearable device data interface to analyze exercise frequency, heart rate changes and sleep stage data; train behavioral risk assessment models; build a time series analysis algorithm to identify long-term trends and abnormal fluctuations in living habits, develop an early warning feedback mechanism, generate a behavioral intervention priority list based on the patient's health portrait, and provide a dynamic threshold calibration function.

[0113] (305) Microservice packaging:

[0114] Each tool is encapsulated as an independent service module, providing a standardized API endpoint based on the HTTP protocol. Input and output examples: the medication tool receives the patient ID and medication list, and returns conflicting drug pairs and safe medication recommendations.

[0115] Dependency and environment isolation: Use containerization technology to package the tool runtime environment (including dependency libraries and configuration files); define resource quotas to avoid resource competition between services;

[0116] Clustered deployment and operations: Deploy to a cloud-native platform, automatically scaling the number of instances based on request load; configuring health checks and failover policies to ensure high service availability; generating interactive interface documentation that automatically annotates request parameter examples, response field descriptions, and error code tables.

[0117] This step transforms task requirements into tool functions through domain capability abstraction. Detection rules and computational models are built based on a knowledge base, ultimately achieving tool standardization and flexibility through a microservices architecture. Each tool's input and output strictly adhere to clinical semantic specifications, ensuring that analysis results can be directly used in decision-making processes. Containerized deployment and automated operations ensure system stability in large-scale concurrent scenarios.

[0118] (4) Data analysis and suggestion generation module

[0119] Tools are called in parallel to perform drug safety assessments, calculate nutritional gaps, and identify habitual risks. Clinical rules are combined to generate graded recommendations, and the template engine is used to dynamically build recommendation content. Standardized results containing basis and priority are output, and intermediate data is cached to support subsequent processes.

[0120] Specifically, (401) encapsulates task routing and parameters:

[0121] Encapsulate input data according to subtask type:

[0122] Drug analysis tasks: package medication lists, patient liver and kidney function indicators, and allergy history;

[0123] Nutritional assessment tasks: encapsulate daily dietary composition tables and nutritional standards for target diseases;

[0124] Habit analysis task: integrating movement frequency data, sleep stage records, and recent physiological indicators;

[0125] Generates a standardized API request body, including the patient ID, task type, and input parameter set.

[0126] (402) Distributed Task Scheduling:

[0127] Submit three types of subtasks to the asynchronous execution queue to trigger parallel processing:

[0128] Drug safety assessment: Call the drug conflict detection service, input the medication list and metabolic data, and output high-risk drug pairs and dosage adjustment suggestions;

[0129] Nutrition compliance assessment: request nutrition calculation services, compare intake with disease standards, and generate calorie / sodium content difference reports;

[0130] Habit risk identification: Submit sleep pattern data to the behavior analysis service to identify abnormal periods and correlate with heart rate variability data to verify health impacts.

[0131] Real-time results monitoring: Monitor the execution status of each task and capture timeouts or exception errors; if a task fails, retry or transfer it to manual processing based on the strategy.

[0132] (403) The drug conflict report, nutritional difference table, and behavioral risk analysis results obtained from the above parallel processing are used as inputs to perform the following processing:

[0133] a) Data alignment and conflict detection: Align multi-dimensional results by time window; the rule engine verifies the consistency of results.

[0134] b) Dynamic content generation: Define a clinical recommendation template library, including:

[0135] Medication recommendation template: <High-risk drug> is recommended to be replaced with <Safe alternative drug>, and the dose is adjusted to <Threshold> mg;

[0136] Dietary advice template: Daily <nutrient> intake needs to be reduced by <difference>%, and <recommended ingredients> should be given priority;

[0137] Habit suggestion template: <behavior name> frequency exceeds <threshold> times / week, associated with <indicator abnormality> risk;

[0138] Match the template according to the tool output results and inject specific values ​​and sources.

[0139] c) Recommendation classification and priority calculation:

[0140] Categorize recommendations based on rules: emergency intervention, optimization and adjustment, and long-term monitoring;

[0141] The comprehensive priority was calculated according to the risk level (high / medium / low) and the strength of evidence (clinical guideline weight > model inference).

[0142] It may also include:

[0143] (404) Standardized result encapsulation:

[0144] Generate a JSON format report, including:

[0145] Recommended content: specific adjustment plan; Source: related knowledge base entries; Priority: numerical score (0-100) and classification label (urgent / optimization / monitoring); Data traceability: mark the original data fields that the dependency depends on.

[0146] Intermediate data cache:

[0147] Storing analytical process data in a low-latency storage system;

[0148] Set cache invalidation strategy to support quick call of subsequent multi-agent debates.

[0149] Output and Audit:

[0150] Push the final report to the doctor's workstation and patient terminal;

[0151] Record task execution logs for quality traceability and system optimization.

[0152] This system accelerates multi-dimensional analysis through an asynchronous parallel architecture, ensures interpretable recommendations through rule-based templating, and balances real-time performance with data integrity through a hierarchical caching mechanism. The entire process, from task dispatch to report generation, is auditable, and key decision nodes are linked to clinical evidence, providing transparent and actionable chronic disease management solutions for doctors and patients.

[0153] (5) Multi-agent debate and conclusion correction module

[0154] The rules engine identifies inconsistencies in preliminary recommendations, organizes multi-domain executors for three rounds of debate and evidence updates, and uses a weighting algorithm to calculate priorities. This triggers a manual review mechanism to resolve contentious scenarios, ultimately reaching a consensus conclusion with authoritative evidence, and storing the structured conclusion for retrospective analysis.

[0155] Specifically, (501) input data:

[0156] Preliminary recommendations (JSON data of medication adjustments, nutrition plans, and behavioral interventions).

[0157] (502) Conflicting rule base matching:

[0158] Define clinical conflict rules; the rule engine scans the suggested combinations and marks the conflicts;

[0159] Evidence tracing and correlation:

[0160] Extract supporting evidence from the analysis results: evidence of medication conflicts; citations of nutritional standards; behavioral association data;

[0161] Output a marked list of contradictions and the associated chain of evidence.

[0162] Analyze the list of contradictions, original analysis data, and external medical knowledge base.

[0163] (503) Organized three rounds of debate and evidence updates by multi-disciplinary actors, including:

[0164] a) Executor role allocation:

[0165] Medication executor: responsible for drug contraindication revalidation and dosage logic defense;

[0166] Dietary Enforcer: Maintains compliance with nutritional standards and provides alternatives;

[0167] Habit enforcers: demonstrating the clinical benefit-risk balance of behavioral recommendations;

[0168] b) Three-round debate mechanism:

[0169] First round of presentation: Each executor submits claims and evidence;

[0170] Cross-examination: executors raise questions to each other;

[0171] Evidence update: Use specialized tools in the field to re-verify controversial points;

[0172] Dynamic Knowledge Enhancement:

[0173] Real-time query of the medical semantic network to obtain the latest clinical evidence;

[0174] If external evidence overturns the original conclusion, it triggers a recommendation for real-time correction.

[0175] (504) Weighted algorithm calculates priority:

[0176] Weight rule definition:

[0177] The authority of clinical guidelines (based on the latest clinical data) accounts for 50%; the confidence of tool test results accounts for 30%; and historical data statistics account for 20%;

[0178] Comprehensive score calculation:

[0179] The priority of each suggestion is calculated using the formula:

[0180] Priority = (Guideline Weight × Matching Degree) + (Tool Weight × Risk Value) + (Historical Weight × Success Rate)

[0181] Output a list of suggestions sorted by scores;

[0182] Manual intervention trigger:

[0183] If the difference in scores between multiple suggestions is less than 5%, it is considered a "low confidence dispute";

[0184] Freeze automated decision-making and push conflicting points, evidence chains, and debate records to the manual review platform;

[0185] Record the review operations in the audit log and mark the basis for revisions.

[0186] It may also include: (505) structured consensus encapsulation:

[0187] Integrate the debate results and manual review conclusions to generate the final set of recommendations;

[0188] Label each suggested decision path:

[0189] Supporting evidence; summary of the debate process; scoring details (guideline match 40% + tool results 27% + historical data 18% = total score 85%);

[0190] Cache and traceability management:

[0191] Store the consensus conclusions and original debate data in a high-performance cache system, indexed by patient ID and timestamp;

[0192] To set a tiered storage policy:

[0193] Hot data (conclusions within 7 days) retains memory-level cache and supports real-time access;

[0194] Historical data is archived in a columnar repository for long-term trend analysis;

[0195] (506) Audit and iterative optimization:

[0196] Regularly analyze dispute cases, extract high-frequency conflict rules and feed back into the knowledge base;

[0197] Adjust the weight model based on physician feedback.

[0198] This step identifies risk points through rule-driven conflict detection, leverages a multi-role debate mechanism to achieve cross-domain decision-making collaboration, combines dynamic knowledge retrieval to enhance evidence reliability, and ultimately quantifies the priority of recommendations through a hierarchical scoring model. The entire process, from conflict identification to consensus implementation, is transparent and traceable, with manual review and automated decision-making complementing each other to ensure that the output is both clinically rigorous and practically feasible.

[0199] (6) Report generation and semantic description module

[0200] We merge and prioritize recommendations from multiple fields, enhance semantic descriptions with JSON-LD, and automatically generate standardized medical documents that include data traceability and reasoning paths. The output is a machine-readable PDF report that complies with clinical standards, completing the entire closed-loop process.

[0201] Specifically, (601) Analyze the consensus conclusions (including modification suggestions and priority scores for medication, diet, and habits)

[0202] (602) Data aggregation and conflict handling:

[0203] Call the data integration module to associate three types of recommendations: medication, diet, and habits, based on the patient's unique identifier;

[0204] Conflict detection: If there are multiple recommendations for the same health indicator, the highest-weighted recommendation is retained based on priority score and clinical urgency, and the rest are converted into annotations;

[0205] Output: Structured merge suggestion table, fields include: suggestion type, specific content, effective time, conflict resolution flag.

[0206] (603) Dynamic Priority Sorting:

[0207] Define the sorting rule level:

[0208] Priority level 1: involving life safety or contraindication correction;

[0209] Secondary priority: optimization suggestions that affect disease control;

[0210] Priority level 3: long-term health management solutions;

[0211] Generate sorting numbers by priority, sort by scores in descending order within the same level, and output a list of suggestions with graded labels.

[0212] Data source annotation:

[0213] Attach multiple layers of evidence to each suggestion:

[0214] Raw data layer: health records that are labeled for dependency;

[0215] Analytical process layer: associative reasoning paths;

[0216] Knowledge basis layer: cites clinical guidelines, research literature or knowledge base entries.

[0217] (604) Machine-readable semantic modeling:

[0218] a) Build a disease management semantic model and define core relationships:

[0219] Patient-recommendation relationship: Patient [perform] → Recommendation [based on] → Evidence [association] → Knowledge Entity;

[0220] Cross-suggestion association: mark the cooperation or exclusion logic between suggestions;

[0221] Outputs a semantically enhanced report intermediate format that supports direct parsing and reasoning by clinical systems.

[0222] b) Clinical document template design:

[0223] Define the reporting framework:

[0224] Cover page: patient name, creation date, medical institution logo;

[0225] Executive Summary: Summary of urgent recommendations;

[0226] Detailed plan: Medication, diet, and habit recommendations are presented in chapters, with priority icons and effective dates.

[0227] Appendix: Complete evidence index, dispute resolution records, data charts.

[0228] c) Multi-language support: Automatically match the template language version according to the patient's native language.

[0229] This process builds a comprehensive recommendation system through multi-dimensional data integration, utilizes semantic modeling to break down barriers to human-machine collaboration, and ultimately outputs authoritative conclusions through standardized clinical documentation. The entire transformation process, from raw data to actionable solutions, is traceable, with urgent recommendations prioritized and the evidence chain transparent and traceable, ensuring rigorous medical decision-making and actionable patient actions.

[0230] The embodiment of the present invention provides a method for detecting the safety of chronic disease intervention based on large-scale model multi-agent collaboration, comprising:

[0231] Acquire and clean multi-source health data to generate patient health profiles;

[0232] Verify data integrity based on patient health profiles, decompose complex tasks into subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification through chain reasoning, validate task logic using a medical knowledge base, and resolve conflicts through a multi-agent negotiation mechanism to generate a standardized task list.

[0233] Build specialized tools for drugs, diet, and habits for subtasks, integrate drug interaction maps and nutritional databases, encapsulate them as containerized microservices, and deploy them to the cloud platform;

[0234] Parallel use of dedicated tools for drug safety assessment, nutritional gap calculation, and habit risk identification, combined with clinical rules to generate preliminary recommendations;

[0235] Identify contradictions in preliminary proposals, resolve conflicts through multi-agent debate, and generate structured consensus conclusions;

[0236] Based on the consensus conclusions, multi-field recommendations are merged and prioritized, and standardized medical documents containing data traceability and reasoning paths are generated through semantic modeling, and output as machine-readable PDF reports.

[0237] The specific implementation process of each step of the method of the present invention is based on the above-mentioned system implementation, which will not be repeated here. Please refer to the specific implementation process of the system.

[0238] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0239] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A chronic disease intervention safety detection system based on large-scale multi-agent collaboration, characterized by: include: Data collection and standardization processing module: used to obtain and clean multi-source health data and generate patient health portraits; Task decomposition and chain reasoning module: Based on the patient's health profile, data integrity is verified. Through chain reasoning, complex tasks are decomposed into subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification. The task logic is verified using a medical knowledge base, and conflicts are resolved through a multi-agent negotiation mechanism to generate a standardized task list. Dedicated tool construction and knowledge base integration module: used to build dedicated tools for drugs, diet, and habits based on the standardized task list, integrate drug interaction maps and nutrition databases, encapsulate them into containerized microservices, and deploy them to the cloud platform; Data analysis and recommendation generation module: This module concurrently calls the dedicated tools to perform drug safety assessment, nutritional gap calculation, and habit risk identification, and generates preliminary recommendations based on clinical rules; Multi-agent debate and conclusion revision module: used to identify contradictions in the preliminary suggestions, resolve conflicts through multi-agent debate, and generate structured consensus conclusions; Report generation and semantic description module: Based on the consensus conclusions, multi-field recommendations are merged and prioritized, and standardized medical documents containing data traceability and reasoning paths are generated through semantic modeling, and output as machine-readable PDF reports.

2. The chronic disease intervention safety detection system based on large-scale model multi-agent collaboration according to claim 1 is characterized in that: Data collection and standardization processing module, the specific processing process is as follows: Real-time access to medication records through standardized interfaces; Use optical character recognition technology to parse user-uploaded food diary images and combine them with a nutrition database to quantify nutritional content; Synchronize exercise and sleep data from wearable devices through the application program interface and integrate user-entered lifestyle records; Perform data cleaning, missing value filling, and outlier detection on the collected medication records, nutritional composition, and lifestyle records; After pre-processing, a hash algorithm is used to desensitize the patient's identity information and construct a semantic document containing labels for medication rules, dietary patterns, and lifestyle habits; After desensitization, the information is aggregated and stored in a hybrid database according to patient dimensions, and a dynamically updated patient health profile is generated.

3. The chronic disease intervention safety detection system based on large-scale model multi-agent collaboration according to claim 1 is characterized in that: Resolve conflicts through a multi-agent negotiation mechanism, specifically including: Conflict detection: The rules engine flags contradictory instructions; Pre-debate: Based on the conflict detection results, agents in each field submit evidence and vote; Task reconstruction: Generate a new priority-weighted task list based on the pre-debate results.

4. The chronic disease intervention safety detection system based on large-scale model multi-agent collaboration according to claim 1 is characterized in that: The special tools include: Drug analysis tools: realize drug contraindication detection, dosage safety range verification and medication timing optimization; Dietary assessment tools: Calculate the difference between nutritional intake and disease management goals, screen for allergens, and generate personalized recipes; Habit monitoring tools: assessing the compatibility of exercise intensity with disease states, analyzing sleep quality patterns, and identifying unhealthy behaviors.

5. The chronic disease intervention safety detection system based on large-scale model multi-agent collaboration according to claim 1 is characterized in that: The data analysis and suggestion generation module is also used to dynamically fill in standardized suggestion templates and cache intermediate data.

6. The chronic disease intervention safety detection system based on large model multi-agent collaboration according to claim 1 is characterized in that: Multi-agent debate includes: organizing multi-domain executors to conduct three rounds of debate and evidence update, combining weight algorithm to calculate priority, and triggering manual review mechanism.

7. The chronic disease intervention safety detection system based on large-scale model multi-agent collaboration according to claim 6 is characterized in that: Organize multi-domain executors to conduct three rounds of debate and evidence updates, combine weighting algorithms to calculate priorities, and trigger a manual review mechanism, including: First round of presentation: Each executor submits their claims and supporting evidence; Cross-examination: executors question and refute each other's claims; Evidence update: Re-verify dispute points using specialized tools in the field and dynamically update evidence content; Among them, during the debate process, the medical semantic network is connected in real time to obtain the latest clinical evidence to supplement or correct the original evidence; The priority of each suggestion is evaluated using the evidence collected above. When the difference in scores between multiple suggestions is less than the preset threshold, the automatic decision is frozen and a manual review is triggered.

8. The chronic disease intervention safety detection system based on large-scale model multi-agent collaboration according to claim 1 is characterized in that: The semantic modeling includes: Use JSON-LD to define the patient-recommendation relationship and cross-recommendation association logic to generate structured data; Design a clinical document template that includes an executive summary, detailed protocol, and appendices; Convert structured data into clinical documents based on clinical document templates, and support automatic matching of multi-language versions and machine-readable parsing.

9. A method for detecting the safety of chronic disease intervention based on large-scale multi-agent collaboration, characterized by: include: Acquire and clean multi-source health data to generate patient health profiles; Based on the patient's health profile, data integrity is verified. Chain reasoning is used to decompose complex tasks into subtasks such as medication contraindication analysis, dietary compatibility assessment, and adverse behavior identification. The task logic is verified using a medical knowledge base, and conflicts are resolved through a multi-agent negotiation mechanism to generate a standardized task list. Build dedicated tools for drugs, diet, and habits based on the standardized task list, integrate drug interaction maps and nutrition databases, encapsulate them as containerized microservices, and deploy them to the cloud platform; The dedicated tools are used in parallel to perform drug safety assessment, nutritional gap calculation, and habit risk identification, and preliminary recommendations are generated in combination with clinical rules; Identify contradictions in the initial proposals, resolve conflicts through multi-agent debate, and generate structured consensus conclusions; Based on the consensus conclusions, multi-field recommendations are merged and prioritized, and standardized medical documents containing data traceability and reasoning paths are generated through semantic modeling and output as machine-readable PDF reports.

Images