A clinical pharmacology analysis system and method based on a multi-agent model architecture, an electronic device, a storage medium and a program product

Abstract

The application discloses a clinical pharmacological analysis system and method based on a multi-agent model architecture, an electronic device, a storage medium and a program product, and belongs to the technical field of pharmacological analysis. The technical solution points are as follows: the system comprises a natural language interface, a workflow engine, a data cleaning module, a clinical pharmacological analysis module library, an automatic report module, a multi-agent collaborative arrangement layer, an execution feedback module, a manual review door and an independent review module; information is transmitted to the workflow engine and the multi-agent collaborative arrangement layer through the natural language interface; the multi-agent collaborative arrangement layer sequentially or in parallel dispatches the data cleaning module, the clinical pharmacological analysis module library, the automatic report module, the independent review module and the execution feedback module according to an execution path; the data processed by the data cleaning module is output to the clinical pharmacological analysis module library; and the data processed by the clinical pharmacological analysis module library is output to the independent review module and the automatic report module. The application can improve the structural degree and stability of a clinical pharmacological task processing link.

Description

Technical Field

[0001] This invention relates to the field of pharmacological analysis technology, and in particular to a clinical pharmacological analysis system, method, electronic device, storage medium and program product based on a multi-agent model architecture. Background Technology

[0002] Clinical pharmacology work spans multiple stages of drug development, typically involving data identification and processing, pharmacokinetic and pharmacodynamic analysis strategy development, population pharmacokinetic modeling, exposure-response relationship analysis, physiological mechanism pharmacokinetic modeling, concentration-QTc analysis, quantitative systems pharmacology modeling, interpretation of safety and efficacy endpoints, graph generation, and output of conclusion reports. Existing pharmacokinetic and population pharmacokinetic modeling methods are relatively mature and widely used in drug development. Meanwhile, exposure-response studies have clear importance in drug development decisions, dosage selection, and regulatory submissions, and related research designs, data analysis, and reporting requirements are standardized by publicly available guidance documents and research literature.

[0003] However, in practical applications, clinical pharmacology tasks are typically not a single model calculation process, but rather consist of multiple stages including data processing, scientific analysis, results review, and report writing. Especially in real-world projects, input information often includes heterogeneous content from multiple sources, such as clinical research data, subject information, dosing and concentration records, efficacy endpoints, safety events, and research background information. Differences in field naming, time representation, and business semantics between different data sources lead to high costs for data cleaning and standardization, and the analysis process generally relies on experienced professionals manually connecting multiple tools. Therefore, existing clinical pharmacology work methods generally suffer from fragmented processes, heavy manual workloads, insufficient consistency of results, and difficulty in reusing experience across projects.

[0004] In recent years, technologies related to large language models have provided new pathways for automating complex tasks. For example, retrieval-enhanced generation techniques can enhance factual support in knowledge-intensive tasks by introducing external knowledge sources, thereby improving the shortcomings of purely parametric models in terms of knowledge updates and traceability. Model alignment methods based on human feedback show that introducing human preferences and feedback can improve the model's ability to follow user intent and the reliability of its output. Multi-agent collaborative frameworks further provide a technical approach to break down complex tasks into multiple collaborative roles and allow models, humans, and tools to participate in solving the problem.

[0005] Other solutions attempt to use artificial intelligence, large language models, or multi-agent technologies to assist in drug discovery, quantitative clinical pharmacology, or model development, such as generating code through natural language, assisting in data retrieval, organizing modeling workflows, or generating report drafts. However, these solutions are usually more geared towards general assistance or localized modeling tasks, and lack a clear data transmission relationship for a closed-loop clinical pharmacology workflow between data processing, model execution, review, reporting, and experience accumulation.

[0006] However, existing technologies still lack a unified AI architecture approach truly geared towards clinical pharmacology assistant scenarios. Existing solutions often focus either on automating a single analysis process or on general question-answering or general assistant capabilities, lacking a unified architecture that can simultaneously integrate natural language interaction, multi-agent collaborative orchestration, multi-source data cleaning, dynamic combination of clinical pharmacology modules, knowledge augmentation, tool invocation, independent review, human review, report generation, and experience feedback.

[0007] Therefore, there is an urgent need to propose a clinical pharmacology analysis system based on a multi-agent model architecture, which can form an executable intelligent collaborative link around clinical pharmacology tasks, maintain human control at key scientific decision-making nodes, and continuously reuse historical experience in subsequent tasks, thereby improving the automation level, reliability of results, and system evolution capability of clinical pharmacology work. Summary of the Invention

[0008] The purpose of this invention is to provide: A clinical pharmacology analysis system, method, electronic device, storage medium, and program product based on a multi-agent model architecture is disclosed. The system includes a natural language interface, a workflow engine, a data cleaning module, a clinical pharmacology analysis module library, an automatic reporting module, a multi-agent collaborative orchestration layer, an execution feedback module, a human review gate (HITL), an independent review module, and a containerized deployment implementation layer. This system elevates clinical pharmacology tasks from a fragmented toolchain to a unified orchestrated intelligent system, addressing existing problems such as process fragmentation, heavy manual workload, insufficient result consistency, and difficulty in reusing cross-project experience, or combinations thereof.

[0009] Terminology Explanation: Unless otherwise defined, all technical terms used herein have the same meanings as commonly understood by one of ordinary skill in the art. Unless otherwise stated, all patents, patent inventions, and publications cited in this document are incorporated herein by reference in their entirety. If multiple definitions exist for terms, the definition provided in this chapter shall prevail.

[0010] It should be understood that the above brief description and the following detailed description are exemplary and for illustrative purposes only, and do not limit the subject matter of the invention in any way. In this invention, the singular is used in conjunction with the plural unless otherwise specifically stated. It should also be noted that, unless otherwise stated, the use of “or” or “or” means “and / or”. Furthermore, the use of the term “comprising” and other forms such as “including,” “containing,” and “contains” are not limiting.

[0011] Unless specifically defined, the software modules, model services, databases, message queues, containers, plugins, interfaces, and computing devices described herein can all be implemented using software engineering methods commonly used in the art. Data transmission between modules can be accomplished through function calls, application programming interfaces (APIs), message queues, file exchange, database read / write, or event buses, and can be implemented using interface specifications, software engineering methods, or the description of this invention known in the art. Generally, the above techniques and methods can be implemented according to conventional methods well-known in the art, based on the descriptions in the various general and more specific documents cited and discussed in this specification.

[0012] The terms “optional / arbitrary” or “optionally / arbitrarily” mean that the event or situation described below may or may not occur, including both the occurrence and non-occurrence of the event or situation.

[0013] The term “retrieval-enhanced generation” used in this paper, or RAG for short, combines language modeling and information retrieval techniques. It retrieves relevant information from a large collection of documents and then uses this retrieved information to guide text generation in order to improve the quality and accuracy of predictions.

[0014] The term "HITL" used in this article, Human-Machine Loop or Human-Machine Loop, is a way of interacting between humans and machines. It involves integrating the judgments of individuals (experts, employees, business personnel) into the AI ​​system process to form a closed-loop system that repeats with relatively fixed steps and states.

[0015] The term "primary agent" (or "root agent") used in this article refers to the core agent in a multi-agent system that is responsible for high-level intent understanding, task decomposition, and scheduling decisions. It does not directly execute specific tasks but coordinates specialized subagents to achieve the goal.

[0016] The term "ER" used in this article stands for Exposure-Response, which studies the relationship between drug exposure in the body (such as blood drug concentration AUC, Cmax) and efficacy or safety indicators, and is used to optimize drug dosage and assess benefit-risk.

[0017] The term “PopPK” used in this article stands for Population Pharmacokinetics. It is a modeling and analysis technique that combines pharmacokinetic principles with statistical methods to quantify the variations in the absorption, distribution, metabolism, and excretion (ADME) of drugs in heterogeneous populations and to identify covariates (such as weight, age, renal function, and gene polymorphism) that affect these variations.

[0018] The term "PBPK" used in this article stands for Physiologically Based Pharmacokinetics; it is a mathematical modeling technique that simulates blood flow in the body's circulatory system based on physiological, anatomical, and biochemical knowledge, and follows the principle of mass conservation.

[0019] The term "LLM" used in this article stands for Large Language Model; it refers to a machine learning model that can understand, generate, reason, or invoke tools based on natural language context.

[0020] The term "CQT" used in this article stands for Concentration-QTc analysis; it refers to a clinical pharmacological analysis technique used to analyze the relationship between drug concentration and changes in the QTc interval.

[0021] The term “MBMA” used in this article stands for Model-Based Meta-Analysis. It refers to an analytical technique that uses modeling methods to synthesize multiple studies or data sources to compare dose, exposure, efficacy, or safety outcomes.

[0022] The term "QSP" used in this article stands for Quantitative Systems Pharmacology. It refers to an analytical technique that combines systems biology, pharmacology, and mathematical modeling methods to quantitatively simulate drug action mechanisms, disease pathways, and clinical responses.

[0023] As used in this article, the term "knowledge plugin" refers to a functional component used to retrieve, organize, and provide agents with professional background, regulatory requirements, project information, historical cases, or other external knowledge.

[0024] The term "Skills plugin" as used in this article refers to a functional component used to encapsulate and invoke data cleaning programs, modeling programs, statistical analysis programs, report generation programs, or other external tools.

[0025] The term "Lessons plugin" as used in this article refers to a functional component used to receive feedback from human review or task execution and to condense that feedback into experience rules, templates, tags, vector representations, or other reusable knowledge units.

[0026] In a first aspect, the present invention provides a clinical pharmacology analysis system based on a multi-agent model architecture, including a natural language interface, a workflow engine, a data cleaning module, a clinical pharmacology analysis module library, an automatic reporting module, a multi-agent collaborative orchestration layer, an execution feedback module, a manual review gate, and an independent review module; The system receives input information through a natural language interface and transmits it to the workflow engine for task intent parsing, data completeness judgment, task routing score calculation, and execution path generation. The workflow engine then transmits the standardized task description to the multi-agent collaborative orchestration layer. The multi-agent collaborative orchestration layer schedules the data cleaning module, clinical pharmacology analysis module library, automatic reporting module, independent review module and execution feedback module sequentially or in parallel according to the execution path; The standardized data packets, quality control results, and anomaly alerts output by the data cleaning module are transmitted to the clinical pharmacology analysis module library. The model results, charts, and scientific explanations output by the clinical pharmacology analysis module library are transmitted to the independent review module and the automatic reporting module. The verification results from the independent review module, the human decision-making from the manual review panel, and the final report results are transmitted to the execution feedback module for updating knowledge entries, skill call records, and experience rules.

[0027] Based on further solutions to the technical problems of the present invention, or simultaneous solutions to multiple technical problems, the preferred solution in the technical solution provided in the first aspect of the present invention includes: First preferred option: The workflow engine is implemented by a master agent as a unified orchestration hub. After receiving a user request, the master agent parses the task intent and calculates the task routing score based on the intent category, data completeness, and historical experience matching degree.

[0028] Second preferred option: The task routing score includes task intent score, data completeness score, historical experience matching score, and audit risk score. When the task intent is exposure-response analysis and exposure indicators and endpoint data are available, the agent selects the ER analysis mode. When the request involves the complete data chain of dosing, concentration, covariates and endpoints, the main agent selects a combined mode of data cleaning - PopPK module - ER module - automated reporting; When the task objective involves mechanism simulation, ECG safety, meta-analysis, or systems pharmacology modeling, the agent selects PBPK, CQT, MBMA, or QSP extended modes. When data is missing, the score is below the preset threshold, or the review risk reaches the preset level, the main AI agent triggers the manual review process or requests the user to supplement information.

[0029] Third preferred option: The data cleaning module includes a scanning and recognition submodule, a standardization submodule, a pharmacokinetic logic cleaning submodule, an efficacy cleaning submodule, a safety cleaning submodule, a quality control submodule, and a merging and exporting submodule. The scanning and identification submodule transmits the file path, table roles, and field list to the standardization submodule; The standardization submodule outputs unified fields, time variables, and encoding results; the pharmacokinetic logic cleaning submodule, efficacy cleaning submodule, and safety cleaning submodule generate pharmacokinetic datasets, efficacy datasets, and safety datasets respectively based on the unified fields. The quality control submodule performs cross-table consistency checks on each generated dataset and generates an anomaly list. The merged export submodule generates a standardized data package based on the quality control results, which can then be used by subsequent analysis modules.

[0030] Fourth preferred option: The clinical pharmacology analysis module library includes the ER module, PopPK module, PBPK module, CQT module, MBMA module, and QSP module.

[0031] Fifth preferred option: The ER module employs a two-stage analysis mechanism; The first stage is the full-scale automatic scanning stage, which is used to perform a comprehensive analysis of all endpoint-exposure indicator combinations. The second stage is the AI-driven refinement stage, which is used to filter, sort, and analyze the full set of results based on the user's natural language commands.

[0032] Sixth preferred option: The automatic reporting module includes a template parser, a data interpreter, an arranger, a chapter generator, a consistency maintainer, and a document assembler; The template parser transmits the report structure, chapter requirements, and terminology to the editor; The data interpreter transforms model parameters, chart files, statistical summaries, and audit results into structured fact tables; The editor schedules the chapter generator to generate corresponding chapters based on the fact table and template structure; The consistency maintainer verifies the values, variable names, and conclusion statements in different chapters and feeds back any inconsistencies to the chapter generator for correction. After the consistency check passes, the document assembler merges text, tables, and image placeholders and outputs a report file.

[0033] Seventh preferred option: In the multi-agent collaborative orchestration layer, the execution order is master agent, data cleaning agent, scientific analysis agent, review agent, reporting agent, and memory management agent; The data cleaning agent is responsible for executing the data cleaning module; The scientific analysis agent is responsible for calling the clinical pharmacology analysis module library and generating analysis interpretations; The reviewing agent is responsible for independently verifying the scientific analysis output; The reporting agent is responsible for organizing the final output; The memory management agent is responsible for maintaining episodic memory, semantic memory, and checkpoint states; The main intelligent agent is responsible for unified orchestration and process gating.

[0034] Eighth preferred option: The execution feedback module includes a knowledge plugin, a skill plugin, and an experience plugin; The knowledge plugin, skill plugin, and experience plugin respectively undertake the functions of knowledge enhancement, external tool execution, and feedback learning.

[0035] Ninth preferred option: The input information includes raw data files from clinical studies, dosing records, concentration records, efficacy endpoints, safety events, study protocols, statistical analysis plans, historical reports, literature, and regulatory guidelines.

[0036] The tenth preferred option: The clinical pharmacology analysis system is deployed in a private computing environment or cloud environment and uses Docker Compose for containerized orchestration. The services running in the container include application services, database services and queue services; the host machine runs modeling job execution units, modeling engine, data cleaning services and local embedding services.

[0037] Secondly, the present invention provides a clinical pharmacology analysis method based on a multi-agent model architecture, comprising the following steps: Step 1: Receive clinical pharmacology analysis requests, project background information, raw data files and output constraints submitted by users through the natural language interface, and generate a standardized task description that includes task objectives, data scope, analysis type and review requirements. Step 2: The workflow engine parses the standardized task description and calculates the task routing score based on the task intent, data completeness, historical experience matching degree and audit risk. Based on the task routing score, the ER analysis mode, the data cleaning-PopPK module-ER module-automated report combination mode, or the PBPK, CQT, MBMA, QSP extended mode are selected. Step 3: Call the data cleaning module to perform scanning and recognition, field standardization, pharmacokinetic logic cleaning, efficacy cleaning, safety cleaning, quality control, and merging and exporting on the original data file, generating standardized data packages, quality control results, and abnormal prompts. Step 4: The multi-agent collaborative orchestration layer schedules the data cleaning agent, scientific analysis agent, review agent, reporting agent, and memory management agent according to the execution path, and transmits the standardized data packets to the clinical pharmacology analysis module library; Step 5: Perform the corresponding analysis using at least one of the ER module, PopPK module, PBPK module, CQT module, MBMA module, or QSP module from the Clinical Pharmacology Analysis Module Library to generate model parameters, charts, statistical summaries, and scientific interpretations; Step 6: Provide the agent with professional background, regulatory requirements, historical data and project context through the knowledge plugin, call external analysis tools through the skills plugin, and inject historical human correction experience through the experience plugin; Step 7: The independent review module or review agent cross-verifies the model parameters, charts, statistical summaries and scientific explanations in Step 5 based on the benchmark fact source, and triggers the manual review gate when the review risk score reaches the preset level, so that the user can perform the approval, modification, rejection, overwrite acceptance or rollback and rerun operation. Step 8: After the review is approved, the automatic reporting module generates a report file based on the report template, structured fact table, model output, charts and review conclusions, and writes the human feedback, final conclusions and task status into the execution feedback module to form experience rules or knowledge items that can be reused for subsequent similar tasks.

[0038] Thirdly, the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a program or instructions running on the processor, and the program or instructions, when executed by the processor, implement the steps of a clinical pharmacology analysis method based on a multi-agent model architecture.

[0039] Fourthly, the present invention provides a readable storage medium on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of a clinical pharmacology analysis method based on a multi-agent model architecture are implemented.

[0040] Fifthly, the present invention provides a computer program product, the program product being stored in a storage medium, and the program product being executed by at least one processor to implement the steps of a clinical pharmacological analysis method based on a multi-agent model architecture.

[0041] Compared with the prior art, the present invention has at least the following beneficial effects: 1. Compared with existing methods that use manual concatenation tools or single models to directly process complex tasks, this invention can improve the structuring and stability of the clinical pharmacology task processing chain. Furthermore, this invention adopts a unified arrangement of the main intelligent agent and the division of labor and cooperation among multiple professional sub-intelligent agents. Data processing, scientific analysis, independent review, report generation, and memory management are no longer mixed in the same processing unit, but are executed separately under unified control, which can reduce the risks of contextual confusion, inference drift, and process loss of control caused by task coupling.

[0042] 2. This invention further improves the factual consistency and scientific reliability of the system output; specifically, by introducing knowledge plugins, review agents, and independent review mechanisms, the system can obtain more comprehensive professional context through retrieval enhancement during the analysis phase, and cross-validate the analysis conclusions based on benchmark fact sources during the review phase, thereby reducing the risk of unfounded speculation and content distortion, and enhancing the credibility and traceability of the results.

[0043] 3. This invention also combines natural language interaction capabilities with professional analysis and execution capabilities. Users do not need to pre-write complex scripts or manually switch between multiple independent software tools. They only need to express their natural language requirements related to clinical pharmacology tasks. The system can then use the main intelligent agent to complete task recognition and call the corresponding data processing services, analysis programs, and reporting modules through skill plugins, transforming the language-level task intent into an executable analysis chain.

[0044] 4. This invention achieves controllable key scientific decisions by introducing HITL into the architecture. The system automatically pauses and awaits human decision-making at points such as basic model confirmation, ER module confirmation, important interpretation conclusions, and final report output. This preserves the efficiency of automated execution while ensuring the leading role of clinical pharmacology experts in key scientific judgments. The experience plugin further ensures that human corrections not only apply to the current task but are transformed into knowledge assets that can be inherited by subsequent tasks.

[0045] 5. This invention has good business adaptability and scalability. Because this invention uses PopPK and ER as the preferred main links, while retaining extension interfaces for modules such as PBPK, MBMA, CQT, and QSP under a unified architecture, the system can focus on core clinical pharmacology analysis tasks, and can also be expanded to more clinical pharmacology-related capabilities according to specific R&D scenarios and user needs.

[0046] 6. This invention improves the portability, persistence, and compatibility with local execution resources of the system deployment by adopting a combined deployment method of Docker containerized services and host machine execution services.

[0047] 7. This invention clarifies the data transmission relationships between the natural language interface, workflow engine, data cleaning module, clinical pharmacology analysis module library, independent review module, automatic reporting module, and execution feedback module, enabling task requests, standardized data packages, model outputs, review conclusions, report content, and empirical rules to flow in a unified chain. Compared with existing solutions that only provide local tool calls or general Q&A, this invention can improve the traceability, verifiability, and cross-project reusability of clinical pharmacology analysis results. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall architecture of the multi-agent-based clinical pharmacology analysis system of the present invention; Figure 2 This is a schematic diagram of the main business process of the clinical pharmacology assistant of the present invention; Figure 3 This is a schematic diagram of the data cleaning and QC process of the present invention; Figure 4 This is a schematic diagram of the clinical pharmacology AI assistant based on a multi-agent architecture according to the present invention. Figure 5 This is a schematic diagram illustrating the deployment of the clinical pharmacology assistant based on the Docker environment of this invention. Detailed Implementation

[0049] The following non-limiting embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way. The following content is merely an exemplary description of the scope of protection claimed by the present invention, and those skilled in the art can make various changes and modifications to the present invention based on the disclosed content, and such changes should also fall within the scope of protection claimed by the present invention.

[0050] The present invention will be further described below by way of specific embodiments. Unless otherwise specified, all instruments, devices, equipment, products, etc., used in the embodiments of the present invention are obtained through conventional commercial means.

[0051] This embodiment provides a clinical pharmacology analysis system based on a multi-agent model architecture, comprising a natural language interface, a workflow engine, a data cleaning module, a clinical pharmacology analysis module library, an automatic reporting module, a multi-agent collaborative orchestration layer, an execution feedback module, a manual review gate, and an independent review module. The core of the manual review gate is HITL. Specifically, the received input information is transmitted to the workflow engine via the natural language interface, and the workflow engine transmits the standardized task description to the multi-agent collaborative orchestration layer. The information received by the natural language interface includes, but is not limited to, clinical pharmacology analysis requests, project background information, raw data files, and output constraints. The standardized task description includes, but is not limited to, task structure such as task objectives, data scope, analysis type, and review requirements. The multi-agent collaborative orchestration layer schedules the data cleaning module, clinical pharmacology analysis module library, automatic reporting module, independent review module, and execution feedback module sequentially or in parallel according to the execution path. The data cleaning module outputs a standardized... Standardized data packages, quality control results, and anomaly alerts are transmitted to the clinical pharmacology analysis module library. The model results, charts, and scientific interpretations output by the clinical pharmacology analysis module library are transmitted to the independent review module and the automatic reporting module. The standardized data package refers to the data results processed by the data cleaning module and output for subsequent use by the clinical pharmacology analysis module. The verification results of the independent review module, the manual decisions of the human review gate, and the final report results are transmitted to the execution feedback module for updating knowledge entries, skill call records, and experience rules, thus forming a closed-loop data transmission process from task reception, data processing, model analysis, review and confirmation, report output to experience feedback. Furthermore, the input information may include raw data files from clinical studies, dosing records, concentration records, efficacy endpoints, safety events, study protocols, statistical analysis plans, historical reports, literature, and regulatory guidelines; the corresponding output information may include standardized data packages, quality control results, anomaly lists, model parameters, charts, statistical summaries, scientific interpretations, automatically generated reports, audit trails, rules of thumb, and knowledge entries.

[0052] In this embodiment, the natural language interface is used to receive analysis tasks, background descriptions, and constraints submitted by users in natural language.

[0053] In this embodiment, the workflow engine is used to identify task intent, organize execution paths, and schedule each processing unit; Specifically, the workflow engine is implemented by a master agent as the unified orchestration hub. Upon receiving a user request, the master agent first performs task intent parsing, data completeness assessment, historical experience matching assessment, and review risk assessment to obtain evaluation factors such as task intent category, data completeness, historical experience matching, and review risk level. Then, based on these evaluation factors, it calculates a task routing score and generates an execution path or selects an appropriate working mode based on the score. The master agent selects between different working modes based on the score. These modes include a mode that performs data cleaning and ER analysis, a combined mode of data cleaning - PopPK module - ER module - automated report, and an extended mode that calls PBPK, CQT, MBMA, and QSP modules. Furthermore, the task routing score includes, but is not limited to, task intent score, data completeness score, historical experience matching score, and review risk score. When the task intent points to exposure-response analysis and exposure indicators and endpoint data are available, the main agent selects the ER analysis mode. When the request involves a complete data chain including dosing, concentration, covariates, and endpoints, the main agent selects a combination mode of data cleaning - PopPK module - ER module - automated reporting. When the task objective involves mechanism simulation, ECG safety, meta-analysis, or systems pharmacology modeling, the main agent selects the PBPK, CQT, MBMA, or QSP extended mode. When data is missing, the score is lower than a preset threshold, or the review risk reaches a preset level, the main agent triggers a manual review gate or requests the user to supplement information. By matching and selecting, the clinical pharmacology assistant is no longer limited to a single ER analysis process, but can call all clinical pharmacology modules in a unified workflow according to task type and data conditions. That is, through the overall arrangement of the main intelligent agent and the division of labor among multiple professional intelligent agents, a collaborative relationship of "unified planning - division of labor - independent verification - result organization - memory reuse" is formed in the structure, so that data processing, scientific analysis, result verification and report generation are no longer mixed in the same processing unit.

[0054] In this embodiment, the data cleaning module is used to complete the identification, standardization, and quality control of multi-source heterogeneous clinical data; Specifically, the data cleaning module performs phased processing on uploaded project data. Specifically, the data cleaning module includes a scan recognition submodule, a standardization submodule, a pharmacokinetic logic cleaning submodule, an efficacy cleaning submodule, a safety cleaning submodule, a quality control submodule, and a merge and export submodule. The scan recognition submodule recursively scans the project directory, reads column information, and identifies data table roles; the standardization submodule performs field mapping, renaming, time conversion, and unified encoding; the pharmacokinetic logic cleaning submodule constructs observation and dosing records and processes pharmacokinetic-related fields such as BLQ, EVID, MDV, and CMT; the efficacy cleaning submodule extracts efficacy endpoints, performs visit-based aggregation, and consistency verification; the safety cleaning submodule processes adverse event, laboratory test, and vital sign data; the quality control submodule performs cross-validation, LLM anomaly identification, and summary generation; and the merge and export submodule generates standardized output results. Furthermore, the data transmission logic between the above sub-modules is as follows: the scanning and identification sub-module transmits the file path, table roles, and field list to the standardization sub-module; the standardization sub-module outputs unified fields, time variables, and encoding results; the pharmacokinetic logic cleaning sub-module, efficacy cleaning sub-module, and safety cleaning sub-module generate pharmacokinetic datasets, efficacy datasets, and safety datasets respectively based on the unified fields; the quality control sub-module performs cross-table consistency checks on the generated datasets and forms an anomaly list; the merging and exporting sub-module generates a standardized data package for subsequent analysis modules to call based on the quality control results.

[0055] In this embodiment, the clinical pharmacology analysis module library is a complete module library that organizes clinical pharmacology analysis capabilities, rather than being designed solely around a single ER scenario. It is used to support different types of analysis capabilities, including ER module, PopPK module, PBPK module, CQT module, MBMA module, QSP module, etc. Specifically, the ER module is used to perform descriptive exposure-response analysis, graphical exposure-response analysis, statistical exposure-response modeling, covariate exploration, and safety exposure-response analysis; the PopPK module is used to perform population pharmacokinetic model building, covariate screening, and diagnostic analysis; the PBPK module is used to perform pharmacokinetic modeling based on physiological mechanisms; the CQT module is used to perform concentration-QTc correlation analysis; the MBMA module is used to perform model-based meta-analysis; and the QSP module is used to perform quantitative systems pharmacology modeling. The ER, PopPK, PBPK, CQT, MBMA, and QSP modules can all be dynamically combined and invoked by the workflow engine based on task scenarios and data conditions, and can also be dynamically combined according to user natural language commands, thus enabling the clinical pharmacology assistant to possess complete and scalable analytical capabilities.

[0056] In a preferred embodiment, the ER module employs a two-stage analysis mechanism; the first stage is a full-scale automatic scanning stage, which is used to perform a comprehensive analysis of all possible endpoint-exposure indicator combinations to ensure that no potentially important signals are missed; the second stage is an AI-driven refinement stage, which is used to filter, sort, and further analyze the full results based on the user's natural language commands.

[0057] It should be noted that this system can execute the complete PopPK and ER joint analysis process; for other implementations, the system can also execute only PopPK module analysis, only ER analysis, or only call some steps according to a specific research purpose; or, modules other than PopPK and ER modules such as PBPK, MBMA, CQT, QSP, etc. can be connected as optional extension modules, which can be trimmed, inserted or replaced by the main agent according to the user task, data type and analysis goal.

[0058] In this embodiment, the automatic reporting module is used to generate structured analysis results and report files; the automatic reporting module includes a template parser, a data interpreter, an arranger, a chapter generator, a consistency maintainer, and a document assembler; making the automatic report no longer just simple text output, but a dedicated module with structured fact organization and consistency control capabilities; The template parser is used to parse the reference report structure, writing style, and terminology; the data interpreter is used to read chart files and modeling output and extract key statistics and parameters; the orchestrator is used to manage task breakdown, state transitions, and exception handling; the chapter generator is used to generate the main text content in segments; the consistency maintainer is used to maintain the global fact table and check the consistency between values ​​and terminology; and the document assembler is used to merge text and image placeholders and output Word, PDF, Markdown, or structured data files. Furthermore, the sequential transmission and feedback verification relationships among the above components are as follows: the template parser transmits the report structure, chapter requirements, and terminology library to the editor; the data interpreter converts model parameters, figure files, statistical summaries, and audit results into a structured fact table; the editor schedules the chapter generator to generate corresponding chapters based on the fact table and template structure; the consistency maintainer verifies the values, variable names, and conclusion statements in different chapters and feeds back inconsistencies to the chapter generator for correction; and the document assembler merges text, tables, and image placeholders and outputs the report file after the consistency verification is passed.

[0059] In this embodiment, the multi-agent collaborative orchestration layer is used to assign tasks to multiple specialized agents for execution. Specifically, in the multi-agent collaborative orchestration layer, the execution order is: master agent, data cleaning agent, scientific analysis agent, review agent, reporting agent, and memory management agent. The data cleaning agent is responsible for executing the data cleaning module; the scientific analysis agent is responsible for calling the clinical pharmacology analysis module library and generating analysis interpretations; the review agent is responsible for independently verifying the scientific analysis output; the reporting agent is responsible for organizing the final output; the memory management agent is responsible for maintaining contextual memory, semantic memory, and checkpoint states; and the master agent is responsible for unified orchestration and process gating. Furthermore, by coordinating and arranging the main intelligent agent and having multiple specialized intelligent agents perform the tasks, a collaborative relationship of "unified planning - division of labor - independent verification - result organization - memory reuse" is formed structurally, so that data processing, scientific analysis, result verification and report generation are no longer mixed in the same processing unit.

[0060] In this embodiment, the execution feedback module includes a knowledge plugin, a skills plugin, and a lessons plugin. These plugins respectively handle knowledge enhancement, external tool execution, and feedback learning functions. It should be noted that this embodiment does not limit the three types of plugins to using a single implementation method. The knowledge plugin can be implemented using vector retrieval, keyword retrieval, knowledge graph retrieval, or a combination of multiple retrieval methods. The skills plugin can call local programs, remote services, command-line tools, microservice interfaces, or third-party analysis platforms. The lessons plugin can extract human feedback into rule text, or convert it into tags, templates, embedded vectors, editable prompts, or other reusable knowledge units. Specifically, knowledge plugins provide each agent with professional background, regulatory requirements, historical data, and project context, and select knowledge items to enter the context through comprehensive relevance scoring; knowledge plugins can be used to enhance retrieval during the analysis phase, and can also serve as a source of baseline facts for independent review units to verify facts during the review phase.

[0061] Furthermore, this Skills plugin encapsulates the professional processing capabilities involved in the execution of clinical pharmacology tasks and can be uniformly invoked by the main agent and workflow engine. This Skills plugin includes at least the NONMEM Skill for performing population pharmacokinetic modeling and model running, the PK-Data-Clean Skill for performing project data identification, field mapping, standardization cleaning and quality control, the ER Analysis Skill for performing exposure-response analysis and related statistical modeling, and the ReportSkill for performing result integration, chapter generation and document export. Each Skill is connected to the multi-agent collaborative architecture through a unified interface, and the main agent selects and schedules them according to the task stage, intermediate results and user instructions. Thus, this invention not only has the ability to understand tasks and plan paths, but also drives the actual execution of key steps such as data cleaning, model analysis and report output. Furthermore, the Lessons plugin is used to receive human feedback. When a user provides suggestions for improvement in HITL, the Lessons plugin extracts these suggestions into empirical rules and stores them quantitatively in the database. Before subsequent similar tasks begin, these rules are injected into the context of the main agent or scientific analysis agent through semantic retrieval, thus forming a continuous learning loop of "task execution - human correction - experience accumulation - subsequent reuse". In other words, the Knowledge plugin, Skills plugin, and Lessons plugin are fixed as enhancement layers in the architecture, respectively undertaking the functions of knowledge injection, tool execution, and feedback learning, thereby forming a plugin collaboration mechanism of "knowledge enhancement - tool execution - experience feedback".

[0062] In this embodiment, HITL and the independent audit unit are used to form a two-layer control for key nodes; Specifically, regarding manual review and independent review, this embodiment sets a manual review risk score to determine whether to trigger the manual review gate. When the review risk score reaches a preset threshold, the system automatically triggers the manual review gate, and the user performs operations such as approval, modification, rejection, acceptance overwrite, or rollback and rerun. At the same time, the independent review unit and the review agent verify the results from different paths, forming a two-layer control structure of cross-validation within the agent and final human decision. Furthermore, HITL and independent review mechanisms are directly incorporated into the multi-agent model architecture of the clinical pharmacology analysis system. Human review is triggered by risk scoring, and cross-validation is carried out through independent review paths, thereby ensuring that key scientific decisions are always under control.

[0063] It should be noted that this system does not require a single implementation for the Hidden Hierarchical Link (HITL) mechanism, state persistence method, and underlying model routing method. In other implementations, manual review can take different levels, such as mandatory approval, release after review, or information confirmation; in other implementations, the system can also support different interaction methods such as pause, rollback, rerun, and overwrite acceptance. State management can be implemented through database checkpoints or by combining session logs, task queues, and event buses; different agents can either call the same large language model or have a unified model gateway select different model paths based on task type.

[0064] This embodiment is preferably deployed in a private computing environment and containerized using Docker Compose. The application service, database service, and queue service run within containers. The application service hosts the web interface, API interfaces, and multi-agent orchestration logic. The database service preferably uses a combination of PostgreSQL and pgvector to store session state, checkpoints, knowledge entries, empirical rules, and application metadata. The queue service preferably uses Redis for task queuing and publish-subscribe communication. The host machine runs the modeling execution unit, modeling engine, data cleaning service, and local embedding service to directly access authorized modeling software and local resources. The application service submits tasks to the host machine via queues and receives completion events, thereby resuming the multi-agent process. Furthermore, Docker containerization deployment and host machine execution services are layered and coordinated, enabling the multi-agent platform logic, state persistence, and vector retrieval capabilities to run decoupled from the clinical pharmacology modeling execution environment, data cleaning services, and local embedded services.

[0065] It should be noted that although this system is preferably used in clinical pharmacology assistant scenarios, its architecture is not limited to a specific disease area, a specific drug project, a specific deployment platform, or a specific model vendor. While maintaining the characteristics of the clinical pharmacology task and the overall architecture of this invention, the user interface, tool execution environment, model access method, and report output format can all be replaced with equivalent alternatives. Therefore, the focus of this invention is not on a single analysis software or a single fixed script, but rather on a multi-agent AI architecture method for clinical pharmacology assistants and the various equivalent implementations derived therefrom.

[0066] In other embodiments, a clinical pharmacology analysis method based on a multi-agent model architecture is also provided, which includes the following steps: Step 1: Receive clinical pharmacology analysis requests, project background information, raw data files and output constraints submitted by users through the natural language interface, and generate a standardized task description that includes task objectives, data scope, analysis type and review requirements. Step 2: The workflow engine parses the standardized task description and calculates the task routing score based on the task intent, data completeness, historical experience matching degree and audit risk. Based on the task routing score, the ER analysis mode, the data cleaning-PopPK module-ER module-automated report combination mode, or the PBPK, CQT, MBMA, QSP extended mode are selected. Step 3: Call the data cleaning module to perform scanning and identification, field standardization, pharmacokinetic logic cleaning, efficacy cleaning, safety cleaning, quality control, and merging and exporting on the original data file, generating a standardized data package, quality control results, and anomaly prompts. The standardized data package refers to the data results generated by the data cleaning module after completing the scanning and identification, field standardization, pharmacokinetic logic cleaning, efficacy cleaning, safety cleaning, quality control, and merging and exporting on the original data file, which are used by the clinical pharmacology analysis module library.

[0067] Step 4: The multi-agent collaborative orchestration layer schedules the data cleaning agent, scientific analysis agent, review agent, reporting agent, and memory management agent according to the execution path, and transmits the standardized data packets to the clinical pharmacology analysis module library; Step 5: Perform the corresponding analysis using at least one of the ER module, PopPK module, PBPK module, CQT module, MBMA module, or QSP module from the Clinical Pharmacology Analysis Module Library to generate model parameters, charts, statistical summaries, and scientific interpretations; Step 6: Provide the agent with professional background, regulatory requirements, historical data and project context through the Knowledge plugin, call external analysis tools through the Skills plugin, and inject historical manual correction experience through the Lessons plugin. Step 7: The independent review module or review agent cross-verifies the model parameters, charts, statistical summaries and scientific explanations in Step 5 based on the benchmark fact source, and triggers the manual review gate when the review risk score reaches the preset level, so that the user can perform the approval, modification, rejection, overwrite acceptance or rollback and rerun operation. Step 8: After the review is approved, the automatic reporting module generates a report file based on the report template, structured fact table, model output, charts and review conclusions, and writes the human feedback, final conclusions and task status into the execution feedback module to form experience rules or knowledge items that can be reused for subsequent similar tasks.

[0068] In other embodiments, an electronic device is also provided, including a processor and a memory, the memory storing a program or instructions that run on the processor, the program or instructions being executed by the processor to implement steps of a clinical pharmacology analysis method based on a multi-agent model architecture.

[0069] In other embodiments, a readable storage medium is also provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps of a clinical pharmacology analysis method based on a multi-agent model architecture.

[0070] In other embodiments, a computer program product is also provided, which is stored in a storage medium and executed by at least one processor to implement the steps of a clinical pharmacological analysis method based on a multi-agent model architecture.

[0071] Example 1 Based on the above implementation methods, referring to Figures 1-5 As shown below, the architecture of this system will be explained using a clinical pharmacology assistant to perform population pharmacokinetics and ER analysis as an example.

[0072] For example, a user first submits a clinical pharmacology analysis request through a natural language interaction portal, uploading project data files, background information, and output requirements. This request can be a complete PopPK and ER joint analysis request, or a separate request targeting only the PopPK module or only the ER module. Upon receiving the user request, the main agent parses the user's intent, identifies whether the current task belongs to the complete workflow mode, PopPK mode, or ER mode, and generates an execution path based on the completeness of the task data, historical experience, and knowledge retrieval results. In the complete workflow mode, the main agent first activates the data cleaning agent, then sequentially organizes exploratory analysis, basic PopPK module establishment, covariate screening, visual prediction testing, Bootstrap, ER modeling, and report generation; in the separate mode, the main agent automatically tailors the workflow.

[0073] After the data enters the system, the data cleaning agent identifies and preprocesses the raw files. First, the system performs format checks on various common data files used in clinical pharmacology analyses, and identifies, maps, and standardizes fields, time variables, dosing records, concentration records, efficacy endpoints, and safety information within the files. The cleaning process includes recursively scanning the project directory, reading file column information, performing field role determination, renaming fields, time conversion, BLQ processing, standardization of pharmacokinetic-related fields such as EVID / MDV / CMT, extraction of efficacy endpoints, organization of safety anomaly information, and final deduplication, sorting, and exporting, resulting in standardized data packages suitable for subsequent analysis. If unknown formats, missing key fields, or cross-table logical conflicts are found during the cleaning process, the main agent can immediately trigger manual review upon completion of the cleaning.

[0074] After data cleaning, the main agent transmits standardized data packages to the scientific analysis agent, simultaneously performing context enhancement through the Knowledge and Lessons plugins. The Knowledge plugin preferably employs a two-layer structure of "knowledge base + vector retrieval," semantically recalling literature, standard operating procedures, project materials, and historical cases, injecting relevant knowledge fragments into the context of the scientific analysis agent to enhance the model's professional understanding of the current task. In the subsequent review phase, the review agent can directly extract authoritative texts from the knowledge base as benchmark fact sources to cross-validate the output of the scientific analysis agent. The Lessons plugin performs historical experience retrieval at the beginning of the task, injecting previously manually corrected similarity rules into the context of the main agent or the scientific analysis agent, allowing the current analysis path to be prioritized based on historical experience.

[0075] In a preferred embodiment for population pharmacokinetic analysis, after obtaining standardized data packages, retrieving knowledge, and historical experience, the scientific analysis agent first performs exploratory analysis to preliminarily summarize the distribution characteristics of different subjects, dosage groups, and covariates, and then generates basic PopPK module suggestions. These suggestions may include key elements such as compartment structure, absorption model, residual error structure, and inter-individual variation structure. Subsequently, the scientific analysis agent calls external modeling programs or execution engines through skill plugins, submits model execution jobs, and reads the model output results after the job is completed, interpreting the convergence, parameter estimation, reduction degree, condition number, and diagnostic plots. After the basic model is completed, the system preferably forces a high-level manual review process, where the user decides on the model structure, diagnostic results, and whether to proceed to covariate screening. If approved, the main agent continues to schedule the scientific analysis agent to proceed to covariate screening, VPC, and Bootstrap stages; if not approved, remodeling or rerunning can be triggered based on user feedback.

[0076] In another preferred embodiment for exposure-response analysis, when the user request only includes the ER task, or when the system has completed the PopPK phase and obtained individual exposure parameters, the main agent passes these parameters, along with the endpoint variable, to the scientific analysis agent, which selects an appropriate ER module form, such as Emax, linear, log-linear, or other applicable forms. When performing model selection and fitting interpretation, the scientific analysis agent can invoke the Knowledge plugin to retrieve clinical pharmacology background information and endpoint interpretation criteria, and use the Skills plugin to call the corresponding modeling tools to generate fitting results, graphs, and statistical summaries. After the ER module is completed, the system re-enters a high-level human review process, requiring the user to confirm the model type, parameter interpretation, and result presentation. If the user requests corrections, these corrections can immediately affect the current analysis and can also be refined by the Lessons plugin into reusable rules for subsequent tasks.

[0077] In this embodiment, the review agent does not share the exact same judgment path as the scientific analysis agent. Instead, it preferably uses independent model instances or verification paths to cross-verify the output of the scientific analysis agent. Upon receiving analysis suggestions, model explanations, or intermediate conclusions, the review agent compares key statements against baseline fact sources in the Knowledge plugin. It also identifies any logical inconsistencies, factual biases, unfounded inferences, or non-compliant statements in the results. If the review agent confirms the results are acceptable, it returns the verified information to the main agent, which then proceeds to the report generation stage. If problems are found, the main agent triggers a manual review process for user approval.

[0078] Once the analysis results are approved, the main agent invokes the reporting agent to generate the final deliverables. The reporting agent preferably organizes the method descriptions, data summaries, basic model results, covariate analysis results, VPC interpretation, Bootstrap summary, ER results, and final conclusions in a structured manner according to clinical pharmacology reporting guidelines, and outputs charts, tables, narrative text, and audit trails. Report output formats can include PDF, Word, Markdown, JSON, or other formats suitable for archiving and review. Furthermore, the reporting agent can perform consistency checks using a global fact table and terminology database to ensure consistency in the descriptions of parameters, variables, and conclusions across different sections.

[0079] In another preferred embodiment, the system also features session persistence, rollback and rerun capabilities, and asynchronous job processing. When a modeling task is time-consuming or needs to be executed sequentially, the main agent can delegate model execution to asynchronous work units via a job queue, and restore the context to continue the process after the job is completed. When a user wants to return to a specific analysis node to retry different strategies, the system can restore to the specified stage based on the saved session state and checkpoints, and re-execute from that stage without repeating all previous steps. Simultaneously, the rejections, modifications, and additional conditions provided by the user at the manual review node are automatically extracted into experience rules by the Lessons plugin and stored, and are preferentially retrieved and injected into the context at the start of the next similar task.

[0080] Regarding model selection, this invention does not limit all agents to using the same underlying large language model. Preferably, the master agent or a unified model gateway can route different models to different agents based on task characteristics. For example, the scientific analysis agent prioritizes models with strong reasoning capabilities, the reporting agent prioritizes models with strong narrative generation capabilities, and the review agent preferably uses a model path different from that of the scientific analysis agent to enhance independent verification. Through this multi-model routing strategy, the system can flexibly adapt to the capabilities of different underlying models without changing the upper-layer multi-agent architecture.

[0081] In a preferred deployment embodiment, the present invention is deployed in a private computing environment or a cloud environment and containerized using Docker Compose. Services running within the containers include application services, database services, and queue services; the host machine runs a modeling job execution unit, a modeling engine, a data cleaning service, and a local embedding service. The application service submits tasks to the host machine execution unit via a queue and receives completion events through a publish-subscribe mechanism; the database service preferably uses PostgreSQL+pgvector to simultaneously store relational data and vector indexes; the queue service preferably uses Redis for task queuing and asynchronous notifications; and the local embedding service is preferably used to generate vector representations of knowledge entries and empirical rules. Through this layered deployment approach, the application logic, state persistence, and vector retrieval capabilities of the multi-agent platform can be decoupled from the authorized modeling software, local execution resources, and data cleaning service.

[0082] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A clinical pharmacology analysis system based on a multi-agent model architecture, characterized in that, It includes a natural language interface, workflow engine, data cleaning module, clinical pharmacology analysis module library, automatic reporting module, multi-agent collaborative orchestration layer, execution feedback module, manual review gate and independent review module; The system receives input information through a natural language interface and transmits it to the workflow engine for task intent parsing, data completeness judgment, task routing score calculation, and execution path generation. The workflow engine then transmits the standardized task description to the multi-agent collaborative orchestration layer. The multi-agent collaborative orchestration layer schedules the data cleaning module, clinical pharmacology analysis module library, automatic reporting module, independent review module and execution feedback module sequentially or in parallel according to the execution path; The standardized data packets, quality control results, and anomaly alerts output by the data cleaning module are transmitted to the clinical pharmacology analysis module library. The model results, charts, and scientific explanations output by the clinical pharmacology analysis module library are transmitted to the independent review module and the automatic reporting module. The verification results from the independent review module, the human decision-making from the manual review panel, and the final report results are transmitted to the execution feedback module for updating knowledge entries, skill call records, and experience rules.

2. The clinical pharmacology analysis system based on a multi-agent model architecture according to claim 1, characterized in that, The workflow engine is implemented by a master agent as a unified orchestration hub. After receiving a user request, the master agent parses the task intent and calculates the task routing score based on the intent category, data completeness, and historical experience matching degree.

3. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 2, characterized in that, The task routing score includes task intent score, data completeness score, historical experience matching score, and audit risk score. When the task intent is exposure-response analysis and exposure indicators and endpoint data are available, the agent selects the ER analysis mode. When the request involves the complete data chain of dosing, concentration, covariates and endpoints, the main agent selects a combined mode of data cleaning - PopPK module - ER module - automated reporting; When the task objective involves mechanism simulation, ECG safety, meta-analysis, or systems pharmacology modeling, the agent selects PBPK, CQT, MBMA, or QSP extended modes. When data is missing, the score is below the preset threshold, or the review risk reaches the preset level, the main AI agent triggers the manual review process or requests the user to supplement information.

4. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 1, characterized in that, The data cleaning module includes a scanning and recognition submodule, a standardization submodule, a pharmacokinetic logic cleaning submodule, an efficacy cleaning submodule, a safety cleaning submodule, a quality control submodule, and a merging and exporting submodule. The scanning and identification submodule transmits the file path, table roles, and field list to the standardization submodule; The standardization submodule outputs standardized fields, time variables, and encoding results; The pharmacokinetic logic cleaning submodule, efficacy cleaning submodule, and safety cleaning submodule generate pharmacokinetic datasets, efficacy datasets, and safety datasets respectively based on a unified field. The quality control submodule performs cross-table consistency checks on each generated dataset and generates an anomaly list. The merged export submodule generates a standardized data package based on the quality control results, which can then be used by subsequent analysis modules.

5. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 3, characterized in that, The clinical pharmacology analysis module library includes the ER module, PopPK module, PBPK module, CQT module, MBMA module, and QSP module.

6. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 5, characterized in that, The ER module employs a two-stage analysis mechanism; The first stage is the full-scale automatic scanning stage, which is used to perform a comprehensive analysis of all endpoint-exposure indicator combinations. The second stage is the AI-driven refinement stage, which is used to filter, sort, and analyze the full set of results based on the user's natural language commands.

7. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 1, characterized in that, The automatic reporting module includes a template parser, a data interpreter, an arranger, a chapter generator, a consistency maintainer, and a document assembler; The template parser transmits the report structure, chapter requirements, and terminology to the editor; The data interpreter transforms model parameters, chart files, statistical summaries, and audit results into structured fact tables; The editor schedules the chapter generator to generate corresponding chapters based on the fact table and template structure; The consistency maintainer verifies the values, variable names, and conclusion statements in different chapters and feeds back any inconsistencies to the chapter generator for correction. After the consistency check passes, the document assembler merges text, tables, and image placeholders and outputs a report file.

8. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 2, characterized in that, In the multi-agent collaborative orchestration layer, the execution order is master agent, data cleaning agent, scientific analysis agent, review agent, reporting agent, and memory management agent; The data cleaning agent is responsible for executing the data cleaning module; The scientific analysis agent is responsible for calling the clinical pharmacology analysis module library and generating analysis interpretations; The reviewing agent is responsible for independently verifying the scientific analysis output; The reporting agent is responsible for organizing the final output; The memory management agent is responsible for maintaining episodic memory, semantic memory, and checkpoint states; The main intelligent agent is responsible for unified orchestration and process gating.

9. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 1, characterized in that, The execution feedback module includes a knowledge plugin, a skill plugin, and an experience plugin; The knowledge plugin, skill plugin, and experience plugin respectively undertake the functions of knowledge enhancement, external tool execution, and feedback learning.

10. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 1, characterized in that, The input information includes raw data files from clinical studies, dosing records, concentration records, efficacy endpoints, safety events, study protocols, statistical analysis plans, historical reports, literature, and regulatory guidelines.

11. A clinical pharmacology analysis system based on a multi-agent model architecture according to claim 1, characterized in that, The clinical pharmacology analysis system is deployed in a private computing environment or cloud environment and uses Docker Compose for containerized orchestration. The services running in the container include application services, database services and queue services. The host machine runs the modeling job execution unit, modeling engine, data cleaning service, and local embedding service.

12. A clinical pharmacology analysis method based on a multi-agent model architecture, running on a clinical pharmacology analysis system based on a multi-agent model architecture as described in any one of claims 1-11, characterized in that, Includes the following steps: Step 1: Receive clinical pharmacology analysis requests, project background information, raw data files and output constraints submitted by users through the natural language interface, and generate a standardized task description that includes task objectives, data scope, analysis type and review requirements. Step 2: The workflow engine parses the standardized task description and calculates the task routing score based on the task intent, data completeness, historical experience matching degree and audit risk. Based on the task routing score, the ER analysis mode, the data cleaning-PopPK module-ER module-automated report combination mode, or the PBPK, CQT, MBMA, QSP extended mode are selected. Step 3: Call the data cleaning module to perform scanning and recognition, field standardization, pharmacokinetic logic cleaning, efficacy cleaning, safety cleaning, quality control, and merging and exporting on the original data file, generating standardized data packages, quality control results, and abnormal prompts. Step 4: The multi-agent collaborative orchestration layer schedules the data cleaning agent, scientific analysis agent, review agent, reporting agent, and memory management agent according to the execution path, and transmits the standardized data packets to the clinical pharmacology analysis module library; Step 5: Perform the corresponding analysis using at least one of the ER module, PopPK module, PBPK module, CQT module, MBMA module, or QSP module from the Clinical Pharmacology Analysis Module Library to generate model parameters, charts, statistical summaries, and scientific interpretations; Step 6: Provide the agent with professional background, regulatory requirements, historical data and project context through the knowledge plugin, call external analysis tools through the skills plugin, and inject historical human correction experience through the experience plugin; Step 7: The independent review module or review agent cross-verifies the model parameters, charts, statistical summaries and scientific explanations in Step 5 based on the benchmark fact source, and triggers the manual review gate when the review risk score reaches the preset level, so that the user can perform the approval, modification, rejection, overwrite acceptance or rollback and rerun operation. Step 8: After the review is approved, the automatic reporting module generates a report file based on the report template, structured fact table, model output, charts and review conclusions, and writes the human feedback, final conclusions and task status into the execution feedback module to form experience rules or knowledge items that can be reused for subsequent similar tasks.

13. An electronic device, characterized in that, It includes a processor and a memory, the memory storing programs or instructions that run on the processor, and when the programs or instructions are executed by the processor, they implement the steps of the clinical pharmacology analysis method based on a multi-agent model architecture as described in claim 12.

14. A readable storage medium, characterized in that, A program or instruction is stored on a readable storage medium, which, when executed by a processor, implements the steps of the clinical pharmacology analysis method based on a multi-agent model architecture as described in claim 12.

15. A computer program product, characterized in that, The program product is stored in a storage medium and executed by at least one processor to implement the steps of the clinical pharmacology analysis method based on a multi-agent model architecture as described in claim 12.

Images