Multi-agent collaborative clinical research implementation method and device and computer equipment

Abstract

The application relates to a multi-agent collaborative clinical scientific research implementation method and device and computer equipment. The method comprises the following steps: in response to a user's clinical scientific research text input operation, generating demand information, obtaining a search instruction through dialogue interaction hub analysis, and obtaining a first preset number of candidate data sets through a medical data set management agent; after the user selects, a training data set is determined, a data set instruction is obtained by combining the search instruction and the training data set, a second preset number of candidate algorithms are generated by an algorithm management agent, the user selects and determines a training algorithm and generates a configuration file; based on the training algorithm and the configuration file, a third preset number of candidate algorithm power schemes are generated through an algorithm power scheduling module, the user selects and determines a training scheme and generates a training configuration file; a clinical scientific research algorithm is obtained through iterative training based on the file, an inference configuration file is generated in response to an inference text input, and finally an inference report is generated. The method can automatically implement clinical scientific research.

Description

Technical Field

[0001] This application relates to the field of intelligent medical research technology, and in particular to a method, apparatus and computer equipment for realizing clinical research through multi-agent collaboration. Background Technology

[0002] With the deep integration of medical informatics and artificial intelligence technologies, clinical research has shifted towards a dual-driven approach of data and algorithms, making the demand for high-quality data, compatible algorithms, and efficient computing power increasingly urgent. Currently, mainstream research methods still rely heavily on decentralized tool operation and manual intervention: researchers need to manually switch platforms to search for resources, select algorithms based on experience, write code and configure parameters using specialized tools, depending entirely on technical background and experience. There is a lack of automated linkage between different stages, and core resource matching still requires manual completion.

[0003] These existing methods have cumbersome interaction modes, high learning costs for non-technical personnel, and are prone to slowing down research progress due to operational errors. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, and computer equipment for automating clinical research by simplifying the operation process and achieving automated linkage of the entire clinical research process, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for implementing multi-agent collaborative clinical research, including:

[0006] In response to the user's clinical research text input, a requirement information is generated; the requirement information is parsed through a pre-trained dialogue interaction center to obtain a search instruction; based on the search instruction, a candidate dataset is retrieved through a medical dataset management agent to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that the user can browse and view; based on the first preset number of candidate datasets, a candidate dataset in the form of entries is generated through the dialogue interaction center and fed back to the user;

[0007] In response to the user's selection of a candidate dataset, the selected candidate dataset is determined as the training dataset; the retrieval command and the training dataset are combined to obtain a dataset command; based on the dataset command, a second preset number of candidate algorithms are generated by the algorithm management agent; each candidate algorithm has detailed algorithm information that the user can browse and view; based on the second preset number of candidate algorithms, a dialogue interaction center generates candidate algorithms in the form of entries and provides feedback to the user; in response to the user's selection of a candidate algorithm, the selected candidate algorithm is determined as the training algorithm, and a configuration file for the training algorithm is generated; based on the training algorithm and the configuration file, a third preset number of candidate computing power schemes are generated by the computing power scheduling module; each candidate computing power scheme has detailed computing power information that the user can browse and view; based on the third preset number of candidate computing power schemes, a dialogue interaction center generates candidate computing power schemes in the form of entries and provides feedback to the user.

[0008] In response to the user's selection of candidate computing power schemes, the selected candidate computing power scheme is determined as the training scheme; based on the training dataset, training algorithm, and training scheme, a training configuration file is generated; based on the training configuration file, the training algorithm is iteratively trained to obtain the trained clinical research algorithm;

[0009] In response to the user's inference text input operation for the clinical research algorithm, an inference configuration file is generated; based on the inference configuration file and the clinical research algorithm, an inference report of the clinical research algorithm is generated.

[0010] In one embodiment, the method further includes:

[0011] When a user's comparison instruction for clinical research algorithms is received, a second preset number of candidate comparison algorithms are generated through the algorithm management agent; in response to the user's selection operation for the candidate comparison algorithms, the candidate comparison algorithms are determined as the comparison algorithms; based on the comparison algorithms, a first preset number of candidate comparison datasets are generated through the medical dataset management agent, and a third preset number of candidate comparison computing power schemes are generated through the computing power scheduling module; in response to the user's selection operation for the candidate comparison datasets and candidate comparison computing power schemes, the comparison datasets and comparison schemes are obtained; based on the comparison datasets and comparison schemes, the comparison algorithms and clinical research algorithms are trained and compared to obtain the comparison results.

[0012] In one embodiment, based on a comparison dataset and a comparison scheme, a comparison algorithm and a clinical research algorithm are trained and compared to obtain comparison results, including:

[0013] According to the comparison scheme, the clinical research algorithm and the comparison algorithm are trained separately on the training set of the comparison dataset to obtain the first clinical research algorithm and the first comparison algorithm. The first clinical research algorithm and the first comparison algorithm are tested on the test set of the comparison dataset to obtain medical test results and comparison test results. Based on the medical test results and comparison test results, algorithm information, multiple evaluation indicators, visual bar charts, and recommendation results are generated for the clinical research algorithm and the comparison algorithm. The recommendation results are either a recommendation for the comparison algorithm or a recommendation for the clinical research algorithm. The algorithm information, multiple evaluation indicators, visual comparison charts, and recommendation results together constitute the comparison results.

[0014] In one embodiment, in response to a user's inference text input operation for a clinical research algorithm, an inference profile is generated, including:

[0015] In response to the user's input of inference text for the clinical research algorithm, a first preset number of candidate inference datasets are generated through the medical dataset management agent; in response to the user's selection of candidate inference datasets, the candidate inference datasets are determined as inference datasets; based on the inference datasets, a third preset number of candidate inference computing power schemes are generated through the computing power scheduling module; in response to the user's selection of candidate inference computing power schemes, an inference scheme is generated; based on the inference datasets and inference schemes, an inference configuration file is generated.

[0016] In one embodiment, based on the inference profile and the clinical research algorithm, an inference report for the clinical research algorithm is generated, including:

[0017] The inference dataset is preprocessed to obtain a preprocessed inference dataset; the preprocessed inference dataset is input into a clinical research algorithm to obtain inference results; based on the inference results, an inference report is generated, which includes: information about the inference dataset, algorithm information of the clinical research algorithm, inference accuracy, and score.

[0018] In one embodiment, the training algorithm is iteratively trained based on a training configuration file to obtain a trained clinical research algorithm, including:

[0019] For any given iteration of training, the training algorithm is trained on the training set in the training dataset to obtain the training model for the current iteration; the training model for the current iteration is tested on the test set in the training dataset to obtain the test accuracy for the current iteration; if the test accuracy for the current iteration is greater than the test accuracy for the previous iteration, then the next iteration of training is performed; if the test accuracy for the current iteration is not greater than the test accuracy for the previous iteration, then the early stopping count for the previous iteration is incremented by 1 to obtain the early stopping count for the current iteration; if the early stopping count for the current iteration is not less than a preset value, then the iteration training is stopped, and the training model for the current iteration is used as the completed clinical research algorithm.

[0020] In one embodiment, the method further includes:

[0021] The system responds to user input of medical literature and generates literature reading requirements. It then analyzes these requirements through a pre-trained dialogue interaction center to obtain reading instructions. These instructions are then analyzed to obtain literature reading information. If the reading information is a file in a preset format, the file is parsed to obtain file information. If the reading information is a link to a preset PDF document, the file information is obtained based on the link. Information is extracted from the file information to generate literature tags and content. The literature content includes an abstract, research objective, methods, results, and conclusions.

[0022] In one embodiment, in response to a user's selection of a candidate dataset, determining the user-selected candidate dataset as the training dataset includes:

[0023] In response to the user's selection of a candidate dataset, a dataset instruction is generated; the dataset instruction is parsed to obtain dataset selection information; if the dataset selection information is a dataset uploaded by the user, then the dataset uploaded by the user is used as the training dataset selected by the user; if the dataset selection information is selection information for a candidate dataset, then the target dataset is selected from the candidate datasets as the training dataset selected by the user.

[0024] Secondly, this application also provides a multi-agent collaborative clinical research implementation device, comprising:

[0025] The requirement input module is used to respond to the user's clinical research text input operation and generate requirement information; the requirement information is parsed through a pre-trained dialogue interaction center to obtain retrieval instructions; based on the retrieval instructions, candidate datasets are retrieved through the medical dataset management agent to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that can be browsed and viewed by the user; based on the first preset number of candidate datasets, the candidate datasets in the form of entries are generated through the dialogue interaction center and fed back to the user.

[0026] The scheme selection module is used to respond to the user's selection operation on the candidate dataset, determine the candidate dataset selected by the user as the training dataset; combine the retrieval command and the training dataset to obtain the dataset command; based on the dataset command, generate a second preset number of candidate algorithms through the algorithm management agent; each candidate algorithm has detailed algorithm information that can be viewed by the user; based on the second preset number of candidate algorithms, generate candidate algorithms in the form of entries through the dialogue interaction center and feed them back to the user; respond to the user's selection operation on the candidate algorithm, determine the candidate algorithm selected by the user as the training algorithm, and generate the configuration file of the training algorithm; based on the training algorithm and the configuration file, generate a third preset number of candidate computing power schemes through the computing power scheduling module; each candidate computing power scheme has detailed computing power information that can be viewed by the user; based on the third preset number of candidate computing power schemes, generate candidate computing power schemes in the form of entries through the dialogue interaction center and feed them back to the user.

[0027] The training module is used to respond to the user's selection of candidate computing power schemes, determine the candidate computing power scheme selected by the user as the training scheme, generate a training configuration file based on the training dataset, training algorithm and training scheme, and iteratively train the training algorithm based on the training configuration file to obtain the trained clinical research algorithm.

[0028] The inference module is used to generate an inference configuration file in response to the user's inference text input operation for the clinical research algorithm; and to generate an inference report for the clinical research algorithm based on the inference configuration file and the clinical research algorithm.

[0029] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0030] In response to the user's clinical research text input, a requirement information is generated; the requirement information is parsed through a pre-trained dialogue interaction center to obtain a search instruction; based on the search instruction, a candidate dataset is retrieved through a medical dataset management agent to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that the user can browse and view; based on the first preset number of candidate datasets, a candidate dataset in the form of entries is generated through the dialogue interaction center and fed back to the user;

[0031] In response to the user's selection of a candidate dataset, the selected candidate dataset is determined as the training dataset; the retrieval command and the training dataset are combined to obtain a dataset command; based on the dataset command, a second preset number of candidate algorithms are generated by the algorithm management agent; each candidate algorithm has detailed algorithm information that the user can browse and view; based on the second preset number of candidate algorithms, a dialogue interaction center generates candidate algorithms in the form of entries and provides feedback to the user; in response to the user's selection of a candidate algorithm, the selected candidate algorithm is determined as the training algorithm, and a configuration file for the training algorithm is generated; based on the training algorithm and the configuration file, a third preset number of candidate computing power schemes are generated by the computing power scheduling module; each candidate computing power scheme has detailed computing power information that the user can browse and view; based on the third preset number of candidate computing power schemes, a dialogue interaction center generates candidate computing power schemes in the form of entries and provides feedback to the user.

[0032] In response to the user's selection of candidate computing power schemes, the selected candidate computing power scheme is determined as the training scheme; based on the training dataset, training algorithm, and training scheme, a training configuration file is generated; based on the training configuration file, the training algorithm is iteratively trained to obtain the trained clinical research algorithm;

[0033] In response to the user's inference text input operation for the clinical research algorithm, an inference configuration file is generated; based on the inference configuration file and the clinical research algorithm, an inference report of the clinical research algorithm is generated.

[0034] The aforementioned method for automating clinical research first generates requirement information in response to the user's input of clinical research text. This requirement information is then parsed by a pre-trained dialogue interaction center to obtain a search instruction. Based on the search instruction, a medical dataset management agent retrieves candidate datasets, resulting in a first preset number of candidate datasets. Each candidate dataset has detailed data available for the user to browse. Based on the first preset number of candidate datasets, the dialogue interaction center generates candidate datasets in the form of entries and provides feedback to the user. Next, in response to the user's selection of a candidate dataset, the user-selected dataset is used as the training dataset. The search instruction and the training dataset are combined to obtain a dataset instruction. Based on the dataset instruction, an algorithm management agent generates a second preset number of candidate algorithms. Each candidate algorithm has detailed algorithm data available for the user to browse. Based on the second preset number of candidate algorithms, the dialogue interaction center generates candidate algorithms in the form of entries. The algorithm selection process is as follows: Algorithm selection feedback is sent to the user; then, in response to the user's selection of a candidate algorithm, the selected algorithm is chosen as the training algorithm, and a configuration file for the training algorithm is generated; based on the training algorithm and configuration file, a third preset number of candidate computing power schemes are generated through the computing power scheduling module; each candidate computing power scheme has detailed computing power information that the user can view; based on the third preset number of candidate computing power schemes, a term-based candidate computing power scheme is generated and sent back to the user through the dialogue interaction center; in response to the user's selection of a candidate computing power scheme, the selected scheme is chosen as the training scheme; based on the training dataset, training algorithm, and training scheme, a training configuration file is generated; based on the training configuration file, the training algorithm is iteratively trained to obtain the trained clinical research algorithm; finally, in response to the user's input of inference text for the clinical research algorithm, an inference configuration file is generated; based on the inference configuration file and the clinical research algorithm, an inference report for the clinical research algorithm is generated. This application breaks down the entire scientific research process into standardized and automated sequential steps by using natural language interaction as the core and linking intelligent agents such as dataset management, algorithm management, and computing power scheduling. This greatly simplifies the operation process by eliminating the need for users to manually switch platforms or write code. By replacing decentralized manual operations with fully automated linkage, it effectively reduces the learning cost and operational error rate for researchers without technical backgrounds, significantly shortens the research cycle, and provides efficient, convenient, and accurate intelligent support for clinical research. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a method for automating clinical research in one embodiment;

[0037] Figure 2 This is a flowchart illustrating the comparison of trained algorithms in one embodiment;

[0038] Figure 3 This is a structural block diagram of a device for automating clinical research in one embodiment;

[0039] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0041] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0042] In one embodiment, such as Figure 1 As shown, a method for implementing multi-agent collaborative clinical research is provided. This embodiment illustrates the method using a terminal as an example. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0043] Step 102: In response to the user's clinical research text input operation, generate demand information; parse the demand information through a pre-trained dialogue interaction center to obtain retrieval instructions; based on the retrieval instructions, retrieve candidate datasets through the medical dataset management agent to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that the user can browse and view; based on the first preset number of candidate datasets, generate candidate datasets in the form of entries through the dialogue interaction center and feed them back to the user.

[0044] Before starting, users need to log in to their accounts and manage the permissions of different accounts through the permission management module. Specifically, this includes: Data permissions: granting access to different levels of datasets based on user level; Operation permissions: restricting high-risk operations such as algorithm modification and data export for ordinary users, allowing only administrators and authorized research leaders to perform these operations.

[0045] Optionally, this application includes an interactive result feedback module, which contains a dialogue interaction hub. The interactive result feedback module uses the dialogue interaction hub to convert the processing results of each functional module (such as dataset recommendation list, algorithm matching results, model training progress, computing power scheme recommendation, reasoning conclusion, etc.) into intuitive and easy-to-understand natural language text or visual charts, and provides accurate feedback to the user in the form of real-time dialogue. For example, when the dataset management module returns a maximum of 5 matching datasets, the module will generate the following feedback text: "Based on your input research needs, the following suitable datasets are recommended for you. You can directly select them through the dialog (e.g., 'Select the second one'), or click 'More Datasets' to view other resources, or create a private dataset: 1. Glaucoma + Medical Imaging + 2D Image Classification + RIM-ONE; 2. Glaucoma + Medical Imaging + 2D Image Classification + LAG; 3. Glaucoma + Medical Imaging + 2D Image Classification + G1020; 4. Glaucoma + Medical Imaging + 2D Image Classification + ACRIMA; 5. Glaucoma + Medical Imaging + 2D Image Classification + ORIGA-light"; If it is an algorithm matching result, the feedback text will be: "Based on the features of the dataset you selected, we recommend..." When choosing an algorithm, you can directly specify your selection (e.g., 'Choose DenseNet'), or click 'More Algorithms' to expand your choices: 1. 2D Image Classification + Medical Imaging + gMLP; 2. 2D Image Classification + Medical Imaging + DenseNet; 3. 2D Image Classification + Medical Imaging + VGG; 4. 2D Image Classification + Medical Imaging + AugViT; 5. 2D Image Classification + Medical Imaging + GoogLeNet. During model training, real-time progress feedback will be pushed: "Model training has completed 30 / 100 rounds, current training loss is 0.32, validation set AUC value is 0.88, estimated remaining time is 45 minutes." All feedback allows users to perform further operations through natural language dialogue (such as adjusting parameters, switching resources, pausing tasks, etc.), achieving a closed-loop interaction.

[0046] Optionally, the dialogue interaction center is a joint model that integrates a medical-specific natural language processing (NLP) model with the Qwen-14B large model. The medical-specific NLP model is pre-trained on a medical research corpus containing core elements such as disease names, research terms, algorithm names, and data types, and has the ability to accurately extract key information for research needs. During the parsing process, the natural language research text input by the user (such as "constructing a glaucoma prediction model based on fundus images") is first subjected to semantic analysis and keyword extraction to filter out core dimension information such as disease type, target task, and data type. Then, it is structured and transformed into retrieval instructions containing retrieval dimensions and matching conditions (retrieval instructions such as: glaucoma + medical imaging + two-dimensional image classification) to ensure the accuracy of the retrieval direction. The medical dataset management agent incorporates a distributed encrypted storage architecture layered by "disease name - data type - target task - sensitivity level" and a keyword cluster classification system in the medical field. During retrieval, it first matches keyword clusters based on the core keywords in the search command (such as "glaucoma", "fundus imaging", "2D image classification") to filter out candidate datasets with consistent labels. Then, it prioritizes the candidate datasets through a multi-factor weighted scoring model. The multi-factors include dataset sample size, historical usage frequency, update time, matching degree with research objectives, and user's historical usage preferences. Finally, it outputs a first preset number (preferably 5) of candidate datasets, and each candidate dataset is associated with a dataset details entry, including complete information such as data structure, purpose, examples, and references, for user decision-making reference.

[0047] Step 104: In response to the user's selection of a candidate dataset, the user-selected candidate dataset is determined as the training dataset; the retrieval command and the training dataset are combined to obtain a dataset command; based on the dataset command, a second preset number of candidate algorithms are generated by the algorithm management agent; each candidate algorithm has detailed algorithm information that the user can browse and view; based on the second preset number of candidate algorithms, candidate algorithms in the form of entries are generated and fed back to the user through the dialogue interaction center; in response to the user's selection of a candidate algorithm, the user-selected candidate algorithm is determined as the training algorithm, and a configuration file for the training algorithm is generated; based on the training algorithm and the configuration file, a third preset number of candidate computing power schemes are generated by the computing power scheduling module; each candidate computing power scheme has detailed computing power information that the user can browse and view; based on the third preset number of candidate computing power schemes, candidate computing power schemes in the form of entries are generated and fed back to the user through the dialogue interaction center.

[0048] This application includes a memory unit module that employs a session state management mechanism to automatically store the historical interaction records of the current research session, including information such as the user-selected dataset, algorithm, parameter configuration, and computing power scheme. When the user pauses the research task or adjusts their needs midway, the task can be quickly resumed based on the historical records without repeating the operation.

[0049] Optionally, the dataset in this application is stored in a data storage module. This module employs a distributed, partitioned, encrypted storage architecture, storing medical data in a hierarchical manner according to four dimensions: "disease name - data type - target task - sensitivity level." The disease name hierarchy follows the International Classification of Diseases (ICD-10) standard, dividing data into major categories such as "circulatory system diseases," "nervous system diseases," and "tumor diseases," with further subdivisions within each category (e.g., "tumor diseases" includes "lung cancer," "breast cancer," and "colorectal cancer"). The data type hierarchy divides data into structured data (e.g., clinical indicators, laboratory test results), semi-structured data (e.g., electronic medical record text), and unstructured data (e.g., medical images, basic medical records, etc.). The data is stored using sequencing data and adapted from relational databases (MySQL), document databases (MongoDB), and object storage services (OSS). The target task is stratified according to research objectives, including sleep analysis, epileptic seizure identification, cohort studies, clinical classification, clinical prediction, two-dimensional image classification, two-dimensional image segmentation, protein-ligand set affinity prediction, and property prediction classification. Sensitivity level stratification divides the data into public data (such as de-identified public research datasets) and private data (such as individual patient data obtained by the user). Public data is stored using standard encryption, while private data is stored with independent encryption and a dedicated access channel to ensure data storage security and compliance.

[0050] Optionally, user selection of candidate datasets includes text-based interactive selection (such as typing "select the second one") or interface point selection. After the system parses the user's command through the dialogue interaction center, it locks the target dataset and marks it as the training dataset, while recording the core features of the dataset (including sample size, data type, target task adaptation label, sensitivity level, etc.). The dataset command (e.g., glaucoma + medical imaging + 2D image classification + dataset) is a structured combination of the retrieval command and the features of the training dataset. In addition to containing information such as disease type and target task from the original retrieval command, it adds specific attribute parameters of the training dataset (e.g., "data type: medical imaging (256×256 pixels), sample size: 7714 cases, task adaptation: 2D image classification"), providing accurate basis for algorithm matching. The algorithm management agent has a built-in library of medical research algorithms categorized by "target task - data type". Based on dataset instructions, it extracts core matching dimensions (target task + data type) and filters candidate algorithms whose suitability meets a preset threshold. Then, it performs a secondary sorting based on indicators such as historical application accuracy, running efficiency, and resource consumption, finally outputting a second preset number (ideally 5) of candidate algorithms. Each candidate algorithm is associated with an algorithm details entry, including algorithm function, principle, references, etc. The configuration file for training algorithms contains a list of adjustable core parameters, specifically annotating the default value, adjustment range, parameter meaning, and impact on the model for each parameter (e.g., "learning rate: default 0.001, adjustment range 0.0001-0.01, affecting model convergence speed and stability"). Users can adjust these parameters themselves to generate the final configuration file or directly use the default parameter configuration. The computing power scheduling module first collects the data characteristics (sample size, data size, data type), training algorithm type, and parameters (batch size, training rounds, etc.) in the dataset instructions, and calculates the required basic computing power specifications through the computing power evaluation model. Then, combined with the real-time status of the platform's computing power resource pool, it generates a third preset number (preferably 3) of candidate computing power solutions, namely a normal version that meets basic requirements (low time consumption), a recommended version that balances efficiency and cost, and a high-performance, short-duration intelligent version. Each solution clearly indicates the CPU / GPU specifications, memory / video memory capacity, and expected running time, allowing users to choose according to their needs.

[0051] Step 106: In response to the user's selection of candidate computing power schemes, determine the candidate computing power scheme selected by the user as the training scheme; generate a training configuration file based on the training dataset, training algorithm and training scheme; iteratively train the training algorithm based on the training configuration file to obtain the trained clinical research algorithm.

[0052] Optionally, users can select candidate computing power schemes through text interaction (e.g., typing "select recommended version") or interface clicks. The system interprets the command through the dialogue interaction center, locks the target scheme, identifies it as the training scheme, and simultaneously records the core resource parameters of the scheme (including CPU / GPU core count, memory / video memory capacity, computing node identifier, etc.). The training configuration file is a standardized file that integrates the core information of the training dataset, training algorithm, and training scheme. Specifically, it includes: dataset configuration (storage path, data preprocessing rules, training / validation set partition ratio), algorithm configuration (algorithm core code path, parameter configuration file, dependency environment list), computing power configuration (computing node allocation information, container creation parameters, resource usage threshold), and task metadata (task name, user ID, estimated training duration, log storage path), ensuring the traceability of the training process and the consistency of the environment. During the iterative training of the training algorithm, the system automatically activates a training monitoring mechanism: real-time collection of training progress data (current training round, percentage completed, remaining time), and calculation of performance metrics on the validation set after each iteration (accuracy, AUC (Area Under Curve), and F1 score for classification tasks; MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and R² (R-Squared / Coefficient of Error) for regression tasks). The system performs determination (coefficient of determination) and generates performance curves. If the performance metric shows no improvement or even declines after five consecutive iterations, an early stopping mechanism is automatically triggered to avoid overfitting. Simultaneously, it monitors hardware resource status (CPU utilization, GPU memory usage, and overall memory usage). If resource shortages or hardware failures occur, training is immediately paused, and an error message with solutions is pushed (e.g., "GPU memory overflow; it is recommended to adjust the batch size to 16 or upgrade the computing power"). After training, the system stores the clinical research algorithm in a standard format (.pth format for PyTorch, .h5 format for TensorFlow) in the user's dedicated space. The file contains the complete model structure, post-training weight parameters, optimizer status, and records of the best-performing iterations. A model metadata report is also generated, clearly annotating key information during training (dataset used, algorithm parameters, computing power configuration, final performance metrics, and training log path), providing a complete basis for subsequent model inference, reuse, and comparison.

[0053] Step 108: In response to the user's inference text input operation for the clinical research algorithm, generate an inference configuration file; based on the inference configuration file and the clinical research algorithm, generate an inference report for the clinical research algorithm.

[0054] Optionally, in the inference report generation stage, the trained clinical research algorithm is first loaded and its integrity and environmental compatibility are verified. After verification, the inference environment is initialized, and the preprocessed inference data is input into the model to perform prediction calculations. Subsequently, the original results output by the model (such as the category probability of the classification task and the image annotation results of the segmentation task) are transformed into natural language explanations that medical researchers can understand (such as "The probability of glaucoma corresponding to the patient's fundus image is 82%, with high confidence, and it is recommended to conduct further pathological examination for confirmation"). At the same time, core evaluation indicators are calculated (AUC value, accuracy, and F1 value for classification tasks, and MAE and RMSE for regression tasks), and an inference summary of about 200 words is generated (covering data scale, core inference conclusions, and performance). Finally, the data details, algorithm principle summary, evaluation indicator values, natural language explanations, and inference summary are integrated to generate a standardized inference report, which supports export in PDF format. The entire inference process record (including inference data, configuration parameters, results, and timestamps) will be stored synchronously and can be traced by keyword retrieval.

[0055] The aforementioned method for automating clinical research first generates requirement information in response to the user's input of clinical research text. This requirement information is then parsed by a pre-trained dialogue interaction center to obtain a search instruction. Based on the search instruction, a medical dataset management agent retrieves candidate datasets, resulting in a first preset number of candidate datasets. Each candidate dataset has detailed data available for the user to browse. Based on the first preset number of candidate datasets, the dialogue interaction center generates candidate datasets in the form of entries and provides feedback to the user. Next, in response to the user's selection of a candidate dataset, the user-selected dataset is used as the training dataset. The search instruction and the training dataset are combined to obtain a dataset instruction. Based on the dataset instruction, an algorithm management agent generates a second preset number of candidate algorithms. Each candidate algorithm has detailed algorithm data available for the user to browse. Based on the second preset number of candidate algorithms, the dialogue interaction center generates candidate algorithms in the form of entries. The algorithm selection process is as follows: Algorithm selection feedback is sent to the user; then, in response to the user's selection of a candidate algorithm, the selected algorithm is chosen as the training algorithm, and a configuration file for the training algorithm is generated; based on the training algorithm and configuration file, a third preset number of candidate computing power schemes are generated through the computing power scheduling module; each candidate computing power scheme has detailed computing power information that the user can view; based on the third preset number of candidate computing power schemes, a term-based candidate computing power scheme is generated and sent back to the user through the dialogue interaction center; in response to the user's selection of a candidate computing power scheme, the selected scheme is chosen as the training scheme; based on the training dataset, training algorithm, and training scheme, a training configuration file is generated; based on the training configuration file, the training algorithm is iteratively trained to obtain the trained clinical research algorithm; finally, in response to the user's input of inference text for the clinical research algorithm, an inference configuration file is generated; based on the inference configuration file and the clinical research algorithm, an inference report for the clinical research algorithm is generated. This application breaks down the entire scientific research process into standardized and automated sequential steps by using natural language interaction as the core and linking intelligent agents such as dataset management, algorithm management, and computing power scheduling. This greatly simplifies the operation process by eliminating the need for users to manually switch platforms or write code. By replacing decentralized manual operations with fully automated linkage, it effectively reduces the learning cost and operational error rate for researchers without technical backgrounds, significantly shortens the research cycle, and provides efficient, convenient, and accurate intelligent support for clinical research.

[0056] In one exemplary embodiment, such as Figure 2 As shown, the method also includes:

[0057] When a user's comparison instruction for clinical research algorithms is received, a second preset number of candidate comparison algorithms are generated through the algorithm management agent; in response to the user's selection operation for the candidate comparison algorithms, the candidate comparison algorithms are determined as the comparison algorithms; based on the comparison algorithms, a first preset number of candidate comparison datasets are generated through the medical dataset management agent, and a third preset number of candidate comparison computing power schemes are generated through the computing power scheduling module; in response to the user's selection operation for the candidate comparison datasets and candidate comparison computing power schemes, the comparison datasets and comparison schemes are obtained; based on the comparison datasets and comparison schemes, the comparison algorithms and clinical research algorithms are trained and compared to obtain the comparison results.

[0058] For example, a user inputs a comparison command via natural language (e.g., "Compare the performance of this glaucoma prediction model with two other algorithms"). After parsing by the dialogue interaction center, the system accurately extracts the core comparison requirements (disease type: glaucoma, target task: 2D image classification, number of comparison algorithms: 2), triggering the model comparison process. The algorithm management agent, based on the core characteristics of the original clinical research algorithm (target task: 2D image classification, suitable data type: medical imaging), selects a second preset number (preferably 5) of candidate comparison algorithms from the algorithm library stored according to "target task - data type" classification. The text returned to the user is: "We recommend the following suitable comparison algorithms. You can select 2 or more for comparison (e.g., 'Select DenseNet and GoogLeNet'): 1. 2D image classification + medical imaging + gMLP; 2. 2D image classification + medical imaging + DenseNet; 3. 2D image classification + medical imaging + VGG; 4. 2D image classification + medical imaging + AugViT; 5. 2D image classification + medical imaging + GoogLeNet." After the user selects an algorithm through a dialog, the system identifies it as the comparison algorithm and retains the standard parameter configurations of each algorithm by default. A pop-up window prompts, "To ensure fairness in the comparison, it is recommended to use the default parameters; if adjustments are needed, please ensure that the parameter modification rules for all comparison algorithms are consistent." Subsequently, the medical dataset management agent generates a first preset number (preferably 5) of candidate comparison datasets based on the common adaptation requirements of the comparison algorithms and the original clinical research algorithms. The feedback text is: "Based on the adaptability of the selected algorithms, we recommend the following comparison datasets (all algorithms will be retrained and tested on the same dataset to ensure fairness in the comparison): 1. Glaucoma + Medical Imaging + 2D Image Classification + RIM-ONE; 2. Glaucoma + Medical Imaging + 2D Image Classification + LAG; 3. Glaucoma + Medical Imaging + 2D Image Classification + G1020; 4. Glaucoma + Medical Imaging + 2D Image Classification + ACRIMA; 5. Glaucoma + Medical Imaging + 2D Image Classification + ORIGA-light." Users can select or confirm the comparison dataset through a dialog, or click "More Datasets" or upload a private dataset. Simultaneously, the computing power scheduling module collects and compares the characteristics of the dataset (sample size, data size), all algorithm types and parameter configurations, and generates a third preset number (preferably 3) of candidate computing power solutions for comparison after comprehensive evaluation. The feedback text is: "Based on the scale of the comparison task, the following computing power solutions are recommended for you: 1. Standard version (16 CPU cores + 20GB GPU memory + 16GB RAM): estimated 3 hours; 2. Recommended version (16 CPU cores + 24GB GPU memory + 24GB RAM): estimated 1.5 hours; 3. Intelligent version (32 CPU cores + 32GB GPU memory + 32GB RAM): estimated 40 minutes." After the user selects, the comparison solution is confirmed.Based on the comparison dataset and comparison scheme, the system automatically allocates independent Docker containers for the original clinical research algorithm and each comparison algorithm, pre-installs a unified dependency environment, and synchronously starts the training process on the same comparison dataset. During training, it monitors the training progress and performance indicators (accuracy, AUC, F1 score, etc.) of each algorithm in real time, automatically triggers an early stopping mechanism to avoid overfitting, and records a complete training log. After training, it generates comparison results through multi-dimensional comparative analysis, including numerical comparisons of the core performance indicators of each algorithm, visualization charts of loss curves and accuracy curves, and statistics on training time and resource consumption. At the same time, it recommends the optimal model based on the user's potential needs (such as prioritizing accuracy or running efficiency), and finally generates a standardized comparison report containing all the above information. It supports PDF export and separate download of charts, which can be used by researchers for paper writing or results presentation.

[0059] In this embodiment, by taking the AI ​​dialogue interaction hub as the core and linking seven functional modules, namely dataset management, algorithm management, model training, model inference, model comparison, literature reading and computing power scheduling, a fully automated collaborative system is constructed, which can realize zero-code intelligent support for medical research.

[0060] In an exemplary embodiment, based on a comparison dataset and a comparison scheme, a comparison algorithm and a clinical research algorithm are trained and compared to obtain comparison results, including:

[0061] According to the comparison scheme, the clinical research algorithm and the comparison algorithm are trained separately on the training set of the comparison dataset to obtain the first clinical research algorithm and the first comparison algorithm. The first clinical research algorithm and the first comparison algorithm are tested on the test set of the comparison dataset to obtain medical test results and comparison test results. Based on the medical test results and comparison test results, algorithm information, multiple evaluation indicators, visual bar charts, and recommendation results are generated for the clinical research algorithm and the comparison algorithm. The recommendation results are either a recommendation for the comparison algorithm or a recommendation for the clinical research algorithm. The algorithm information, multiple evaluation indicators, visual bar charts, and recommendation results together constitute the comparison results.

[0062] For example, following the recommended computing power configuration of "16 CPU cores + 24GB GPU memory + 24GB RAM" in the comparison scheme, the system allocated separate Docker containers for the original clinical research algorithm (DenseNet) and the comparison algorithms (GoogLeNet, VGG), pre-installing a unified dependency environment (Python 3.8, PyTorch 1.12, CUDA 11.3) to ensure consistency in the training environment. Subsequently, the comparison dataset (RIM-ONE) was divided into training and test sets in a 7:3 ratio. Training of the three algorithms was started simultaneously on the training set—default parameters were used during training (learning rate 0.001, batch size 32, training epochs 50), and an early stopping mechanism was enabled (training was terminated if there was no performance improvement for 5 consecutive epochs). Data such as the number of training epochs, loss value, and validation set performance metrics of each algorithm were recorded in real time. After training, the optimized first clinical research algorithm (DenseNet optimized version) and the first comparison algorithm (GoogLeNet optimized version and VGG optimized version) were obtained. The system input the test set of the comparison dataset into the three algorithms for performance testing: For the glaucoma two-dimensional image classification task, the core evaluation index was calculated. The accuracy of the DenseNet optimized version was 92%, the AUC value was 0.91, the F1 value was 0.90, and the training time was 1.1 hours; the accuracy of the GoogLeNet optimized version was 88%, the AUC value was 0.87, the F1 value was 0.86, and the training time was 1.3 hours; the accuracy of the VGG optimized version was 85%, the AUC value was 0.83, the F1 value was 0.84, and the training time was 1.5 hours. Based on the above test results, the system generates a complete comparison result: the algorithm information section clearly lists the name of each algorithm, the applicable data type (medical imaging), the target task (2D image classification), and a summary of the core principles (such as the dense connection mechanism of DenseNet); multiple evaluation indicators are presented in tabular form, showing the specific values ​​and rankings of accuracy, AUC, F1 score, and training time; visual bar charts show the horizontal comparison of each indicator, intuitively presenting the differences in algorithm performance; the recommendation result, combined with the user's potential needs (with a default focus on comprehensive performance), generates a natural language conclusion: "The original clinical research algorithm (DenseNet optimized version) is recommended, as it performs best in all three core indicators—accuracy, AUC, and F1 score—and its comprehensive performance is more suitable for glaucoma prediction scenarios; if a balance between performance and computational cost is required, the GoogLeNet optimized version can be selected." The above algorithm information, evaluation indicator tables, visual bar charts, and recommendation results together constitute the final comparison result, supporting users to export a complete report and download the charts separately.

[0063] In this embodiment, the operation threshold is reduced through natural language interaction, the whole process is automated through module collaboration, resource utilization is optimized through intelligent adaptation, data compliance is ensured through security management, and results traceability is supported through standardized reports. This effectively solves the pain points of complex interaction, poor collaboration, and waste of resources in traditional scientific research, and shortens the research cycle of building glaucoma prediction models from several weeks in traditional methods to several hours, significantly improving the efficiency and accuracy of medical scientific research.

[0064] In one exemplary embodiment, in response to a user's inference text input operation for a clinical research algorithm, an inference profile is generated, including:

[0065] In response to the user's input of inference text for the clinical research algorithm, a first preset number of candidate inference datasets are generated through the medical dataset management agent; in response to the user's selection of candidate inference datasets, the candidate inference datasets are determined as inference datasets; based on the inference datasets, a third preset number of candidate inference computing power schemes are generated through the computing power scheduling module; in response to the user's selection of candidate inference computing power schemes, an inference scheme is generated; based on the inference datasets and inference schemes, an inference configuration file is generated.

[0066] For example, a user inputs a reasoning text command in natural language (such as "Use this model to reason about fundus image datasets and analyze the probability of glaucoma in patients") for a trained glaucoma prediction clinical research algorithm (DenseNet optimized version). After the system is parsed by the dialogue interaction center, it triggers the reasoning dataset recommendation process. The medical dataset management agent, based on the core adaptation features of this clinical research algorithm (target task: 2D image classification, data type: medical imaging, applicable disease: glaucoma), selects a first preset number (preferably 5) of candidate inference datasets with matching labels from a dataset library stored hierarchically in "disease name-data type-target task". The text returned to the user is: "The following adaptation inference datasets are recommended for you. You can directly select (e.g., 'select the third one') or click 'More datasets' / upload a private dataset: 1. Glaucoma + Medical Imaging + 2D Image Classification + RIM-ONE; 2. Glaucoma + Medical Imaging + 2D Image Classification + LAG; 3. Glaucoma + Medical Imaging + 2D Image Classification + ORIGA-light; 4. Glaucoma + Medical Imaging + 2D Image Classification + ACRIMA; 5. Glaucoma + Medical Imaging + 2D Image Classification + G1020." After the user selects "ORIGA-light" through the dialogue, the system determines that the dataset is an inference dataset and simultaneously collects its core features (sample size 584 cases, image size 256×256 pixels, including glaucoma and normal labels). Subsequently, the computing power scheduling module receives the information of "clinical research algorithm (DenseNet) + inference dataset features (sample size 584 + image size 256×256)," evaluates the basic computing power requirements of the agent (8 CPU cores + 16GB GPU memory + 16GB RAM), generates a third preset number (preferably 3) of candidate inference computing power solutions, and the feedback text is: "Based on the scale of the inference task, the following computing power solutions are recommended for you: 1. Standard version (8 CPU cores + 16GB GPU memory + 16GB RAM): estimated 40 minutes; 2. Recommended version (16 CPU cores + 20GB GPU memory + 24GB RAM): estimated 20 minutes; 3. Intelligent version (32 CPU cores + 32GB GPU memory + 32GB RAM): estimated 10 minutes." After the user selects the "Recommended Version", the system determines that the solution is the inference solution and integrates the storage path of the inference dataset, preprocessing rules, sample features, as well as the CPU / GPU specifications, memory / video memory capacity, and computing power node allocation information of the inference solution. At the same time, it supplements key information such as model loading path, dependent environment list, and inference result output format, and generates a standardized inference configuration file to ensure that the inference process is standardized and controllable, and to provide a complete basis for subsequent inference calculations.

[0067] In this embodiment, through intelligent recommendation of inference datasets and computing power solutions, natural language interaction selected by users, and automated integration of configuration files, the configuration process for inference tasks is simplified, reducing the operational difficulty for researchers without technical backgrounds. Furthermore, precise resource adaptation ensures inference efficiency and result accuracy, fully demonstrating the practicality and superiority of this method in medical research inference scenarios.

[0068] In one exemplary embodiment, based on the inference profile and the clinical research algorithm, an inference report for the clinical research algorithm is generated, including:

[0069] The inference dataset is preprocessed to obtain a preprocessed inference dataset; the preprocessed inference dataset is input into a clinical research algorithm to obtain inference results; based on the inference results, an inference report is generated, which includes: information about the inference dataset, algorithm information of the clinical research algorithm, inference accuracy, and score.

[0070] For example, the system first performs a standardized preprocessing procedure based on the features of the inference dataset (ORIGA-light) in the inference configuration file: First, it verifies the integrity of the dataset (e.g., checks whether the image files are corrupted or the labels are missing), and removes 3 invalid data. Then, it uniformly scales all fundus images to 256×256 pixels and performs grayscale normalization and noise removal. Finally, according to the input requirements of the clinical research algorithm (DenseNet optimized version), it converts the dataset to tensor format and completes batch partitioning to obtain a preprocessed inference dataset (581 valid samples, including 273 glaucoma samples and 308 normal samples). The preprocessed inference dataset is then input into the loaded clinical research algorithm to perform inference calculations and obtain the inference results: the algorithm outputs the probability of glaucoma and classification label for each image, correctly identifying 249 glaucoma samples and 287 normal samples, while also generating a confidence score (range 0-1) for each sample. Based on the above inference results, a structured inference report is generated, and the contents of each core module are as follows:

[0071] Information about the inference dataset: including dataset name (ORIGA-light), original sample size (584 cases), effective sample size (581 cases), data type (medical imaging - fundus imaging), label distribution (glaucoma 273 cases / normal 308 cases), preprocessing operations (size normalization, grayscale calibration, noise removal), and data source (publicly available de-identified research dataset); Algorithm information for the clinical research algorithm: algorithm name (DenseNet optimized version), core principle (dense connection mechanism, enhancing feature propagation and reuse), adapted task (2D image classification), training parameters (learning rate 0.005, batch size 16, training epochs 35), and model storage path ( / user / XXX / models / glaucoma analysis-D (enseNet-20250930.pth); Inference Accuracy: The overall accuracy is (249+287) / 581≈92.26%, with glaucoma sample recall of 249 / 273≈91.21% and precision of 249 / 265≈93.96%, and normal sample recall of 287 / 308≈93.18% and precision of 287 / 316≈90.82%; Scoring: It consists of two parts: first, the overall model performance score (calculated based on a weighted average of accuracy, recall, and precision, with a maximum score of 10 and a score of 9.2); second, the single-sample confidence score (each inference result is labeled with a corresponding confidence level, such as the image with the number IMG_0123 having a glaucoma probability of 89% and a confidence score of 0.87). This inference report is generated in PDF format, supports online viewing and download, and includes detailed tables of inference results and visualization charts of performance indicators, facilitating researchers to quickly verify the model's inference performance and data reliability.

[0072] In this embodiment, through standardized data preprocessing, precise reasoning calculation, and structured report generation, not only are the core results of glaucoma prediction efficiently output, but researchers are also provided with complete reasoning basis and performance evaluation, which greatly reduces the workload of organizing research results and fully demonstrates the efficiency and practicality of this method in medical research reasoning scenarios.

[0073] In one exemplary embodiment, the training algorithm is iteratively trained based on a training configuration file to obtain a trained clinical research algorithm, including:

[0074] For any given iteration of training, the training algorithm is trained on the training set in the training dataset to obtain the training model for the current iteration; the training model for the current iteration is tested on the test set in the training dataset to obtain the test accuracy for the current iteration; if the test accuracy for the current iteration is greater than the test accuracy for the previous iteration, then the next iteration of training is performed; if the test accuracy for the current iteration is not greater than the test accuracy for the previous iteration, then the early stopping count for the previous iteration is incremented by 1 to obtain the early stopping count for the current iteration; if the early stopping count for the current iteration is not less than a preset value, then the iteration training is stopped, and the training model for the current iteration is used as the completed clinical research algorithm.

[0075] For example, the training dataset (LAG dataset, divided into a 7:3 ratio of 490 training cases and 210 test cases), training algorithm (DenseNet), and parameter configuration (learning rate 0.005, batch size 16, training epochs 100, preset early stopping value 5) are specified in the training configuration file, and the iterative training process is started. For each iteration, the training algorithm is trained based on the training set in the training dataset to obtain the training model for the current iteration; the training model for the current iteration is tested based on the test set in the training dataset to obtain the test accuracy for the current iteration; if the test accuracy for the current iteration is greater than the test accuracy for the previous iteration, the next iteration of training is performed; if the test accuracy for the current iteration is not greater than the test accuracy for the previous iteration, the early stopping count for the previous iteration is incremented by 1 to obtain the early stopping count for the current iteration; if the early stopping count for the current iteration is not less than the preset value, the iterative training is stopped, and the training model for the current iteration is used as the completed clinical research algorithm; the core information of the model, such as the training epochs, weight parameters, and optimizer status, is recorded synchronously and stored in the user's dedicated space.

[0076] In this embodiment, the dynamic iterative judgment and early stopping mechanism with test accuracy as the core avoids the waste of computing power and time caused by invalid iterations, and accurately locks the model with the best performance, effectively balancing training efficiency and model accuracy, and fully demonstrating the scientific nature and efficiency of this training method in the construction of medical research models.

[0077] In one exemplary embodiment, the method further includes:

[0078] The system responds to user input of medical literature and generates literature reading requirements. It then analyzes these requirements through a pre-trained dialogue interaction center to obtain reading instructions. These instructions are then analyzed to obtain literature reading information. If the reading information is a file in a preset format, the file is parsed to obtain file information. If the reading information is a link to a preset PDF document, the file information is obtained based on the link. Information is extracted from the file information to generate literature tags and content. The literature content includes an abstract, research objective, methods, results, and conclusions.

[0079] For example, a user inputs a literature reading request: "Analyze the core content of literature related to deep learning diagnosis of glaucoma," and uploads a local PDF document (pre-set format file). The pre-trained dialogue interaction center analyzes this request, obtaining the literature reading instructions: "Topic: Glaucoma + Fundus Imaging + Deep Learning, Operation: Extract Core Information, Source: Local PDF." After parsing the instructions, the system automatically analyzes the file structure and text content, obtaining file information including full-text text and figure annotations. Subsequently, it extracts information using medical-specific text mining technology, generating literature tags such as "glaucoma," "fundus imaging," "deep learning," and "disease diagnosis." Simultaneously, it structurally breaks down the literature content: the abstract summarizes the overall research, clearly stating the research objective as "improving the accuracy of early glaucoma imaging diagnosis," detailing the datasets, algorithms, and experimental design used, presenting core indicators in the results, and summarizing the research value and limitations in the conclusions. Finally, a structured list is provided to the user.

[0080] In this embodiment, users do not need complicated operations. They only need to input their reading needs and upload a PDF file or provide a PDF link. The system can automatically connect to resources, parse the file, and generate accurate literature tags and structured content to help researchers quickly grasp the core value of the literature. At the same time, the tagging process facilitates subsequent retrieval and management, provides knowledge support for research design and algorithm selection, further improves the automated system of the entire clinical research process, and enhances research efficiency.

[0081] In an exemplary embodiment, in response to a user's selection operation on a candidate dataset, determining the user-selected candidate dataset as the training dataset includes:

[0082] In response to the user's selection of a candidate dataset, a dataset instruction is generated; the dataset instruction is parsed to obtain dataset selection information; if the dataset selection information is a dataset uploaded by the user, then the dataset uploaded by the user is used as the training dataset selected by the user; if the dataset selection information is selection information for a candidate dataset, then the target dataset is selected from the candidate datasets as the training dataset selected by the user.

[0083] For example, after the system recommends five candidate datasets (Dataset 1, Dataset 2, Dataset 3, etc.) to the user, the user triggers the dataset selection operation by entering "Select the second dataset" in the dialog, clicking the "Dataset 2" option on the interface, or entering "Upload local glaucoma image dataset". The system then generates the corresponding dataset instruction. After parsing the instruction, the dataset selection information is obtained: if it is "Upload local glaucoma image dataset", the system verifies the integrity of the uploaded compressed file (including data samples and label files) and then determines the uploaded dataset as the training dataset; if it is "Select the second dataset", then Dataset 2 is matched from the candidate datasets and determined as the training dataset selected by the user.

[0084] In this embodiment, a dataset instruction is generated and the selection information is parsed based on the user's operation. This can flexibly adapt to two core scenarios: candidate dataset selection and private dataset upload, and accurately locate the training dataset required by the user.

[0085] In an exemplary embodiment, a multi-agent collaborative clinical research implementation method includes: generating requirement information in response to a user's clinical research text input operation; parsing the requirement information through a pre-trained dialogue interaction center to obtain a retrieval instruction; based on the retrieval instruction, retrieving candidate datasets through a medical dataset management agent to obtain a first preset number of candidate datasets; in response to a user's selection operation for a candidate dataset, determining the user-selected candidate dataset as a training dataset; combining the retrieval instruction and the training dataset to obtain a dataset instruction; based on the dataset instruction, generating a second preset number of candidate algorithms through an algorithm management agent; in response to a user's selection operation for a candidate algorithm, determining the user-selected candidate algorithm as a training algorithm and generating a configuration file for the training algorithm; based on the training algorithm and the configuration file, generating a third preset number of candidate computing power schemes through a computing power scheduling module; responding to... Upon the user's selection of candidate computing power schemes, the selected scheme is determined as the training scheme. A training configuration file is generated based on the training dataset, training algorithm, and training scheme. The training algorithm is iteratively trained based on the configuration file. For each iteration, the algorithm is trained on the training set in the training dataset to obtain the training model for the current iteration. The training model is tested on the test set in the training dataset to obtain the test accuracy for the current iteration. If the test accuracy for the current iteration is greater than that for the previous iteration, the next iteration begins. If the test accuracy for the current iteration is not greater than that for the previous iteration, the early stopping count for the previous iteration is incremented by 1 to obtain the early stopping count for the current iteration. If the early stopping count for the current iteration is not less than a preset value, the iterative training stops, and the training model for the current iteration is used as the completed clinical research algorithm. In response to the user's input of inference text for the clinical research algorithm, a first preset number of candidate inference datasets are generated through the medical dataset management agent. In response to the user's selection of a candidate inference dataset, the candidate inference dataset is selected as the inference dataset. Based on the inference dataset, a third preset number of candidate inference computing power schemes are generated through the computing power scheduling module. In response to the user's selection of a candidate inference computing power scheme, an inference scheme is generated. Based on the inference dataset and the inference scheme, an inference configuration file is generated. The inference dataset is preprocessed to obtain a preprocessed inference dataset. The preprocessed inference dataset is input into the clinical research algorithm to obtain inference results. Based on the inference results, an inference report is generated, which includes: information about the inference dataset, algorithm information of the clinical research algorithm, inference accuracy, and score.Upon receiving a user's comparison instruction for clinical research algorithms, the algorithm management agent generates a second preset number of candidate comparison algorithms. In response to the user's selection of a candidate comparison algorithm, the algorithm is selected as the comparison algorithm. Based on the comparison algorithm, the medical dataset management agent generates a first preset number of candidate comparison datasets, and the computing power scheduling module generates a third preset number of candidate comparison computing power schemes. In response to the user's selection of the candidate comparison datasets and candidate comparison computing power schemes, a comparison dataset and a comparison scheme are obtained. According to the comparison scheme, the clinical research algorithm and the comparison algorithm are trained on the training set of the comparison dataset, respectively, to obtain a first clinical research algorithm and a first comparison algorithm. The first clinical research algorithm and the first comparison algorithm are tested on the test set of the comparison dataset to obtain medical test results and comparison test results. Based on the medical test results and comparison test results, algorithm information, multiple evaluation indicators, a visual bar chart, and a recommendation result are generated for the clinical research algorithm and the comparison algorithm. The recommendation result is either a recommended comparison algorithm or a recommended clinical research algorithm. The algorithm information, multiple evaluation indicators, the visual bar chart, and the recommendation result together constitute the comparison result.

[0086] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0087] In one exemplary embodiment, such as Figure 3 As shown, a multi-agent collaborative clinical research implementation device is provided, including: a demand input module 301, a scheme selection module 302, a training module 303, and an inference module 304, wherein:

[0088] The requirement input module is used to respond to the user's clinical research text input operation and generate requirement information; the requirement information is parsed through a pre-trained dialogue interaction center to obtain retrieval instructions; based on the retrieval instructions, candidate datasets are retrieved through the medical dataset management agent to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that can be browsed and viewed by the user; based on the first preset number of candidate datasets, the candidate datasets in the form of entries are generated through the dialogue interaction center and fed back to the user.

[0089] The scheme selection module is used to respond to the user's selection operation on the candidate dataset, determine the candidate dataset selected by the user as the training dataset; combine the retrieval command and the training dataset to obtain the dataset command; based on the dataset command, generate a second preset number of candidate algorithms through the algorithm management agent; each candidate algorithm has detailed algorithm information that can be viewed by the user; based on the second preset number of candidate algorithms, generate candidate algorithms in the form of entries through the dialogue interaction center and feed them back to the user; respond to the user's selection operation on the candidate algorithm, determine the candidate algorithm selected by the user as the training algorithm, and generate the configuration file of the training algorithm; based on the training algorithm and the configuration file, generate a third preset number of candidate computing power schemes through the computing power scheduling module; each candidate computing power scheme has detailed computing power information that can be viewed by the user; based on the third preset number of candidate computing power schemes, generate candidate computing power schemes in the form of entries through the dialogue interaction center and feed them back to the user.

[0090] The training module is used to respond to the user's selection of candidate computing power schemes, determine the candidate computing power scheme selected by the user as the training scheme, generate a training configuration file based on the training dataset, training algorithm and training scheme, and iteratively train the training algorithm based on the training configuration file to obtain the trained clinical research algorithm.

[0091] The inference module is used to generate an inference configuration file in response to the user's inference text input operation for the clinical research algorithm; and to generate an inference report for the clinical research algorithm based on the inference configuration file and the clinical research algorithm.

[0092] In one exemplary embodiment, it further includes: a comparison module, used for:

[0093] When a user's comparison instruction for clinical research algorithms is received, a second preset number of candidate comparison algorithms are generated through the algorithm management agent; in response to the user's selection operation for the candidate comparison algorithms, the candidate comparison algorithms are determined as the comparison algorithms; based on the comparison algorithms, a first preset number of candidate comparison datasets are generated through the medical dataset management agent, and a third preset number of candidate comparison computing power schemes are generated through the computing power scheduling module; in response to the user's selection operation for the candidate comparison datasets and candidate comparison computing power schemes, the comparison datasets and comparison schemes are obtained; based on the comparison datasets and comparison schemes, the comparison algorithms and clinical research algorithms are trained and compared to obtain the comparison results.

[0094] In one exemplary embodiment, the comparison module is further configured to:

[0095] According to the comparison scheme, the clinical research algorithm and the comparison algorithm are trained separately on the training set of the comparison dataset to obtain the first clinical research algorithm and the first comparison algorithm. The first clinical research algorithm and the first comparison algorithm are tested on the test set of the comparison dataset to obtain medical test results and comparison test results. Based on the medical test results and comparison test results, algorithm information, multiple evaluation indicators, visual bar charts, and recommendation results are generated for the clinical research algorithm and the comparison algorithm. The recommendation results are either a recommendation for the comparison algorithm or a recommendation for the clinical research algorithm. The algorithm information, multiple evaluation indicators, visual bar charts, and recommendation results together constitute the comparison results.

[0096] In one exemplary embodiment, the inference module is further configured to:

[0097] In response to the user's input of inference text for the clinical research algorithm, a first preset number of candidate inference datasets are generated through the medical dataset management agent; in response to the user's selection of candidate inference datasets, the candidate inference datasets are determined as inference datasets; based on the inference datasets, a third preset number of candidate inference computing power schemes are generated through the computing power scheduling module; in response to the user's selection of candidate inference computing power schemes, an inference scheme is generated; based on the inference datasets and inference schemes, an inference configuration file is generated.

[0098] In one exemplary embodiment, the inference module is further configured to:

[0099] The inference dataset is preprocessed to obtain a preprocessed inference dataset; the preprocessed inference dataset is input into a clinical research algorithm to obtain inference results; based on the inference results, an inference report is generated, which includes: information about the inference dataset, algorithm information of the clinical research algorithm, inference accuracy, and score.

[0100] In one exemplary embodiment, the training module is further configured to:

[0101] For any given iteration of training, the training algorithm is trained on the training set in the training dataset to obtain the training model for the current iteration; the training model for the current iteration is tested on the test set in the training dataset to obtain the test accuracy for the current iteration; if the test accuracy for the current iteration is greater than the test accuracy for the previous iteration, then the next iteration of training is performed; if the test accuracy for the current iteration is not greater than the test accuracy for the previous iteration, then the early stopping count for the previous iteration is incremented by 1 to obtain the early stopping count for the current iteration; if the early stopping count for the current iteration is not less than a preset value, then the iteration training is stopped, and the training model for the current iteration is used as the completed clinical research algorithm.

[0102] In one exemplary embodiment, the demand input module is further configured to:

[0103] The system analyzes the demand information through a pre-trained dialogue interaction center to obtain document reading instructions; it then analyzes these instructions to obtain document reading information; if the document reading information is a file in a preset format, it analyzes the file to obtain file information; if the document reading information is a link to a preset document PDF, it obtains file information based on the link; it then extracts information from the file information to generate document tags and document content; the document content includes an abstract, research objective, methods, results, and conclusions.

[0104] In one exemplary embodiment, the training module is further configured to:

[0105] In response to the user's selection of a candidate dataset, a dataset instruction is generated; the dataset instruction is parsed to obtain dataset selection information; if the dataset selection information is a dataset uploaded by the user, then the dataset uploaded by the user is used as the training dataset selected by the user; if the dataset selection information is selection information for a candidate dataset, then the target dataset is selected from the candidate datasets as the training dataset selected by the user.

[0106] The modules in the aforementioned automated clinical research device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0107] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a multi-agent collaborative clinical research method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0108] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0109] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0110] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0112] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for implementing multi-agent collaborative clinical research, characterized in that, The method includes: In response to the user's clinical research text input, a requirement information is generated; the requirement information is parsed through a pre-trained dialogue interaction center to obtain a search instruction; based on the search instruction, a medical dataset management agent is used to search for candidate datasets to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that the user can browse and view; based on the first preset number of candidate datasets, a candidate dataset in the form of entries is generated through the dialogue interaction center and fed back to the user; In response to the user's selection of the candidate dataset, the selected candidate dataset is determined as the training dataset; the retrieval instruction and the training dataset are combined to obtain a dataset instruction; based on the dataset instruction, a second preset number of candidate algorithms are generated by the algorithm management agent; each candidate algorithm has detailed algorithm information that the user can browse and view; based on the second preset number of candidate algorithms, a dialogue interaction center generates candidate algorithms in the form of entries and feeds them back to the user; in response to the user's selection of the candidate algorithms, the selected candidate algorithm is determined as the training algorithm, and a configuration file for the training algorithm is generated; based on the training algorithm and the configuration file, a third preset number of candidate computing power schemes are generated by the computing power scheduling module; each candidate computing power scheme has detailed computing power information that the user can browse and view; based on the third preset number of candidate computing power schemes, a candidate computing power scheme in the form of entries is generated by the dialogue interaction center and fed back to the user. In response to the user's selection of the candidate computing power scheme, the selected candidate computing power scheme is determined as the training scheme; a training configuration file is generated based on the training dataset, the training algorithm, and the training scheme; the training algorithm is iteratively trained based on the training configuration file to obtain the trained clinical research algorithm; In response to a user's input of inference text for the clinical research algorithm, an inference configuration file is generated; based on the inference configuration file and the clinical research algorithm, an inference report for the clinical research algorithm is generated.

2. The method according to claim 1, characterized in that, The method further includes: When a user's comparison instruction for the clinical research algorithm is received, a second preset number of candidate comparison algorithms are generated through the algorithm management agent; In response to the user's selection operation for the candidate comparison algorithm, a candidate comparison algorithm is determined as the comparison algorithm; Based on the comparison algorithm, a first preset number of candidate comparison datasets are generated by the medical dataset management agent, and a third preset number of candidate comparison computing power schemes are generated by the computing power scheduling module. In response to the user's selection operation for the candidate comparison dataset and the candidate comparison computing power scheme, the comparison dataset and comparison scheme are obtained; Based on the comparison dataset and the comparison scheme, the comparison algorithm and the clinical research algorithm are trained and compared to obtain the comparison results.

3. The method according to claim 2, characterized in that, The process of training and comparing the comparison algorithm and the clinical research algorithm based on the comparison dataset and the comparison scheme to obtain comparison results includes: According to the comparison scheme, the clinical research algorithm and the comparison algorithm are trained on the training set of the comparison dataset respectively to obtain the first clinical research algorithm and the first comparison algorithm. The first clinical research algorithm and the first comparison algorithm are tested based on the test set of the comparison dataset to obtain medical test results and comparison test results; Based on the medical test results and the comparative test results, algorithm information, multiple evaluation indicators, a visual comparison chart, and recommendation results are generated for the clinical research algorithm and the comparative algorithm; the recommendation results are either a recommendation for the comparative algorithm or a recommendation for the clinical research algorithm; the algorithm information, the multiple evaluation indicators, the visual comparison chart, and the recommendation results together constitute the comparison results.

4. The method according to claim 1, characterized in that, The step of generating an inference configuration file in response to a user's inference text input operation for the clinical research algorithm includes: In response to the user's inference text input operation for the clinical research algorithm, a first preset number of candidate inference datasets are generated through the medical dataset management agent; In response to the user's selection operation for the candidate inference dataset, the selected candidate inference dataset is determined as the inference dataset; based on the inference dataset, a third preset number of candidate inference computing power schemes are generated through the computing power scheduling module; In response to the user's selection of the candidate inference computing power scheme, an inference scheme is generated; Based on the inference dataset and the inference scheme, an inference configuration file is generated.

5. The method according to claim 4, characterized in that, The step of generating an inference report for the clinical research algorithm based on the inference configuration file and the clinical research algorithm includes: The inference dataset is preprocessed to obtain a preprocessed inference dataset; The preprocessed inference dataset is input into the clinical research algorithm to obtain the inference result; Based on the reasoning results, a reasoning report is generated, which includes: information about the reasoning dataset, algorithmic information of the clinical research algorithm, reasoning accuracy, and score.

6. The method according to claim 1, characterized in that, The step of iteratively training the training algorithm based on the training configuration file to obtain the trained clinical research algorithm includes: For any iteration of the training process, the training algorithm is trained based on the training set in the training dataset to obtain the training model for the current iteration; The training model for the current iteration is tested based on the test set in the training dataset to obtain the test accuracy for the current iteration. If the test accuracy of the current iteration is greater than the test accuracy of the previous iteration, then proceed to the next iteration of training. If the test accuracy of the current iteration is not greater than the test accuracy of the previous iteration, then the early stopping count of the previous iteration is incremented by 1 to obtain the early stopping count of the current iteration. If the number of early stops in the current iteration is not less than a preset value, then the iterative training is stopped, and the training model of the current iteration is used as the completed clinical research algorithm.

7. The method according to claim 1, characterized in that, The method further includes: Responding to users' input of medical literature, generating literature reading requirements; The document reading needs are analyzed by a pre-trained dialogue interaction center to obtain document reading instructions; The document reading instruction is parsed to obtain document reading information; If the document reading information is a file in a preset format, then the file is parsed to obtain file information; if the document reading information is a link to a preset document PDF, then file information is obtained based on the link. Information is extracted from the document to generate document tags and document content; the document content includes abstract, research purpose, methods, results and conclusions.

8. The method according to claim 1, characterized in that, The step of determining the candidate dataset selected by the user as the training dataset in response to the user's selection operation on the candidate dataset includes: In response to the user's selection of the candidate dataset, a dataset instruction is generated; Parse the dataset instructions to obtain dataset selection information; If the dataset selection information is a dataset uploaded by the user, then the dataset uploaded by the user will be used as the training dataset selected by the user. If the dataset selection information is for the candidate dataset, then the target dataset is selected from the candidate dataset as the training dataset selected by the user.

9. A multi-agent collaborative clinical research implementation device, characterized in that, The device includes: The requirement input module is used to respond to the user's clinical research text input operation and generate requirement information; the requirement information is parsed through a pre-trained dialogue interaction center to obtain a search instruction; based on the search instruction, the medical dataset management agent is used to search for candidate datasets to obtain a first preset number of candidate datasets; each candidate dataset has detailed dataset information that can be viewed by the user; based on the first preset number of candidate datasets, the dialogue interaction center generates candidate datasets in the form of entries and feeds them back to the user. The scheme selection module is used to respond to the user's selection operation on the candidate dataset, determine the candidate dataset selected by the user as the training dataset; combine the retrieval instruction and the training dataset to obtain a dataset instruction; based on the dataset instruction, generate a second preset number of candidate algorithms through the algorithm management agent; each candidate algorithm has detailed algorithm information that can be viewed by the user; based on the second preset number of candidate algorithms, generate candidate algorithms in the form of entries through the dialogue interaction center and feed them back to the user; respond to the user's selection operation on the candidate algorithms, determine the candidate algorithm selected by the user as the training algorithm, and generate a configuration file for the training algorithm; based on the training algorithm and the configuration file, generate a third preset number of candidate computing power schemes through the computing power scheduling module; each candidate computing power scheme has detailed computing power information that can be viewed by the user; based on the third preset number of candidate computing power schemes, generate candidate computing power schemes in the form of charts through the dialogue interaction center and feed them back to the user. The training module is used to respond to the user's selection operation for the candidate computing power scheme, determine the candidate computing power scheme selected by the user as the training scheme; generate a training configuration file based on the training dataset, the training algorithm and the training scheme; and perform iterative training on the training algorithm based on the training configuration file to obtain the trained clinical research algorithm. The inference module is used to generate an inference configuration file in response to a user's inference text input operation for the clinical research algorithm; and to generate an inference report for the clinical research algorithm based on the inference configuration file and the clinical research algorithm.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images