Gynecological tumor consultation processing method based on LLMs and multidisciplinary agent team

Abstract

The invention provides a gynecological tumor consultation processing method based on LLMs and a multidisciplinary agent team, a multidisciplinary agent team framework is specially designed for gynecological tumor consultation, and a deep multidisciplinary collaborative decision process designed by the invention is realized through a large language model. Verification results show that the collaborative artificial intelligence system has great potential in supporting complex medical decision, can provide operable, convenient and accurate guidance in clinical application, and has good application prospects.

Description

Technical Field

[0001] This application relates to the field of medical technology, specifically to a method for gynecological tumor consultation and treatment based on LLMs and multidisciplinary intelligent agent teams. Background Technology

[0002] Gynecologic oncology is a particularly complex and challenging field, often requiring multidisciplinary teams due to the combined complexity of cancer management and the specialized nature of women's reproductive health. Cancer cases inherently require multidisciplinary collaboration, involving pathologists, radiologists, medical oncologists, and surgeons to develop optimal treatment plans. In a gynecologic context, this complexity is further amplified, often necessitating the involvement of reproductive medicine specialists for fertility preservation assessments and interventions; if pregnancy is involved, collaboration with the obstetrics team is also required to formulate maternal and fetal management decisions. Therefore, a multidisciplinary team approach is widely considered the gold standard for comprehensive patient assessment in oncology.

[0003] In the field of Artificial Intelligence (AI), while large language models (LLMs) have demonstrated significant potential in various medical areas, they still face numerous challenges in handling clinical scenarios requiring complex diagnostic reasoning and multidisciplinary collaboration. Recently, large language models have shown promising results in medical knowledge acquisition and basic diagnostic tasks, but their performance in real-world, complex clinical environments remains limited, especially in specialties that heavily rely on collaborative expertise. Although current large language models possess vast medical knowledge bases, they still exhibit significant limitations when faced with the nuanced decision-making processes required in complex medical situations.

[0004] The gap between theoretical medical knowledge and actual clinical application remains one of the major challenges facing current medical artificial intelligence systems. Therefore, based on the introduction of large-scale language models, how to construct a highly targeted and accurate processing model for gynecological tumor consultation is still a problem that urgently needs to be solved. Summary of the Invention

[0005] This application provides a gynecological oncology consultation and treatment method based on LLMs and a multidisciplinary agent team. A multidisciplinary agent team (MDAT) framework is designed specifically for gynecological oncology consultation. The deep multidisciplinary collaborative decision-making process designed in this application is realized through a large language model. The verification results show that this collaborative artificial intelligence system has great potential in supporting complex medical decisions. In clinical applications, it can provide operable, convenient and accurate guidance and has good application prospects.

[0006] Firstly, this application provides a method for gynecological tumor consultation and treatment based on LLMs and a multidisciplinary intelligent agent team, the method comprising: Multidisciplinary sample data were collected for different types of gynecological tumors, including different clinical problems corresponding to different types of gynecological tumors. The processing model configured under the multidisciplinary intelligent agent team framework designed for gynecologic oncology clinics is trained using training samples configured with multidisciplinary sample data. The processing model adopts a multi-model fusion architecture, specifically including a first large-scale language model, a second large-scale language model, and a third large-scale language model. The working process of configuring the processing model includes: taking a virtual senior gynecologic oncology expert as the core, automatically determining the required professional field based on the complexity of the case, generating individual experts covering multiple disciplines as different intelligent agents to respond, and then resolving the conflict of opinions among experts through a consensus building mechanism to integrate and form the final diagnosis and treatment recommendations. Based on the current case data requiring gynecological tumor consultation and treatment, a trained processing model is used to process the data and obtain the corresponding gynecological tumor consultation and treatment results.

[0007] Secondly, this application provides a gynecological tumor consultation and treatment device based on LLMs and a multidisciplinary intelligent agent team, the device comprising: The data acquisition unit is used to collect multidisciplinary sample data of different types of gynecological tumors. The multidisciplinary sample data involves different clinical problems corresponding to different types of gynecological tumors. The training unit is used to configure the processing model under the multidisciplinary intelligent agent team framework designed for gynecologic oncology consultation clinics. It is trained using training samples configured with multidisciplinary sample data. The processing model adopts a multi-model fusion architecture, which specifically includes a first large-scale language model, a second large-scale language model, and a third large-scale language model. The working process of configuring the processing model includes: taking a virtual senior gynecologic oncology expert as the core, automatically determining the required professional field based on the complexity of the case, generating individual experts covering multiple disciplines as different intelligent agents to respond, and then resolving the conflict of opinions among experts through a consensus building mechanism to integrate and form the final diagnosis and treatment recommendations. The application unit is used to process the case data required for gynecological tumor consultation and treatment using a trained processing model, and obtain the corresponding gynecological tumor consultation and treatment results.

[0008] Thirdly, this application provides a processing device, including a processor and a memory, wherein a computer program is stored in the memory, and the processor executes the method provided in the first aspect of this application when it invokes the computer program in the memory.

[0009] Fourthly, this application provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the method provided in the first aspect of this application.

[0010] From the above, it can be concluded that this application has the following beneficial effects: This application designs a multidisciplinary intelligent agent team framework specifically for gynecologic oncology consultation. It uses a large language model to realize the deep multidisciplinary collaborative decision-making process designed in this application. The verification results show that this collaborative artificial intelligence system has great potential in supporting complex medical decisions. In clinical applications, it can provide operable, convenient and accurate guidance and has good application prospects. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team, as described in this application. Figure 2 This is a schematic diagram of a gynecological tumor consultation and treatment device based on LLMs and a multidisciplinary intelligent agent team, as described in this application. Figure 3 This is a schematic diagram of one type of processing equipment used in this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. The naming or numbering of steps appearing in this application does not imply that the steps in the method flow must be performed in the chronological / logical order indicated by the naming or numbering. The execution order of named or numbered process steps can be changed according to the desired technical purpose, as long as the same or similar technical effect is achieved.

[0015] The module division described in this application is a logical division. In practical applications, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection between modules shown or discussed may be through some interfaces, and the indirect coupling or communication connection between modules may be electrical or other similar forms, none of which are limited in this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules. Some or all of the modules may be selected to achieve the purpose of the solution in this application according to actual needs.

[0016] Before introducing the gynecologic oncology consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams provided in this application, we will first introduce the background content involved in this application.

[0017] The gynecological oncology consultation and treatment method, device, and computer-readable storage medium based on LLMs and multidisciplinary intelligent agent teams provided in this application can be applied to processing equipment. A multidisciplinary intelligent agent team framework is designed specifically for gynecological oncology consultation. The deep multidisciplinary collaborative decision-making process designed in this application is realized through a large language model. The verification results show that this collaborative artificial intelligence system has great potential in supporting complex medical decisions. In clinical applications, it can provide operable, convenient, and accurate guidance and has good application prospects.

[0018] The gynecological tumor consultation and processing method based on LLMs and multidisciplinary intelligent agent teams mentioned in this application can be implemented by a gynecological tumor consultation and processing device based on LLMs and multidisciplinary intelligent agent teams, or by different types of processing devices such as servers, physical hosts, or user equipment (UEs) that integrate such devices. The gynecological tumor consultation and processing device based on LLMs and multidisciplinary intelligent agent teams can be implemented in hardware or software. Specifically, the UE can be a smartphone, tablet, laptop, desktop computer, or personal digital assistant (PDA) or other terminal device. The processing devices can be configured in a device cluster.

[0019] It is understandable that the solution proposed in this application is mainly based on existing data. Therefore, in practical applications, the processing equipment that implements the gynecological tumor consultation and processing method based on LLMs and multidisciplinary intelligent agent teams, or that is equipped with the corresponding application service of the gynecological tumor consultation and processing method based on LLMs and multidisciplinary intelligent agent teams, usually only needs to meet the required data processing capabilities. The specific equipment type and equipment deployment form are quite flexible and can be flexibly configured according to the actual situation.

[0020] If the direct collection of existing data (mainly case data) mentioned above is involved, it is necessary to integrate the existing data collection equipment into the processing equipment, or to include the existing data collection equipment in the equipment cluster of the processing equipment, or to provide the processing equipment with access through third-party equipment.

[0021] In addition, if there are corresponding content display requirements for the solution processing results or process, the processing device can also be equipped with a corresponding display screen (including touch screen) to display the content, or it can be connected to an external display device or call other devices with display screens to display the content.

[0022] The following section introduces the gynecological tumor consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams provided in this application.

[0023] First, refer to Figure 1 , Figure 1 This paper illustrates a flowchart of a gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team, as described in this application. The gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team provided in this application may specifically include the following steps S101 to S103: Step S101: Collect multidisciplinary sample data of different types of gynecological tumors, wherein the multidisciplinary sample data involves different clinical problems corresponding to different types of gynecological tumors; Understandably, this application aims to achieve multidisciplinary gynecological tumor consultation based on a large-scale language model. To meet the training requirements of the corresponding model, it is necessary to first configure multidisciplinary sample data of different types of gynecological tumors.

[0024] It is important to note that the multidisciplinary sample data here itself encompasses individual experts from multiple disciplines on various issues that may arise in current clinical consultations for gynecological oncology.

[0025] The collection and processing of multidisciplinary sample data here usually involves the extraction and processing of existing / pre-collected data. This can be done through manual entry, local reading, or extraction from data sources such as online systems. Of course, the possibility of conducting the collection in a practical situation cannot be ruled out.

[0026] Specifically, as an exemplary embodiment here, in this application, on the one hand, the different types of gynecological tumors involved can specifically cover ovarian cancer, cervical cancer, vulvar cancer, uterine tumors, gestational trophoblastic tumors and other types of cancer; On the other hand, multidisciplinary sample data can specifically include two types of data: cases generated by AI models and cases recorded in the PubMed medical database.

[0027] As an example, in the process of generating medical records using relevant AI models, taking another large language model as an example, the following processing can be performed: First, a target dataset containing 100 synthetic cases (5 cancer types, 20 cases per type) was constructed. Based on clinical guidelines, a Retrieval Augmented Generation (RAG) knowledge base was built. Subsequently, a large language model generated preliminary cases (n=118). After review and elimination of unqualified samples (n=18) by a panel of three medical experts, 100 validated gynecological tumor case data generated by the AI ​​model were finally retained, covering ovarian cancer (n=20), cervical cancer (n=20), vulvar cancer (n=20), uterine tumors (n=20), and gestational trophoblastic tumors (n=20).

[0028] The inclusion criteria for AI-generated cases may include: 1) Complete and comprehensive clinical presentation information, including patient demographics, chief symptoms, diagnostic tests, imaging results, and pathological diagnosis; 2) Strict adherence to current NCCN guidelines and established clinical treatment guidelines for gynecologic malignancies; 3) Sufficient clinical complexity, requiring multidisciplinary team assessment, including comprehensive consideration of surgery, medical oncology, and radiotherapy; 4) Disease staging and risk stratification consistent with real-world scenarios in standard clinical practice; 5) Incorporation of contemporary diagnostic and treatment methods, reflecting current clinical standards.

[0029] Exclusion criteria may include: 1) Synthetic cases lacking key clinical elements, such as staging information, histological diagnosis, or missing treatment decision-making processes; 2) Cases whose clinical manifestations are significantly inconsistent with existing disease spectrum or diagnostic criteria; 3) Cases containing factual errors in medical terminology, drug dosage, or operational descriptions; 4) Cases exhibiting artificial intelligence generation bias or "hallucination" phenomena, resulting in unprofessional content, falsified medical literature, or inappropriate medical advice that contradicts evidence-based practice; 5) Cases containing contradictory information or inconsistent timelines, affecting the clinical credibility of the case.

[0030] Three qualified gynecologic oncology experts independently evaluate each AI-generated case, assessing aspects such as clinical accuracy, diagnostic complexity, and the appropriateness of the treatment plan. A case is considered approved when at least two experts agree that it meets the inclusion criteria; if opinions differ, a third expert makes the final decision. Cases that fail to meet the multidisciplinary consultation assessment criteria are excluded and regenerated iteratively until they meet the established clinical quality standards and the target sample size is achieved. In cases of disagreement regarding case quality assessment or clinical interpretation, the expert panel reaches a consensus through focused discussion, ensuring consistency in medical content and logic, and clinical credibility across all cases.

[0031] As another example, the process of retrieving medical records from the PubMed medical database can include the following steps: A systematic literature search of the PubMed database (n=251) was conducted, and a rigorous process of expert selection and manual screening was followed by initial screening, exclusion, and full-text evaluation. Ultimately, 82 gynecological tumor MDT case reports were included, covering ovarian cancer (n=28), cervical cancer (n=19), other types of cancer (n=8), uterine tumors (n=21), and gestational trophoblastic tumors (n=6).

[0032] For the PubMed medical database, the inclusion criteria for the acquired data can include: 1) Complete clinical data on gynecological malignancies; 2) Detailed information to support multidisciplinary team (MDT) assessment; 3) Complex case presentation requiring multidisciplinary decision-making; 4) Clear treatment outcomes and follow-up data.

[0033] Exclusion criteria may include: 1) Studies that are not open access or whose full text is not available, making it impossible to complete a full evaluation; 2) Studies published before 2020, or from non-peer-reviewed publications, lacking quality assurance; 3) Review articles, conference abstracts, case series lacking individual case details, or duplicate publications with overlapping patients; 4) Studies not published in English.

[0034] Understandably, the specific types of gynecological tumors that this application can cover are limited in the scheme settings here. In practice, this application has shown very good intelligent multidisciplinary intelligent team consultation effect for these types of gynecological tumors.

[0035] Step S102: The processing model configured under the multidisciplinary intelligent agent team framework designed for the gynecologic oncology consultation clinic is trained using training samples configured with multidisciplinary sample data. The processing model adopts a multi-model fusion architecture, specifically including a first large-scale language model, a second large-scale language model, and a third large-scale language model. The working process of configuring the processing model includes: taking a virtual senior gynecologic oncology expert as the core, automatically determining the required professional field based on the complexity of the case, generating individual experts covering multiple disciplines as different intelligent agents to respond, and then resolving the conflict of opinions among experts through a consensus building mechanism to integrate and form the final diagnosis and treatment recommendations. It is understandable that the multidisciplinary sample data of different types of gynecological tumors obtained above can be considered as relatively raw data samples. However, the model training involved in this processing step needs to be configured to adapt to the specific model training scheme adopted or designed.

[0036] As an example, for the training samples, this application generated 34,848 question-answer pairs, covering 182 cases and 33 model-hint combinations, all of which were scored using an automated evaluation system. Simultaneously, three clinical experts independently evaluated 100 responses, constructing a robust gold standard for comparison. The scoring consistency among the three experts was extremely high, demonstrating excellent scoring consistency (see Table 1 below). The high consistency between the automated evaluation system and expert consensus indicates that the system can reliably achieve expert-level evaluation standards. The extremely low correlation between scores and token quantity (r=-0.0226) indicates that response length is essentially unrelated to quality scores, verifying that the automated evaluation system effectively eliminates length bias in scoring.

[0037] Table 1 - Examples of Scoring Consistency As another example, for the 182 finalized clinical cases, each case had multiple clinically significant questions independently designed by medical experts in the field of gynecologic oncology. These questions addressed decision-making issues that might arise during multidisciplinary consultations. The content of these questions covered multiple key aspects of patient management, including: diagnostic procedures, staging examinations, treatment plan development, prognostic assessment, and follow-up strategies. The question-building process generated a total of 1,056 independent assessment items, achieving comprehensive coverage across multiple dimensions of clinical competence and providing a solid foundation for subsequent model performance evaluation.

[0038] All questions are categorized according to specific clinical subfields, covering the following seven dimensions: 1) Diagnostic reasoning and differential diagnosis, 2) Staging assessment and risk stratification, 3) Surgical strategy development and surgical procedure selection, 4) Selection of systemic treatment (including chemotherapy, targeted therapy, and immunotherapy) regimens, 5) Indications and technical considerations for radiotherapy, 6) Supportive care and complication management, and 7) Prognostic assessment and patient communication. This multidimensional classification system helps to comprehensively evaluate the model's performance in all key aspects of gynecologic oncology clinical practice, ensuring coverage of the complete clinical decision-making process from diagnosis to treatment, and from technical decisions to humanistic communication.

[0039] Furthermore, to ensure a comprehensive assessment of case complexity and to validate the applicability of the dataset in multidisciplinary consultation assessment, this application also scored all 1056 independent assessment items on six key dimensions to quantify their overall complexity.

[0040] Specifically, in order to achieve a consistent and structured scoring standard, this application has also developed a dedicated scoring scale, the details of which are shown in Table 2 below. This scoring scale covers six core assessment dimensions: 1) diagnostic complexity, 2) staging assessment difficulty, 3) treatment decision complexity, 4) multidisciplinary collaboration requirements, 5) prognostic assessment challenges, and 6) clinical decision risks.

[0041] Each dimension is scored by experts based on the complexity of the problem itself, with scores ranging from 1 (lowest difficulty / complexity) to 5 (highest difficulty / complexity).

[0042] The scores from the six dimensions are added together to form the final quantitative difficulty score for each question, which objectively reflects the consensus judgment of experts on the inherent complexity of the question, so as to better assess and demonstrate the complexity of the training samples, i.e., question-answer pairs, involved in the configuration and processing model of this application.

[0043] To verify the reliability and internal consistency of the manual scoring process, the consistency of scoring among the three gynecologic oncology experts was assessed. Specifically, the intraclass correlation coefficient (ICC) for all question scores was calculated. This indicator is a standard method for measuring the degree of consistency among multiple reviewers and can provide a robust quantitative basis for the consistency and reproducibility of the assigned difficulty scores.

[0044] Table 2 - Examples of Quantitative Scoring of Problem Difficulty In addition, it is worth noting that the multidisciplinary sample data of different types of gynecological tumors obtained above are not necessarily the most original data samples. They may involve corresponding preprocessing operations to further enhance the data quality, or the preprocessing operations can also be performed in step S102 here.

[0045] In terms of details, the preprocessing operations involved in this application may include specific operations such as abnormal data removal, blank data filling, and normalization. Considering that these fall within the scope of existing technology, they will not be elaborated on here. However, it should be understood that in actual situations, it is also possible to adopt specific preprocessing operations that are improved from existing technology or even novel.

[0046] It is important to note that the focus of the model framework construction here is the processing model configured under the Multi-Disciplinary Agent Team (MDAT) framework designed for gynecologic oncology clinics in this application. This model is scalable and consists of three different types of large language models. Based on the first, second, and third large language models, the working process configured under this application's design can specifically include the following: Centered around a virtual senior gynecologic oncology expert, the system automatically determines the required professional field based on the complexity of the case and generates individual experts covering multiple disciplines as different intelligent agents to respond. Then, a consensus-building mechanism is used to resolve conflicts of opinion among experts and integrate them to form the final diagnosis and treatment recommendations.

[0047] In the aforementioned work, several points are noteworthy. This application, involving three large language models, also introduces a multi-agent generation mechanism for each model. This allows each model to virtually / fict a corresponding number of different agents – “multidisciplinary experts” – utilizing the collective intelligence of multiple AI agents, each employing different prompting strategies, to form a virtual multidisciplinary, multi-expert team discussion scenario. Then, through a correspondingly configured consensus-building mechanism, consensus is built on the diverse professional knowledge and viewpoints obtained, eliminating conflicts of opinion among the “multidisciplinary experts.” Finally, the opinions are integrated / fused to obtain the final diagnostic and treatment recommendations for output. Under this setting, these recommendations clearly demonstrate a comprehensive, in-depth, and integrated collaborative question-and-answer processing effect, encompassing all aspects of the complexities inherent in gynecologic oncology.

[0048] Thus, within the framework of the multidisciplinary intelligent agent team in this application, the interpretability and reliability of large language models in medical reasoning are enhanced through an open and transparent expert agent collaborative reasoning process. This enables clinicians to track and verify the decision-making path behind each recommendation, laying the foundation for a trustworthy AI clinical assistance system.

[0049] Furthermore, this application also provides a set of supporting solutions for the three large language models involved in the multidisciplinary intelligent agent team framework.

[0050] Specifically, as an exemplary embodiment here, the first large-scale language model of this application, when configured to use a closed-source model, can specifically be the ChatGPT-4o model; When configured to use open-source models, the second and third large-scale language models in this application can be DeepSeek-R1 and Llama-4 models, respectively.

[0051] It is understandable that the three major models—ChatGPT-4o, DeepSeek-R1, and Llama-4—cover different scales and architectures. The ChatGPT-4o (i.e., gpt-4o-2024-11-20) model represents the most advanced closed-source business model currently available and has strong reasoning capabilities. The DeepSeek-R1 (DeepSeek-R1-0528) and Llama-4 (meta-llama / Llama-4-Maverick-17B-128E-Original) models are representative of open-source models, supporting reproducible research and having wider accessibility.

[0052] It is important to understand that although the above three models are currently the most powerful large-scale language models on the market, the high-performance multidisciplinary intelligent agent team consultation processing effect achieved by the multidisciplinary intelligent agent team framework of this application is not directly obtained from the model performance of these three models themselves. It may achieve better processing results compared to other specific large-scale language models.

[0053] In other words, by selecting the three specific models mentioned above, this application serves as an application example to conveniently evaluate the effectiveness of the multidisciplinary intelligent agent team framework of this application on different model architectures and computational scales. This ensures that the research results of the inventors of this application have universal applicability beyond single model types and confirms the effectiveness of the solution of this application.

[0054] Furthermore, in terms of validation, this application specifically compared and evaluated six carefully selected baseline methods, covering a variety of strategies for cue optimization and response generation: 1) Zero-shot uses the standard prompt method as the basic performance benchmark without any modifications; 2) Zero-shot-CoT introduces a step-by-step reasoning mechanism; 3) Self-refine, which achieves iterative optimization through an automatic feedback loop; 4) Universal Self-consistency (USC): Selects the optimal output through multiple rounds of response sampling; 5) Expert Prompting: Utilizing expertise in a specific field; 6) Multi-agent Debate: Competitive reasoning through opposing viewpoints.

[0055] The selection of the above six baselines ensures a comprehensive evaluation of the multidisciplinary intelligent agent team framework of this application from multiple dimensions, including basic prompting methods, enhanced reasoning strategies, iterative optimization techniques, consistency screening methods, adversarial reasoning and expert-driven strategies, thereby effectively verifying its comprehensive performance under the existing prompting method system.

[0056] To explore the relationship between model performance and problem difficulty, this application conducted a correlation analysis for each model-hint combination during the specific verification process. The analysis revealed a key result: the multidisciplinary agent team framework and its fixed agent variants showed a significant positive correlation (p < 0.001) in all models (including DeepSeekR1, Llama4, and ChatGPT4o models). That is, as the problem difficulty increases, the performance of these specific hint strategies actually improves.

[0057] At the same time, the inventors of this application have also conducted further research on the specific number of intelligent agents generated by large-scale language models as experts in different disciplines, or in other words, on the composition of the optimal intelligent agent team.

[0058] To identify the aforementioned issues, this application also employed a dual-strategy approach for a systematic analysis.

[0059] Firstly, in all three large-scale language models, four preset agent configurations (3, 5, 7, and 9 agents) were evaluated to achieve complementary medical expertise among different specialties in gynecologic oncology by progressively expanding team size while maintaining consistency in expert roles.

[0060] That is, as an exemplary embodiment here, under the condition that a fixed agent configuration is introduced in advance, the process of generating different agents for each large language model can also be configured to follow the constraint of 3, 5, 7 or 9 agents.

[0061] Secondly, this application also introduces an autonomous selection mechanism, in which each large language model independently determines the optimal number of agents required in each consultation case without any preset constraints.

[0062] That is, as an exemplary embodiment here, under the condition that an autonomous selection mechanism is introduced in advance, each large language model can also be configured to independently determine the optimal number of agents during the process of generating different agents.

[0063] Understandably, the systematic evaluation of different agent team configurations in this application shows a significant consistency between the empirically optimal performance range and the team structure autonomously chosen by the model. When tested with fixed team sizes (3, 5, 7, and 9 members), all models exhibited the best and most stable performance in team configurations of 5 to 7 experts, forming a clear performance plateau.

[0064] More importantly, when the models were given the task of autonomously selecting team size, their selection results were highly consistent with the above empirical findings: the ChatGPT-4o model selected an average of 5.70 people, the DeepSeek-R1 model selected an average of 5.98 people, and the Llama-4 model selected an average of 5.68 people. This result strongly indicates that large language models have the inherent ability to self-regulate and build effective team structures without explicit guidance, and can be used to deal with complex medical consultation tasks. This finding not only verifies the robustness of a 5 to 7 person configuration, but also reflects the model's advanced decision-making ability in multi-agent collaboration.

[0065] Corresponding to this discovery, as an exemplary embodiment, under the autonomous selection mechanism, the constraint on the optimal number of agents can be specifically configured to generate a range of 5 to 7 agents.

[0066] Furthermore, the comparison results between the multi-agent configuration group of this application and traditional single-cue strategies (such as zero-shot, COT, and self-refine) almost all showed highly significant differences (p < 0.001). This result strongly indicates that, regardless of the underlying model, the multi-agent method of this application significantly outperforms traditional cue strategies in overall performance.

[0067] First, the results show a clear and consistent performance hierarchy among the three large-scale language models. DeepSeek-R1 performs best, dominating the top positions. Its leading configurations, especially the fixed multi-agent cue strategy (ranked 3rd, 5th, 7th, and 9th), all achieved extremely high average scores, ranging from 97.84 to 97.93. Their 95% confidence intervals are narrow and highly overlapping, indicating that these configurations are indistinguishable in performance and represent the peak of model performance. Llama4 performs steadily, generally in the middle tier. Its best results also come from the multi-agent cue strategy, but its highest score of 93.72 is significantly lower than DeepSeek-R1. ChatGPT-4o ranks third, with its best-performing configuration (ranked 5th) achieving an average score of 88.78, significantly lower than the other two models.

[0068] Secondly, the data further demonstrates that the multi-agent framework of this application significantly outperforms traditional prompting strategies in all tested models. In three large-scale language models, fixed-agent prompting methods (e.g., 9-agent and 7-agent) consistently outperform single-response methods, including Chain-of-Thought (COT), zero-shot prompting, and Universal Self-Consistent (USC). This model highlights the inherent advantages of using structured, multi-perspective collaborative mechanisms when dealing with complex clinical problems—regardless of the underlying model used, the multi-agent framework of this application consistently improves model performance.

[0069] Furthermore, within the framework of a multidisciplinary intelligent agent team, the process of handling model configuration, namely the collaborative decision-making process, can be specifically divided into the following three stages as an exemplary embodiment.

[0070] The first phase involves dynamically assigning expert roles based on case complexity, which may include the following processing steps: 1.1) Input the patient's basic information, examination results, preliminary diagnosis, and treatment questions in the clinical case; 1.2) Based on the content of 1.1), suggestions are made for senior gynecologic oncology experts to analyze the complexity of cases; 1.3) Based on the processing results of 1.2), automatic expert selection determines the required species, generates role descriptions, and assigns expert identifiers.

[0071] The second phase involves parallel execution of multi-expert consultations through prompting, which may include the following processing steps: 2.1) Based on the processing results of 1.3), generate individual experts covering multiple disciplines as corresponding to the different intelligent agents; 2.2) Based on the content of 1.1), the individual experts covering multiple disciplines generated in 2.1) perform prompt word segmentation; 2.3) Based on the prompt words obtained from the segmentation in 2.2, individual experts covering multiple disciplines generated in 2.1) conduct consultations to obtain corresponding original answer summaries.

[0072] The third phase implements a seven-step consensus-building process, which includes the following steps: 3.1) Consistent fact identification and processing; 3.2) Conflict detection and handling; 3.3) Conflict resolution and handling; 3.4) Unique fact identification processing; 3.5) Integrate and process all facts; 3.6) Answer merging process; 3.7) Best answer selection process.

[0073] As can be seen, in this embodiment, the specific working logic of the processing model of this application, or the specific working logic of the multidisciplinary intelligent agent team framework of this application, is based on the three stages and provides a more specific and refined configuration scheme.

[0074] Meanwhile, in order to better and more accurately evaluate the model performance / response quality during and after model training, this application also constructs a comprehensive automated evaluation framework to evaluate all cases, response quality, and clinical appropriateness of responses. This framework adopts a standardized six-dimensional evaluation scheme to ensure the systematicness and professionalism of the evaluation process.

[0075] Specifically, to support consistent quantitative scoring of response quality, this application designed a structured six-dimensional Likert scale, which evaluates each response from six core dimensions, including: Inaccuracy, Omissions, Potential Harm, Bias, Clarity, and Hallucination.

[0076] Each dimension is rated on a scale of 1 (mildest / lowest frequency) to 5 (most severe / highest frequency), with a total score range of 5 to 30.

[0077] An automated assessment prompt was customized based on the Likert scale, and the model returned results in a structured JSON format for easy subsequent quantitative analysis.

[0078] For the complete definition and assessment prompts of the Likert scale, please see Tables 3-5 below.

[0079] Table 3 - Complete Definitions of Likert Scales 1 Table 4 - Complete Definitions of the Likert Scale 2 Table 5 - Components of the cue words used for assessment This application also employs Spearman's rank correlation coefficient to assess the consistency among three automated ratings. This method is suitable for detecting monotonic relationships and ranking consistency between rating results, especially when automated assessment systems may have differences in rating scales or biases, effectively determining whether their relative judgments of response quality remain consistent. Since only one set of automated ratings is ultimately selected as the basis for the results, Spearman's rank correlation coefficient is suitable for assessing "inter-raterreliability" between systems.

[0080] In the analysis of human-machine consistency, this application uses Pearson's product-moment correlation coefficient to measure the numerical fit between the automated scoring results and expert consensus. Pearson's product-moment correlation coefficient was chosen because it can measure the linear relationship between continuous variables and is suitable for assessing the extent to which automated scoring approximates the "gold standard" determined by expert review.

[0081] Since expert ratings are considered the gold standard, they focus more on the numerical accuracy of the ratings than just the consistency of the rankings. To ensure the statistical power of the assessment, this application selected 100 samples for consistency and correlation analysis. Based on previous power analysis (α=0.05, power=0.90), the minimum sample size required to detect a moderate effect size correlation (r=0.50) is approximately 36. Therefore, the 100 samples used are sufficient in terms of sample size and are statistically robust.

[0082] The evaluation process also identified several types of anomalous responses, requiring systematic processing to ensure a comprehensive assessment of the multidisciplinary agent team framework. The most common problem was model refusal to answer, where the model, due to its built-in safety mechanisms, occasionally refused to answer clinical questions to avoid potential medical liability. To address this, this application explicitly states its application scenario as educational and research use and regenerates the responses. Another challenge was inconsistent formatting and role confusion in multi-agent interactions, where individual expert agents provided advice outside their area of ​​expertise. To address this, this application deployed an automated verification system and corrected for errors with enhanced professionally targeted prompts, ultimately achieving a 99.8% success rate for structured responses.

[0083] In addition, a very small number of cases of “illusionary” medical information were found, in which the model generated clinical details that did not conform to the facts. Such errors can be identified through the accuracy judgment mechanism in the automated evaluation system. This phenomenon highlights that even in a collaborative framework designed with consensus mechanisms to improve reliability, robust verification mechanisms are still needed to ensure the safety of clinical AI applications.

[0084] Finally, the specific performance of the processing model configured under the multidisciplinary intelligent agent team framework of this application can be summarized as follows: 1) A systematic evaluation of the multidisciplinary intelligent agent team framework shows that collaborative artificial intelligence systems have significant potential to improve clinical decision-making capabilities in complex medical specialties. A systematic analysis of three advanced large-scale language models and six types of prompting methods shows that the DeepSeek-R1 model consistently outperforms other models in gynecologic oncology applications, while multi-agent collaborative strategies generally outperform traditional prompting methods in clinical reasoning tasks. 2) All models autonomously converged to select a team configuration of 5 to 7 experts without human intervention. This not only verified the optimality of this range in multidisciplinary consultation for gynecological oncology, but also demonstrated the high level of understanding of effective team collaboration mechanisms by modern large-scale language models. This finding provides an operational guideline for the implementation of AI-assisted multidisciplinary team systems in clinical practice, and highlights the potential ability of models to achieve self-regulation and structural optimization in complex collaborative scenarios. 3) The constructed automated evaluation framework has been rigorously validated and shows a high correlation with expert consensus. It establishes a scalable clinical artificial intelligence performance evaluation method. This progress effectively alleviates the key bottleneck problem that has long existed in medical artificial intelligence evaluation, making it possible to comprehensively evaluate the system capabilities in diverse clinical scenarios. 4) The research findings have deepened the understanding of the optimization path of collaborative artificial intelligence framework in the field of medical specialties. The multidisciplinary intelligent agent team method effectively reproduces the key elements of multidisciplinary clinical decision-making and shows its application potential in supporting medical workers to handle complex cases. 5) The research findings have deepened the understanding of the optimization path of collaborative artificial intelligence framework in the field of medical specialties. The multidisciplinary intelligent agent team approach effectively reproduces the key elements of multidisciplinary clinical decision-making and demonstrates its application potential in supporting medical workers in handling complex cases.

[0085] Step S103: Based on the case data for gynecological tumor consultation and treatment, the trained processing model is used to process the data and obtain the corresponding gynecological tumor consultation and treatment results.

[0086] Understandably, once the configuration of the processing model (mainly model training based on training samples) is completed within the framework of the multidisciplinary intelligent agent team in this application, it can be put into practical use and applied accordingly to the current gynecological oncology consultation task.

[0087] It is easy to understand that the case data required for gynecological tumor consultation and treatment can be obtained from a variety of sources in practice. The acquisition methods of the cases involved in the training samples can be referenced. These case data are usually real patient / case data. Of course, it is also possible that in some cases, case data obtained through secondary processing or even directly fabricated can be used for the corresponding research.

[0088] Thus, by inputting the case data that currently requires gynecological tumor consultation and treatment into the processing model and processing it accordingly, the corresponding gynecological tumor consultation and treatment results can be obtained, thus meeting the current needs for gynecological tumor consultation and treatment.

[0089] The results of this gynecological tumor consultation may involve further data application and processing.

[0090] For example, the results of this gynecological tumor consultation can be processed through local storage, off-site storage, result display, output completion notification, result push, or further data analysis.

[0091] Understandably, specific data application processing can be flexibly adjusted according to pre-configured and real-time data application strategies.

[0092] In conclusion, this application presents a multidisciplinary intelligent agent team framework specifically designed for gynecologic oncology consultations. It utilizes a large-scale language model to realize the deep multidisciplinary collaborative decision-making process designed in this application. Validation results demonstrate that this collaborative artificial intelligence system has great potential in supporting complex medical decisions and can provide operable, convenient, and precise guidance in clinical applications, showing promising application prospects.

[0093] The above is an introduction to the gynecological tumor consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams provided in this application. In order to facilitate better implementation of the gynecological tumor consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams provided in this application, this application also provides a gynecological tumor consultation and treatment device based on LLMs and multidisciplinary intelligent agent teams from the perspective of functional modules.

[0094] See Figure 2 , Figure 2 This is a schematic diagram of a gynecological tumor consultation and treatment device based on LLMs and a multidisciplinary intelligent agent team, as described in this application. Specifically, the gynecological tumor consultation and treatment device 200 based on LLMs and a multidisciplinary intelligent agent team may include the following structure: The acquisition unit 201 is used to collect multidisciplinary sample data of different types of gynecological tumors, wherein the multidisciplinary sample data involves different clinical problems corresponding to different types of gynecological tumors; Training unit 202 is used to configure the processing model under the multidisciplinary intelligent agent team framework designed for gynecologic oncology consultation clinics. It is trained using training samples configured with multidisciplinary sample data. The processing model adopts a multi-model fusion architecture, which specifically includes a first large-scale language model, a second large-scale language model, and a third large-scale language model. The working process of configuring the processing model includes: taking a virtual senior gynecologic oncology expert as the core, automatically determining the required professional field based on the complexity of the case, generating individual experts covering multiple disciplines as different intelligent agents to respond, and then resolving the conflict of opinions among experts through a consensus building mechanism to integrate and form the final diagnosis and treatment recommendations. Application unit 203 is used to process the case data that currently requires gynecological tumor consultation and treatment using a trained processing model, and obtain the corresponding gynecological tumor consultation and treatment results.

[0095] In one exemplary embodiment, the different types of gynecological tumors include ovarian cancer, cervical cancer, vulvar cancer, uterine tumors, gestational trophoblastic tumors, and other types of cancer; The multidisciplinary sample data includes cases generated by AI models and data from cases recorded in the PubMed medical database.

[0096] In yet another exemplary embodiment, the first large language model, when configured to use a closed-source model, specifically selects the ChatGPT-4o model; When configured to use open-source models, the second and third large-scale language models are DeepSeek-R1 and Llama-4 models, respectively.

[0097] In yet another exemplary embodiment, under the condition of a pre-introduced fixed agent configuration, each large language model is configured to follow a constraint of 3, 5, 7, or 9 agents during the generation of different agents.

[0098] In yet another exemplary embodiment, with a pre-introduced autonomous selection mechanism, each large language model is configured to independently determine the optimal number of agents during the generation of different agents.

[0099] In yet another exemplary embodiment, the optimal number of agents is specifically configured to be generated in the range of 5 to 7.

[0100] In yet another exemplary embodiment, the process of handling model configuration is specifically divided into three stages; The first phase involves dynamically assigning expert roles based on case complexity, and specifically includes the following processing steps: 1.1) Input the patient's basic information, examination results, preliminary diagnosis, and treatment questions in the clinical case; 1.2) Based on the content of 1.1), suggestions are made for senior gynecologic oncology experts to analyze the complexity of cases; 1.3) Based on the processing results of 1.2), automatic expert selection determines the required species, generates role descriptions, and assigns expert identifiers; The second phase involves parallel execution of multi-expert consultations through prompting, specifically including the following processing steps: 2.1) Based on the processing results of 1.3), generate individual experts covering multiple disciplines as corresponding intelligent agents; 2.2) Based on the content of 1.1), the prompt words are segmented by individual experts covering multiple disciplines generated by 2.1); 2.3) Based on the prompt words obtained from the segmentation in 2.2, individual experts covering multiple disciplines generated in 2.1) conduct consultations to obtain corresponding original answer summaries; The third phase implements a seven-step consensus-building process, which includes the following steps: 3.1) Consistent fact identification and processing; 3.2) Conflict detection and handling; 3.3) Conflict resolution and handling; 3.4) Unique fact identification processing; 3.5) Integrate and process all facts; 3.6) Answer merging process; 3.7) Best answer selection process.

[0101] This application also provides a processing device from a hardware architecture perspective, see [link / reference]. Figure 3 , Figure 3 This diagram illustrates a structural schematic of the processing device of this application. Specifically, the processing device may include a processor 301, a memory 302, and an input / output device 303. The processor 301 executes the computer program stored in the memory 302 to implement, for example... Figure 1 The steps of the gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team in the corresponding embodiment; or, when the processor 301 executes the computer program stored in the memory 302, it implements as follows: Figure 2 Corresponding to the functions of each unit in the embodiment, the memory 302 is used to store the functions executed by the processor 301 as described above. Figure 1 The corresponding embodiment includes the computer program required for the gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team.

[0102] For example, a computer program may be divided into one or more modules / units, one or more of which are stored in memory 302 and executed by processor 301 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a computer device.

[0103] The processing device may include, but is not limited to, processor 301, memory 302, and input / output device 303. Those skilled in the art will understand that the illustrations are merely examples of the processing device and do not constitute a limitation on the processing device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the processing device may also include network access devices, buses, etc., and processor 301, memory 302, input / output device 303, etc., are connected via a bus.

[0104] Processor 301 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the processing device, connecting various parts of the device through various interfaces and lines.

[0105] The memory 302 can be used to store computer programs and / or modules. The processor 301 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 302 and by calling data stored in the memory 302. The memory 302 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function, etc.; the data storage area may store data created according to the use of the processing device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0106] When processor 301 executes a computer program stored in memory 302, it can specifically perform the following functions: Multidisciplinary sample data were collected for different types of gynecological tumors, including different clinical problems corresponding to different types of gynecological tumors. The processing model configured under the multidisciplinary intelligent agent team framework designed for gynecologic oncology clinics is trained using training samples configured with multidisciplinary sample data. The processing model adopts a multi-model fusion architecture, specifically including a first large-scale language model, a second large-scale language model, and a third large-scale language model. The working process of configuring the processing model includes: taking a virtual senior gynecologic oncology expert as the core, automatically determining the required professional field based on the complexity of the case, generating individual experts covering multiple disciplines as different intelligent agents to respond, and then resolving the conflict of opinions among experts through a consensus building mechanism to integrate and form the final diagnosis and treatment recommendations. Based on the current case data requiring gynecological tumor consultation and treatment, a trained processing model is used to process the data and obtain the corresponding gynecological tumor consultation and treatment results.

[0107] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the gynecological tumor consultation and treatment device, processing equipment, and its corresponding units based on LLMs and multidisciplinary intelligent agent teams described above can be found in the following reference: Figure 1 The description of the gynecological tumor consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams in the corresponding embodiment will not be repeated here.

[0108] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0109] Therefore, this application provides a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute the present application. Figure 1 The steps of the gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team in the corresponding embodiment can be referred to as follows for specific operations. Figure 1 The description of the gynecological tumor consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams in the corresponding embodiments will not be repeated here.

[0110] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0111] Because of the instructions stored in the computer-readable storage medium, the present application can be executed as described above. Figure 1 The steps of the gynecological tumor consultation and treatment method based on LLMs and a multidisciplinary intelligent agent team in the corresponding embodiment can therefore achieve the results of this application. Figure 1 The beneficial effects of the gynecological tumor consultation and treatment method based on LLMs and multidisciplinary intelligent agent teams in the corresponding embodiments are detailed in the preceding description and will not be repeated here.

[0112] The foregoing has provided a detailed description of the gynecological tumor consultation and treatment method, apparatus, processing equipment, and computer-readable storage medium based on LLMs and multidisciplinary intelligent agent teams provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for gynecological tumor consultation processing based on LLMs and a team of multi-disciplinary intelligent agents, characterized in that, The method comprises: Collecting multi-disciplinary sample data of different types of gynecological tumors, wherein the multi-disciplinary sample data relates to different clinical problems corresponding to the different types of gynecological tumors; Training a processing model configured under a multi-disciplinary intelligent agent team framework designed for a gynecological tumor consultation center using training samples configured by the multi-disciplinary sample data, wherein the processing model adopts a multi-model fusion architecture, and the processing model specifically includes a first large language model, a second large language model, and a third large language model, and the working process of the processing model includes: taking a virtual senior gynecological tumor expert as the core, automatically determining the required professional field according to the case complexity, and generating individual experts covering multiple disciplines as different intelligent agents to respond, and then solving the conflicts between experts through a consensus building mechanism to integrate the final diagnosis and treatment recommendations; Based on the current case data required for gynecological tumor consultation processing, the trained processing model is used for processing, and the corresponding gynecological tumor consultation processing result is obtained.

2. The method of claim 1, wherein, The different types of gynecological tumors include ovarian cancer, cervical cancer, vulvar cancer, uterine tumors, gestational trophoblastic tumors, and other types of cancer. The multi-disciplinary sample data includes AI model generated cases and PubMed medical database recorded case data.

3. The method of claim 1, wherein, The first large language model specifically selects ChatGPT-4o model when configured to use a closed-source model. The second large language model and the third large language model specifically select DeepSeek-R1 model and Llama-4 model respectively when configured to use an open-source model.

4. The method of claim 1, wherein, Under the condition of pre-introducing fixed intelligent agent configuration, each large language model generates the different intelligent agents while being configured to follow the intelligent agent number constraints of 3, 5, 7, or 9.

5. The method of claim 1, wherein, Under the condition of pre-introducing an autonomous selection mechanism, each large language model generates the different intelligent agents while being configured to independently determine the optimal intelligent agent number.

6. The method of claim 5, wherein, The optimal intelligent agent number constraint is specifically configured to generate a range of 5-7.

7. The method of claim 1, wherein, The working process of the processing model is specifically divided into three stages; The first stage dynamically allocates expert roles based on case complexity, specifically including the following processing content: 1.1) Input the basic information, examination and inspection, preliminary diagnosis, and treatment problems of the clinical case patient; 1.2) Based on the content of 1.1), prompt the senior gynecological tumor expert to analyze the case complexity; 1.3) Based on the processing result of 1.2), automatically select the required species, generate role descriptions, and assign expert identifiers; The second stage divides and executes multiple expert consultations in parallel by prompting, specifically including the following processing content: 2.1) Based on the processing result of 1.3), generate the individual experts covering multiple disciplines as the different intelligent agents; 2.2) Based on the content of 1.1), divide the prompt words by the individual experts covering multiple disciplines generated by 2.1); 2.3) Based on the prompt words obtained in 2.2, the individual experts covering multiple disciplines generated in 2.1 are consulted to obtain the corresponding original answer summaries respectively; The third stage implements a seven-step consensus construction process, which specifically includes the following processing contents: 3.1) Consistent fact identification processing; 3.2) Conflict detection processing; 3.3) Conflict resolution processing; 3.4) Unique fact identification processing; 3.5) All fact integration processing; 3.6) Answer merging processing; 3.7) Best answer selection processing.

8. A gynecological tumor consultation processing apparatus based on LLMs and a team of multi-disciplinary intelligent agents, characterized by, The device comprises: A collection unit for collecting multi-disciplinary sample data of different types of gynecological tumors, wherein the multi-disciplinary sample data relates to different clinical problems corresponding to the different types of gynecological tumors; A training unit for training a processing model configured under a multi-disciplinary intelligent agent team framework designed for gynecological tumor consultation, using training samples configured from the multi-disciplinary sample data, wherein the processing model adopts a multi-model fusion architecture, and the processing model specifically includes a first large language model, a second large language model, and a third large language model. The working process of the processing model includes: taking a virtual senior gynecological tumor expert as the core, automatically determining the required professional field according to the complexity of the case, and generating individual experts covering multiple disciplines as different intelligent agents to respond, and then solving the conflicts between experts through a consensus construction mechanism to integrate and form the final diagnosis and treatment recommendations; An application unit for processing the case data currently required for gynecological tumor consultation processing using the trained processing model and obtaining the corresponding gynecological tumor consultation processing results.

9. A processing device, characterized by A processor and a memory, the memory storing a computer program, the processor executing the computer program in the memory to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a plurality of instructions adapted to be loaded by the processor to execute the method of any one of claims 1 to 7.

Images