Medical material intelligent auditing system and method based on multi-Agent collaboration and dynamic decision

Abstract

The invention relates to a medical material intelligent auditing system based on multi-Agent collaboration and dynamic decision making. The medical material intelligent auditing system comprises a multi-modal data acquisition and preprocessing module, an Agent intelligent auditing module, a risk decision making and early warning module and a result feedback and optimization module which are connected in sequence. Wherein the multi-modal data acquisition and preprocessing module is used for receiving and processing input medical materials and outputting standardized structured data; the Agent intelligent auditing module is used for receiving the standardized structured data and outputting a preliminary auditing conclusion through cooperative work of a plurality of agents in the Agent intelligent auditing module; the risk decision and early warning module is used for receiving the preliminary auditing conclusion and dynamically distributing an auditing strategy based on a risk assessment result; and the result feedback and optimization module is used for outputting a final auditing result and collecting feedback data to optimize the Agent intelligent auditing module.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and computer application technology, in particular to a medical material intelligent auditing system and method based on multi-agent cooperation and dynamic decision-making. BACKGROUND

[0002] In the industry of delivery riders, logistics distribution, etc., the platform party usually purchases employer's liability insurance for practitioners to cover personal injury or third-party loss caused by accidents (such as car accidents or people) in work. In the claim handling process of such insurance, the auditing of the medical materials submitted by the riders (including but not limited to medical expense invoices, hospital expense itemized lists, outpatient medical records, etc.) is the core link to confirm the scope, amount and reasonableness of the claim, which is directly related to the efficiency, cost control and risk prevention of the claim.

[0003] Currently, the auditing of such medical materials mainly relies on the following two modes with the following defects: First, the manual auditing mode has the following technical defects: The auditing personnel rely on personal experience to manually review, compare and judge the submitted medical materials. This mode requires the auditors to verify the authenticity of the bills one by one, the consistency of the medical information, whether the expense items are within the medical insurance reimbursement range, whether the expense amount is reasonable, etc. This way has the following significant defects: (1) The manual auditing mode is low in efficiency: The average time spent on auditing a single case is as long as 15 to 30 minutes, which cannot cope with high concurrency and large quantities of claim scenarios, resulting in a long claim period.

[0004] (2) The manual auditing mode is highly subjective and has different standards: The auditing results are heavily dependent on the personal experience and knowledge level of the auditors, and different auditors may have different judgments on the same material, making it difficult to ensure the consistency of the auditing standards.

[0005] (3) The manual auditing mode is high in cost: A large amount of human resources needs to be invested for repetitive labor, which is high in operating cost.

[0006] Second, the traditional automated auditing system has the following technical defects: To improve efficiency, some systems use an automated solution combining optical character recognition (OCR) technology and rule engine. The basic process is: first extract the text information in the medical material image through OCR, and then match the extracted text with the pre-set, fixed rule library (such as the medical insurance drug directory) to make a pass or fail judgment. However, such systems have the following key technical limitations: (1) The multi-modal and unstructured data processing capability of the traditional automated auditing system is weak: Medical materials have various formats, including structured electronic invoices, unstructured paper scans, and various formats of hospitalization lists. The traditional OCR model has low recognition accuracy for unstructured, low-quality, or handwritten content, leading to frequent errors in information extraction, forming a "garbage in, garbage out" dilemma, which seriously affects the accuracy of subsequent audits.

[0007] (2) The rule base of the traditional automated auditing system is rigid and has poor adaptability: The rule base it relies on is usually static and cannot quickly adapt to the dynamic changes and differences of medical insurance policies in different regions and hospitals. For example, a certain drug is covered by medical insurance in Province A, but is self-paid in Province B. The fixed rule base cannot perceive the context and dynamically adjust it, and frequent manual maintenance is required, which lacks system flexibility.

[0008] (3) The traditional automated auditing system lacks intelligent decision-making and risk perception capabilities: The system can only perform simple "yes / no" rule matching and cannot perform in-depth evaluation of the reasonableness of expenses (such as identifying over-medical treatment, expense fraud, etc.), nor can it dynamically adjust the auditing strategy according to the complexity of the case (such as involving multiple departments, special medical equipment, etc.). This leads to waste of auditing resources for simple cases, insufficient auditing for complex cases, and potential risks that cannot be identified and warned in advance. SUMMARY

[0009] The present application designs an intelligent medical material auditing system and method based on multi-Agent collaboration and dynamic decision-making, which solves the four core problems of efficiency, accuracy, adaptability, and risk control in existing medical material auditing technology.

[0010] To solve the above technical problems, the present application adopts the following scheme: An intelligent medical material auditing system based on multi-Agent collaboration and dynamic decision-making, characterized by: a multi-modal data acquisition and preprocessing module, an Agent intelligent auditing module, a risk decision and warning module, and a result feedback and optimization module are connected in turn; wherein the multi-modal data acquisition and preprocessing module is used to receive and process the input medical materials, and output standardized structured data; the Agent intelligent auditing module is used to receive the standardized structured data and output the preliminary audit conclusion through the collaborative work of multiple intelligent agents in it; the risk decision and warning module is used to receive the preliminary audit conclusion and dynamically allocate the audit strategy based on the risk assessment result; the result feedback and optimization module is used to output the final audit result and collect feedback data to optimize the Agent intelligent auditing module.

[0011] Preferably, the multi-modal data acquisition and preprocessing module specifically comprises: an image processing unit for enhancing, correcting and optimizing images of medical materials; a text extraction unit using a multi-model combination strategy for extracting text information from medical materials of different formats; a data standardization unit for converting the extracted text information into a unified structured data format.

[0012] Preferably, the text extraction unit is configured to use a standardized API to extract data for electronic invoices, use a PP-OCRv4 model for text recognition for paper invoice scans, and use a dedicated medical text parsing model to extract key information for hospitalization lists.

[0013] Preferably, the Agent intelligent auditing module internally includes: a basic information extraction Agent for extracting patient information, medical information and expense details from the standardized structured data; a medical insurance rule matching Agent for judging the compliance of the expense details according to a preset medical insurance rule library; a cost reasonableness evaluation Agent for evaluating the reasonableness of the expense details based on historical data to identify abnormal expense patterns; wherein the basic information extraction Agent, the medical insurance rule matching Agent and the cost reasonableness evaluation Agent interact and make collaborative decisions through multiple rounds of dialogue mechanism to form the preliminary audit conclusion.

[0014] Preferably, the basic information extraction Agent uses a large language model fine-tuned for medical text; the medical insurance rule matching Agent uses a combination of rule engines and machine learning; and the cost reasonableness evaluation Agent is trained based on historical claims data of delivery riders and industry benchmark data.

[0015] Preferably, the risk decision and early warning module is specifically used for: according to the medical expenses, the patient's condition and the type of accident, the risk level of the case is evaluated, and it is divided into low risk, medium risk or high risk; based on the risk level, the audit strategy is dynamically allocated for the case: low-risk cases are automatically passed, medium-risk cases trigger manual review and prompt key points, and high-risk cases trigger manual in-depth review and generate a risk warning report.

[0016] Preferably, the result feedback and optimization module is specifically used for: outputting the audit conclusion and risk prompt to the auditor in a visual form; collecting the final decision data of the auditor on the preliminary audit conclusion and using the data as feedback data; using the feedback data to continuously train and optimize the models in the Agent intelligent auditing module.

[0017] An intelligent medical material auditing method based on multi-agent cooperation and dynamic decision-making, characterized by being applied to the intelligent medical material auditing system based on multi-agent cooperation and dynamic decision-making, and comprising the following steps: a multi-modal data acquisition and preprocessing step: receiving and processing input medical materials, and outputting standardized structured data; an agent intelligent auditing step: receiving the standardized structured data, and outputting a preliminary auditing conclusion through the cooperative work of multiple intelligent agents; a risk decision and early warning step: receiving the preliminary auditing conclusion, and dynamically allocating auditing strategies based on risk assessment results; and a result feedback and optimization step: outputting a final auditing result, and collecting feedback data to optimize subsequent intelligent auditing processes.

[0018] Preferably, the agent intelligent auditing step specifically comprises: extracting patient information, medical information, and expense details through a basic information extraction agent; judging the compliance of the expense details through a medical insurance rule matching agent; evaluating the rationality of the expense details and identifying abnormal expense patterns through a cost rationality evaluation agent; and forming the preliminary auditing conclusion through the multi-round dialogue mechanism between the multiple intelligent agents, based on the extracted information, compliance judgment, and rationality evaluation results.

[0019] Preferably, in the result feedback and optimization step, the process of optimizing the intelligent agent model using the collected feedback data comprises at least one of transfer learning, dynamic rule base updating, and continuous learning.

[0020] The intelligent medical material auditing system and method based on multi-agent cooperation and dynamic decision-making have the following beneficial effects: (1) The present application greatly shortens the auditing time of a single medical material from an average of 20 minutes in the manual mode to less than 5 minutes through automatic preprocessing of multi-modal data and parallel cooperative auditing of multiple agents, and improves the overall auditing efficiency by more than 75%, which can easily cope with high-concurrency and large-batch scenarios such as take-out delivery riders and claim settlement.

[0021] (2) The intelligent agent based on a large language model has strong semantic understanding and context reasoning capabilities, overcoming the low accuracy of traditional OCR systems in identifying unstructured data. At the same time, the system eliminates the subjective differences of manual auditing through the combination of preset rules and machine learning models, and the auditing accuracy is stable at more than 95%, which is more than 15% higher than traditional automated systems.

[0022] (3) The medical insurance rule matching agent of the present application has a dynamic rule updating mechanism, which can automatically identify and adapt to the differences in medical insurance policies of different regions and different hospitals, without the need for frequent manual adjustment of the rule base, greatly reducing the system maintenance cost and solving the core pain point of the rigid rule base of traditional systems.

[0023] (4) This invention introduces for the first time a cost reasonableness assessment agent and dynamic decision engine based on historical data, which can proactively identify potential risks such as "over-medicalization" and "inflated costs" and intelligently allocate review resources (automatic approval, manual review, and in-depth review) according to the risk level. This capability reduces the subsequent dispute rate of claims by more than 30% and achieves optimal allocation of review resources.

[0024] (5) Through the result feedback and optimization module, the auditor's manual decisions can be transformed into high-quality feedback data for continuous training and optimization of each agent. This gives the system a self-learning and continuous evolution capability that traditional static systems do not have, and the audit accuracy continues to increase with the time of use.

[0025] (6) For auditors, the system can handle a large amount of simple and repetitive work and provide clear decision support for focusing on high-risk and complex cases, effectively reducing workload and improving job satisfaction and professionalism. For insurance platforms and riders, the shortened claims cycle directly improves service satisfaction and platform competitiveness, while automated review also significantly reduces human resource operating costs. Attached Figure Description

[0026] Figure 1 : A schematic diagram of the overall architecture of the system of this invention; Figure 2 : A schematic diagram of the processing flow of the multimodal data acquisition and preprocessing module of this invention; Figure 3 : A schematic diagram of the internal collaborative workflow of the Agent intelligent review module of this invention; Detailed Implementation

[0027] The following is combined Figures 1 to 3 The present invention will be further described as follows: like Figure 1 As shown, this invention provides an intelligent medical material review system based on multi-agent collaboration and dynamic decision-making. Its overall architecture mainly comprises four core modules connected in sequence: a multimodal data acquisition and preprocessing module, an agent-based intelligent review module, a risk decision-making and early warning module, and a result feedback and optimization module. The modules interact and communicate with each other through a standardized application programming interface.

[0028] The specific workflow of the system can be illustrated using a typical food delivery rider claims scenario: Suppose that food delivery rider A collided with a pedestrian during a delivery, causing injury to the pedestrian. After the pedestrian received medical treatment, the delivery rider submitted the pedestrian's medical invoices and hospitalization expense list through the platform to apply for compensation.

[0029] Step 1: Multimodal Data Collection and Preprocessing: As shown in the figure, the starting point of the process is to receive various medical materials submitted by delivery riders or related personnel through the platform. These materials come in various forms, mainly including: photographed or scanned images of paper medical invoices. Hospital system-derived inpatient expense list PDF documents. Other formats of medical credential images. Figure 2 The system first starts the multimodal data collection and preprocessing module. This module receives the medical material images uploaded by Rider Li (such as photographed screenshots of paper invoices, hospital system-derived inpatient lists in PDF format), and performs the following operations:

[0030] Image processing includes enhancing and correcting images that may be blurred, tilted, or unevenly lit due to shooting, ensuring that the image quality meets the high-precision recognition requirements. Text extraction is the best extraction strategy for different materials. The text extraction unit uses a multi-model combination strategy to call the most suitable model or interface for information extraction for materials of different sources and formats to achieve the highest accuracy. For electronic inpatient lists with relatively standard formats, call the interface to directly parse; for paper invoice scans with complex backgrounds and various formats, use advanced OCR models such as PP-OCRv4 for text recognition to ensure that key items such as western medicine fees, examination fees, and medical treatment fees and their amounts are accurately captured.

[0031] Data standardization includes unifying the recognized disorganized text information, such as converting "2024-01-01" to "2024-01-01" date format, and integrating scattered drug names, unit prices, quantities, and total amounts into a structured expense detail list, and finally packaging into a unified, standardized medical data package.

[0032] The end point of the process is to output the standardized medical data package generated above. This clean and neat data package will be directly sent to the next stage - the Agent intelligent review module for analysis by various agents.

[0033] Step 2: Multi-Agent Collaborative Intelligent Review:

[0034] As shown in the figure, the preprocessed standardized medical data package is sent into the Agent intelligent review module in parallel. This module is divided into three professional agents, which simulate the "multi-round dialogue" mechanism of expert consultation to conduct in-depth analysis of the case. Figure 3

[0035] ​​The basic information extraction agent: first, the data packet is deeply semantically parsed. It can accurately identify that the patient is "Mr. Wang" (not the rider A himself), the hospital is "XX Hospital", the date of treatment is logically consistent with the time of the accident, and all the cost details such as "CT examination fee 350 yuan", "medicine A 2 times a day, total 150 yuan" are extracted, providing complete context for subsequent review.

[0036] The medical insurance rule matching agent: it runs synchronously and accesses a dynamic medical insurance rule library covering all cities in China. The agent will check the extracted cost details one by one. For example, it finds that "medicine A" is in the medical insurance directory of "XX City", which belongs to the reimbursable range; but at the same time, it identifies that "a certain rehabilitation therapy fee" is a self-financing project in this area. It will output these compliance and non-compliance judgments as preliminary conclusions.

[0037] The cost reasonableness evaluation agent: the agent is trained based on a large amount of historical claims data of the rider's employer's liability insurance. It will intelligently analyze the cost pattern. For example, it finds that the "CT examination" in this case is repeated twice in the case of soft tissue injury without fracture, which may be unreasonable examination; or it finds that the total amount of "medicine A" is far more than the regular course of treatment, which may be suspected of overuse. It will mark these reasonable doubts as risk signals.

[0038] As shown in Figure 3 , the preliminary conclusions of the three agents are collected into a collaborative decision-making hub. Here, they exchange information and make comprehensive judgments. For example, the cost reasonableness evaluation agent questions the "repeated CT examination", the medical insurance rule matching agent supplements that "even if the CT examination is compliant, repeated examination may not be fully reimbursed", and the basic information extraction agent provides "evidence that the medical record does not indicate the need for repeated examination". Through multiple rounds of such "dialogues", they finally reach a consensus and form a comprehensive review conclusion, pointing out that in this claims application, "medicine A" is reimbursable, but "a certain rehabilitation therapy fee" is self-financing, "repeated CT examination" is unreasonable, and the overall case risk level needs to be evaluated.

[0039] To ensure the high precision and high adaptability of the system in the specific scenario of delivery rider claims, the training data, model selection and definition of the core engine of each intelligent agent are described in detail as follows.

[0040] The basic information extraction agent, medical insurance rule matching agent, and cost reasonableness evaluation agent of the system are selected from the Qwen series of open source large language models and multi-modal large models (such as the Qwen2-VL series), and are optimized for medical audit tasks. The model parameter size can be selected among 7B (700 million), 14B, and other different specifications according to the actual deployment of computing power and response requirements. The training process follows the "pre-training-fine-tuning" paradigm. In the early stage, the open source model parameters are used for zero-sample or small-sample reasoning. In the later stage, the high-quality dataset constructed by the following is used for supervised fine-tuning, reinforcement learning, and further optimization.

[0041] The training data of the basic information extraction agent: integrate the internal desensitization claim data from cooperating insurance companies and public medical bill data sets. Contains medical invoices, hospital lists and other images and corresponding texts in historical delivery rider employer liability insurance cases, the key is that professional personnel have annotated the reasoning thinking process to guide the model to locate and extract key entities from complex texts. Including the bill subset in the CBLUE medical text understanding data set, ICDAR2019 medical document OCR data, etc., to improve the model's generalization ability to various bill styles.

[0042] The training data of the medical insurance rule matching agent: integrate official policies, historical cases, and simulation data. The core data is the policy documents published by the national and provincial medical insurance bureaus, as well as the closed case claim records provided by the cooperating parties. These records not only contain the "whether it meets the medical insurance regulations" label, but also contain the judgment basis and thinking process during manual review. Based on real drug and diagnosis item codes, a large number of compliant and non-compliant sample pairs are artificially constructed to enhance the model's judgment ability for boundary cases.

[0043] The training data of the cost reasonableness evaluation agent: mainly relies on industry benchmarks and historical claim data. Using the desensitized delivery rider historical claim database, the complete cost details in the case are annotated with exceptions or irrelevant items (such as basic disease drugs unrelated to the accident), and the annotation reasons are noted. The "Clinical Path Fee Reference Standard" released by the National Health Commission and the third-party health insurance claim white paper are integrated to provide objective references for the model to evaluate the cost reasonableness.

[0044] The training of the model is a complex process with many parameters, involving dozens or even hundreds of different super parameter configurations. For different training stages and targets, the system uses multiple training paradigms, including supervised fine-tuning and reinforcement learning based on human feedback. The specific loss function, optimizer selection, and training rounds of super parameters will be dynamically adjusted and optimally configured according to the selected model architecture, training dataset size, and specific scene requirements to maximize the model performance.

[0045] Step 3: Risk decision and dynamic early warning: The risk decision and early warning module receives the comprehensive audit conclusion described above. This module is the dispatch center of the system and executes the final decision: the risk decision and dynamic early warning module includes risk level assessment and dynamic audit strategy allocation.

[0046] The risk level assessment is a module that automatically assesses the case as “medium risk” based on the high total amount of the case and the identification of multiple risk points such as “repeat inspection” and “self-funded project”.

[0047] Dynamic audit strategy allocation is based on the “medium risk” level. The system does not automatically pass, but starts the manual review process. The system pushes the audit interface to the insurance company's auditor and clearly highlights “self-funded project: certain rehabilitation physiotherapy fee” and “suspected unreasonable fee: second CT examination fee 350 yuan” on this interface, and attaches the agent's analysis reasons, greatly improving the efficiency and pertinence of manual review.

[0048] Step 4: Result feedback and optimization: The result feedback and optimization module is responsible for the end and learning. The result feedback and optimization module includes result visualization and feedback and closed-loop learning two functions.

[0049] Result visualization and feedback is the final decision made by the auditor based on the system's prompt to “deduct the repeated CT examination fee”. This final decision and reason are recorded and confirmed through this module.

[0050] Closed-loop learning is the difference between the auditor's operation and the system's initial suggestion (for example, the system only suggests “unreasonable”, but the auditor finally approves the specific deduction amount), which is collected by the system as valuable feedback data. These data will be used to regularly fine-tune the models of the basic information extraction agent, medical insurance rule matching agent and fee reasonableness assessment agent, so that the system can directly give more accurate “suggested deduction XX yuan” conclusion when encountering similar “repeated CT examination” cases next time, thereby realizing continuous self-optimization.

[0051] Through the above process, the invention successfully completes the automated preliminary audit and risk positioning of a rider claim case that originally required manual review for more than 20 minutes in 5 minutes, and intelligently guides the manual audit resources to focus on the core controversial points, thereby achieving overall improvement in efficiency, accuracy and risk control ability.

[0052] The application is described above with reference to the drawings, and it is obvious that the implementation of the application is not limited by the above manner, as long as various improvements are made by using the method concept and technical scheme of the application, or the concept and technical scheme of the application is directly applied to other occasions without improvement, which are all within the protection scope of the application.

Claims

1. A medical material intelligent auditing system based on multi-agent cooperation and dynamic decision-making, characterized in that: The system comprises a multimodal data acquisition and preprocessing module, an Agent intelligent auditing module, a risk decision and early warning module, and a result feedback and optimization module connected in sequence. The multimodal data acquisition and preprocessing module is configured to receive and process input medical materials and output standardized structured data. The Agent intelligent auditing module is configured to receive the standardized structured data and output a preliminary audit conclusion through the collaborative work of multiple intelligent agents. The risk decision and early warning module is configured to receive the preliminary audit conclusion and dynamically allocate audit strategies based on risk assessment results. The result feedback and optimization module is configured to output a final audit result and collect feedback data to optimize the Agent intelligent auditing module. 2.The medical material intelligent auditing system based on multi-agent cooperation and dynamic decision-making according to claim 1, characterized in that: The multimodal data acquisition and preprocessing module comprises: An image processing unit configured to enhance, correct, and optimize images of medical materials. A text extraction unit configured to extract text information from medical materials in different formats using a multi-model combination strategy. A data standardization unit configured to convert the extracted text information into a unified structured data format.

3. The medical material intelligent auditing system based on multi-Agent collaboration and dynamic decision-making according to claim 2, wherein: The text extraction unit is configured to extract data from electronic invoices using a standardized API, recognize text from paper invoice scans using a PP-OCRv4 model, and extract key information from hospitalization lists using a dedicated medical text analysis model.

4. The medical material intelligent auditing system based on multi-Agent collaboration and dynamic decision-making according to claim 1, wherein: The Agent intelligent auditing module comprises: A basic information extraction Agent configured to extract patient information, medical information, and expense details from the standardized structured data. A medical insurance rule matching Agent configured to determine the compliance of the expense details based on a pre-set medical insurance rule library. A cost reasonableness evaluation Agent configured to evaluate the reasonableness of the expense details based on historical data and identify abnormal expense patterns. The basic information extraction Agent, the medical insurance rule matching Agent, and the cost reasonableness evaluation Agent interact and make collaborative decisions through multiple rounds of dialogue to form the preliminary audit conclusion.

5. The medical material intelligent auditing system based on multi-Agent collaboration and dynamic decision-making according to claim 4, wherein: The basic information extraction Agent uses a large language model fine-tuned for medical text; the medical insurance rule matching Agent uses a combination of rule engines and machine learning; and the cost reasonableness evaluation Agent is trained based on historical claims data of delivery riders and industry benchmark data. 6.The medical material intelligent auditing system based on multi-agent cooperation and dynamic decision-making according to claim 1, characterized in that: The risk decision and early warning module is specifically configured to: According to medical expenses, patient conditions and accident types, the risk level of the case is evaluated, and the case is divided into low risk, medium risk or high risk; based on the risk level, the audit strategy is dynamically allocated for the case: low-risk cases are automatically passed, medium-risk cases trigger manual review and prompt focus, and high-risk cases trigger manual deep review and generate a risk warning report. 7.The medical material intelligent auditing system based on multi-agent cooperation and dynamic decision-making according to claim 1, characterized in that: The result feedback and optimization module is specifically used for: outputting the audit conclusion and risk prompt to the auditor in a visual form; collecting the final decision data of the auditor on the preliminary audit conclusion and taking the data as feedback data backflow; continuously training and optimizing the model in the Agent intelligent audit module using the feedback data.

8. A medical material intelligent auditing method based on multi-agent cooperation and dynamic decision-making, characterized in that, The method is applied to the medical material intelligent audit system based on multi-Agent collaboration and dynamic decision-making as claimed in any one of claims 1-7, and comprises the following steps: a multi-modal data acquisition and preprocessing step: receiving and processing the input medical material, and outputting standardized structured data; an Agent intelligent audit step: receiving the standardized structured data, and outputting a preliminary audit conclusion through the collaborative work of multiple intelligent agents; a risk decision and warning step: receiving the preliminary audit conclusion, and dynamically allocating an audit strategy based on the risk assessment result; a result feedback and optimization step: outputting the final audit result, and collecting feedback data to optimize the subsequent intelligent audit process. 9.The medical material intelligent auditing method based on multi-agent cooperation and dynamic decision-making according to claim 8, characterized in that: The Agent intelligent audit step specifically comprises: extracting patient information, medical information and expense details by a basic information extraction Agent; judging the compliance of the expense details by a medical insurance rule matching Agent; evaluating the reasonableness of the expense details by a cost reasonableness evaluation Agent to identify abnormal expense patterns; through the multi-round dialogue mechanism between the multiple intelligent agents, the extracted information, the compliance judgment and the reasonableness evaluation result are collaborated and integrated to form the preliminary audit conclusion. 10.The medical material intelligent auditing method based on multi-agent cooperation and dynamic decision-making according to claim 8, characterized in that: In the result feedback and optimization step, the process of optimizing the intelligent agent model using the collected feedback data comprises at least one of transfer learning, dynamic rule base updating and continuous learning.

Images