A rehabilitation assistive device evaluation method based on multi-agent and RAG technology

Abstract

The application discloses a rehabilitation auxiliary tool evaluation method based on multi-agent and RAG technology. The method comprises the following steps: constructing a multi-agent system, including a rehabilitation doctor agent, an auxiliary tool engineer agent, a nursing staff agent and a user interaction agent; building a dynamic knowledge base based on RAG technology, including an auxiliary tool product database, a rehabilitation medical knowledge base, a clinical case database and an industry standard database; interacting with the user, and dynamically adjusting the questioning method, language style and question order according to the user type and answer content; retrieving in the dynamic knowledge base according to the obtained information, and respectively outputting professional evaluation opinions; and integrating the professional evaluation opinions of the rehabilitation doctor agent, the auxiliary tool engineer agent and the nursing staff agent according to a collaborative decision-making model to generate a multi-dimensional auxiliary tool evaluation report and a recommended scheme. The above method can make the evaluation process of the auxiliary tool more efficient, and the evaluation result more accurate and effective.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of assistive devices for the disabled or the elderly, in particular to a rehabilitation assistive device evaluation method. BACKGROUND

[0002] At present, assistive device fitting evaluation mainly relies on two ways of expert evaluation and traditional questionnaire. In the process of expert evaluation, the evaluator needs to communicate with the user to understand their physical condition, use demand and living scene and other information by relying on their own professional knowledge, and then recommends appropriate assistive devices for the user in combination with the parameters of assistive device products. The whole process requires the evaluator to have rich knowledge of medicine, rehabilitation and assistive devices. The traditional questionnaire system collects the basic information of the user through pre-set evaluation questionnaire and algorithm, and then performs simple matching to output the assistive device recommendation result. Some systems will access a small amount of assistive device product database, but the data updating speed is slow, and only the single function of information collection and preliminary matching can be realized. However, the current assistive device fitting evaluation method has at least one of the following shortcomings:

[0003] 1. Low efficiency of expert evaluation: The evaluation process relies on one-on-one communication by artificial, and the evaluation period of each user usually takes 1-3 days, which cannot meet the evaluation needs of a large number of users, and the evaluation result is easily affected by the subjective experience, fatigue degree and other factors of the evaluator, so that the evaluation results of different evaluators for the same user may be different, and the accuracy is difficult to guarantee.

[0004] 2. Limitations of questionnaire evaluation function: The current single intelligent system can only realize information collection and preliminary matching, the operation process is complex, the fitting scene is extremely limited, it is difficult to simulate the collaborative evaluation of multiple professional roles (such as rehabilitation physicians, assistive device engineers and nursing staff), and it is difficult to comprehensively analyze the user demand from multiple dimensions. For example, it is difficult to conduct in-depth evaluation combined with the daily activity scene, family environment layout and other detailed information of the user, so that the recommended assistive device may not meet the actual use scene of the user.

[0005] 3. Weak information retrieval and updating ability: The database of the existing intelligent system is mostly local fixed database, and the assistive device product information, rehabilitation medical guidelines and industry standards are not updated in time, so that the latest assistive device technology, clinical research results and other information cannot be obtained in time. When encountering the evaluation needs of special users (such as those with rare diseases or complex functional disorders), the system cannot quickly retrieve relevant professional knowledge and cases, and it is difficult to give scientific and reasonable evaluation suggestions.

[0006] 4. Poor user interaction experience: Most of the existing systems collect information in the form of fixed questionnaires, and cannot adjust the questioning content in real time according to the user's answers. For special groups such as elderly users and users with cognitive disorders, complex questionnaire process and professional terms will increase their use difficulty, resulting in incomplete user information collection, and thus affecting the accuracy of the evaluation result. SUMMARY

[0007] The main purpose of the present application is to provide a rehabilitation assistive device evaluation method based on multi-agent and RAG technology to solve the problems of low evaluation efficiency and inaccurate evaluation results in the prior art.

[0008] One embodiment of the present application provides a rehabilitation assistive device evaluation method based on multi-agent and RAG technology, comprising the following steps:

[0009] A multi-agent system is constructed, which includes a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent, and a user interaction agent;

[0010] A dynamic knowledge base is built based on RAG technology, which includes an assistive device product database, a rehabilitation medical knowledge base, a clinical case database, and an industry standard database;

[0011] The user interaction agent is used to interact with the user, and the user interaction agent adopts an adaptive interaction mechanism to dynamically adjust the questioning method, language style, and question order according to the user type and answer content to improve the completeness and accuracy of user information collection;

[0012] According to the information obtained by the user interaction agent, the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent retrieve in the dynamic knowledge base and output professional evaluation opinions respectively;

[0013] A multi-agent collaborative decision-making model is constructed, and the professional evaluation opinions of the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent are integrated according to the collaborative decision-making model to generate a multi-dimensional assistive device evaluation report and a recommended solution.

[0014] In one embodiment, the adaptive interaction mechanism includes dynamically adjusting the questioning method, language style, and question order according to the completeness of user information collection, and the completeness of user information collection is calculated by the following method:

[0015] ;

[0016] Wherein, P is the completeness of user information collection; n is the number of core information items; xi is the effective collection mark of the ith information; ωi is the dynamic difficulty coefficient of the ith information; γ is the real-time collection quality correction factor; δ is the personalized correction coefficient. i i auto i u, i

[0017] ​​​​In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0018] The threshold model trained according to the historical data determines the integrity threshold;

[0019] If the user information meets the condition of the integrity threshold, x i = 1;

[0020] If the user information is missing or the error exceeds the integrity threshold, the user interaction agent starts the automatic follow-up system for completion;

[0021] If the automatic follow-up system is started, and the user information meets the condition of the integrity threshold, x i = 1;

[0022] If the automatic follow-up system is started, and the user information is still missing or the error exceeds the integrity threshold, x i = 0.

[0023] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0024] Obtain the average collection success rate R i of the i-th information in the historical database;

[0025] Obtain the importance score I i assigned to the i-th information in the evaluation result influence degree model;

[0026] Determine the dynamic difficulty coefficient ω i of the i-th information according to the average collection success rate R i of the i-th information and the importance score I i auto of the i-th information.

[0027] In one embodiment, the dynamic difficulty coefficient ω i auto of the i-th information is calculated according to the following method:

[0028] .

[0029] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0030] Obtain the time length data T i of the user's answer to the i-th information;

[0031] Obtain the average time length Tˉ i of the same type of user's answer to the i-th information in the historical database;

[0032] According to the time length data T of the user answering the i-th information i and the average time length T of the same type of users answering the i-th information i Determine the real-time collection quality correction factor γ of the i-th information i .

[0033] In one embodiment, the real-time collection quality correction factor γ of the i-th information is calculated according to the following manner i :

[0034] ;

[0035] Wherein, min (T i / Tˉ i , 2) is the smaller value of T i / Tˉ i and 2.

[0036] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0037] Obtain the number of user portrait labels k;

[0038] Obtain the weight information β t of the t-th label;

[0039] Obtain the influence coefficient σ t,i of the t-th label on the collection of the i-th information;

[0040] Calculate the individualization correction coefficient δ of the i-th information according to the following manner u, i :

[0041] .

[0042] In one embodiment, the adaptive interaction mechanism includes dynamically adjusting the questioning manner, language style and question order according to the interaction adaptation degree, and the interaction adaptation degree is calculated by the following manner:

[0043] ;

[0044] Wherein, Q is the interaction adaptation degree; α auto is the interaction adaptation degree weight; Sim′(U, M) is the user- interaction manner weighted cosine similarity; S auto is the question order rationality score; λ is the real-time feedback correction factor.

[0045] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0046] Contribution mean of question order adaptation to overall adaptation degree in historical database Q seq ;

[0047] Contribution mean of interaction mode adaptation to overall adaptation degree in historical database Q mode ;

[0048] Contribution mean of question order adaptation to overall adaptation degree Q seq and contribution mean of interaction mode adaptation to overall adaptation degree Q mode Determine interaction adaptation degree weight α auto .

[0049] In one embodiment, the interaction adaptation degree weight α auto is calculated according to the following manner:

[0050] .

[0051] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0052] A plurality of interaction parameters are preset, including age, cognitive level and device type;

[0053] Different weight values ω d are given to the interaction parameters based on dimension influence degree;

[0054] Obtain the standardized value U d of user characteristics;

[0055] Obtain the standardized value M d of interaction mode characteristics;

[0056] Calculate the user-interaction mode weighted cosine similarity Sim'(U, M) according to the following manner:

[0057] ;

[0058] Where H u is the user historical interaction acceptance correction term.

[0059] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0060] Obtain the information collection efficiency E seq of the current question order;

[0061] Obtain the average information collection efficiency E seq of the same type of users in the historical database;

[0062] Predicting the current cognitive load L of the user based on the answer duration and error rate u ;

[0063] Setting the cognitive load threshold L of the same type of user based on historical data max ;

[0064] The information collection efficiency E according to the current question order seq The average information collection efficiency E of the same type of user seq The current cognitive load L u The cognitive load threshold L max Determine the question order rationality score S auto .

[0065] In one embodiment, the question order rationality score S is calculated according to the following way auto :

[0066] .

[0067] In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0068] Obtain the number of times of user active confirmation ф1;

[0069] Obtain the number of times of user skipping or re-listening ф2;

[0070] Determine the real-time feedback correction factor λ according to the number of times of user active confirmation ф1 and the number of times of user skipping or re-listening ф2.

[0071] In one embodiment, the real-time feedback correction factor λ is calculated according to the following way:

[0072] .

[0073] In one embodiment, in the process of generating a multi-dimensional assistive device evaluation report and recommended scheme:

[0074] The collaborative decision-making module calculates the comprehensive evaluation score of a plurality of different assistive devices;

[0075] According to the comprehensive evaluation score of the assistive device, select the highest score of 1-3 assistive devices for recommendation;

[0076] The comprehensive evaluation score is determined by the following way:

[0077] ;

[0078] Wherein, S total is the comprehensive evaluation score; j is the agent identifier; S ja score of a sub-report for the agent; γ i a credibility correction factor for the sub-report; ω j final a final weight of the agent.

[0079] In one of the embodiments, the final weight of the agent ω j final comprises a first value ω j final,0 and a second value ω j final,1 ;

[0080] In the case of a non-conflict scenario, the final weight of the agent ω j final comprises a first value ω j final,0 determined by:

[0081] ;

[0082] wherein, a scenario base weight; a historical accuracy rate of the agent;

[0083] In the case of a conflict scenario, the final weight of the agent ω j final comprises a second value ω j final,1 determined by:

[0084] ;

[0085] wherein, a conflict correction coefficient; a user demand correction coefficient.

[0086] In one of the embodiments, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0087] obtaining the sub-reports output by the rehabilitation physician agent, the assistive device engineer agent and the nursing staff agent respectively;

[0088] calculating the cosine similarity between two sub-reports;

[0089] determining the value of the divergence degree according to the cosine similarity between two sub-reports;

[0090] if the value of the divergence degree is less than a preset value, it is determined that there is a conflict scenario between the sub-reports;

[0091] If the value of the degree of divergence is greater than or equal to the preset value, it is determined that there is no conflict scenario between the sub-reports;

[0092] wherein the degree of divergence D is calculated by the following manner:

[0093]

[0094] wherein Sim is the cosine similarity of the sub-reports.

[0095] In one of the embodiments, the multi-agent system further comprises a psychological assessment agent, and the dynamic knowledge base further comprises a psychological intervention case library;

[0096] The psychological assessment agent is configured to assess the psychological acceptance and use willingness of the user to the assistive device.

[0097] The collaborative decision-making model integrates the professional assessment opinions of the rehabilitation physician agent, the assistive device engineer agent, the nursing staff agent, and the psychological assessment agent to generate a multi-dimensional assistive device evaluation report and a recommended solution.

[0098] In one of the embodiments, the rehabilitation assistive device evaluation method based on the multi-agent and RAG technology further comprises the following steps:

[0099] According to the collected user information, the user identity features are determined.

[0100] According to the user identity features, it is determined whether the assistive device evaluation report and the recommended solution have been provided for the user within a preset time range.

[0101] If yes, the new requirements are determined according to the difference between the current user requirements and the previous user requirements; the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent retrieve in the dynamic knowledge base according to the new requirements and supplement to the previous professional assessment opinions respectively.

[0102] In one of the embodiments, the user interaction agent further comprises:

[0103] The video shooting module is configured to shoot a home environment video according to the user requirements.

[0104] The nursing staff agent performs scene assessment according to the content of the shot home environment video to improve the accuracy of the scene information collection.

[0105] Another embodiment of the present application further provides a rehabilitation assistive device evaluation system based on the multi-agent and RAG technology, comprising a multi-agent module, a RAG technology module, a dynamic knowledge base module, and a collaborative decision-making module.

[0106] ​The multi-agent module comprises a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent and a user interaction agent;

[0107] The dynamic knowledge base module comprises an assistive device product database, a rehabilitation medicine knowledge base, a clinical case database and an industry standard database.

[0108] The RAG technology module is used for connecting the multi-agent module and the dynamic knowledge base module, so as to realize retrieval and information extraction of the dynamic knowledge base module and provide information support for the multi-agent module.

[0109] The user interaction agent adopts an adaptive interaction mechanism to interact with the user, dynamically adjusts the questioning mode, language style and question order according to the user type and answer content, so as to improve the completeness and accuracy of user information collection.

[0110] The rehabilitation physician agent, the assistive device engineer agent and the nursing staff agent retrieve in the dynamic knowledge base module according to the information obtained by the user interaction agent, and respectively output professional evaluation opinions.

[0111] The collaborative decision-making module is used for integrating the professional evaluation opinions of the rehabilitation physician agent, the assistive device engineer agent and the nursing staff agent, so as to generate a multi-dimensional assistive device evaluation report and a recommended scheme.

[0112] The rehabilitation assistive device evaluation method based on multi-agent and RAG technology has the following advantages and beneficial effects:

[0113] 1. In the rehabilitation assistive device evaluation method based on multi-agent and RAG technology provided in the application, a multi-agent system is constructed, a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent and a user interaction agent are set, each agent performs its own function and cooperates with each other, a multi-professional team collaborative evaluation process is simulated, and user demand is analyzed from multiple dimensions. With the help of multi-agent collaborative work, the one-on-one artificial evaluation process is replaced, the single user evaluation period is compressed to within 10 minutes, and the interference of artificial subjective factors is avoided. At the same time, based on the multi-agent task dynamic allocation algorithm, the evaluation process is optimized, the task amount of each agent is dynamically allocated, the evaluation resources are more reasonably utilized, and the accuracy of the evaluation result is improved by more than 30%.

[0114] 2. In the rehabilitation assistive device evaluation method based on multi-agent and RAG technology provided in the application, the multi-agent system respectively evaluates from the dimensions of rehabilitation medicine, assistive device technology, nursing scene, etc. Combined with the multi-source information retrieved by the RAG technology, the user's physical condition, use scene, home environment, and assistive device adaptability can be covered. According to the evaluation dimension weight precise calculation algorithm, the evaluation weight of each dimension is accurately calculated and adjusted, which solves the problem of single system evaluation dimension limitation, and ensures that the evaluation result is more comprehensive and accurate.

[0115] 3. In the rehabilitation assistive device evaluation method based on multi-agent and RAG technology provided in the application, by integrating RAG technology, a dynamic knowledge base including assistive device product database, rehabilitation medicine knowledge base, clinical case database, industry standard database, etc. is built, and real-time retrieval of the latest information is realized to provide data support for multi-agent evaluation. At the same time, based on the RAG technology, the dynamic knowledge base is also built, which enhances the real-time and richness of the information of the rehabilitation assistive device evaluation method in the application. RAG technology can connect to the Internet and professional databases in real time, realize real-time updating and accurate retrieval of assistive device product information, medical guidelines, and clinical cases. Combined with the multi-agent task dynamic allocation algorithm, the information retrieval strategy is optimized, and for special user needs, rare disease assistive device adaptation cases and the latest assistive device technology parameters can be quickly obtained, ensuring the scientificity of the evaluation suggestion.

[0116] 4. In the rehabilitation assistive device evaluation method based on multi-agent and RAG technology provided in the application, by designing an adaptive interaction mechanism, the user interaction agent can dynamically adjust the questioning method, language style and question order according to the user type (such as elderly users, cognitive impairment users) and the answer content, improving the completeness and accuracy of user information collection. The adaptive interaction mechanism can adjust the interaction method according to the user characteristics, for example, for elderly users, voice interaction and simplified text and picture display are used, and for cognitive impairment users, step-by-step guided questioning is used. The evaluation dimension weight precise calculation algorithm is used to analyze the user interaction feedback data, and the interaction strategy is dynamically optimized accordingly, reducing the user's use difficulty and improving the completeness of information collection to more than 95%.

[0117] 5. In the rehabilitation auxiliary assessment method based on multi-agent and RAG technology provided in the application, the multi-agent system comprises a rehabilitation physician agent, an auxiliary engineer agent, a nursing staff agent and a user interaction agent. The dynamic knowledge base comprises an auxiliary product database, a rehabilitation medical knowledge base, a clinical case library and an industry standard library. The above-mentioned modular design process can conveniently add agent roles and expand the knowledge base range. For example, in one of the embodiments, the multi-agent system further comprises a psychological assessment agent, and the dynamic knowledge base further comprises a psychological intervention case library. In the above manner, the model expansion is very convenient. The system expansion model can be constructed based on the multi-agent task dynamic allocation algorithm and the evaluation dimension weight precise calculation algorithm, so as to ensure that the system can adapt to the auxiliary assessment requirements in different scenes and has a wider application range. BRIEF DESCRIPTION OF DRAWINGS

[0118] The drawings described herein are used to provide further understanding of the application, and form a part of the application. The illustrative embodiments of the application and their descriptions serve to explain the application, and do not constitute an improper limitation on the application. In the drawings:

[0119] Figure 1 A flowchart of the rehabilitation auxiliary assessment method based on multi-agent and RAG technology disclosed in one of the embodiments of the application;

[0120] Figure 2 For Figure 1 In the application, the effective collection mark x of the i-th item of information when determining the integrity of user information i The flowchart of setting;

[0121] Figure 3 For Figure 1 In the application, the dynamic difficulty coefficient ω of the i-th item of information when determining the integrity of user information i auto The flowchart of setting;

[0122] Figure 4 For Figure 1 In the application, the real-time collection quality correction factor γ of the i-th item of information when determining the integrity of user information i The flowchart of setting;

[0123] Figure 5 For Figure 1 In the application, the individualized correction coefficient δ of the i-th item of information when determining the integrity of user information u, i The flowchart of setting;

[0124] Figure 6 For Figure 1 In the application, the interaction adaptation degree weight α when determining the interaction adaptation degree auto The flowchart of setting;

[0125] Figure 7 For Figure 1 the flow chart of setting the cosine similarity Sim'(U, M) of the user-interactive mode when determining the interactive adaptation degree;

[0126] Figure 8 For Figure 1 the flow chart of setting the question order rationality score S auto when determining the interactive adaptation degree;

[0127] Figure 9 For Figure 1 the flow chart of setting the question order rationality score S auto when determining the interactive adaptation degree;

[0128] Figure 10 The module schematic diagram of the assistive device evaluation system based on multi-agent and RAG technology disclosed in one of the embodiments of the present application;

[0129] Figure 11 The module schematic diagram of the computer device disclosed in one of the embodiments of the present application. DETAILED DESCRIPTION

[0130] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0131] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form, unless the context clearly indicates otherwise, and it should also be understood that the terms "comprise" and / or "include" as used in the specification indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0132] Unless specifically stated otherwise, the relative arrangements of the components and steps illustrated in these embodiments and the numerical expressions and values set forth herein are not limiting of the scope of the present application. It should be understood that the various parts of the drawings are not necessarily drawn to scale, and that, for the purpose of convenience and clarity, not all components can be shown to scale in the drawings. Technical, methods, and apparatus known to the skilled artisan can not be discussed in detail, but should be considered as part of the specification, where appropriate. In all examples shown and discussed herein, any specific value should be interpreted as merely an example, and not as a limitation. Thus, other examples of the exemplary embodiments can have different values. It should be noted that like reference numerals and letters refer to like items in the following drawings, and thus, once an item is defined in one drawing, it need not be discussed further in subsequent drawings.

[0133] See Figure 1 In one embodiment of the present application, a rehabilitation assistive device evaluation method based on multi-agent and RAG technology is provided. The rehabilitation assistive device evaluation method comprises the following steps:

[0134] A multi-agent system is constructed, which comprises a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent, and a user interaction agent;

[0135] A dynamic knowledge base is built based on RAG technology, which comprises an assistive device product database, a rehabilitation medical knowledge base, a clinical case database, and an industry standard database;

[0136] The user interaction agent interacts with the user according to an adaptive interaction mechanism, and dynamically adjusts the questioning method, language style, and question order according to the user type and answer content, so as to improve the completeness and accuracy of user information collection;

[0137] The rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent retrieve in the dynamic knowledge base according to the information obtained by the user interaction agent, and respectively output professional evaluation opinions;

[0138] A multi-agent collaborative decision-making model is constructed, and the professional evaluation opinions of the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent are integrated according to the collaborative decision-making model to generate a multi-dimensional assistive device evaluation report and a recommended scheme.

[0139] In the present embodiment, the rehabilitation physician agent judges the type and degree of user dysfunction based on user physical condition information (including limb activity ability, perception function, and disease diagnosis result), in combination with rehabilitation medical guidelines and clinical evaluation standards retrieved by RAG (Retrieval-augmented Generation) technology, and outputs rehabilitation targets and medical requirements for assistive device adaptation. Specifically, in the working process, the rehabilitation physician agent receives user physical information transmitted by the user interaction agent, retrieves relevant medical knowledge through the RAG module, analyzes the user's functional level by using a medical evaluation algorithm, and finally generates a medical evaluation sub-report. The medical evaluation algorithm can be a FIM (Functional Independence Measure) algorithm, etc.

[0140] In this embodiment, the assistive device engineer agent screens the assistive device types that meet the conditions and analyzes the technical adaptability of the assistive device according to the medical requirements of the rehabilitation physician agent, in combination with the assistive device product parameters (such as size, load capacity, adjustment function) and technical standards (such as GB / T 14730-2000 “Hand-propelled wheelchairs for disabled persons”) retrieved by the RAG technology. Specifically, in the working process, the assistive device engineer agent receives the rehabilitation medical evaluation sub-report, retrieves the adaptive assistive device product library through the RAG module, evaluates the degree of fit between the assistive device and the user's needs by using the technical adaptation algorithm (such as the parameter matching degree algorithm), and finally generates the technical evaluation sub-report.

[0141] In this embodiment, the caregiver agent evaluates the convenience, safety and nursing operation difficulty of the assistive device in actual use by retrieving nursing scene cases and assistive device use safety specifications through the RAG technology in combination with the user's living scene information (such as the family living environment, daily activity habits, and nursing staff allocation situation). Specifically, in the working process, the caregiver agent receives the living scene information delivered by the user interaction agent, retrieves the nursing case library through the RAG module, analyzes the use feasibility of the assistive device by using the scene adaptation algorithm, and finally generates the nursing evaluation sub-report.

[0142] In this embodiment, the user interaction agent is responsible for interacting with the user and collecting user information. The user information includes physical condition, living scene, use preference and the like. After the information collection is completed, the user interaction agent will deliver the corresponding information to other agents, receive the feedback information of each agent, and show the evaluation progress and preliminary results to the user. Specifically, in the working process, the user interaction agent has a self-adaptive interaction mechanism, which can dynamically adjust the interaction mode through the user portrait algorithm. The user portrait algorithm constructs the user portrait based on the user's age, education level, health status and the like. For example, for elderly users, the user interaction agent can automatically switch to voice interaction to reduce text input. For example, for users who answer vaguely, the user interaction agent will automatically ask for supplementary information. For example, the user only states that “walking is difficult”, the user interaction agent will ask supplementary questions such as “whether there is unilateral limb weakness” and “whether support is needed when walking”.

[0143] In the rehabilitation assistive device evaluation method based on multi-agent and RAG technology provided in the embodiment, a multi-agent system is constructed, a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent and a user interaction agent are set, each agent performs its own function and cooperates, a multi-professional team collaborative evaluation process is simulated, and user demand is analyzed from multiple dimensions. With the help of multi-agent collaborative work, the artificial one-on-one evaluation process is replaced, the single user evaluation period is compressed to within 10 minutes, and the interference of artificial subjective factors is avoided. At the same time, based on the multi-agent task dynamic allocation algorithm, the evaluation process is optimized, the task amount of each agent is dynamically allocated, the evaluation resources are more reasonably utilized, and the accuracy of the evaluation result is improved by more than 30%.

[0144] In addition, the multi-agent system respectively evaluates from the dimensions of rehabilitation medicine, assistive device technology, nursing scene, etc., and combines the multi-source information retrieved by the RAG technology, which can cover the user's physical condition, use scene, family environment, assistive device adaptability and other aspects of demand. According to the evaluation dimension weight precise calculation algorithm, the evaluation weight of each dimension is accurately calculated and adjusted, the problem of single system evaluation dimension limitation is solved, and the evaluation result is more comprehensive and accurate.

[0145] Moreover, by integrating the RAG technology, a dynamic knowledge base including an assistive device product database, a rehabilitation medicine knowledge base, a clinical case database, an industry standard database and the like is built, real-time retrieval of the latest information is realized, and data support is provided for multi-agent evaluation. At the same time, based on the RAG technology, the dynamic knowledge base is built, and the real-time and richness of the rehabilitation assistive device evaluation method in the application are enhanced. The RAG technology can be connected to the Internet and professional databases in real time, realize real-time updating and accurate retrieval of assistive device product information, medical guidelines and clinical cases. Combined with the multi-agent task dynamic allocation algorithm, the information retrieval strategy is optimized, and for special user needs, rare disease assistive device adaptation cases, the latest assistive device technology parameters and the like can be quickly obtained, and the scientificity of the evaluation suggestion is ensured.

[0146] In addition, by designing an adaptive interaction mechanism, the user interaction agent can dynamically adjust the questioning method, language style and question order according to the user type (such as an elderly user, a cognitive impairment user) and the answer content, and improve the completeness and accuracy of user information collection. The adaptive interaction mechanism can adjust the interaction method according to the characteristics of the user, for example, voice interaction and simplified text and image display are used for elderly users, and step-by-step guided questioning is used for cognitive impairment users. The evaluation dimension weight precise calculation algorithm is used to analyze user interaction feedback data, and the interaction strategy is dynamically optimized accordingly, the user use difficulty is reduced, and the information collection completeness rate is improved to more than 95%.

[0147] In addition, in the rehabilitation assistive device evaluation method based on the multi-agent and RAG technology provided in the embodiment, the multi-agent system includes a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent, and a user interaction agent. The dynamic knowledge base includes an assistive device product database, a rehabilitation medical knowledge base, a clinical case library, and an industry standard library. The above modular design process can conveniently add agent roles and expand the knowledge base range. For example, in one of the embodiments, the multi-agent system further includes a psychological evaluation agent, and the dynamic knowledge base further includes a psychological intervention case library. In the above manner, the model can be easily expanded. Later, a system expansion model can be constructed based on a multi-agent task dynamic allocation algorithm and an evaluation dimension weight precise calculation algorithm, so as to ensure that the system can adapt to the assistive device evaluation requirements in different scenarios and has a wider application range.

[0148] In one of the embodiments, the adaptive interaction mechanism includes dynamically adjusting the questioning method, language style, and question order according to the completeness of the user information collection, and the completeness of the user information collection is calculated in the following manner:

[0149] ;

[0150] wherein P is the completeness of the user information collection; n is the number of core information items; xi is the effective collection mark of the ith information item; ωi is the dynamic difficulty coefficient of the ith information item; γi is the real-time collection quality correction factor; and δi is the individualized correction coefficient. i i auto i u, i

[0151] In the embodiment, by setting the effective collection mark, the dynamic difficulty coefficient, the real-time collection quality correction factor, and the individualized correction coefficient, the completeness and accuracy of the user information collection can be effectively improved. Specifically, the completeness P of the user information collection can be output automatically by the system, and the precision can be selected to retain 2 decimal places.

[0152] In one of the embodiments, the adaptive interaction mechanism includes the following steps:

[0153] Obtaining the information provided by the user;

[0154] Calculating the completeness of the user information according to the information provided by the user;

[0155] Dynamically adjusting the questioning method, language style, and question order according to the completeness of the user information.

[0156] ​​​​In the embodiment, n is the number of core information items. Specifically, the number of core information items is automatically matched by the system based on the evaluation scene. For example, n = 12 for postoperative rehabilitation scene and n = 10 for elderly home care scene. According to the need, a scene-information mapping table can be set in the dynamic knowledge base, and then automatically called in the process of assistive device evaluation. According to the need, an assistive device evaluation scene rule sub-library can be set in the dynamic knowledge base. The assistive device evaluation scene rule sub-library is generated by training historical scene data to provide the content of the scene-information mapping in the process of assistive device evaluation.

[0157] Please see Figure 2 In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0158] The threshold model trained according to the historical data determines the integrity threshold;

[0159] If the user information meets the condition of the integrity threshold, x i = 1;

[0160] If the user information is missing or the error exceeds the integrity threshold, the user interaction agent starts the automatic follow-up system for completion;

[0161] If the automatic follow-up system is started, the user information meets the condition of the integrity threshold, x i = 1;

[0162] If the automatic follow-up system is started, the user information is still missing or the error exceeds the integrity threshold, x i = 0.

[0163] In the embodiment, by setting the integrity threshold, when the user information meets the condition of the integrity threshold, the value of the effective collection mark x i of the i-th information is set to 1. When the user information is missing or the error of the user information exceeds the integrity threshold, the automatic follow-up system is enabled to complete by supplementing questions. If the completion is still invalid, the value of the effective collection mark x i of the i-th information is set to 0. For example, when asking the question of "family stair height", the user's answer of the height range should be accurate within ±2cm. Otherwise, the system will supplement the follow-up question to determine an actual family stair height. It can be understood that the user information is obtained through the communication process between the user interaction agent and the user. The integrity threshold is obtained according to the threshold model trained from the historical data, so that the completeness of the user information is more accurate and effective.

[0164] Please see Figure 3In one embodiment, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0165] obtaining the average collection success rate R of the ith information in the historical database i ;

[0166] obtaining the importance score I given to the ith information in the evaluation result influence degree model i ;

[0167] determining the dynamic difficulty coefficient ω of the ith information according to the average collection success rate R of the ith information i and the importance score I of the ith information i i auto .

[0168] In one embodiment, the dynamic difficulty coefficient ω of the ith information is calculated according to the following manner i auto :

[0169] .

[0170] In this embodiment, by obtaining the average collection success rate R of the ith information i and the importance score I given to the ith information i , the dynamic difficulty coefficient ω of the ith information can be effectively obtained i auto For example, for the question of “muscle strength level”, the average collection success rate of this information in the historical database is 92%. Therefore, for the information of “muscle strength level”, the average collection success rate R i = 92%. For another example, for the question of “density of home obstacles”, the average collection success rate of this information in the historical database is 68%. Therefore, for the information of “density of home obstacles”, the average collection success rate R i = 68%. The importance score I i is automatically given by the system based on the evaluation result influence degree model. The value of I i ranges from 1 to 5. For example, for the information of “disease diagnosis code”, the value of the importance score I i is 5. For the information of “assistive device color preference”, the value of the importance score I i is 1. In this embodiment, the historical database contains more than 100,000 evaluation cases to provide a more accurate value of the average collection success rate R i of the ith information. The evaluation result influence degree model is obtained by training based on historical recommendation effects, so that the importance score I i ​It has more practical significance.

[0171] Please see also Figure 4 In one embodiment, the rehabilitation assistive device assessment method based on multi-agent and RAG technology further includes the following steps:

[0172] Obtain the time T for the user to answer the i-th item of information. i ;

[0173] Obtain the average time Tˉ for similar users to answer the i-th item of information in the historical database. i ;

[0174] Based on the time T of the user answering the i-th item of information i And the average time Tˉ for similar users to answer the i-th item of information. i Determine the real-time collection quality correction factor γ for the i-th information item. i .

[0175] In one embodiment, the real-time collection quality correction factor γ of the i-th piece of information is calculated according to the following method. i :

[0176] ;

[0177] Wherein, min (T) i / Tˉ i ,2) is T i / Tˉ i The smaller of the two.

[0178] Understandably, the duration of a user's response to the i-th item of information also significantly impacts the completeness of user information collection. In this embodiment, the duration T of the user's response to the i-th item of information is obtained. i And the average time Tˉ for obtaining the i-th item of information from similar users in the historical database. i Real-time collection of quality correction factor γ i This can be achieved through duration data T i and average duration Tˉ i Calculations are performed to make the data on the completeness P of the collected user information more accurate. For example, for elderly users, the average time to answer the question "Home floor" is 15 seconds. Understandably, the time T for a user to answer the i-th piece of information... i This can be derived from the interaction process between the user interaction agent and the user.

[0179] Please see also Figure 5 In one embodiment, the rehabilitation assistive device assessment method based on multi-agent and RAG technology further includes the following steps:

[0180] Obtaining the number of user portrait tags k;

[0181] Obtaining the weight information β of the tth tag t ;

[0182] Obtaining the influence coefficient σ of the tth tag on the i information collection t,i ;

[0183] Calculating the personalized correction coefficient δ of the i information according to the following method u, i :

[0184] .

[0185] It can be understood that the user's age, education level, and health status and other information also have important influence on the completeness of user information collection. In the embodiment, by obtaining the number of user portrait tags, the weight information β of the tth tag t , the influence coefficient σ of the tth tag on the i information collection t,i and other information, the personalized correction coefficient δ of the i information can be effectively calculated u, i . According to the needs, the user interaction intelligent agent can also include a user portrait module, which can output user portrait tag data through an automated classification algorithm. For example, for a certain user, it includes 3 tags of "old user", "cognitive impairment" and "voice interaction". Then the number of user portrait tags k of this user is equal to 3. The weight information β t is automatically calculated by the system based on the influence degree of the user portrait tag on information collection. For example, for a user with a "cognitive impairment" tag, β t is set to 0.15. The influence coefficient σ t,i is obtained by correlation analysis of historical tags and collection effects. For example, for a user with a "cognitive impairment" tag, the value of the influence coefficient σ t,i is set to 0.2. According to the needs, the dynamic knowledge base also includes a tag influence degree database. The tag influence degree database is used to store the correlation information of historical tags and collection effects. In the embodiment, the user portrait algorithm used by the system is a classification algorithm based on decision tree. In combination with the characteristics of user age, education level, health status (such as whether there is cognitive impairment), operation device type and the like, the system can output user interaction type tags, such as "old voice type", "ordinary text type", "cognitive guidance type" and the like. Similarly, in the field of precision marketing, by analyzing the user's consumption habits and purchase behavior, enterprises can use decision tree algorithm to provide more accurate marketing strategies and target customer groups.

[0186] It can be seen that, on the basis of obtaining the completeness P data of user information collection, the dynamic difficulty coefficient ωi auto , γ i , δ u, i The parameters are automatically generated by historical data statistics + real-time data analysis + algorithm model without manual setting. The effective collection mark x i of the ith information is determined based on the integrity threshold + automatic follow-up completion, and the whole process does not need manual intervention. The above method greatly speeds up the evaluation time of the disabled assistive device, and effectively improves the accuracy of the evaluation result.

[0187] In one of the embodiments, the adaptive interaction mechanism includes dynamically adjusting the questioning method, language style and question order according to the interaction adaptation degree, and the interaction adaptation degree is calculated by the following method:

[0188] ;

[0189] Wherein, Q is the interaction adaptation degree; a auto is the interaction adaptation degree weight; Sim'(U, M) is the user-interactive method weighted cosine similarity; S auto is the question order rationality score; λ is the real-time feedback correction factor.

[0190] In this embodiment, by setting the interaction adaptation degree weight a auto , the user-interactive method weighted cosine similarity Sim'(U, M), the question order rationality score S auto and the real-time feedback correction factor λ, the value of the interaction adaptation degree Q can be effectively calculated. The value of the interaction adaptation degree Q represents the adaptation degree of the assistive device and the user. The value of the interaction adaptation degree Q ranges from 0 to 1, and the precision can be kept to 3 decimal places.

[0191] In one of the embodiments, the adaptive interaction mechanism further includes the following steps:

[0192] Obtaining the information provided by the user;

[0193] Calculating the interaction adaptation degree according to the information provided by the user;

[0194] According to the interaction adaptation degree, dynamically adjusting the questioning method, language style and question order.

[0195] Please see Figure 6 In one of the embodiments, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further includes the following steps:

[0196] Obtaining the average contribution Q of the question order adaptation to the overall adaptation degree in the historical database seq ;

[0197] Obtain the average contribution Qˉ of interaction method adaptation to overall adaptation in the historical database. mode ;

[0198] The interaction fit weight α is calculated as follows: auto :

[0199] .

[0200] Understandably, the order of questions and the interaction method affect the overall interaction fit Q. In this embodiment, the average contribution Qˉ of question order fit to the overall fit is obtained from the historical database. seq And obtain the average contribution Qˉ of interaction method adaptation to the overall adaptation in the historical database. mode Then the interaction adaptation weight α can be calculated. auto In this embodiment, Qˉ seq and Qˉ mode The results were derived from the system's adaptation analysis based on a large number of (over 100,000) interaction cases.

[0201] Please see also Figure 7 In one embodiment, the rehabilitation assistive device assessment method based on multi-agent and RAG technology further includes the following steps:

[0202] Multiple preset interaction parameters are provided, including age, cognitive level, and device type.

[0203] The interaction parameters are assigned different weight values ​​ω based on the dimensionality influence. d ;

[0204] Obtain the standardized value U of user features d ;

[0205] Obtain the standardized value M of the interaction method characteristics d ;

[0206] The user-interaction weighted cosine similarity Sim'(U, M) is calculated as follows:

[0207] ;

[0208] Among them, H u This is a correction item for user history interaction acceptance.

[0209] Understandably, cosine similarity refers to evaluating the similarity between two vectors by calculating the cosine of the angle between them. In this embodiment, multiple interaction parameters are preset, and different weight values ​​ω are assigned to these interaction parameters based on dimensional influence. dAt this time, the user-interaction mode weighted cosine similarity Sim'(U, M) can be easily obtained by the standardized value U d of the user feature and the standardized value M d of the interaction mode feature.

[0210] In the embodiment, the weight value ω d is a value automatically given by the system according to the dimension influence degree. For example, the weight value ω1 of the interaction parameter of "cognitive level" is 0.45. The weight value ω2 of the interaction parameter of "age" is 0.3. The weight value ω2 of the interaction parameter of "device type" is 0.25.

[0211] In the embodiment, the standardized value U d of the user feature is extracted by the system from the input information of the user and is standardized. For example, assuming that the age of the user is 68 years old, the standardized value U d of the user feature is set to 0.85.

[0212] In the embodiment, the standardized value M d of the interaction mode feature is automatically called by the system from the interaction mode feature library. For example, the standardized value M d of the interaction mode feature of "voice interaction" is set to 0.3.

[0213] According to the need, a user historical interaction acceptance correction term parameter H u may also be set. The system counts the acceptance rate of the user to the same type of mode in the past 3 interactions, such as an acceptance rate of 80%, and the user historical interaction acceptance correction term H u is set to 0.08.

[0214] In the embodiment, the dynamic knowledge base further includes an interaction mode feature library. The interaction mode feature library stores parameters of various interaction mode features for the multi-agent collaborative decision-making model to call.

[0215] Please refer to Figure 8 , in one of the embodiments, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further includes the following steps:

[0216] Obtain the information collection efficiency E seq of the current problem order;

[0217] Obtain the average information collection efficiency Eˉ seq of the same type of user in the historical database;

[0218] Predict the current cognitive load L u of the user based on the answer duration and the error rate;

[0219] Setting the cognitive load threshold L of the same type of user based on historical data max ;

[0220] The problem order rationality score S is calculated according to the following manner auto :

[0221] .

[0222] It can be understood that the problem order rationality also has an impact on the interaction adaptation degree Q of the assistive device. In the present embodiment, by acquiring the information collection efficiency E seq , the average information collection efficiency Eˉ seq , the current cognitive load L u , and the cognitive load threshold Lˉ max , etc., the problem order rationality score S auto can be calculated. In the present embodiment, the system will calculate the ratio of the number of questions to the collection time in real time, thereby obtaining the information collection efficiency E seq of the current question order. Then, the system queries the average information collection efficiency Eˉ seq of the same type of user from the historical database, so as to obtain the ratio of the information collection efficiency E seq of the current question order to the average information collection efficiency Eˉ seq . On the other hand, based on the response time of the user and the error rate of the user's response, the current cognitive load L u of the user can be calculated. For example, if the error rate of the response is 10%, the current cognitive load L u of the user can be set to 0.15. The cognitive load threshold Lˉ max of the same type of user is set based on historical data. For example, for elderly users, the cognitive load threshold Lˉ max can be set to 0.3.

[0223] In one of the embodiments, the rehabilitation assistive device evaluation method based on multi-agent and RAG technology further comprises the following steps:

[0224] Acquiring the number of times ф1 of the user's active confirmation;

[0225] Acquiring the number of times ф2 of the user's skipping or re-listening;

[0226] The real-time feedback correction factor λ is calculated according to the following manner:

[0227] .

[0228] In this embodiment, the feedback correction factor λ can be effectively calculated by acquiring the number of times the user actively confirms (ф1) and the number of times the user skips or re-listens (ф2). Specifically, if the user clicks "Question clear," the number of times the user actively confirms (ф1) can be set to 1. If the user selects "Skip" or "Re-listen," the value of ф2 can be set accordingly. In this embodiment, the system automatically records user clicks, voice commands, and other behaviors to obtain a real-time interaction behavior log.

[0229] In one embodiment, the agent's final weight ω j final Including the first value ω j final,0 Second value ω j final,1 ;

[0230] In a conflict-free scenario, the agent's final weight ω j final The first value ω j final,0 Determined in the following ways:

[0231] ;

[0232] in, As the basic weight for the scenario; For the historical accuracy of the intelligent agent;

[0233] In conflict scenarios, the final weight ω of the agent j final The second value ω j final,1 Determined in the following ways:

[0234] ;

[0235] in, This is the conflict correction factor; Adjustment coefficients based on user needs.

[0236] In one embodiment, the rehabilitation aid assessment method based on multi-agent and RAG technology further includes the following steps:

[0237] Obtain the sub-reports output by the rehabilitation physician intelligent agent, the assistive device engineer intelligent agent, and the nursing staff intelligent agent respectively;

[0238] Calculate the cosine similarity between the two sub-reports;

[0239] The degree of divergence is determined based on the cosine similarity between the two sub-reports;

[0240] If the value of the divergence degree is less than the preset value, it is determined that there is a conflict scenario between the sub-reports.

[0241] If the value of the divergence degree is greater than or equal to the preset value, it is determined that there is no conflict scenario between the sub-reports.

[0242] The divergence degree D is calculated in the following manner:

[0243]

[0244] The Sim is a cosine similarity of the sub-reports.

[0245] According to needs, the preset value can be set to 0.3. At this time, if the value of the divergence degree D is less than 0.3, it is determined that there is a conflict scenario between the sub-reports. If the value of the divergence degree D is greater than or equal to 0.3, it is determined that there is no conflict scenario between the sub-reports.

[0246] In one of the embodiments, the multi-agent system further comprises a psychological assessment agent, and the dynamic knowledge base further comprises a psychological intervention case library.

[0247] The psychological assessment agent is configured to assess the psychological acceptance and use willingness of the user to the assistive device.

[0248] The collaborative decision-making model integrates the professional assessment opinions of the rehabilitation physician agent, the assistive device engineer agent, the nursing staff agent, and the psychological assessment agent to generate a multi-dimensional assistive device evaluation report and a recommended scheme.

[0249] In the present embodiment, by setting the psychological assessment agent to assess the psychological acceptance and use willingness of the user to the assistive device, the dimensions considered by the output assistive device evaluation report and recommended scheme are more, and the pertinence is stronger.

[0250] Please refer to Figure 9 In one of the embodiments, the rehabilitation assistive device evaluation method based on the multi-agent and RAG technology further comprises the following steps:

[0251] According to the collected user information, the user identity feature is determined.

[0252] According to the user identity feature, it is determined whether an assistive device evaluation report and a recommended scheme have been provided for the user within a preset time range.

[0253] If so, the new demand is determined according to the difference between the present user demand and the last user demand; the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent retrieve in the dynamic knowledge base according to the new demand, and supplement to the last professional assessment opinion respectively.

[0254] ​In the embodiment, if the assistive device evaluation report and the recommended scheme have been provided for the user before, the new needs can be determined according to the difference between the current user needs and the last user needs. At this time, the rehabilitation physician agent, the assistive device engineer agent and the nursing staff agent retrieve in the dynamic knowledge base according to the new needs and supplement to the last professional evaluation opinions respectively. That is, in the embodiment, the incremental retrieval mode can be used instead of the full retrieval, and only the new data related to the current user needs (such as the assistive device products updated in the last 3 months) are retrieved, so as to shorten the retrieval time and improve the system response speed. This mode is suitable for the scene with poor network environment.

[0255] In one of the embodiments, the user interaction agent further comprises:

[0256] a video shooting module, configured to shoot a home environment video according to the user needs;

[0257] The nursing staff agent performs scene evaluation according to the content of the shot home environment video, so as to improve the accuracy of scene information collection.

[0258] In the embodiment, the video shooting module is added and the home environment video is shot according to the user needs, at this time, the nursing staff agent performs scene evaluation according to the content of the shot home environment video, so as to improve the accuracy of scene information collection. This mode is suitable for the user in complex home environment.

[0259] See Figure 10 Another embodiment of the present application further provides an assistive device evaluation system 100 based on multi-agent and RAG technology. The assistive device evaluation system 100 comprises a multi-agent module 110, a RAG technology module 120, a dynamic knowledge base module 130 and a collaborative decision-making module 140.

[0260] The multi-agent module 110 comprises a rehabilitation physician agent 111, an assistive device engineer agent 112, a nursing staff agent 113 and a user interaction agent 114.

[0261] The dynamic knowledge base module 130 comprises an assistive device product database 131, a rehabilitation medical knowledge base 132, a clinical case database 133 and an industry standard database 134.

[0262] The RAG technology module 120 is configured to connect the multi-agent module 110 and the dynamic knowledge base module 130, so as to realize retrieval and information extraction of the dynamic knowledge base module 130 and provide information support for the multi-agent module 110.

[0263] The user interaction agent 114 interacts with the user according to an adaptive interaction mechanism, and dynamically adjusts the questioning method, language style and question order according to the user type and answer content, so as to improve the completeness and accuracy of user information collection;

[0264] The rehabilitation physician agent 111, the assistive device engineer agent 112 and the nursing staff agent 113 retrieve information in the dynamic knowledge base module 130 according to the information obtained by the user interaction agent 114, and respectively output professional evaluation opinions;

[0265] The collaborative decision-making module 140 is used for integrating the professional evaluation opinions of the rehabilitation physician agent 111, the assistive device engineer agent 112 and the nursing staff agent 113, so as to generate a multi-dimensional assistive device evaluation report and a recommended scheme.

[0266] In the embodiment, the system is built based on computer hardware devices (including servers, terminal devices such as computers, tablets and smart phones), and the software part is composed of a button multi-agent module, an RAG technology module, a dynamic knowledge base module, an adaptive interaction module, a collaborative decision-making module and an evaluation report generation module. The modules realize bidirectional data interaction through data interfaces. Specifically, the server uses a high-performance cloud computing server, which is configured with an 8-core CPU, 32 GB of memory and 1 TB of SSD storage. These configurations ensure excellent CPU performance and memory capacity, as well as fast data read-write speed. The server is designed for running multi-agent algorithms, RAG retrieval models and storing dynamic knowledge base data, supports multiple users to access online at the same time, and ensures the stability and response speed of the system. The terminal devices used by users support computers (Windows / macOS system), tablets (iOS / Android system) and smart phones (iOS / Android system). Users can access the system through the browser or a special APP of the terminal device, fill in the information and view the evaluation results; evaluation personnel can monitor the evaluation process through the terminal device and adjust the system parameters.

[0267] In the information retrieval process, each agent such as the rehabilitation physician agent 111, the assistive device engineer agent 112, the nursing staff agent 113, and the user interaction agent 114, etc. first sends a retrieval request to the RAG technology module 120. For example, the rehabilitation physician agent 111 requests “upper limb assistive device adaptation guidelines for stroke patients”. The RAG technology module 120 analyzes the request keywords, uses a vector retrieval algorithm to match related information in the dynamic knowledge base module 130, and returns the retrieval results to the agent after structured processing. The vector retrieval algorithm can be the FAISS (Facebook AI Similarity Search) algorithm, etc. The retrieval results can be guideline document fragments, related cases, etc.

[0268] According to the needs, the RAG technology module 120 can also perform the operation of updating the dynamic knowledge base module 130 periodically. For example, the latest information of authoritative websites such as the official website of the National Rehabilitation Assistive Device Research Center and the official website of the International Rehabilitation Engineering and Assistive Technology Association, etc. can be captured by a timing crawler program to obtain assistive device product update information, medical guideline revision information, etc. The crawler program can be set to run automatically at dawn every day. In combination with manual review, the effective information is stored in the dynamic knowledge base to ensure the timeliness of the data. The evaluation personnel review the updated content every week to ensure that the updated content is accurate and meets the specifications.

[0269] In the present embodiment, the dynamic knowledge base module 130 includes four sub-libraries: the assistive device product database 131, the rehabilitation medical knowledge base 132, the clinical case library 133, and the industry standard library 134. Each sub-library uses a mixed storage mode of a relational database (MySQL) and a non-relational database (MongoDB).

[0270] The assistive device product database 131 is used to store domestic and foreign assistive device product information, including product name, model, manufacturer, technical parameters, price, certification, user evaluation, etc.

[0271] The rehabilitation medical knowledge base 132 is used to store rehabilitation medical guidelines, clinical evaluation standards, disease and assistive device adaptation correspondence, rehabilitation training programs, etc.

[0272] The clinical case library 133 is used to store assistive device evaluation cases of users with different disease types and degrees of dysfunction. The assistive device evaluation cases include user basic information, evaluation process, recommended assistive devices, and use effect feedback, etc.

[0273] The industry standard library 134 is used to store national and international assistive device industry standards, quality detection specifications, policies and regulations such as assistive device subsidy policies, etc.

[0274] As needed, the dynamic knowledge base module 130 employs data encryption technology (AES-256 encryption) to protect user privacy data security and sets access permissions. For example, ordinary users can only view publicly available assistive device product information, while evaluators can access the complete knowledge base.

[0275] The collaborative decision-making module 140 is used to receive assessment sub-reports output by rehabilitation physician agent 111, assistive device engineer agent 112, and nursing staff agent 113, and integrate the opinions of each sub-report through a multi-dimensional weighting algorithm to generate a comprehensive assessment result and assistive device recommendation scheme.

[0276] In this embodiment, the collaborative decision-making module 140 determines the comprehensive evaluation score of the assistive device using the following formula:

[0277] ;

[0278] Among them, S total The overall evaluation score; j is the agent identifier; S j For agent evaluation sub-report score; γ i ω is the sub-report credibility correction factor; j final The final weights of the agent.

[0279] In this embodiment, j=1 represents a rehabilitation physician; j=2 represents an assistive device engineer; and j=3 represents a nursing staff. The corresponding roles can be predefined in the role library of the multi-agent module 110.

[0280] S j This refers to the agent assessment sub-report score. Specifically, the agent assessment sub-report score ranges from 0 to 100. For example, for a rehabilitation physician, their agent assessment sub-report score S1 = 0.4F FIM + 0.3D 适配 +0.3R 康复 .

[0281] Sub-report credibility correction factor γ i The value ranges from 0.8 to 1.2. In this embodiment, the sub-report credibility correction factor is calculated using the following formula:

[0282] ;

[0283] Where, N j Q represents the number of information retrieved by RAG. j Information quality score; N max Q represents the highest number of information retrieved by RAG. max This is the highest score for information quality.

[0284] ωj final The maximum weight of the agent. In the embodiment, ω j final The numerical value is calculated separately in the conflict-free / conflict scenario.

[0285] In the conflict-free scenario, the maximum weight of the agent ω j final,0 is calculated by the following formula:

[0286] ;

[0287] Wherein, is the scene basic weight; is the historical accuracy of the agent. For example, in the postoperative rehabilitation scenario, ω = 0.4. In the embodiment, the scene basic weight is automatically matched by the scene classification model. The historical accuracy of the agent ranges from 0.7 to 1.0. Specifically, the consistency rate of the same type of evaluation in the past 1000 times is counted, and then the historical accuracy of the agent is obtained.

[0288] It should be noted that in the conflict-free scenario, one is that the number of coincident assistive device styles in the recommended top 3 assistive devices in each sub-report is greater than or equal to 2; the other is that the value of the divergence D is less than 0.3. Wherein, the divergence D is calculated by the following way:

[0289] ;

[0290] Wherein, Sim is the cosine similarity of the sub-report.

[0291] In the conflict scenario, the maximum weight of the agent ω j final,1 is calculated by the following formula:

[0292] ;

[0293] Wherein, is the conflict correction coefficient; is the user demand correction coefficient.

[0294] In the embodiment, the conflict correction coefficient is calculated by the following way:

[0295] ;

[0296] Wherein, is the number of conflict retrieval information; is the quality score of the conflict retrieval information.

[0297] In the present embodiment, the user demand correction coefficient is calculated in the following manner:

[0298] ;

[0299] wherein, is the user demand priority; is the agent demand influence coefficient. In the present embodiment, the user demand priority and the agent demand influence coefficient are provided by the NLP demand extraction module.

[0300] In the present embodiment, the weights of the opinions of the agents are set according to the evaluation scenarios (such as postoperative rehabilitation assistive device evaluation, home assistive device evaluation for the elderly). In the present embodiment, the multi-agent weight algorithm adopts the analytic hierarchy process (AHP). By constructing a hierarchical structure of the evaluation target (such as “postoperative rehabilitation assistive device evaluation”) - the criterion layer (medical adaptability, technical adaptability, scene adaptability) - the scheme layer (the opinions of the agents), the weights of the criteria are determined by pairwise comparison, and the final weights of the opinions of the agents are calculated.

[0301] When the opinions of the agents conflict, for example: the rehabilitation physician recommends a high-support assistive device, and the nursing staff believes that the operation difficulty of the assistive device is too high, the system automatically retrieves the relevant cases and standards in the RAG module, combines the user priority demand (such as whether the user pays more attention to safety or convenience), adjusts the weights, and outputs the optimal scheme.

[0302] In the present embodiment, the working process of the collaborative decision-making module 140 is as follows:

[0303] Receiving the evaluation sub-reports output by the rehabilitation physician agent 111, the assistive device engineer agent 112, and the nursing staff agent 113, calculating the agent evaluation sub-report score S j and the sub-report credibility correction factor γ i .

[0304] According to the cosine similarity Sim of the sub-reports, the degree of divergence D is calculated, so as to determine whether there is a conflict scenario;

[0305] In the case of no conflict scenario, the final weight ω j of the agent is calculated; final,0 In the case of conflict scenario, the final weight ω j of the agent is calculated; final,1 ;

[0306] According to the agent evaluation sub-report score S j , the sub-report credibility correction factor γ i , and the numerical value of the final weight of the agent, the comprehensive evaluation score Stotal and select 1-3 auxiliary production recommendation programs with the highest scores.

[0307] According to needs, the assistive device evaluation system 100 further comprises an evaluation report generation module 150. The evaluation report generation module 150 automatically generates a structured assistive device evaluation report according to the comprehensive evaluation result of the collaborative decision-making module 140. At the same time, the generated assistive device evaluation report also supports PDF format download and online viewing. In this embodiment, the generated assistive device evaluation report covers user basic information, evaluation conclusions of each dimension (including body function evaluation, assistive device technology adaptation evaluation, and scene use evaluation), recommended assistive device list (including product parameters, adaptation reasons, and use suggestions), and subsequent follow-up plan, etc.

[0308] In one of the embodiments, the multi-agent module 110 further comprises a psychological evaluation agent 115. The dynamic knowledge base module 130 further comprises a psychological intervention case library 135.

[0309] The psychological evaluation agent 115 is used to evaluate the psychological acceptance and use willingness of the user to the assistive device;

[0310] The collaborative decision-making module 140 integrates the professional evaluation opinions of the rehabilitation physician agent 111, the assistive device engineer agent 112, the nursing staff agent 113, and the psychological evaluation agent 115 to generate a multi-dimensional assistive device evaluation report and a recommended scheme.

[0311] In the assistive device evaluation system 100 provided in the above embodiments, by constructing a multi-agent system composed of rehabilitation physicians, assistive device engineers, nursing staff, and user interactions, each agent has clear division of labor, and efficient collaborative work is achieved through standardized data interfaces. This architecture simulates the evaluation process of a multi-specialty team to represent the evaluation contribution of each agent, effectively solving the problem of limited evaluation dimension of a single role in the prior art. For example, in a smart home system, data sharing and service interaction are achieved through multi-agent collaborative technology, improving the efficiency and energy management of home control; in the intelligent management of industrial equipment, multi-agent collaborative technology realizes effective network collaboration of equipment, improving the level of autonomous intelligent control.

[0312] In addition, the assistive device evaluation system 100 also deeply applies RAG technology to the assistive device evaluation scene, and constructs a dynamic knowledge base containing multiple types of sub-libraries. This knowledge base not only supports real-time retrieval of the latest assistive device information, medical knowledge, and cases, but also has an automatic updating function, effectively solving the problems of information lag and low retrieval efficiency in existing systems.

[0313] Moreover, the assistive device evaluation system 100 also provides an adaptive interaction mechanism for different user groups. Through the user portrait algorithm, the assistive device evaluation system 100 can dynamically adapt the interaction mode, such as voice interaction, step-by-step guided questioning, etc., to meet the needs of special groups such as elderly users and users with cognitive impairment. This mechanism effectively solves the problem of single interaction mode and complex operation of existing systems by improving the completeness of information collection and user experience.

[0314] In generating the comprehensive evaluation result and the assistive device recommendation scheme, the assistive device evaluation system 100 also sets the weight of each agent's opinion according to the evaluation scene using the analytic hierarchy process (AHP) weight processing method, and combines the case and standard retrieved by RAG and the weight solving model of information entropy theory to solve the opinion conflict and optimize the decision-making process. By synthesizing the evaluation opinions of each agent, a comprehensive evaluation scheme is generated, which ensures the scientificity and individualization of the evaluation result and effectively solves the problem of single decision-making and lack of flexible adjustment in the prior art.

[0315] The following are typical application scenario cases of the rehabilitation assistive device evaluation method and system based on multi-agent and RAG technology provided by the embodiments of the present application:

[0316] Case 1: For a user after cerebral apoplexy surgery, who has right limb weakness and a staircase in the home environment, the system evaluates the user's motor function through the rehabilitation physician agent, recommends a walking aid with a handrail through the assistive device engineer agent, and suggests installing a non-slip pad in combination with the staircase scene through the nursing staff agent, and finally generates an evaluation report containing the walking aid recommendation and home modification suggestions.

[0317] Case 2: Alzheimer's disease user evaluation, the user's cognitive ability is declining and it is difficult to operate electronic devices, the system automatically switches to voice interaction + picture selection mode to collect information, the rehabilitation physician agent evaluates the user's cognitive level, and the assistive device engineer agent recommends a simple-to-operate smart bracelet.

[0318] See Figure 11 Another embodiment of the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method of any one of the above embodiments when executing the computer program.

[0319] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0320] The above only is the preferred embodiment of the present application, and is not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A rehabilitation assistive device evaluation method based on multi-agent and RAG technology, characterized in that, The method comprises the following steps: A multi-agent system is constructed, which comprises a rehabilitation physician agent, an assistive device engineer agent, a nursing staff agent, and a user interaction agent; A dynamic knowledge base is built based on RAG technology, which comprises an assistive device product database, a rehabilitation medicine knowledge base, a clinical case database, and an industry standard database; The user interaction agent is used to interact with the user, and the user interaction agent adopts an adaptive interaction mechanism to dynamically adjust the questioning method, language style, and question order according to the user type and answer content, so as to improve the completeness and accuracy of user information collection; According to the information obtained by the user interaction agent, the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent retrieve in the dynamic knowledge base and output professional evaluation opinions respectively; A multi-agent collaborative decision-making model is constructed, and the professional evaluation opinions of the rehabilitation physician agent, the assistive device engineer agent, and the nursing staff agent are integrated according to the collaborative decision-making model to generate a multi-dimensional assistive device evaluation report and a recommended scheme; The adaptive interaction mechanism comprises dynamically adjusting the questioning method, language style, and question order according to the completeness of user information collection, and the completeness of user information collection is calculated by the following method: ； wherein, P for the integrity of the collection of user information; n for the number of core information items; x i for the first i effective collection marker of the item of information; ω i auto for the first i dynamic difficulty coefficient of the item of information; γ i for the real-time collection quality correction factor; δ u, i for the personalization correction coefficient.

2. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 1, characterized in that, The method further comprises the following steps: A threshold model trained according to historical data is used to determine a completeness threshold; If the user information satisfies the condition of the integrity threshold, x i = 1; If the user information is missing or the error exceeds the completeness threshold, the user interaction agent starts an automatic follow-up system to complete it; If the automatic follow-up system is activated, the user information meets the condition of the completeness threshold, x i = 1; If the automatic follow-up system is activated, the user information is still missing or the error exceeds the integrity threshold, x i = 0.

3. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 1, characterized in that, The method further comprises the following steps: acquiring an average collection success rate of item information in the historical database i acquiring an average collection success rate of item information in the historical database R i ; acquiring the importance score assigned to the item of information in the evaluation result influence degree model i I i ;​ According to the first i Average collection success rate of the item of information R i And the importance score of the first i Item of information I i Determine the dynamic difficulty coefficient ω of the first i Item of information i auto .

4. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 1, characterized in that, The method further comprises the following steps: acquiring a time length data of the user's answer to the i item information T i ; Retrieve answers from similar users in the historical database i Average duration of information item Tˉ i ; According to the user's answer to the first i item of information, the time length data T i of the same kind of user's answer to the first i item of information Tˉ i The real-time collection quality correction factor γ of the first i item of information is determined i .

5. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 1, characterized in that, The method further comprises the following steps: Obtaining user profile tag quantity k ; acquiring weight information of the first tag t β t ;​ acquiring a first t influence coefficient of a second i item of information collection σ t,i ; The personalized correction coefficient δ of the item information is calculated according to the following manner i u, i :​ 。 6. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 1, characterized in that, The adaptive interaction mechanism comprises dynamically adjusting the questioning method, language style, and question order according to the interaction adaptation degree, and the interaction adaptation degree is calculated by the following method: ； wherein, Q is the interaction fitness; α auto is the interaction fitness weight; Sim'(U, M) is the user-interaction mode weighted cosine similarity; S auto is the question order rationality score; λ is the real-time feedback correction factor.

7. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 6, characterized in that, The method further comprises the following steps: Obtaining the mean value of the contribution of the question order adaptation in the history database to the overall adaptation degree Qˉ seq ; Obtaining a mean value of a contribution of an interaction mode adaptation to an overall adaptation degree in a history database Qˉ mode ; Contribution mean of the problem order adaptation to the overall adaptation degree Qˉ seq Contribution mean of the interaction mode adaptation to the overall adaptation degree Qˉ mode Determining the interaction adaptation degree weight α auto .

8. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 6, characterized in that, The method further comprises the following steps: A plurality of interaction parameters are preset, which include age, cognitive level, and device type; assigning different weight values to the interaction parameters based on a dimension influence degree ω d ; Obtaining standardized values of user features U d ; Standardized values of interaction mode features are obtained M d ; The user-interactive method weighted cosine similarity Sim'(U, M) is calculated according to the following method: ； wherein, H u is a user history interaction acceptance modifier.

9. The rehabilitation assistive device evaluation method based on multi-agent and RAG technology according to claim 6, characterized in that, The method further comprises the following steps: Acquiring information collection efficiency of current problem order E seq ; Obtaining average information collection efficiency of same kind of users in history database Eˉ seq ; Predicting a user's current cognitive load based on response length and error rate L u ; Setting a cognitive load threshold for a same type of user based on historical data Lˉ max ; information collection efficiency according to current problem order E seq average information collection efficiency of similar users Eˉ seq current cognitive load L u cognitive load threshold Lˉ max determining problem order reasonableness score S auto .

Images