Intelligent matching system for multi-disciplinary collaborative remote consultation experts based on multi-dimensional profiles

Abstract

This invention discloses an intelligent matching system for multi-disciplinary collaborative remote consultation experts based on multi-dimensional profiling, belonging to the field of medical remote consultation technology. This invention solves the problem that existing systems struggle to accurately construct expert profiles, leading to matched experts that may not meet actual consultation needs. By employing a medical knowledge graph, it quickly identifies departments and experts related to the current case, comprehensively considers experts' time preferences and calendar information, and automatically recommends multiple optimal consultation time slots. It uses a neural network model to comprehensively analyze expert profiles, generating a personalized list of recommended consultation experts. Furthermore, it flexibly adjusts the system based on actual circumstances to ensure timely and efficient consultations. This fully leverages the advantages of multidisciplinary collaboration, improving diagnostic accuracy and treatment outcomes, and achieving efficient and accurate matching of consultation experts as well as optimized consultation scheduling.

Description

Technical Field

[0001] This invention relates to the field of medical telemedicine technology, specifically to a multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling. Background Technology

[0002] In modern healthcare systems, multidisciplinary remote consultations are crucial for the diagnosis and treatment of complex diseases. However, current expert matching systems suffer from several problems, such as: firstly, the difficulty in comprehensively and accurately constructing expert profiles, resulting in matched experts who may not meet actual consultation needs; secondly, insufficient in-depth exploration and utilization of medical knowledge, failing to effectively construct complex networks of relationships between diseases, symptoms, departments, and experts, thus limiting the improvement of consultation quality.

[0003] Therefore, not content with existing needs, we propose a multi-disciplinary collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent matching system for multi-disciplinary collaborative remote consultation experts based on multi-dimensional profiles. It rapidly identifies departments and experts related to the current case using a medical knowledge graph, and automatically recommends multiple optimal consultation time slots by comprehensively considering experts' time preferences and calendar information. A neural network model is used to comprehensively analyze expert profiles and generate a personalized list of recommended consultation experts. The system is flexibly adjusted according to actual circumstances to ensure timely and efficient consultations. It fully leverages the advantages of multidisciplinary collaboration, improves diagnostic accuracy and treatment effectiveness, achieves efficient and accurate matching of consultation experts, and optimizes consultation time scheduling, thus solving the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multi-disciplinary collaborative remote consultation expert intelligent matching system based on multi-dimensional profiles includes: The multi-dimensional profile building unit is configured to construct a multi-dimensional profile consisting of expert profiles, disease profiles, department profiles, and patient profiles based on preprocessed multi-dimensional data. The knowledge graph construction unit is configured to extract entity relationships from multi-dimensional data based on natural language processing technology, including: expert information, disease type, department information and patient information, and to construct a medical knowledge graph by combining multi-dimensional profiles; The intelligent matching unit is configured to incorporate quantum simulation and neural network technology, transforming the expert matching problem into a multi-objective optimization problem through a neural network model, and retrieving the final recommended expert combination that satisfies the multi-objective optimization from the knowledge graph; The consultation recommendation unit is configured to generate a list of recommended consultation experts and optimal consultation time slots based on quantum simulation results; and display them to medical staff through an interactive interface, allowing users to provide feedback and make adjustments to the recommendation results.

[0006] Furthermore, the intelligent matching unit includes: The collaborative association module is configured to identify complications, comorbidities, and potential departments associated with related symptoms in the current case based on a medical knowledge graph. By leveraging the relationships between diseases, symptoms, and departments in the atlas, multiple departments related to the current case can be quickly located. Calculate the correlation coefficients between the symptoms and disease type of the current case and various departments to ensure that the identified departments are highly correlated with the cases and reduce misjudgments; Based on the historical collaboration network in the expert profile, experts from departments with close collaborative relationships with core experts are recommended to form an initial set of candidate experts; The collaboration strength between candidate experts and core experts was calculated using the correlation coefficient statistical method. Define a binary decision variable for each candidate expert to indicate whether the expert is selected into the final consultation team, and construct a Hamiltonian including a professional matching term, a collaboration intensity term, and a penalty term. By optimizing the algorithm to perform selection, crossover, and mutation operations on the candidate expert set, the optimal solution that minimizes the objective function is quickly searched, and the corresponding optimal expert combination is obtained.

[0007] Furthermore, the intelligent matching unit also includes: The time optimization module is configured to comprehensively analyze the time preferences and calendar information of all candidate experts, transforming the consultation scheduling problem into a multi-objective optimization problem. By introducing consultation quality objectives, a neural network model is used to seek the optimal balance in a multi-objective optimization problem and automatically recommend multiple best consultation time periods. The time allocation module is configured to use a probability allocation method based on neural networks to assign probability values ​​to each candidate consultation time period according to the expert's time preferences and calendar information. A dynamic adjustment mechanism is introduced to update the probability distribution of each time period in real time. Based on the expert's schedule, the urgency of the consultation, and the complexity of the patient's condition, the consultation time period is flexibly adjusted.

[0008] Furthermore, the intelligent matching unit also includes: The dynamic collaboration prediction module is configured to incorporate a deep learning model, combining historical and real-time collaboration data from experts to predict potential collaboration relationships between experts in the future.

[0009] Furthermore, the consultation recommendation unit includes: The personalized recommendation module is configured to comprehensively analyze expert profiles through a consultation expert recommendation model, identify the unique advantages of experts in specific disease areas, as well as the complementary nature of collaboration with experts, and generate a personalized consultation expert recommendation list based on the patient's specific condition and consultation needs. The dynamic adjustment module is configured to monitor the preparation status of the consultation and the real-time status of the experts after the recommendation list is generated; if an expert is temporarily unable to participate in the consultation, a suitable expert will be called from the reserve expert pool and the consultation time will be readjusted to ensure that the consultation is carried out on time. The composition of the expert team is dynamically adjusted based on changes in the patient's condition or new needs that arise during the consultation process, and the consultation plan is optimized in real time.

[0010] Furthermore, the multi-dimensional profile construction unit includes: The profile dynamic update module is configured to periodically conduct a comprehensive evaluation of multi-dimensional profiles, and dynamically adjust expert ability scores, disease complexity scores, departmental collaboration scores, and patient satisfaction scores based on real-time feedback data, thereby driving the optimization and improvement of multi-dimensional profiles.

[0011] Furthermore, the knowledge graph construction unit includes: The graph construction module is configured to extract key medical entities from text data such as medical records, papers, and consultation opinions, and identify the relationships between entities; it also uses semantic parsing technology to obtain complex semantic relationships in the text, ensuring that the extracted entity relationships are accurate. By treating experts, diseases, departments, and patients as entity nodes, and extracting entity relationships, diseases and symptoms are linked together through association relationships. Based on the department visited, link the disease with the department; based on affiliation, link the expert with the department; Based on the experts' areas of expertise and consultation experience, the experts are associated with the diseases; through historical collaboration records, the collaborative relationships between experts are linked. By linking diseases with symptoms, departments, and experts in multiple dimensions, and by linking experts with departments and the collaborative relationships between experts in multiple dimensions, a medical knowledge graph is formed. This knowledge graph is then applied to matching consultation experts and optimizing consultation plans.

[0012] Furthermore, the knowledge graph construction unit also includes: The knowledge graph application module is configured to identify departmental experts with close collaborative relationships with core experts by leveraging the relationships between experts and diseases, and between experts and departments in the knowledge graph, thereby forming an efficient multidisciplinary consultation team and accurately matching consultation experts. By leveraging the correlation between diseases and treatment options in the knowledge graph, the optimal treatment plan is recommended; and by combining the disease prognosis data in the knowledge graph, the patient's prognosis is assessed. Based on the feedback data after the consultation, the relationships and entity information in the knowledge graph are dynamically adjusted.

[0013] Furthermore, the multi-dimensional profile construction unit includes: The acquisition module is used to acquire preprocessed multi-dimensional data, which includes patient data, disease data, department data, and expert data. The first construction module is used to construct a four-dimensional association matrix based on the three-dimensional classification of the multi-dimensional data into a subject-attribute-association matrix, wherein the matrix elements of the four-dimensional association matrix are the association strength values ​​between subjects; The generation module is used to generate a three-layer label system for each profile in the multi-dimensional profile, based on the four-dimensional association matrix and multi-dimensional data. The multi-dimensional profile includes expert profiles, disease profiles, department profiles, and patient profiles. The inherent labels are extracted from the main basic data through keyword matching and mapping with a medical terminology dictionary. The association labels are generated based on the association strength of the four-dimensional association matrix through threshold filtering and priority sorting. The dynamic labels are captured based on real-time data triggers and the weight of each label is calculated. The second construction module is used to build a closed-loop verification chain consisting of patient profile labels, disease profile verification, department profile verification, expert profile verification, and feedback correction of patient profile labels. It performs data consistency verification, business rationality verification, and correlation matching degree verification on the three-layer labels. It handles verification conflict labels through a three-source correction mechanism of data-level review, business-level rule verification, and correlation-level strength adjustment. The fusion module is used to fuse tags in the order of inherent tag layer, associated tag layer, and dynamic tag layer. The inherent tag layer retains core tags and important tags and sorts them according to the priority of the main attributes. The associated tag layer filters strongly associated tags. The dynamic tag layer is updated in real time and displayed in association with inherent tags and associated tags. It generates a structured output of three-layer features, association graph, and business adaptation report for each profile, resulting in expert profile, disease profile, department profile, and patient profile. The three-layer features include a weighted set of inherent tags, a set of associated tags with association strength, and a set of dynamic tags with update time.

[0014] Furthermore, the construction method of the expert recommendation model includes: Obtain a training dataset, which includes patient-side data, expert-side data, and patient-expert matching labels. The patient-side data includes textual data of medical conditions, structured data of medical conditions, and data on consultation needs. The expert-side data includes expert profile data, capability data, and dynamic performance data. Multimodal feature engineering is performed on the training dataset to obtain the patient global representation and the expert global representation; The patient global representation and the expert global representation are concatenated into a 512-dimensional vector, which is then input into a 3-layer MLP encoder containing a medical domain gating unit. The output is the mean and variance of the joint distribution. The joint distribution is sampled using a standard reparameterization technique to generate a joint latent vector. The joint latent vector and the expert domain attention vector are then input into a 2-layer MLP decoder to output the matching probability of the patient to the expert, thus obtaining a domain-adaptive multimodal variational autoencoder. Construct a three-dimensional loss function; The domain-adaptive multimodal variational autoencoder is dynamically optimized and trained in stages to obtain an initial consultation expert recommendation model; the stages include a pre-training stage, a fine-tuning stage, and a collaborative optimization stage. Obtain a test dataset, which includes in-distribution samples and out-of-distribution cold-start samples; test the initial consultation expert recommendation model based on the test dataset. When the in-distribution matching accuracy is >80% and the cold-start scenario recommendation accuracy is >70%, the trained consultation expert recommendation model is obtained.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes a medical knowledge graph to quickly identify departments and experts relevant to the current case. Taking into account experts' time preferences and calendar information, it automatically recommends multiple optimal consultation time slots. A neural network model comprehensively analyzes expert profiles, identifying experts' unique strengths in specific disease areas and their complementary collaborations. Combined with the patient's specific condition and consultation needs, a personalized list of recommended consultation experts is generated. This list is flexibly adjusted based on actual circumstances to ensure timely and efficient consultations. This effectively solves the problems of inaccurate expert matching and unreasonable time scheduling in traditional consultation systems, providing an intelligent and personalized solution for multidisciplinary consultations of complex diseases. This improves the level of medical services and enhances the patient's medical experience. Especially in dealing with difficult and rare diseases, it fully leverages the advantages of multidisciplinary collaboration, improving diagnostic accuracy and treatment effectiveness. It achieves efficient and accurate matching of consultation experts and optimized consultation time scheduling, significantly improving the efficiency and quality of multidisciplinary collaborative remote consultations. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling, as described in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To address the technical challenges of existing systems in comprehensively and accurately constructing expert profiles, which may result in matched experts failing to meet actual consultation needs; and the insufficient in-depth mining and utilization of medical knowledge, hindering the effective construction of complex relationship networks between diseases, symptoms, departments, and experts, thus limiting the improvement of consultation quality, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution: A multi-disciplinary collaborative remote consultation expert intelligent matching system based on multi-dimensional profiles includes: The multi-dimensional profile building unit is configured to construct a multi-dimensional profile based on preprocessed multi-dimensional data, consisting of expert profiles, disease profiles, department profiles, and patient profiles. Expert profiles include, but are not limited to, information on professional fields, diseases of expertise, consultation experience, academic achievements, and collaboration history. Disease profiles include, but are not limited to, basic disease information, symptoms, possible complications, comorbidities, and common departments. Department profiles include, but are not limited to, the department's professional focus, expert team composition, and inter-departmental collaboration. Patient profiles include, but are not limited to, the patient's medical records, symptom descriptions, past medical history, and examination and test results, forming a comprehensive patient characteristic description for accurate matching of consultation experts. The multi-dimensional profile building unit includes: The profile dynamic update module is configured to periodically conduct a comprehensive evaluation of multi-dimensional profiles, combining real-time feedback data to dynamically adjust expert competence scores, disease complexity scores, departmental collaboration scores, and patient satisfaction scores, driving the optimization and improvement of multi-dimensional profiles. For example, if an expert proposes an innovative treatment plan in a consultation and it is recognized by peers, their innovation score will increase accordingly; if a disease develops new complications recently, its complexity score will increase accordingly; if a department performs well in a recent multidisciplinary consultation, its collaboration score will increase accordingly; and if a patient has high satisfaction with a consultation, their satisfaction score will increase accordingly.

[0019] The knowledge graph construction unit is configured to extract entity relationships from multi-dimensional data using natural language processing technology, including expert information, disease types, department information, and patient information. Combined with multi-dimensional profiles, a medical knowledge graph is constructed. In the knowledge graph, diseases and symptoms are linked through associations; diseases and departments are associated based on commonly visited departments; experts and departments are associated through affiliation; and experts and diseases are associated based on their areas of expertise and consultation experience. Collaborative relationships among experts are also considered, forming a complex and rich network of associations to provide knowledge support for matching consultation experts. The knowledge graph construction unit includes: The graph construction module is configured to extract key medical entities from textual data such as medical records, papers, and consultation opinions. These entities include disease names, symptom descriptions, examination results, expert names, and department names. It identifies relationships between entities, such as the association between disease and symptoms, disease and complications, the relationship between an expert and their area of ​​expertise, and collaborative relationships between experts. Semantic parsing technology is used to obtain complex semantic relationships within the text, ensuring the accuracy of the extracted entity relationships. Experts, diseases, departments, and patients are treated as entity nodes. Expert nodes include: basic expert information, professional fields, diseases of expertise, consultation experience, academic achievements, and collaborative history. Disease nodes include: basic disease information, symptom presentation, possible complications, comorbidities, and common departments. Department nodes include: the department's professional direction, expert team composition, and collaborative relationships between departments. Patient nodes include: the patient's medical record information and symptoms. The system incorporates descriptions, past medical history, and examination / test results. It extracts entity relationships to connect diseases and symptoms through associations; for example, diabetes is associated with symptoms such as polydipsia, polyphagia, and polyuria. It also associates diseases with departments based on the department visited; for example, cardiology is associated with cardiology. Furthermore, it links experts with departments based on affiliation; for example, an expert belongs to cardiology. Experts are also linked to diseases based on their areas of expertise and consultation experience; for example, an expert has extensive consultation experience in diabetes. Historical collaboration records link collaboration relationships between experts; for example, two experts frequently collaborate in multidisciplinary consultations. This multi-dimensional association of diseases with symptoms, departments, and experts, as well as multi-dimensional association of experts with departments and collaboration relationships, forms a medical knowledge graph. This knowledge graph is then applied to matching consultation experts and optimizing consultation plans.

[0020] The knowledge graph application module is configured to identify departmental experts with close collaborative relationships with core experts by leveraging the relationships between experts and diseases, and experts and departments within the knowledge graph, thereby forming an efficient multidisciplinary consultation team and accurately matching consultation experts; recommend the optimal treatment plan by leveraging the relationships between diseases and treatment plans within the knowledge graph; assess the patient's prognosis by combining disease prognosis data from the knowledge graph; and dynamically adjust the relationships and entity information within the knowledge graph based on feedback data after consultation.

[0021] The intelligent matching unit is configured to incorporate quantum simulation and neural network technologies. It transforms the expert matching problem into a multi-objective optimization problem using a neural network model, and retrieves the final recommended expert combination that satisfies the multi-objective optimization from the knowledge graph. The intelligent matching unit includes: The collaboration and association module is configured based on a medical knowledge graph to identify complications, comorbidities, and potential departments associated with related symptoms in the current case. Through the relationships between diseases, symptoms, and departments in the graph, it quickly locates multiple departments related to the current case, providing comprehensive departmental coverage for multidisciplinary consultations. It calculates the correlation coefficients between the current case's symptoms, disease type, and various departments to ensure a high degree of correlation between the identified departments and the case, reducing misjudgments. For example, for a common cardiovascular disease combined with diabetes, the correlation coefficient is calculated to prioritize cardiology and endocrinology as associated departments. Based on the historical collaboration network in the expert profile, it recommends experts from departments with close collaborative relationships with core experts, forming an initial set of candidate experts. For example, by analyzing data such as collaboration frequency and outcomes in historical consultation records, it identifies experts who have excellent cooperation and collaborative effects with core experts in multidisciplinary consultations.

[0022] The collaboration strength between candidate experts and core experts is calculated using correlation coefficient statistics. The assessment of collaboration strength considers not only the frequency of collaboration but also the quality of collaborative outcomes, such as the accuracy of the consultation diagnosis and the effectiveness of the treatment plan, to ensure that the recommended experts are highly matched with the core experts in terms of collaboration ability, thereby improving the overall collaboration efficiency of the consultation team. A binary decision variable is defined for each candidate expert to indicate whether the expert is selected for the final consultation team, and a Hamiltonian is constructed, including a professional matching term, a collaboration strength term, and a penalty term. The professional matching term is measured by calculating the correlation coefficient between the candidate expert's professional field and the case requirements. The collaboration strength term is based on the collaboration strength assessment results between the candidate expert and the core expert. The penalty term is used to handle constraints such as expert time conflicts and consultation costs. An optimization algorithm performs selection, crossover, and mutation operations on the candidate expert set to quickly search for the optimal solution that minimizes the objective function, thus obtaining the corresponding optimal expert combination.

[0023] The time optimization module is configured to comprehensively analyze the time preferences and calendar information of all candidate experts, transforming the consultation scheduling problem into a multi-objective optimization problem; it introduces the consultation quality objective, and seeks the optimal balance in the multi-objective optimization problem through a neural network model, automatically recommending multiple best consultation time slots.

[0024] The time allocation module is configured to use a probability allocation method based on neural networks. Based on expert time preferences and calendar information, it assigns a probability value to each candidate consultation time slot. This probability value reflects the likelihood of that time slot being selected as the final consultation time. A dynamic adjustment mechanism is introduced to update the probability distribution of each time slot in real time. Consultation time slots are flexibly adjusted according to expert schedules, the urgency of the consultation, and the complexity of the patient's condition. For example, for critically ill patients, the probability distribution of consultation time slots will tilt towards earlier time slots to ensure that patients can receive multidisciplinary expert consultations as quickly as possible. For patients with relatively stable conditions, the probability distribution of consultation time slots will be more balanced, taking into account both expert schedules and consultation quality, thereby achieving personalized and precise consultation time recommendations.

[0025] The dynamic collaboration prediction module is configured to incorporate a deep learning model to analyze historical and real-time collaboration data of experts and predict potential collaboration relationships among them in the future. For example, when a new rare disease emerges, even if there are no current collaboration records among relevant experts, by analyzing the activity trajectories of experts in similar diseases or related research fields, it can predict in advance which expert combinations may need to collaborate, providing forward-looking support for responding to public health emergencies or rare disease consultations, and further enriching the initial candidate expert set.

[0026] The consultation recommendation unit is configured to generate a list of recommended consultation experts and optimal consultation time slots based on quantum simulation results; this is then displayed to medical staff through an interactive interface, allowing users to provide feedback and adjustments to the recommendations. The consultation recommendation unit includes: The personalized recommendation module is configured to comprehensively analyze expert profiles through a consultation expert recommendation model, identify the unique strengths of experts in specific disease areas, and the complementary nature of collaboration with experts. Combining the patient's specific condition and consultation needs, it generates a personalized consultation expert recommendation list, ensuring that the recommended expert teams achieve optimal professional capabilities and collaborative efficiency.

[0027] The dynamic adjustment module is configured to monitor the preparation status and real-time status of experts for consultations after the recommendation list is generated. If an expert is temporarily unable to participate in the consultation, a suitable expert will be called from the reserve expert pool, and the consultation schedule will be readjusted to ensure that the consultation takes place on time. For example, a reserve expert pool is pre-established, which includes experts with extensive experience in specific disease areas who have not been recommended. When an expert is temporarily unable to participate in the consultation, the system can quickly call a suitable expert from the reserve expert pool. Furthermore, the module dynamically adjusts the composition of the expert team based on changes in the patient's condition or new needs that arise during the consultation, optimizing the consultation plan in real time. For example, if the patient has multiple complications, the recommended expert team will include experts with extensive experience in these complications.

[0028] The beneficial effects achieved by the above are as follows: By using a medical knowledge graph, departments and experts related to the current case are quickly identified; considering the experts' time preferences and calendar information, multiple optimal consultation time slots are automatically recommended; through a neural network model, the expert profile is comprehensively analyzed to identify the unique advantages of experts in specific disease areas and the complementary nature of collaboration with experts; combined with the patient's specific condition and consultation needs, a personalized list of recommended consultation experts is generated; and adjustments are made flexibly according to the actual situation to ensure that consultations can be conducted on time and efficiently; this effectively solves the problems of inaccurate expert matching and unreasonable time scheduling in traditional consultation systems, providing an intelligent and personalized solution for multidisciplinary consultations of complex diseases, thereby improving the level of medical services and the patient's medical experience. Especially when dealing with difficult and rare diseases, it can give full play to the advantages of multidisciplinary collaboration, improve diagnostic accuracy and treatment effects, achieve efficient and accurate matching of consultation experts and optimized scheduling of consultation time, and significantly improve the efficiency and quality of multidisciplinary collaborative remote consultations.

[0029] Working principle: Multi-dimensional profiles and medical knowledge graphs are constructed using preprocessed multi-dimensional data. Based on the knowledge graph, departments and experts related to cases are identified, correlation coefficients are calculated to ensure high relevance, and an initial set of candidate experts is formed by combining historical collaboration networks. An optimization algorithm searches for the optimal combination among these candidates. Expert time preferences and calendar information are comprehensively analyzed, transforming the scheduling problem into a multi-objective optimization problem. A neural network model seeks the optimal balance, automatically recommending the best consultation time slots and updating the probability distribution of each time slot in real time, flexibly adjusting consultation times. A personalized consultation expert recommendation list is generated based on the neural network model, and the consultation preparation status and expert status are monitored in real time. Candidate experts are called in as needed, and consultation times are adjusted. Simultaneously, the composition of the expert team is dynamically adjusted based on changes in the patient's condition or new consultation needs, optimizing the consultation plan.

[0030] The multi-dimensional profile building unit includes: The acquisition module is used to acquire preprocessed multi-dimensional data, which includes patient data, disease data, department data, and expert data. The first construction module is used to construct a four-dimensional association matrix based on the three-dimensional classification of the multi-dimensional data into a subject-attribute-association matrix, wherein the matrix elements of the four-dimensional association matrix are the association strength values ​​between subjects; The generation module is used to generate a three-layer label system for each profile in the multi-dimensional profile, based on the four-dimensional association matrix and multi-dimensional data. The multi-dimensional profile includes expert profiles, disease profiles, department profiles, and patient profiles. The inherent labels are extracted from the main basic data through keyword matching and mapping with a medical terminology dictionary. The association labels are generated based on the association strength of the four-dimensional association matrix through threshold filtering and priority sorting. The dynamic labels are captured based on real-time data triggers and the weight of each label is calculated. The second construction module is used to build a closed-loop verification chain consisting of patient profile labels, disease profile verification, department profile verification, expert profile verification, and feedback correction of patient profile labels. It performs data consistency verification, business rationality verification, and correlation matching degree verification on the three-layer labels. It handles verification conflict labels through a three-source correction mechanism of data-level review, business-level rule verification, and correlation-level strength adjustment. The fusion module is used to fuse tags in the order of inherent tag layer, associated tag layer, and dynamic tag layer. The inherent tag layer retains core tags and important tags and sorts them according to the priority of the main attributes. The associated tag layer filters strongly associated tags. The dynamic tag layer is updated in real time and displayed in association with inherent tags and associated tags. It generates a structured output of three-layer features, association graph, and business adaptation report for each profile, resulting in expert profile, disease profile, department profile, and patient profile. The three-layer features include a weighted set of inherent tags, a set of associated tags with association strength, and a set of dynamic tags with update time.

[0031] In this embodiment, the patient data includes electronic medical records, vital sign data, and medical behavior data; the disease data includes pathological data, treatment guideline data, and complication statistics; the department data includes treatment scope data, equipment list data, and scheduling data; and the expert data includes qualification records, treatment case data, and areas of expertise data.

[0032] In this embodiment, the multi-dimensional data is standardized. Based on the ICD-10 disease coding and medical terminology dictionary, the data terminology is unified. The maximum and minimum value normalization formula is used to unify the data dimensions. The associated data is quantified into a 0-1 range value according to the matching strength.

[0033] In this embodiment, the multi-dimensional data is classified into a three-dimensional category of subject-attribute-association to construct a four-dimensional association matrix. The matrix elements of the four-dimensional association matrix are the association strength values ​​between subjects, including: Based on medical business rules, a cross-association matrix is ​​constructed, where the matrix elements represent the correlation strength values ​​between entities. The correlation strength is calculated as follows: ;in Here, m represents the association strength value, M represents the total number of associated features, K represents the medical business fit coefficient, and D represents the data credibility. The business fit coefficient K is set based on medical business priorities (Patient-Disease K=0.95, Disease-Department K=0.9, Department-Expert K=0.85, Expert-Patient K=0.8, Patient-Department K=0.75, Disease-Expert K=0.7). The data credibility D is calculated as follows: structured data (e.g., examination reports) D=0.95, text data (e.g., medical record descriptions) D=0.8, and time-series data (e.g., vital sign trends) D=0.75. The matrix output yields a patient-disease-department-expert association matrix. For example, the association strength between Patient A and Disease B is 0.88, Disease B and Department C is 0.92, Department C and Expert D is 0.86, and Expert D and Patient A are 0.78.

[0034] In this embodiment, based on the four-dimensional correlation matrix and multi-dimensional data, a three-layer label system is generated for each image in the multi-dimensional profile: inherent label, associated label, and dynamic label, as shown in Table 1.

[0035] Inherent label extraction: Based on the main body's basic data, it is extracted through keyword matching and medical terminology dictionary mapping. For example: a patient's 5-year history of hypertension in electronic medical records → inherent label: basic disease: hypertension; an expert's research direction in the profile: coronary heart disease → inherent label: research direction: coronary heart disease.

[0036] Association tag generation: Based on the association strength of the association matrix, combined with threshold filtering and priority sorting, the tags are generated. For example: Patient and disease association strength ≥ 0.7 → Association tag disease fit: this disease; Department and disease association strength ≥ 0.8 → Association tag disease coverage: this disease.

[0037] Real-time capture of dynamic tags: triggered by real-time data, for example: a patient's wearable device uploads a heart rate of 120 beats / min → the current vital sign of the dynamic tag is: abnormal heart rate; the department system reports that the number of patients received has reached the upper limit → the resource load of the dynamic tag is: full load.

[0038] Tag weight calculation: The weight of each tag = data credibility × business priority × correlation strength. The weight value is used to determine the priority when merging profiles in the future (weight ≥ 0.7 is a core tag, 0.5-0.7 is an important tag, and < 0.5 is a secondary tag).

[0039] In this embodiment, a closed-loop verification chain is constructed, consisting of patient profile tags, disease profile verification, department profile verification, expert profile verification, and feedback-corrected patient profile tags. Data consistency verification, business rationality verification, and correlation matching degree verification are performed on the three layers of tags. A three-source correction mechanism—data-level review, business-level rule verification, and correlation-level strength adjustment—is used to handle conflicting tags, including: With data consistency, business rationality, and correlation matching as the core verification criteria, each profile's tags, after generation, must undergo collaborative verification with three other profiles to form a closed-loop verification chain: Verification chain design: Patient profile tags → Disease profile verification → Department profile verification → Expert profile verification → Feedback to correct patient profile tags. Data consistency verification: Compare the consistency of associated data in the four-dimensional profile, e.g., Patient profile symptoms: Does chest pain match the core symptom of the disease profile: Chest pain? Expert profile expertise: Does cardiovascular disease match the specialty of the department profile: Cardiovascular disease? Business rationality verification: Judge the rationality of tags based on medical business rules, e.g., Disease profile diagnosis and treatment standards: Does surgical treatment match the core equipment of the department profile: Does the surgical equipment match? Patient profile allergy history: Does penicillin conflict with the appropriate drugs in the disease profile: Does penicillin conflict? Correlation matching verification: Based on the correlation strength of the correlation matrix. Verify the rationality of label associations. For example, for patient profile disease matching: type 2 diabetes and department profile disease coverage: the association strength of type 2 diabetes is ≥0.8. If it is <0.6, label correction is triggered. Three-source correction mechanism: Data level: review the credibility of conflicting data and remove labels corresponding to data with credibility <0.5; Business level: call the diagnosis and treatment guidelines (disease profile), department diagnosis and treatment specifications (department profile), and expert consensus (expert profile) to verify the rules; Association level: adjust the association strength of conflicting labels. If the association strength is still <0.5 after correction, delete the label.

[0040] In this embodiment, tags are fused based on the inherent tag layer, associated tag layer, and dynamic tag layer in that order, including: The fusion order is as follows: inherent tag layer → associated tag layer → dynamic tag layer. Each fusion layer references cross-portrait verification results to ensure the integrity and consistency of the portrait. Inherent tag layer fusion: All core tags (weight ≥ 0.7) and important tags (0.5-0.7) are retained, sorted by subject attribute priority (e.g., patient: underlying disease > allergy history > genetic disease history; department: specialty area > core equipment > medical staff configuration). Associated tag layer fusion: Based on the association strength of the association matrix, strongly associated tags are selected. For example, the patient portrait retains the 3 disease-matching tags with the highest association strength, and the department portrait retains the 3 expert affiliation tags with the highest association strength. Dynamic tag layer fusion: The current dynamic tags are updated in real time and displayed in association with inherent and associated tags. For example, the patient's current physical sign (abnormal blood glucose) is associated with the disease match (type 2 diabetes). Each profile includes three layers of features, a correlation graph, and a business adaptation report to ensure clinical usability: Three layers of features: an inherent set of labels (with weights), a set of related labels (with correlation strength), and a dynamic set of labels (with update time); A correlation graph: a visual representation of the relationships between the profile and the three profiles (node ​​size = label weight, line thickness = correlation strength); Business adaptation report: For example, the patient profile outputs the optimal matching path for disease-department-expert; the department profile outputs optimization suggestions for disease coverage-equipment configuration-expert matching resources.

[0041] The working principle and beneficial effects of the above technical solution are as follows: It acquires multi-dimensional data on patients, diseases, departments, and experts, and constructs a four-dimensional association matrix to record the strength of associations between subjects. This comprehensively characterizes the relevant subject features from multiple perspectives, providing a rich and accurate data foundation for subsequent profile generation. It also constructs a closed-loop verification chain to perform multi-faceted verification on the three-layer labels and handles conflicting labels through a three-source correction mechanism, ensuring the consistency, business rationality, and association matching degree of the profile data, thus improving the quality and reliability of the profile. Finally, it merges labels in the order of inherent, associated, and dynamic label layers, filters key labels, updates dynamic labels in real time, and displays them in association, ultimately generating a structured output of three-layer features, an association graph, and a business adaptation report. This facilitates understanding and application and effectively supports medical business decisions and services.

[0042] The construction method of the expert consultation recommendation model includes: Obtain a training dataset, which includes patient-side data, expert-side data, and patient-expert matching labels. The patient-side data includes textual data of medical conditions, structured data of medical conditions, and data on consultation needs. The expert-side data includes expert profile data, capability data, and dynamic performance data. Multimodal feature engineering is performed on the training dataset to obtain the patient global representation and the expert global representation; The patient global representation and the expert global representation are concatenated into a 512-dimensional vector, which is then input into a 3-layer MLP encoder containing a medical domain gating unit. The output is the mean and variance of the joint distribution. The joint distribution is sampled using a standard reparameterization technique to generate a joint latent vector. The joint latent vector and the expert domain attention vector are then input into a 2-layer MLP decoder to output the matching probability of the patient to the expert, thus obtaining a domain-adaptive multimodal variational autoencoder. Construct a three-dimensional loss function; The domain-adaptive multimodal variational autoencoder is dynamically optimized and trained in stages to obtain an initial consultation expert recommendation model; the stages include a pre-training stage, a fine-tuning stage, and a collaborative optimization stage. Obtain a test dataset, which includes in-distribution samples and out-of-distribution cold-start samples; test the initial consultation expert recommendation model based on the test dataset. When the in-distribution matching accuracy is >80% and the cold-start scenario recommendation accuracy is >70%, the trained consultation expert recommendation model is obtained.

[0043] In this embodiment, the training dataset undergoes multimodal feature engineering to obtain global patient and expert representations, including: dimensional alignment processing; textual feature representation: the BioBERT medical pre-trained model is used to encode the textual data of the patient's condition and the description of the expert's area of ​​expertise, generating a 768-dimensional vector; this vector is compressed to 256 dimensions through a linear projection layer and then normalized to eliminate distribution differences; structured data representation: discrete features (condition type, expert's area of ​​expertise) are transformed into dense vectors through a 64-dimensional embedding layer, and then mapped to 256 dimensions through a fully connected layer; continuation features (quantified values ​​of condition severity levels 1-5, expert's years of experience) After Min-Max normalization, it is expanded through a 32-dimensional embedding layer and mapped to 256 dimensions through a fully connected layer; Layered needs / capabilities representation: Patient needs layered: core needs (disease type matching) → 64-dimensional representation → projected to 256 dimensions, secondary needs (professional title / region) → 32-dimensional representation → projected to 256 dimensions, additional needs (response speed) → 16-dimensional representation → projected to 256 dimensions; Family capabilities layered: core capabilities (professional matching) → 64-dimensional representation → projected to 256 dimensions, secondary capabilities (consultation experience) → 32-dimensional representation → projected to 256 dimensions, additional capabilities (service quality) → 16-dimensional representation → projected to 256 dimensions.

[0044] Define the core vector in the medical field Generates textual features based on historically highly matched patient-expert pairs, fixed at 256 dimensions; calculates modal attention weights. ,in, For the m-th mode, a 256-dimensional representation is provided. The total number of modes; Attention weights for the m-th modality (text / structured / hierarchical); fusion output: ; The final output is a 256-dimensional global representation of the patient. and 256-dimensional expert global representation ; The cosine similarity function calculates the cosine of the angle between two vectors. This is a core vector for the 256-dimensional medical field. A 256-dimensional global representation of the patient-side fusion; This is a 256-dimensional global representation after expert-side fusion.

[0045] In this embodiment, the reasoning process of the consultation expert recommendation model includes: processing the side data of the patient to be recommended and the side data of the candidate experts through the aforementioned multimodal feature engineering to obtain a global representation of the target patient. Global representation of the target expert Input the trained expert recommendation model and output the basic matching probability. Enhancement layer processing: Cold start scenario determination: newly registered experts (historical consultation volume < 5 cases) or rare diseases (training set percentage < 0.196); Cross-domain collaboration weight adjustment: based on the similarity between the expert's hospital and the patient's hospital. (Based on cosine similarity calculation of historical consultation patterns), adjust the matching score: Cold start dynamic correction: for cold start experts / conditions Make corrections, correction factor Final score ;according to Sort the results in descending order and output a personalized recommendation list of the top 10 experts along with the matching reasons (based on key features derived from attention weights).

[0046] In this embodiment, the expert recommendation model architecture includes: an input layer that receives three core data types: patient condition (text + structured data), consultation needs, and expert profiles; a feature engineering layer that performs dimensional alignment and dynamic weighted fusion on the three data types, outputting a 256-dimensional global representation of the patient / expert; and a core model layer consisting of: an encoder (3-layer MLP (256→512→256 units) + a medical domain gating unit); and a gating unit. , ,in , , The core vector for the medical field is 256-dimensional; the output is the mean of the 512-dimensional latent space. and variance ; It is a 256-dimensional gated vector; : Sigmoid activation function; It is a 256×512 dimensional weight matrix; This is a 512-byte concatenated vector; It is a 256-dimensional bias vector; The 256-dimensional feature vector after gating is used as the encoder input; Reparameterization: , ; It is a 512-dimensional joint latent vector that represents the implicit relationship of patient-expert matching; It is a 512-dimensional mean vector, output by the encoder; It is a 512-dimensional standard deviation vector; It is a 512-dimensional standard normal noise vector; The distribution follows a 512-dimensional standard normal distribution; Decoder: Expert domain attention vector: ,in 256-dimensional representation for the expert's area of ​​expertise. The disease type is represented by 256 dimensions; This is a 256-dimensional expert domain attention vector; The input is a 256×256 dimensional query weight matrix; (512+256=768 dimensions), output matching probability after 2 layers of MLP (768→384→1) ; The probability of matching patient p with expert e; It is a 1×768 dimensional weight vector; It is a 768-dimensional concatenated vector; The scalar bias term is used; the loss optimization layer is trained by constraining the three-dimensional loss function; the enhancement layer includes a cold start generalization module (data augmentation) and a cross-domain collaboration module (dynamic weight adjustment); the output layer generates an expert recommendation list and interpretable matching reasons.

[0047] In this embodiment, the patient-side data preprocessing includes: medical condition text data: cleaning redundant descriptions and using Unified Medical Language System (UMLS) to standardize medical terminology; structured medical condition data: severity is quantified according to clinical guidelines levels 1-5, and treatment history is encoded using multi-hot encoding; consultation requirement data: core requirements (matching medical condition types), secondary requirements (professional title / regional constraints), and additional requirements (response speed threshold).

[0048] In this embodiment, the expert-side data preprocessing includes: expert profile data: professional titles are quantified according to the National Health Commission standards (resident physician = 1, chief physician = 5), and the areas of expertise are coded using ICD-11; ability data: diagnosis and treatment success rate and handling of difficult cases are standardized using Z-score; dynamic performance data: consultation satisfaction in the past 3 months is calculated using an index-weighted moving average.

[0049] In this embodiment, the three-dimensional loss function is constructed, including: ; , , Weighting coefficients; Matching accuracy loss: ;in, To match labels, KL divergence constrains the alignment of patient-expert representation distributions; Here is the KL divergence regularization coefficient; The Gaussian distribution approximation for patient characteristics; The Gaussian distribution approximation for expert representation; two-dimensional fairness loss: ; ; The number of elements in the set; This results in a loss of fairness on the expert side. This results in a loss of fairness on the patient's side. For the patient set assigned to expert e; ; This is a set of experts who can be recommended to patient p; Number of elements in the set; Medical risk constraint loss: ;in RiskScore(e) is the expert's medical risk coefficient (calculated based on professional title / dispute history, ranging from 0.7 to 1.0). Let e ​​be the safety threshold for expert e to patient p; LoadFactor(e) is the expert load factor. In this embodiment, the phased dynamic optimization training includes: pre-training phase: freezing , Only optimization Learning rate 0.001, Adam optimizer, early stopping strategy (verification: no improvement in loss after 10 rounds); Fine-tuning phase: dynamically updated weights. , ; Let be the weight of the loss of the k-th term in the t-th iteration; The learning rate parameter; The partial derivative of the total loss with respect to the weights; the cold start generalization module is triggered every 20 rounds: New expert simulation: For experts with a historical consultation volume of <10, synthetic samples are generated based on the characteristic distribution of experts with the same title in the same hospital (VAE resampling); Rare disease enhancement: For disease types with a proportion of <0.5%, CTGAN is used to generate adversarial network augmentation samples; Collaborative optimization stage: Global weight calculation: Where Sim is the JS divergence of the patient-expert distribution among hospitals; Let i be the global weight of hospital i in the collaborative optimization; Features of the i-th hospital; For reference hospitals; expert personalized weighting updates: , Global and personalized weights are updated every 50 rounds; Convergence determination: validation set. The decline rate is less than 0.1% for five consecutive rounds or the total number of rounds is greater than or equal to 500. Personalized weights updated for expert e; The current personalized weight for expert e; The coefficients are retained for historical weights; Weighting of experts in the same field average value; This is the set of expert weights that are in the same professional field as expert e.

[0050] The working principle and beneficial effects of the above technical solution are as follows: Acquiring a training set containing multi-type patient data and multi-dimensional expert data allows for comprehensive consideration of various factors, providing rich information for accurate recommendations; obtaining global representations of patients and experts through multimodal feature engineering enables more comprehensive capture of data features; employing a 3-layer MLP encoder and a 2-layer MLP decoder with medical domain gating units, combined with standard reparameterization techniques, enables the learning of joint latent vectors for patient-expert matching, outputting matching probabilities and improving recommendation accuracy; dynamically optimizing the model through three stages: pre-training, fine-tuning, and collaborative optimization, gradually improving model performance; and using a test set containing both in-distribution and out-of-distribution cold-start samples for testing ensures model performance in different scenarios, completing training only when certain accuracy requirements are met, guaranteeing the model's reliability and practicality.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or high-voltage switchgear that comprises a list of elements includes not only those elements but also elements not expressly listed, or elements inherent to such process, method, article, or high-voltage switchgear.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling, characterized in that: include: The multi-dimensional profile building unit is configured to construct a multi-dimensional profile consisting of expert profiles, disease profiles, department profiles, and patient profiles based on preprocessed multi-dimensional data. The knowledge graph construction unit is configured to extract entity relationships from multi-dimensional data based on natural language processing technology, including: expert information, disease type, department information and patient information, and to construct a medical knowledge graph by combining multi-dimensional profiles; The intelligent matching unit is configured to incorporate quantum simulation and neural network technology, transforming the expert matching problem into a multi-objective optimization problem through a neural network model, and retrieving the final recommended expert combination that satisfies the multi-objective optimization from the knowledge graph; The consultation recommendation unit is configured to generate a list of recommended consultation experts and the best consultation time slots based on quantum simulation results; and display them to medical staff through an interactive interface, allowing users to provide feedback and make adjustments to the recommendation results; The multi-dimensional profile construction unit includes: The acquisition module is used to acquire preprocessed multi-dimensional data, which includes patient data, disease data, department data, and expert data. The first construction module is used to construct a four-dimensional association matrix based on the three-dimensional classification of the multi-dimensional data into a subject-attribute-association matrix, wherein the matrix elements of the four-dimensional association matrix are the association strength values ​​between subjects; The generation module is used to generate a three-layer label system for each profile in the multi-dimensional profile, based on the four-dimensional association matrix and multi-dimensional data. The multi-dimensional profile includes expert profiles, disease profiles, department profiles, and patient profiles. The inherent labels are extracted from the main basic data through keyword matching and mapping with a medical terminology dictionary. The association labels are generated based on the association strength of the four-dimensional association matrix through threshold filtering and priority sorting. The dynamic labels are captured based on real-time data triggers and the weight of each label is calculated. The second construction module is used to build a closed-loop verification chain consisting of patient profile labels, disease profile verification, department profile verification, expert profile verification, and feedback correction of patient profile labels. It performs data consistency verification, business rationality verification, and correlation matching degree verification on the three-layer labels. It handles verification conflict labels through a three-source correction mechanism of data-level review, business-level rule verification, and correlation-level strength adjustment. The fusion module is used to fuse tags in the order of inherent tag layer, associated tag layer, and dynamic tag layer. The inherent tag layer retains core tags and important tags and sorts them according to the priority of the main attributes. The associated tag layer filters strongly associated tags. The dynamic tag layer is updated in real time and displayed in association with inherent tags and associated tags. It generates a structured output of three-layer features, association graph, and business adaptation report for each profile, resulting in expert profile, disease profile, department profile, and patient profile. The three-layer features include a weighted set of inherent tags, a set of associated tags with association strength, and a set of dynamic tags with update time.

2. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 1, characterized in that, The intelligent matching unit includes: The collaborative association module is configured to identify complications, comorbidities, and potential departments associated with related symptoms in the current case based on a medical knowledge graph. By leveraging the relationships between diseases, symptoms, and departments within the atlas, multiple departments related to the current case can be quickly located. Calculate the correlation coefficients between the symptoms and disease type of the current case and various departments to ensure that the identified departments are highly correlated with the cases and reduce misjudgments; Based on the historical collaboration network in the expert profile, experts from departments with close collaborative relationships with core experts are recommended to form an initial set of candidate experts; The collaboration strength between candidate experts and core experts was calculated using the correlation coefficient statistical method. Define a binary decision variable for each candidate expert to indicate whether the expert is selected into the final consultation team, and construct a Hamiltonian including a professional matching term, a collaboration intensity term, and a penalty term. By optimizing the algorithm to perform selection, crossover, and mutation operations on the candidate expert set, the optimal solution that minimizes the objective function is quickly searched, and the corresponding optimal expert combination is obtained.

3. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 2, characterized in that, The intelligent matching unit further includes: The time optimization module is configured to comprehensively analyze the time preferences and calendar information of all candidate experts, transforming the consultation scheduling problem into a multi-objective optimization problem. By introducing consultation quality objectives, a neural network model is used to seek the optimal balance in a multi-objective optimization problem and automatically recommend multiple best consultation time periods. The time allocation module is configured to use a probability allocation method based on neural networks to assign probability values ​​to each candidate consultation time period according to the expert's time preferences and calendar information. A dynamic adjustment mechanism is introduced to update the probability distribution of each time period in real time. Based on the expert's schedule, the urgency of the consultation, and the complexity of the patient's condition, the consultation time period is flexibly adjusted.

4. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 3, characterized in that, The intelligent matching unit further includes: The dynamic collaboration prediction module is configured to incorporate a deep learning model, combining historical and real-time collaboration data from experts to predict potential collaboration relationships between experts in the future.

5. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 4, characterized in that, The consultation recommendation unit includes: The personalized recommendation module is configured to comprehensively analyze expert profiles through a consultation expert recommendation model, identify the unique advantages of experts in specific disease areas, as well as the complementary nature of collaboration with experts, and generate a personalized consultation expert recommendation list based on the patient's specific condition and consultation needs. The dynamic adjustment module is configured to monitor the preparation status of the consultation and the real-time status of the experts after the recommendation list is generated; if an expert is temporarily unable to participate in the consultation, a suitable expert will be called from the reserve expert pool and the consultation time will be readjusted to ensure that the consultation is carried out on time. The composition of the expert team is dynamically adjusted based on changes in the patient's condition or new needs that arise during the consultation process, and the consultation plan is optimized in real time.

6. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 1, characterized in that, The multi-dimensional profile building unit includes: The profile dynamic update module is configured to periodically conduct a comprehensive evaluation of multi-dimensional profiles, and dynamically adjust expert ability scores, disease complexity scores, departmental collaboration scores, and patient satisfaction scores based on real-time feedback data, thereby driving the optimization and improvement of multi-dimensional profiles.

7. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 6, characterized in that, The knowledge graph construction unit includes: The graph construction module is configured to extract key medical entities from text data such as medical records, papers, and consultation opinions, and identify the relationships between entities; it also uses semantic parsing technology to obtain complex semantic relationships in the text, ensuring that the extracted entity relationships are accurate. By treating experts, diseases, departments, and patients as entity nodes, and extracting entity relationships, diseases and symptoms are linked together through association relationships. Based on the department visited, link the disease with the department; based on affiliation, link the expert with the department; Based on the experts' areas of expertise and consultation experience, the experts are associated with the diseases; through historical collaboration records, the collaborative relationships between experts are linked. By linking diseases with symptoms, departments, and experts in multiple dimensions, and by linking experts with departments and the collaborative relationships between experts in multiple dimensions, a medical knowledge graph is formed. This knowledge graph is then applied to matching consultation experts and optimizing consultation plans.

8. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 7, characterized in that, The knowledge graph construction unit also includes: The knowledge graph application module is configured to identify departmental experts with close collaborative relationships with core experts by leveraging the relationships between experts and diseases, and between experts and departments in the knowledge graph, thereby forming an efficient multidisciplinary consultation team and accurately matching consultation experts. By leveraging the correlation between diseases and treatment options in the knowledge graph, the optimal treatment plan is recommended; and by combining the disease prognosis data in the knowledge graph, the patient's prognosis is assessed. Based on the feedback data after the consultation, the relationships and entity information in the knowledge graph are dynamically adjusted.

9. The multi-departmental collaborative remote consultation expert intelligent matching system based on multi-dimensional profiling according to claim 5, characterized in that, The construction method of the expert consultation recommendation model includes: Obtain a training dataset, which includes patient-side data, expert-side data, and patient-expert matching labels. The patient-side data includes textual data of medical conditions, structured data of medical conditions, and data on consultation needs. The expert-side data includes expert profile data, capability data, and dynamic performance data. Multimodal feature engineering is performed on the training dataset to obtain the patient global representation and the expert global representation; The patient global representation and the expert global representation are concatenated into a 512-dimensional vector, which is then input into a 3-layer MLP encoder containing a medical domain gating unit. The output is the mean and variance of the joint distribution. The joint distribution is sampled using a standard reparameterization technique to generate a joint latent vector. The joint latent vector and the expert domain attention vector are then input into a 2-layer MLP decoder to output the matching probability of the patient to the expert, thus obtaining a domain-adaptive multimodal variational autoencoder. Construct a three-dimensional loss function; The domain-adaptive multimodal variational autoencoder is dynamically optimized and trained in stages to obtain an initial consultation expert recommendation model; the stages include a pre-training stage, a fine-tuning stage, and a collaborative optimization stage. Obtain a test dataset, which includes in-distribution samples and out-of-distribution cold-start samples; test the initial consultation expert recommendation model based on the test dataset. When the in-distribution matching accuracy is >80% and the cold-start scenario recommendation accuracy is >70%, the trained consultation expert recommendation model is obtained.

Images