Construction method and construction system of telemedicine expert recommendation model, expert recommendation method and electronic equipment

Abstract

The invention provides a construction method and a construction system of a telemedicine expert recommendation model, an expert recommendation method and electronic equipment. The construction method comprises the following steps: acquiring sample information; extracting entities and relationships among the entities from the preliminary diagnosis information of the historical patients and the specialist specialist information of the experts; constructing a knowledge graph according to the extracted entities and the relationship between the entities; performing feature calculation on the sample information, and inputting the calculated disease relevancy and multi-granularity relevancy between multiple groups of historical patients and experts and basic features of the experts corresponding to the historical patients into a deep neural network for training to obtain an expert recommendation model; the method has the beneficial effects that more adaptive experts are recommended for patients, the accuracy and robustness of recommendation results are effectively improved, the efficiency and quality of remote consultation are improved, and the problem of uneven medical resource allocation is solved; the method is suitable for the technical field of telemedicine expert recommendation.

Description

Technical Field

[0001] The present application relates to the technical field of telemedicine expert recommendation, and specifically to a method for constructing a telemedicine expert recommendation model, a construction system, an expert recommendation method and an electronic device. Background Art

[0003] Different from the traditional medical service model where patients directly face doctors, telemedicine is a service model where doctors at primary hospitals initiate consultation applications based on demand, and after matching by telemedicine management personnel, doctors at primary hospitals and experts at higher-level hospitals discuss the patient's condition at a specific time and come up with the next treatment plan. Compared with other Internet medical models (such as online medical care, etc.), telemedicine overcomes problems such as unclear description of the patient's condition, and the online communication model between doctors and experts through communication technology is often smoother and more efficient than communication between patients and doctors through text.

[0004] Although telemedicine has developed rapidly, a survey conducted at the National Telemedicine Center found that telemedicine operation and management still faces some problems. The National Telemedicine Center currently has about 1,400 telemedicine consultations per month, an increase of about 27% compared with an average of 1,100 per month in previous years. However, it still uses manual arrangements to assign experts for teleconsultation applications at grassroots hospitals. With the gradual increase in the number of consultations, the disadvantages of low timeliness and lag in manual triage are gradually revealed. On the one hand, due to the increase in the number of consultations, manual arrangement of consultation experts is time-consuming and labor-intensive, which may cause patients to wait too long. On the other hand, the huge amount of information makes it impossible for triage personnel to fully understand the condition of each patient and the information of the experts, and thus cannot match the most appropriate expert consultation for the patient. The expert may also match patients who are not in their field of expertise. Therefore, it is necessary to explore how to improve this situation and thus improve the quality and efficiency of telemedicine. How to further realize the intelligent recommendation of experts and improve the efficiency and quality of consultations are issues that need to be resolved in the field of telemedicine. Summary of the invention

[0005] In order to solve one of the above-mentioned technical defects, the present application provides a method for constructing a remote medical expert recommendation model, a construction system, an expert recommendation method and an electronic device.

[0006] According to a first aspect of the present application, a method for constructing a telemedicine expert recommendation model is provided, comprising:

[0007] Acquiring sample information, the sample information including: basic information and preliminary diagnosis information of multiple historical patients, and basic information, professional expertise information and historical diagnosis information of multiple experts;

[0008] Entities and relationships between entities are extracted from the preliminary diagnosis information of historical patients and the professional expertise information of experts; entities include historical patients, experts, and diseases; relationships between entities include the suffering relationship between historical patients and diseases, and the expert treatment relationship between experts and diseases;

[0009] Build a knowledge graph based on the extracted entities and the relationships between them;

[0010] Perform feature calculation on sample information to obtain the disease correlation and multi-granularity correlation between each historical patient and all experts, as well as the basic features of the experts corresponding to the historical patients;

[0011] The calculated disease correlations and multi-granularity correlations between multiple groups of historical patients and experts, as well as the basic features of experts corresponding to historical patients, are input into a deep neural network for training to obtain an expert recommendation model.

[0012] Preferably, feature calculation is performed on the sample information, specifically including:

[0013] Calculate the disease relevance between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts;

[0014] Calculate the multi-granularity correlation between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the historical diagnosis information of multiple experts;

[0015] The basic features of the expert corresponding to each historical patient are calculated based on the basic information of multiple historical patients and the basic information of the experts corresponding to the historical patients.

[0016] More preferably, the disease correlation between each historical patient and all experts is calculated based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts, specifically including:

[0017] The disease relevance is calculated based on the path between historical patients and experts in the knowledge graph; the calculation formula for disease relevance is:

[0018] Let max r∈D (path ir )=min r∈D )path ir )hour,

[0019] In the formula, represents the disease correlation between the i-th historical patient and the j-th expert, path ij represents the number of all paths between the i-th historical patient and the j-th expert, D represents the number of experts, minr∈D (path ir ) represents the minimum number of paths between D experts and the i-th historical patient, max r∈D (path ir ) represents the maximum number of paths between D experts and the i-th historical patient.

[0020] Preferably, the multi-granularity correlation between each historical patient and all experts is calculated based on the preliminary diagnosis information of multiple historical patients and the historical diagnosis information of multiple experts, specifically including:

[0021] Calculate the coarse-grained correlation, including:

[0022] The topic probability distribution in the historical patients' preliminary diagnosis information and the topic probability distribution in the experts' historical diagnosis information are calculated through the LDA topic model;

[0023] Calculate the JS discrete distance value between the topic probability distribution in the preliminary diagnosis information of each historical patient and the topic probability distribution in all historical diagnosis information of all experts; the calculation formula of the JS discrete distance value is:

[0024]

[0025] Where D JS (P a ‖Q b ) represents the JS discrete distance value between the a-th preliminary diagnosis information and the b-th historical diagnosis information, P a (x) represents the topic probability distribution in the a-th preliminary diagnostic information, Q b (x) represents the probability distribution of the topic in the bth historical diagnostic information, n represents the number of all preliminary diagnostic information, m represents the number of all historical diagnostic information, x represents the value of the random variable, and x is P a (x) and Q b (x) is a specific value on the surface of the y-axis. X represents the set of all possible values, i.e., the sample space of the random variable.

[0026] The JS discrete distance value is converted to obtain the coarse-grained correlation, and the conversion formula is:

[0027]

[0028] In the formula, Sim LDA (P a ,Q b ) represents the coarse-grained correlation between the a-th preliminary diagnosis information and the b-th historical diagnosis information, Sim LDA (P a ,Q b )∈[0,1], h represents the adjustment factor;

[0029] Calculate fine-grained relevance, including:

[0030] The preliminary diagnosis information of historical patients and the historical diagnosis information of experts are respectively input into the Sentence-BERT model, and the initial embedding vector of historical patients and the initial embedding vector of experts are extracted through the Sentence-BERT model;

[0031] Perform average pooling feature extraction on the historical patient initial embedding vector and the expert initial embedding vector respectively to obtain the historical patient feature vector u and the expert feature vector v;

[0032] The historical patient feature vector u and the expert feature vector v are calculated through the loss function, and the calculation formula of the loss function is:

[0033] o=softmax[W t (u,v,|uv|)];

[0034] In the formula, o represents the loss function, W t represents the weight matrix, which is used to map the concatenated feature vector to a new space, and softmax represents the normalized exponential function;

[0035] According to the calculation results of the loss function, the fine-grained feature vectors of the preliminary diagnosis information of historical patients and the historical diagnosis information of experts are obtained; among them, the fine-grained feature vector of the a-th preliminary diagnosis information is: F a =(f1,f2,...,f L ), the fine-grained feature vector of the b-th historical diagnostic information is: F b =(f′1,f′2,...,f′ L ), L represents the feature dimension;

[0036] The fine-grained correlation between the historical patient's preliminary diagnosis information and the expert's historical diagnosis information is calculated using the following formula:

[0037]

[0038] By combining the coarse-grained correlation and the fine-grained correlation, we can obtain the multi-grained correlation between the preliminary diagnosis information of each historical patient and the multiple historical diagnosis information of all experts. The calculation formula is:

[0039]

[0040] In the formula, δ represents the proportion of coarse-grained correlation in multi-grained correlation, and 1-δ represents the proportion of fine-grained correlation in multi-grained correlation;

[0041] The multi-granularity correlation between each historical patient and each expert is calculated using the following formula:

[0042]

[0043] In the formula, represents the multi-granularity correlation between the i-th historical patient and the j-th expert, N j represents the number of historical diagnostic information corresponding to the jth expert. When the jth expert corresponds to the bth piece of historical diagnostic information, Y jb =1, otherwise, Y jb =0.

[0044] Preferably, the basic information of the historical patients includes age and gender; the basic information of the experts includes: professional title, and consultation time and number of consultations within a preset time; the basic characteristics of the experts include: age characteristics, gender characteristics, professional title characteristics and activity characteristics.

[0045] More preferably, the basic features of all experts are calculated based on the basic information of multiple experts and the basic information of corresponding historical patients, specifically including:

[0046] Calculate the age characteristics of all experts, specifically including: classify historical patients into infants, children, young adults, and middle-aged and elderly people according to their ages; obtain the corresponding age characteristics of experts according to the classification of historical patients;

[0047] When the current historical patient was an infant, the age characteristics of the specialist were:

[0048] When the current historical patient is a child, the age characteristics of the specialist are:

[0049] When the current historical patients are young adults, the age characteristics of the experts are:

[0050] When the current historical patients are middle-aged or elderly, the age characteristics of the experts are:

[0051] In the formula, represents the age characteristic value, N j ′(Infant) indicates the number of historical patients in the Infant age group corresponding to the expert, N j ′(Teenager) represents the number of historical patients in the Teenager age group corresponding to the expert, N j ′(Youth) represents the number of historical patients in the Youth age group corresponding to the corresponding expert, N j′(Elder) represents the number of patients in the Elder age group corresponding to the expert, N j ′ represents the number of all historical patients corresponding to the corresponding expert, Infant represents infants aged 0-6 years old, Teenager represents children aged 7-17 years old, Youth represents young adults aged 18-45 years old, and Elder represents middle-aged and elderly people over 46 years old;

[0052] Get the gender characteristics of the corresponding experts according to the gender of historical patients;

[0053] When the current historical patient is female, the gender characteristics are:

[0054] When the current historical patient is male, the gender characteristics are:

[0055] In the formula, Indicates the gender characteristic value, N j "(Female)" indicates the number of female historical patients among the historical patients corresponding to the corresponding expert, N j "(Male)" indicates the number of male historical patients among the historical patients corresponding to the corresponding expert, N j ” indicates the number of historical patients corresponding to the corresponding expert;

[0056] Calculate the professional title characteristics of all experts, specifically including: divide the professional titles of experts into chief physicians, deputy chief physicians and attending physicians; then assign a value of 3 to the chief physicians, 2 to the deputy chief physicians, and 1 to the attending physicians among all experts; perform normalization to obtain professional title characteristics;

[0057] According to the time when the historical patients applied for consultation, the activity characteristics of the corresponding experts were obtained;

[0058] Current historical patients are Week c When submitting a consultation request, the activity characteristics of the expert are:

[0059]

[0060] Where c∈[1,7], t j Indicates the time of the last consultation of the corresponding expert, t j (Week c ) indicates the corresponding expert's most recent c The consultation time, N j”′ (Week c ) represents the total number of consultations by the corresponding expert in week c, and z represents the adjustment factor.

[0061] According to a second aspect of the present application, a remote medical expert recommendation method is provided, comprising:

[0062] Obtain the information to be recommended, including: basic information and preliminary diagnosis information of the patient to be recommended, as well as basic information, professional expertise information and historical diagnosis information of multiple experts;

[0063] Extracting entities and relationships between entities from the preliminary diagnosis information of the patient to be recommended and the professional expertise information of multiple experts; wherein the entities include the patient to be recommended, the expert, and the disease, and the relationships between entities include the suffering relationship between the patient to be recommended and the disease, and the expert-treatment relationship between the expert and the disease;

[0064] Build a knowledge graph based on the extracted entities and the relationships between them;

[0065] Perform feature calculation on the recommended information to obtain the disease correlation and multi-granularity correlation between the recommended patient and multiple experts, as well as the basic features of the experts corresponding to the recommended patient;

[0066] Input the calculated disease correlation and multi-granularity correlation between the patient to be recommended and multiple experts, as well as the basic features of the experts corresponding to the patient to be recommended into the expert recommendation model as described in any of the above items to obtain an expert recommendation set;

[0067] Output the expert recommendation set to the patient to be recommended.

[0068] According to a third aspect of the present application, a system for constructing a telemedicine expert recommendation model is provided, comprising a module for implementing a method for constructing a telemedicine expert recommendation model as described in any of the above items.

[0069] According to a fourth aspect of the present application, a device is provided, including:

[0070] Memory;

[0071] Processor; and

[0072] Computer programs;

[0073] Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the method for constructing a remote medical expert recommendation model as described in any of the above contents.

[0074] According to a fifth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored; the computer program is executed by a processor to implement the method for constructing a remote medical expert recommendation model as described in any of the above contents.

[0075] This application is based on the basic information and preliminary diagnosis information of historical patients, the basic information, professional expertise information and historical diagnosis information of experts for feature calculation, and then the obtained disease relevance, multi-granularity relevance and basic features are input into the deep neural network for training to obtain the expert recommendation model. The preliminary diagnosis information of historical patients in this application is the information given by doctors in grassroots hospitals, which avoids the problems of incomprehensibility and low communication efficiency caused by direct communication between historical patients and experts. Not only the professional information that experts are good at treating is taken into account, but also the basic characteristics of experts are taken into account, which can recommend more suitable experts to patients, and can effectively improve the accuracy and robustness of the recommendation results. Then, doctors in grassroots hospitals conduct remote consultations with suitable experts, which improves the efficiency and quality of remote consultations. It solves the problems of delays and inconveniences caused by the need to manually find suitable doctors offline in the prior art, as well as the problem of uneven distribution of medical resources.

[0076] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the contents indicated in the written description and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0078] Figure 1 A flowchart of a method for constructing a remote medical expert recommendation model provided in this application;

[0079] Figure 2 for Figure 1 A schematic diagram of a flow chart of feature calculation of sample information in the embodiment;

[0080] Figure 3 for Figure 2 A schematic diagram of a flow chart for calculating the multi-granularity correlation between each historical patient and all experts in the embodiment;

[0081] Figure 4 for Figure 3 A schematic diagram of a flow chart for calculating coarse-grained correlation in the embodiment;

[0082] Figure 5 for Figure 3 A schematic diagram of a process for calculating fine-grained correlation in the embodiment;

[0083] Figure 6 It is a line graph of HR values ​​corresponding to different multi-granularity correlations;

[0084] Figure 7 It is a line chart of MRR values ​​corresponding to different multi-granularity correlations;

[0085] Figure 8 It is a line graph of PRE values ​​corresponding to different multi-granularity correlations;

[0086] Fig. 9 It is the line graph of HR value corresponding to different models;

[0087] Fig.10 It is a line chart of MRR values ​​corresponding to different models;

[0088] Fig.11 It is the line chart of PRE values ​​corresponding to different models;

[0089] Fig.12 Recommend accuracy result charts for each department;

[0090] Fig.13 This is the result diagram of the influence of feature 1 to feature 6 on the recommendation results;

[0091] Fig.14 A flowchart of a remote medical expert recommendation method provided for this application;

[0092] Fig.15 A partial flow chart of a remote medical expert recommendation method provided for this application;

[0093] Fig.16 A functional structure diagram of a system for building a remote medical expert recommendation model provided in this application;

[0094] Fig.17 for Fig.16 Schematic diagram of the functional structure of the feature calculation module;

[0095] Fig.18 for Fig.17 Schematic diagram of the functional structure of the multi-granularity correlation calculation unit;

[0096] Fig.19 for Fig.17 Schematic diagram of the functional structure of the basic feature calculation unit;

[0097] In the figure:

[0098] 10 is an acquisition module, 20 is an extraction model, 30 is a knowledge graph construction module, 40 is a feature calculation module, 50 is a training module, 401 is a disease correlation calculation unit, 402 is a multi-granularity correlation calculation unit, 403 is a basic feature calculation unit, 4021 is a coarse-grained correlation calculation unit, 4022 is a fine-grained correlation calculation unit, 4023 is a fusion unit, 4024 is a second calculation unit, 4031 is an age feature calculation unit, and 4032 is a gender feature calculation unit. Element, 4033 is a job title feature calculation unit, 4034 is an activity calculation unit, 40211 is a distribution probability calculation unit, 40212 is a discrete distance value calculation unit, 40213 is a conversion unit, 40221 is an initial embedding vector calculation unit, 40222 is a feature extraction unit, 40223 is a loss function calculation unit, 40224 is a fine-grained feature acquisition unit, 40225 is a first calculation unit, 40311 is a classification unit, and 40312 is a third calculation unit. DETAILED DESCRIPTION

[0099] In order to make the technical solutions and advantages in the embodiments of the present application more clearly understood, the exemplary embodiments of the present application are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0100] This application uses the telemedicine consultation data of the National Telemedicine Center from 2019 to 2021 and the professional expertise information of experts crawled from the official website of the First Affiliated Hospital of Zhengzhou University as the data set. Telemedicine consultation data contains the basic information and preliminary diagnosis information of historical patients, the basic information, professional expertise information and historical diagnosis information of experts. Among them, the basic information of historical patients includes name, gender, age, and time of consultation application; the basic information of experts includes consultation department, expert name, and expert title; historical diagnosis information includes consultation diagnosis, consultation time and consultation opinion. Combined with the actual situation of remote consultation, more active experts with a consultation volume greater than or equal to 20 are selected as candidate experts. The final data set includes 213 experts and 253 historical patients. The professional expertise information of 213 experts was crawled from the official website of the First Affiliated Hospital of Zhengzhou University as supplementary data. After preprocessing such as eliminating missing items, 40,052 valid remote consultation data were obtained (including 253 pieces of preliminary diagnosis information of historical patients, which corresponded one-to-one with the preliminary diagnosis information of historical patients, and 39,799 pieces of historical diagnosis information of experts), involving 49 departments in total, including internal medicine, surgery, gynecology, pediatrics, geriatrics, emergency department, medical technology department, etc. According to statistical analysis, the age range of historical patients was between 0 and 105 years old, with an average age of 54 years old, 53.37% of them were male, and 46.63% were female.

[0101] Figure 1 A flowchart of a method for constructing a remote medical expert recommendation model provided in this application is shown in FIG. Figure 1 As shown, in order to solve the above problems, a method for constructing a remote medical expert recommendation model is provided in an embodiment of the present application, including:

[0102] Acquiring sample information, the sample information including: basic information and preliminary diagnosis information of multiple historical patients, and basic information, professional expertise information and historical diagnosis information of multiple experts;

[0103] Preprocessing sample information;

[0104] Entities and relationships between entities are extracted from the preliminary diagnosis information of historical patients and the professional expertise information of experts; entities include historical patients, experts, and diseases; relationships between entities include the suffering relationship between historical patients and diseases, and the expert treatment relationship between experts and diseases;

[0105] Build a knowledge graph based on the extracted entities and the relationships between them;

[0106] Perform feature calculation on sample information to obtain the disease correlation and multi-granularity correlation between each historical patient and all experts, as well as the basic features of the experts corresponding to the historical patients;

[0107] The calculated disease correlations and multi-granularity correlations between multiple groups of historical patients and experts, as well as the basic features of experts corresponding to historical patients, are input into a deep neural network for training to obtain an expert recommendation model.

[0108] This application is based on the basic information and preliminary diagnosis information of historical patients, the basic information, professional expertise information and historical diagnosis information of experts for feature calculation, and then the obtained disease relevance, multi-granularity relevance and basic features are input into the deep neural network for training to obtain the expert recommendation model. The preliminary diagnosis information of historical patients in this application is the information given by doctors in grassroots hospitals, which avoids the problems of incomprehensibility and low communication efficiency caused by direct communication between historical patients and experts. Not only the professional information that experts are good at treating is taken into account, but also the basic characteristics of experts are taken into account, which can recommend more suitable experts to patients, and can effectively improve the accuracy and robustness of the recommendation results. Then, doctors in grassroots hospitals conduct remote consultations with suitable experts, which improves the efficiency and quality of remote consultations. It solves the problems of delays and inconveniences caused by the need to manually find suitable doctors offline in the prior art, as well as the problem of uneven distribution of medical resources.

[0109] Specifically, the sample information is preprocessed, including:

[0110] The sample information is segmented using the Chinese word segmentation library (Jieba) in the computer programming language (Python);

[0111] Based on the stop word list, add custom stop words according to the actual situation to filter out words and symbols that have no practical meaning in the sample information; the stop word list can be Baidu, Harbin Institute of Technology, etc.;

[0112] A user dictionary is constructed through a medical word library, and medical terms in sample information are identified through the user dictionary. The medical word library can be a Sogou input method word library.

[0113] The original data in the sample information in this application is full of noise. In order to reduce the spatial dimension of the text, accurately extract text features, and improve the efficiency of the model operation, the data set needs to be preprocessed by word segmentation, medical terminology recognition, stop word filtering, and other operations.

[0114] Figure 2 for Figure 1 A schematic diagram of a flow chart of feature calculation of sample information in the embodiment is shown in FIG. Figure 2 As shown, further, feature calculation is performed on the sample information, specifically including:

[0115] Calculate the disease relevance between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts;

[0116] Calculate the multi-granularity correlation between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the historical diagnosis information of multiple experts;

[0117] The basic features of the expert corresponding to each historical patient are calculated based on the basic information of multiple historical patients and the basic information of the experts corresponding to the historical patients.

[0118] Furthermore, based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts, the disease relevance between each historical patient and all experts is calculated, specifically including:

[0119] The disease relevance is calculated based on the path between historical patients and experts in the knowledge graph; the calculation formula for disease relevance is:

[0120] Let max r∈D (path ir )=min r∈D (path ir )hour,

[0121] In the formula, represents the disease correlation between the i-th historical patient and the j-th expert, path ij represents the number of all paths between the i-th historical patient and the j-th expert, D represents the number of experts, min r∈D (path ir ) represents the minimum number of paths between D experts and the i-th historical patient, max r∈D (path ir ) represents the maximum number of paths between D experts and the i-th historical patient.

[0122] Specifically, the path between historical patients and experts in this application refers to the situation when there is a path between historical patients and experts for the same disease.

[0123] For example, when there are three paths of "patient 1-disease 1-expert 1", "patient 2-disease 2-expert 2", and "patient 2-disease 1-expert 1", "patient 1-disease 1-patient 2-disease 2-expert 2" will appear in the knowledge graph, but it is obvious that patient 1 and expert 2 are not compatible. In this application, it is deemed that there is no path between patient 1 and expert 2.

[0124] In this application, the expert's professional expertise information can reflect that the expert is good at treating patients in a certain disease field, and the preliminary diagnosis information of historical patients can reflect the diseases or symptoms suffered by historical patients. Entities are extracted from them, and a knowledge graph is constructed based on the relationship between entities. There is a path in the knowledge graph that connects historical patients with experts, that is, Path. Through the path-related information in the knowledge graph, the disease correlation between historical patients and experts can be calculated.

[0125] The disease correlation between historical patients and experts is shown in Table 1:

[0126] Table 1 Disease correlation

[0127] serial number Expert 1 … Expert 184 … Expert 213 History of Patient 1 0 … 0.08333333 … 0 … … … … … … History of Patient 4 0 … 0.27272727 … 0 History of Patient 5 0 … 0.07692308 … 0 … … … … … … History of patients 253 0 … 0 … 0

[0128] Figure 3 for Figure 2 A flow chart of calculating the multi-granularity correlation between each historical patient and all experts in the embodiment is shown in FIG. Figure 3 As shown, further, based on the preliminary diagnosis information of multiple historical patients and the historical diagnosis information of multiple experts, the multi-granularity correlation between each historical patient and all experts is calculated, specifically including:

[0129] Figure 4 for Figure 3 A schematic diagram of the process of calculating the coarse-grained correlation in the embodiment is shown in FIG. Figure 4 As shown, the coarse-grained correlation is calculated, specifically including:

[0130] The topic probability distribution in the historical patient's preliminary diagnosis information and the topic probability distribution in the expert's historical diagnosis information are calculated through the latent Dirichlet allocation (LDA topic model);

[0131] The LDA topic model is an unsupervised clustering algorithm based on Bayesian thinking, which uses probabilistic inference to process input text. In the model, topics are used as the middle layer to connect words and documents, generating three layers of probabilities for documents, topics, and words. Since 49 departments are involved, the disease symptoms between departments may overlap, so the number of topics is determined to be 2 to 50. In this embodiment, the number of topics is 17, the preliminary diagnosis information of historical patients is 253, and the historical diagnosis information of experts is 39,799, for a total of 40,052 pieces of diagnosis information.

[0132] The topic-word distribution in the LDA topic model is shown in Table 2:

[0133] Table 2 Topic-word distribution table

[0134] Topic Number Words 1 Magnetic resonance imaging Lumbar disc nerve Lumbar epilepsy Cervical spondylosis Stenosis Myelencephalitis 2 Chronic acute obstructive disease Gastritis Bronchitis exacerbation Thrombocytopenia Pancreatitis … … 17 Effusion, pleural and abdominal effusion, bilateral pleural effusion, intestinal obstruction, right pericardium, ascites drainage

[0135] The document-topic distribution in the LDA topic model is shown in Table 3:

[0136] Table 3 Document-topic distribution table

[0137] Text Number Theme 1 Theme 2 … Topic 17 1 0.01149577 0.6769135 … 0.01149447 2 0.12461945 0.0073011 … 0.00729927 … … … … … 40052 0.00280117 0.00280125 … 0.03340625

[0138] Calculate the JS discrete distance value between the topic probability distribution in the preliminary diagnosis information of each historical patient and the topic probability distribution in all historical diagnosis information of all experts; the calculation formula of the JS discrete distance value is:

[0139]

[0140] Where D JS (P a ||Q b ) represents the JS discrete distance value between the a-th preliminary diagnosis information and the b-th historical diagnosis information, P a (x) represents the topic probability distribution in the a-th preliminary diagnostic information, Q b (x) represents the probability distribution of the topic in the bth historical diagnostic information, n represents the number of all preliminary diagnostic information, m represents the number of all historical diagnostic information, x represents the value of the random variable, and x is P a (x) and Q b (x) is a specific value on the surface of the y-axis. X represents the set of all possible values, i.e., the sample space of the random variable.

[0141] The JS discrete distance value is converted to obtain the coarse-grained correlation, and the conversion formula is:

[0142]

[0143] In the formula, Sim LDA (P a ,Q b ) represents the coarse-grained correlation between the a-th preliminary diagnosis information and the b-th historical diagnosis information, Sim LDA (P a ,Q b )∈[0,1], h represents the adjustment factor.

[0144] The coarse-grained correlation between the historical patient's preliminary diagnosis information and the expert's historical diagnosis information is shown in Table 4:

[0145] Table 4 Coarse-grained correlation between historical patients’ preliminary diagnosis information and experts’ historical diagnosis information

[0146]

[0147] Figure 5 for Figure 3 A schematic diagram of the process of calculating fine-grained correlation in the embodiment is shown in FIG. Figure 5 As shown, the sentence embedding representation model (Sentence-BERT model) is used to calculate fine-grained relevance, including:

[0148] The preliminary diagnosis information of historical patients and the historical diagnosis information of experts are respectively input into the Sentence-BERT model, and the initial embedding vector of historical patients and the initial embedding vector of experts are extracted through the Sentence-BERT model;

[0149] Perform average pooling feature extraction on the historical patient initial embedding vector and the expert initial embedding vector respectively to obtain the historical patient feature vector u and the expert feature vector v;

[0150] The historical patient feature vector u and the expert feature vector v are calculated through the loss function, and the calculation formula of the loss function is:

[0151] o=softmax[W t (u,v,|uv|)];

[0152] In the formula, o represents the loss function, W t represents the weight matrix, which is used to map the concatenated feature vector to a new space, and softmax represents the normalized exponential function;

[0153] According to the calculation results of the loss function, the fine-grained feature vectors of the preliminary diagnosis information of historical patients and the historical diagnosis information of experts are obtained; among them, the fine-grained feature vector of the a-th preliminary diagnosis information is: F a =(f1,f2,...,f L ), the fine-grained feature vector of the b-th historical diagnostic information is: F b =(f′1,f′2,...,f′ L ), L represents the feature dimension; in this embodiment, L=512;

[0154] The fine-grained correlation between the historical patient's preliminary diagnosis information and the expert's historical diagnosis information is calculated using the following formula:

[0155]

[0156] The fine-grained features of the historical patients’ preliminary diagnosis information and the experts’ historical diagnosis information are shown in Table 5:

[0157] Table 5 Fine-grained feature vectors of historical patients’ preliminary diagnosis information and experts’ historical diagnosis information

[0158] serial number Vectorized Representation 1 [-0.03224770,0.00040633,0.04270793,…,-0.04969637] 2 [-0.05244868,0.00555693,0.11365228,…,0.00741914] … … 40052 [0.03792923,0.07245108,-0.02122660,…,0.00869036]

[0159] The fine-grained correlation between the historical patient's preliminary diagnosis information and the expert's historical diagnosis information is shown in Table 6 below:

[0160] Table 6 Fine-grained correlation between historical patients’ preliminary diagnosis information and experts’ historical diagnosis information

[0161]

[0162] By combining the coarse-grained correlation and the fine-grained correlation, we can obtain the multi-grained correlation between the preliminary diagnosis information of each historical patient and the multiple historical diagnosis information of all experts. The calculation formula is:

[0163]

[0164] In the formula, δ represents the proportion of coarse-grained correlation in multi-grained correlation, and 1-δ represents the proportion of fine-grained correlation in multi-grained correlation;

[0165] The multi-granularity correlation between each historical patient and each expert is calculated using the following formula:

[0166]

[0167] In the formula, represents the multi-granularity correlation between the i-th historical patient and the j-th expert, N jrepresents the number of historical diagnostic information corresponding to the jth expert. When the jth expert corresponds to the bth piece of historical diagnostic information, Y jb =1, otherwise, Y jb =0.

[0168] Specifically, in this application, δ=0.3, such as Figure 6-8 As shown in Table 7, when δ = 0.3 and (1-δ) = 0.7, the evaluation index values ​​of the recommendation results are close to the optimal. At this time, the multi-granularity correlation between each historical patient and each expert is shown in Table 7 below:

[0169] Table 7 Multi-granularity correlation between each historical patient and each expert

[0170] serial number Expert 1 Expert 2 Expert 3 … Expert 213 History of Patient 1 0.10962492 0.13862119 0.11979597 … 0.12741069 History of Patient 2 0.04571596 0.07833831 0.05072149 … 0.18829341 History of Patient 3 0.16500303 0.19174133 0.19063752 … 0.03779729 … … … … … … History of patients 253 0.03594537 0.01039094 0.02285084 … 0.18797766

[0171] In this application, multi-granularity correlation is calculated, which comprehensively considers the preliminary diagnosis information of historical patients and the historical diagnosis information of experts, which is more comprehensive and accurate, including coarse-grained correlation and fine-grained correlation. The coarse-grained correlation is calculated by the LDA topic model. The LDA topic model calculates the "text-topic" probability distribution in a probabilistic inference manner, and then measures the topic distribution distance between historical patients and experts through the JS divergence, and finally converts the obtained JS discrete distance value into a coarse-grained correlation. The fine-grained correlation is calculated by the Sentence-BERT model. The Sentence-BERT model is an improved model based on the BERT model, which enables the vectorized expression of the text to carry more semantic information and is more suitable for the calculation of fine-grained correlation. It can identify the meaning of words in the preliminary diagnosis information of historical patients and the historical diagnosis information of experts, unify the context information and expression, and combine the LDA topic model with the Sentence-BERT model to provide more accurate multi-granularity correlation.

[0172] Furthermore, the basic information of the historical patients includes age and gender; the basic information of the experts includes: professional title, and consultation time and number of consultations within a preset time; the basic characteristics of the experts include: age characteristics, gender characteristics, professional title characteristics and activity characteristics.

[0173] This application fully considers the impact of the expert's basic characteristics on the recommendation results. For example, if an expert is better at treating female patients in a certain disease area, then the expert will be recommended to female patients first.

[0174] Furthermore, the basic features of all experts are calculated based on the basic information of multiple experts and the basic information of corresponding historical patients, including:

[0175] Get the gender characteristics of the corresponding experts according to the gender of historical patients;

[0176] When the current historical patient is female, the gender characteristics are:

[0177] When the current historical patient is male, the gender characteristics are:

[0178] In the formula, Indicates the gender characteristic value, N j "(Female)" indicates the number of female historical patients among the historical patients corresponding to the corresponding expert, N j "(Male)" indicates the number of male historical patients among the historical patients corresponding to the corresponding expert, N j ” indicates the number of historical patients corresponding to the corresponding expert.

[0179] Table 9 Gender characteristics

[0180] serial number Expert 1 Expert 2 Expert 3 … Expert 213 History of Patient 1 0.396825397 0.50862069 0.636363636 … 0.318181818 History of Patient 2 0.603174603 0.49137931 0.363636364 … 0.681818182 History of Patient 3 0.396825397 0.50862069 0.636363636 … 0.318181818 … … … … … … History of patients 253 0.603174603 0.49137931 0.363636364 … 0.681818182

[0181] Calculate the age characteristics of all experts, specifically including: classify historical patients into infants, children, young adults, and middle-aged and elderly people according to their ages; obtain the corresponding age characteristics of experts according to the classification of historical patients;

[0182] When the current historical patient was an infant, the age characteristics of the specialist were:

[0183] When the current historical patient is a child, the age characteristics of the specialist are:

[0184] When the current historical patients are young adults, the age characteristics of the experts are:

[0185] When the current historical patients are middle-aged or elderly, the age characteristics of the experts are:

[0186] In the formula, represents the age characteristic value, N j '(Infant) indicates the number of historical patients in the Infant age group corresponding to the expert, N j '(Teenager) indicates the number of historical patients in the Teenager age group corresponding to the expert, N j ′(Youth) represents the number of historical patients in the Youth age group corresponding to the corresponding expert, N j ′(Elder) represents the number of patients in the Elder age group corresponding to the expert, N j′ represents the number of all historical patients corresponding to the corresponding expert, Infant represents infants aged 0-6 years old, Teenager represents children aged 7-17 years old, Youth represents young adults aged 18-45 years old, and Elder represents middle-aged and elderly people over 46 years old.

[0187] Table 8 Age characteristics

[0188] serial number Expert 1 Expert 2 Expert 3 … Expert 213 History of Patient 1 0.857142857 0.879310345 0.863636364 … 0.954545455 History of Patient 2 0.857142857 0.879310345 0.863636364 … 0.954545455 … … … … … … History of patients 249 0.126984127 0.120689655 0.136363636 … 0.045454545 … … … … … … History of patients 253 0.857142857 0.879310345 0.863636364 … 0.954545455

[0189] According to the time when the historical patients applied for consultation, the activity characteristics of the corresponding experts were obtained;

[0190] Current historical patients are Week c When submitting a consultation request, the activity characteristics of the expert are:

[0191]

[0192] Where c∈[1,7], t j Indicates the time of the last consultation of the corresponding expert, t j (Week c ) indicates the corresponding expert's most recent c The consultation time, N j”′ (Week c ) represents the total number of consultations by the corresponding expert in week c, and z represents the adjustment factor.

[0193] Table 11 Activity characteristics

[0194] serial number Expert 1 Expert 2 Expert 3 … Expert 213 History of Patient 1 0.66666667 0.66666667 0.66666667 … 1 … … … … … … History of patients 253 0.66666667 0.66666667 0.66666667 … 1

[0195] The professional title characteristics of all experts were calculated. The calculation results are shown in Table 10 below, which specifically include: dividing the professional titles of experts into chief physicians, deputy chief physicians and attending physicians; then assigning a value of 3 to chief physicians, 2 to deputy chief physicians, and 1 to attending physicians among all experts; and performing normalization to obtain professional title characteristics.

[0196] Table 10 Characteristics of job titles

[0197] serial number Expert 1 Expert 2 Expert 3 … Expert 213 History of Patient 1 0.10962492 0.13862119 0.11979597 … 0.12741069 History of Patient 2 0.04571596 0.07833831 0.05072149 … 0.18829341 History of Patient 3 0.16500303 0.19174133 0.19063752 … 0.03779729 … … … … … … History of patients 253 0.03594537 0.01039094 0.02285084 … 0.18797766

[0198] In this application, disease relevance is used as feature 1, multi-granularity relevance is used as feature 2, gender feature is used as feature 3, age feature is used as feature 4, activity feature is used as feature 5, and job title feature is used as feature 6, which are input into the input layer of the deep neural network (DNN), then mapped to the hidden layer, and finally the recommendation result is outputted in the output layer. Among them, the hidden layer includes two layers, including 64 neurons and 32 neurons respectively.

[0199] The loss function in the DNN model is set to mean square error (MSE), the batch size is 32, and the maximum number of iterations is set to 200. The adaptive moment estimation (Adam) optimizer is used to train the model parameters, and the learning rate is set to 0.001. The training process uses Keras and Tensorflow frameworks, and the training process is implemented on a computer equipped with an AMD Ryzen 55600H 3.3GHz processor, Windows operating system, and 16GB RAM. The activation function used is the linear rectification (Relu) function.

[0200] This application divides the output recommendation results into three categories: accurate recommendation, appropriate recommendation and non-recommendation. Among them, accurate recommendation means that the recommended expert is consistent with the actual consulting expert, appropriate recommendation means that the recommended expert and the actual consulting expert belong to the same department, and non-recommendation means that the recommended expert and the actual consulting expert are inconsistent and do not belong to the same department.

[0201] Using the common back propagation neural network (BP) and vector space model (VSM) as comparative examples, it is proved that the recommendation results of the expert recommendation model trained with the DNN model are more accurate.

[0202] The accuracy of different models is evaluated through model evaluation index values, which include:

[0203] Accuracy (HR):

[0204] In the formula, |Test∩Top_N| represents the number of intersections between the test problem set and the recommendation set, and |Test| represents the number of test problem sets;

[0205] Mean Reciprocal Rank (MRR):

[0206] In the formula, Rank q represents the position of the doctor recommended by the preliminary diagnosis information of the qth patient in the expert recommendation set; N represents the number of patients;

[0207] Precision (PRE):

[0208] In the formula, |Top_N| represents the number of recommendation sets.

[0209] The model evaluation results are as follows Figure 9-11As shown in the figure, when the number of recommended experts is in the range of 1 to 10, the DNN recommendation results show the best performance in all evaluation indicators. The HR values ​​of the DNN model recommendation results are all above 70%, and when the number of recommended doctors exceeds 5, the HR is stable at above 96%. Compared with the BP neural network, the HR value is increased by 7.20% on average; compared with the SVM algorithm, the HR value is increased by 1.44% on average; the MRR values ​​of the DNN recommendation results are all above 0.70, and are higher than the other two methods. Compared with the BP neural network, the MRR value is increased by 8.76% on average. Compared with the SVM algorithm, the MRR value is increased by 2.13% on average; the PRE values ​​of the DNN recommendation results are not less than 0.455, and are higher than the other two methods, indicating that the number of correct recommendations in the recommendation set is more than the other two groups. Compared with the BP neural network, the PRE value is increased by 6.53% on average. Compared with the SVM algorithm, the PRE value is increased by 2.42% on average. The above results show that the adopted deep neural network model is robust and can generate relatively accurate recommendation results for telemedicine expert recommendation.

[0210] This application also calculates the HR value of the recommendation accuracy of each department. The department recommendation accuracy is as follows: Fig.12 As shown in the figure. Among them, the accuracy of 13 departments reached 100%, and only four departments had an accuracy rate lower than 80%, namely EICU, Radiology, Chest Pain Surgery and Vascular Surgery. Investigating the reason, it was found that the number of consultations involving the four departments in the test set was small, less than 10 cases each, of which EICU had only one case, so there was a certain degree of contingency. Therefore, when the amount of data is large enough, it can be considered that the recommendation results of all departments are relatively accurate.

[0211] This application also uses an interpretable machine learning model (SHAP model) to perform interpretability analysis on the recommendation results, which can intuitively show that features 1 to 6 have an impact on the recommendation results. The degree of influence of the features on the recommendation results is as follows: Fig.13 As shown. Fig.13 It can be seen that features 1 to 6 all contribute to the recommendation results, and they are all positive. The contributions to the recommendation results from large to small are: multi-granularity correlation, disease correlation, age characteristics, activity characteristics, title characteristics, and gender characteristics, with contribution values ​​of 0.26, 0.09, 0.05, 0.025, 0.013, and 0.0125, respectively. It can be seen that multi-granularity correlation is the main reason affecting the recommendation results. The higher the multi-granularity correlation, the more likely it is to recommend the corresponding expert to the patient.

[0212] Fig.14 A flowchart of a remote medical expert recommendation method provided for this application, Fig.15 A partial flow chart of a remote medical expert recommendation method provided for this application, such as Fig.14 ,15 As shown, the present application also provides a remote medical expert recommendation method, comprising:

[0213] Receives and responds to specialist referral requests for patients awaiting referral;

[0214] Obtain the information to be recommended, including: basic information and preliminary diagnosis information of the patient to be recommended, as well as basic information, professional expertise information and historical diagnosis information of multiple experts;

[0215] Extracting entities and relationships between entities from the preliminary diagnosis information of the patient to be recommended and the professional expertise information of multiple experts; wherein the entities include the patient to be recommended, the expert, and the disease, and the relationships between entities include the suffering relationship between the patient to be recommended and the disease, and the expert-treatment relationship between the expert and the disease;

[0216] Build a knowledge graph based on the extracted entities and the relationships between them;

[0217] Perform feature calculation on the recommended information to obtain the disease correlation and multi-granularity correlation between the recommended patient and multiple experts, as well as the basic features of the experts corresponding to the recommended patient;

[0218] Input the calculated disease correlation and multi-granularity correlation between the patient to be recommended and multiple experts, as well as the basic features of the experts corresponding to the patient to be recommended into the expert recommendation model as described in any of the above items to obtain an expert recommendation set;

[0219] Output the expert recommendation set to the patient to be recommended.

[0220] The present application provides a remote medical expert recommendation method that can promptly receive and respond to expert recommendation requests from patients to be recommended, and obtain a corresponding expert recommendation set based on the basic information and preliminary diagnosis information of the patients to be recommended, providing a convenient and fast recommendation method for patients to be recommended.

[0221] The present application also provides a system for constructing a telemedicine expert recommendation model, including a module for implementing a method for constructing a telemedicine expert recommendation model as described in any of the above contents.

[0222] Fig.16 A functional structure diagram of a system for building a remote medical expert recommendation model provided in this application, such as Fig.16 Specifically, the construction system includes:

[0223] The acquisition module 10 is used to acquire sample information, wherein the sample information includes: basic information and preliminary diagnosis information of multiple historical patients, and basic information, professional expertise information and historical diagnosis information of multiple experts;

[0224] The extraction model 20 is used to extract entities and relationships between entities from the preliminary diagnosis information of historical patients and the professional expertise information of experts; wherein the entities include historical patients, experts, and diseases; the relationships between entities include the suffering relationship between historical patients and diseases, and the expert-treatment relationship between experts and diseases;

[0225] A knowledge graph construction module 30, used to construct a knowledge graph based on the extracted entities and the relationships between the entities;

[0226] The feature calculation module 40 is used to perform feature calculation on the sample information to obtain the disease correlation and multi-granularity correlation between each historical patient and all experts, as well as the basic features of the experts corresponding to the historical patients;

[0227] The training module 50 is used to input the calculated disease correlations and multi-granularity correlations between multiple groups of historical patients and experts, as well as the basic features of experts corresponding to the historical patients into a deep neural network for training to obtain an expert recommendation model.

[0228] Fig.17 for Fig.16 The functional structure diagram of the feature calculation module is as follows: Fig.17 As shown, specifically, the feature calculation module 40 includes:

[0229] The disease correlation calculation unit 401 is used to calculate the disease correlation between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts;

[0230] A multi-granularity correlation calculation unit 402, used to calculate the multi-granularity correlation between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the historical diagnosis information of multiple experts;

[0231] The basic feature calculation unit 403 is used to calculate the basic feature of the expert corresponding to each historical patient based on the basic information of multiple historical patients and the basic information of the experts corresponding to the historical patients.

[0232] More specifically, the disease relevance calculation formula is:

[0233] Let max r∈D (path ir )=min r∈D (path ir )hour,

[0234] In the formula, represents the disease correlation between the i-th historical patient and the j-th expert, path ijrepresents the number of all paths between the i-th historical patient and the j-th expert, D represents the number of experts, min r∈D (path ir ) represents the minimum number of paths between D experts and the i-th historical patient, max r∈D (path ir ) represents the maximum number of paths between D experts and the i-th historical patient.

[0235] Fig.18 for Fig.17 The functional structure diagram of the multi-granularity correlation calculation unit is as follows: Fig.18 As shown, specifically, the multi-granularity correlation calculation unit 402 includes:

[0236] A coarse-grained correlation calculation unit 4021, used to calculate the coarse-grained correlation;

[0237] A distribution probability calculation unit 40211 is used to calculate the topic probability distribution in the historical patient's preliminary diagnosis information and the topic probability distribution in the expert's historical diagnosis information through an LDA topic model;

[0238] The discrete distance value calculation unit 40212 is used to calculate the JS discrete distance value between the subject probability distribution in the preliminary diagnosis information of each historical patient and the subject probability distribution in all historical diagnosis information of all experts; the calculation formula of the JS discrete distance value is:

[0239] Topic Probability Distribution Topic Probability Distribution

[0240]

[0241] Where D JS (P a ||Q b ) represents the JS discrete distance value between the a-th preliminary diagnosis information and the b-th historical diagnosis information, P a (x) represents the topic probability distribution in the a-th preliminary diagnostic information, Q b (x) represents the probability distribution of the topic in the bth historical diagnostic information, n represents the number of all preliminary diagnostic information, m represents the number of all historical diagnostic information, x represents the value of the random variable, and x is P a (x) and Q b (x) is a specific value on the surface of the y-axis. X represents the set of all possible values, i.e., the sample space of the random variable.

[0242] The conversion unit 40213 is used to convert the JS discrete distance value to obtain a coarse-grained correlation, and the conversion formula is:

[0243]

[0244] In the formula, Sim LDA (P a ,Q b ) represents the coarse-grained correlation between the a-th preliminary diagnosis information and the b-th historical diagnosis information, Sim LDA (P a ,Q b )∈[0,1], h represents the adjustment factor;

[0245] A fine-grained correlation calculation unit 4022, used to calculate fine-grained correlation;

[0246] An initial embedding vector calculation unit 40221 is used to input the preliminary diagnosis information of historical patients and the historical diagnosis information of experts into the Sentence-BERT model respectively, and extract the initial embedding vector of the historical patients and the initial embedding vector of the experts through the Sentence-BERT model;

[0247] The feature extraction unit 40222 is used to perform average pooling feature extraction on the historical patient initial embedding vector and the expert initial embedding vector respectively to obtain the historical patient feature vector u and the expert feature vector v;

[0248] The loss function calculation unit 40223 is used to calculate the historical patient feature vector u and the expert feature vector v through the loss function. The calculation formula of the loss function is:

[0249] o=softmax[W t (u,v,|uv|)];

[0250] In the formula, o represents the loss function, W t represents the weight matrix, which is used to map the concatenated feature vector to a new space, and softmax represents the normalized exponential function;

[0251] The fine-grained feature acquisition unit 40224 is used to obtain the fine-grained feature vector of the historical patient's preliminary diagnosis information and the expert's historical diagnosis information according to the calculation result of the loss function; wherein the fine-grained feature vector of the a-th preliminary diagnosis information is: F a =(f1,f2,...,f L ), the fine-grained feature vector of the b-th historical diagnostic information is: F b =(f′1,f′2,...,f′ L ), L represents the feature dimension;

[0252] The first calculation unit 40225 is used to calculate the fine-grained correlation between the historical patient's preliminary diagnosis information and the expert's historical diagnosis information, and the calculation formula is:

[0253]

[0254] The fusion unit 4023 is used to fuse the coarse-grained correlation and the fine-grained correlation to obtain the multi-grained correlation between the preliminary diagnosis information of each historical patient and the multiple historical diagnosis information of all experts. The calculation formula is:

[0255]

[0256] In the formula, δ represents the proportion of coarse-grained correlation in multi-grained correlation, and 1-δ represents the proportion of fine-grained correlation in multi-grained correlation;

[0257] The second calculation unit 4024 is used to calculate the multi-granularity correlation between each historical patient and each expert, and the calculation formula is:

[0258]

[0259] In the formula, represents the multi-granularity correlation between the i-th historical patient and the j-th expert, N j represents the number of historical diagnostic information corresponding to the jth expert. When the jth expert corresponds to the bth piece of historical diagnostic information, Y jb =1, otherwise, Y jb =0.

[0260] Fig.19 for Fig.17 The functional structure diagram of the basic feature calculation unit in the figure is as follows: Fig.19 As shown, specifically, the basic feature calculation unit 403 includes:

[0261] An age feature calculation unit 4031, used to calculate the age features of all experts;

[0262] Classification unit 40311 is used to classify historical patients into infants, children, young adults and middle-aged and elderly people according to their ages;

[0263] The third calculation unit 40312 is used to obtain the age characteristics of the corresponding experts according to the classification of historical patients;

[0264] When the current historical patient was an infant, the age characteristics of the specialist were:

[0265] When the current historical patient is a child, the age characteristics of the specialist are:

[0266] When the current historical patients are young adults, the age characteristics of the experts are:

[0267] When the current historical patients are middle-aged or elderly, the age characteristics of the experts are:

[0268] In the formula, represents the age characteristic value, N j ′(Infant) indicates the number of historical patients in the Infant age group corresponding to the expert, N j ′(Teenager) represents the number of historical patients in the Teenager age group corresponding to the corresponding expert, N j ′(Youth) represents the number of historical patients in the Youth age group corresponding to the corresponding expert, N j ′(Elder) represents the number of patients in the Elder age group corresponding to the expert, N j ′ represents the number of all historical patients corresponding to the corresponding expert, Infant represents infants aged 0-6 years old, Teenager represents children aged 7-17 years old, Youth represents young adults aged 18-45 years old, and Elder represents middle-aged and elderly people over 46 years old;

[0269] A gender characteristic calculation unit 4032, used to obtain the gender characteristic of the corresponding expert according to the gender of the historical patient;

[0270] When the current historical patient is female, the gender characteristics are:

[0271] When the current historical patient is male, the gender characteristics are:

[0272] In the formula, Indicates the gender characteristic value, N j "(Female)" indicates the number of female historical patients among the historical patients corresponding to the corresponding expert, N j "(Male)" indicates the number of male historical patients among the historical patients corresponding to the corresponding expert, N j ” indicates the number of historical patients corresponding to the corresponding expert;

[0273] The professional title characteristic calculation unit 4033 is used to calculate the professional title characteristics of all experts, specifically including: dividing the professional titles of experts into chief physicians, deputy chief physicians and attending physicians; then assigning a value of 3 to the chief physicians, 2 to the deputy chief physicians, and 1 to the attending physicians among all experts; and performing normalization processing to obtain professional title characteristics;

[0274] An activity calculation unit 4034 is used to obtain the activity characteristics of the corresponding expert according to the time when the historical patient applied for consultation;

[0275] Current historical patients are Weekc When submitting a consultation request, the activity characteristics of the expert are:

[0276]

[0277] Where c∈[1,7], t j Indicates the time of the most recent consultation with the corresponding expert, t j (Week c ) indicates the corresponding expert's most recent c The consultation time, N j”′ (Week c ) represents the total number of consultations by the corresponding expert in week c, and z represents the adjustment factor.

[0278] The present application also provides a device, comprising:

[0279] Memory;

[0280] Processor; and

[0281] Computer programs;

[0282] Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the method for constructing a remote medical expert recommendation model as described in any of the above contents.

[0283] The present application also provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement the method for constructing a remote medical expert recommendation model as described in any of the above contents.

[0284] The "corresponding" mentioned in this application refers to the historical diagnosis relationship between historical patients and experts.

[0285] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, C language, VHDL language, Verilog language, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0286] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0287] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0288] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0289] In the description of the present application, “plurality” means at least two, for example, two, three, etc., unless otherwise clearly and specifically defined.

[0290] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0291] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for constructing a telemedicine expert recommendation model, characterized in that: include: Acquiring sample information, the sample information including: basic information and preliminary diagnosis information of multiple historical patients, and basic information, professional expertise information and historical diagnosis information of multiple experts; Entities and relationships between entities are extracted from the preliminary diagnosis information of historical patients and the professional expertise information of experts; entities include historical patients, experts, and diseases; relationships between entities include the suffering relationship between historical patients and diseases, and the expert treatment relationship between experts and diseases; Build a knowledge graph based on the extracted entities and the relationships between them; The features of the sample information are calculated to obtain the disease correlation and multi-granularity correlation between each historical patient and all experts, as well as the basic features of the experts corresponding to the historical patients; the calculated disease correlation and multi-granularity correlation between multiple groups of historical patients and experts, as well as the basic features of the experts corresponding to the historical patients are input into the deep neural network for training to obtain the expert recommendation model.

2. The method for constructing a remote medical expert recommendation model according to claim 1, characterized in that: Perform feature calculation on sample information, including: Calculate the disease relevance between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts; Calculate the multi-granularity correlation between each historical patient and all experts based on the preliminary diagnosis information of multiple historical patients and multiple historical diagnosis information of multiple experts; The basic features of the expert corresponding to each historical patient are calculated based on the basic information of multiple historical patients and the basic information of the experts corresponding to the historical patients.

3. The method for constructing a remote medical expert recommendation model according to claim 2, characterized in that: Based on the preliminary diagnosis information of multiple historical patients and the professional expertise information of multiple experts, the disease relevance between each historical patient and all experts is calculated, including: The disease relevance is calculated based on the path between historical patients and experts in the knowledge graph; the calculation formula for disease relevance is: In the formula, represents the disease correlation between the i-th historical patient and the j-th expert, path ij represents the number of all paths between the i-th historical patient and the j-th expert, D represents the number of experts, min r∈D (path ir ) represents the minimum number of paths between D experts and the i-th historical patient, max r∈D (path ir ) represents the maximum number of paths between D experts and the i-th historical patient.

4. The method for constructing a remote medical expert recommendation model according to claim 2, characterized in that: Based on the preliminary diagnosis information of multiple historical patients and multiple historical diagnosis information of multiple experts, the multi-granularity correlation between each historical patient and all experts is calculated, including: Calculate the coarse-grained correlation, including: The topic probability distribution in the historical patients' preliminary diagnosis information and the topic probability distribution in the experts' historical diagnosis information are calculated through the LDA topic model; Calculate the JS discrete distance value between the topic probability distribution in the preliminary diagnosis information of each historical patient and the topic probability distribution in all historical diagnosis information of all experts; the calculation formula of the JS discrete distance value is: Where D JS (P a ||Q b ) represents the JS discrete distance value between the a-th preliminary diagnosis information and the b-th historical diagnosis information, P a (x) represents the topic probability distribution in the a-th preliminary diagnostic information, Q b (x) represents the probability distribution of the topic in the bth historical diagnostic information, n represents the number of all preliminary diagnostic information, m represents the number of all historical diagnostic information, x represents the value of the random variable, and x is P a (x) and Q b (x) is a specific value on the surface of the y-axis. X represents the set of all possible values, i.e., the sample space of the random variable. The JS discrete distance value is converted to obtain the coarse-grained correlation, and the conversion formula is: In the formula, Sim LDA (P a ,Q b ) represents the coarse-grained correlation between the a-th preliminary diagnosis information and the b-th historical diagnosis information, Sim LDA (P a ,Q b )∈[0,1], h represents the adjustment factor; Calculate fine-grained relevance, including: The preliminary diagnosis information of historical patients and the historical diagnosis information of experts are respectively input into the Sentence-BERT model, and the initial embedding vector of historical patients and the initial embedding vector of experts are extracted through the Sentence-BERT model; Perform average pooling feature extraction on the historical patient initial embedding vector and the expert initial embedding vector respectively to obtain the historical patient feature vector u and the expert feature vector v; The historical patient feature vector u and the expert feature vector v are calculated through the loss function, and the calculation formula of the loss function is: o=softmax[W t (u,v,|uv|)]; In the formula, o represents the loss function, W t represents the weight matrix, which is used to map the concatenated feature vector to a new space, and softmax represents the normalized exponential function; According to the calculation results of the loss function, the fine-grained feature vectors of the preliminary diagnosis information of historical patients and the historical diagnosis information of experts are obtained; among them, the fine-grained feature vector of the a-th preliminary diagnosis information is: F a =(f1,f2,...,f L ), the fine-grained feature vector of the b-th historical diagnostic information is: F b =(f1′,f2′,...,f L ′), L represents the feature dimension; The fine-grained correlation between the historical patient's preliminary diagnosis information and the expert's historical diagnosis information is calculated using the following formula: By combining the coarse-grained correlation and the fine-grained correlation, we can obtain the multi-grained correlation between the preliminary diagnosis information of each historical patient and the multiple historical diagnosis information of all experts. The calculation formula is: In the formula, δ represents the proportion of coarse-grained correlation in multi-grained correlation, and 1-δ represents the proportion of fine-grained correlation in multi-grained correlation; The multi-granularity correlation between each historical patient and each expert is calculated using the following formula: In the formula, represents the multi-granularity correlation between the i-th historical patient and the j-th expert, N j represents the number of historical diagnostic information corresponding to the jth expert. When the jth expert corresponds to the bth piece of historical diagnostic information, Y jb =1, otherwise, Y jb =0.

5. The method for constructing a remote medical expert recommendation model according to claim 2, characterized in that: The basic information of the historical patients includes age and gender; The basic information of the expert includes: professional title, and consultation time and number of consultations within a preset time; the basic characteristics of the expert include: age characteristics, gender characteristics, professional title characteristics and activity characteristics.

6. The method for constructing a remote medical expert recommendation model according to claim 5, characterized in that: The basic characteristics of all experts are calculated based on the basic information of multiple experts and the basic information of corresponding historical patients, including: Calculate the age characteristics of all experts, specifically including: classify historical patients into infants, children, young adults, and middle-aged and elderly people according to their ages; obtain the corresponding age characteristics of experts according to the classification of historical patients; When the current historical patient was an infant, the age characteristics of the specialist were: When the current historical patient is a child, the age characteristics of the specialist are: When the current historical patients are young adults, the age characteristics of the experts are: When the current historical patients are middle-aged or elderly, the age characteristics of the experts are: In the formula, represents the age characteristic value, N j '(Infant) indicates the number of historical patients in the Infant age group corresponding to the expert, N j '(Teenager) indicates the number of historical patients in the Teenager age group corresponding to the expert, N j ′(Youth) represents the number of historical patients in the Youth age group corresponding to the corresponding expert, N j ′(Elder) represents the number of patients in the Elder age group corresponding to the expert, N j ′ represents the number of all historical patients corresponding to the corresponding expert, Infant represents infants aged 0-6 years old, Teenager represents children aged 7-17 years old, Youth represents young adults aged 18-45 years old, and Elder represents middle-aged and elderly people over 46 years old; Get the gender characteristics of the corresponding experts according to the gender of historical patients; When the current historical patient is female, the gender characteristics are: When the current historical patient is male, the gender characteristics are: In the formula, Indicates the gender characteristic value, N j "(Female)" indicates the number of female historical patients among the historical patients corresponding to the corresponding expert, N j "(Male)" indicates the number of male historical patients among the historical patients corresponding to the corresponding expert, N j ” indicates the number of historical patients corresponding to the corresponding expert; Calculate the professional title characteristics of all experts, specifically including: divide the professional titles of experts into chief physicians, deputy chief physicians and attending physicians; then assign a value of 3 to the chief physicians, 2 to the deputy chief physicians, and 1 to the attending physicians among all experts; perform normalization to obtain professional title characteristics; According to the time when the historical patients applied for consultation, the activity characteristics of the corresponding experts were obtained; Current historical patients are Week c When submitting a consultation request, the activity characteristics of the expert are: Where c∈[1,7], t j Indicates the time of the most recent consultation with the corresponding expert, t j (Week c ) indicates the corresponding expert's most recent c The consultation time, N j ”'(Week c ) represents the total number of consultations by the corresponding expert in week c, and z represents the adjustment factor.

7. A remote medical expert recommendation method, characterized in that: include: Receives and responds to specialist referral requests for patients awaiting referral; Obtain the information to be recommended, including: basic information and preliminary diagnosis information of the patient to be recommended, as well as basic information, professional expertise information and historical diagnosis information of multiple experts; Extracting entities and relationships between entities from the preliminary diagnosis information of the patient to be recommended and the professional expertise information of multiple experts; wherein the entities include the patient to be recommended, the expert, and the disease, and the relationships between entities include the suffering relationship between the patient to be recommended and the disease, and the expert-treatment relationship between the expert and the disease; Build a knowledge graph based on the extracted entities and the relationships between them; Perform feature calculation on the recommended information to obtain the disease correlation and multi-granularity correlation between the recommended patient and multiple experts, as well as the basic features of the experts corresponding to the recommended patient; Inputting the calculated disease correlation and multi-granularity correlation between the patient to be recommended and the multiple experts, and the basic features of the experts corresponding to the patient to be recommended into the expert recommendation model as claimed in any one of claims 1 to 6 to obtain an expert recommendation set; Output the expert recommendation set to the patient to be recommended.

8. A system for constructing a remote medical expert recommendation model, characterized in that: The invention comprises a module for implementing the method for constructing the remote medical expert recommendation model as described in any one of claims 1 to 6.

9. A device, characterized in that: include: Memory; processor; as well as Computer programs; Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the method for constructing a remote medical expert recommendation model as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon; the computer program is executed by a processor to implement the method for constructing a remote medical expert recommendation model as described in any one of claims 1 to 6.

Images