WO2023272563A1 - 智能分诊方法、装置、存储介质及电子设备 - Google Patents

智能分诊方法、装置、存储介质及电子设备 Download PDF

Info

Publication number
WO2023272563A1
WO2023272563A1 PCT/CN2021/103444 CN2021103444W WO2023272563A1 WO 2023272563 A1 WO2023272563 A1 WO 2023272563A1 CN 2021103444 W CN2021103444 W CN 2021103444W WO 2023272563 A1 WO2023272563 A1 WO 2023272563A1
Authority
WO
WIPO (PCT)
Prior art keywords
doctor
vector
matching
patient
vectors
Prior art date
Application number
PCT/CN2021/103444
Other languages
English (en)
French (fr)
Inventor
张振中
Original Assignee
京东方科技集团股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 京东方科技集团股份有限公司 filed Critical 京东方科技集团股份有限公司
Priority to CN202180001717.6A priority Critical patent/CN115803821A/zh
Priority to PCT/CN2021/103444 priority patent/WO2023272563A1/zh
Priority to US17/907,775 priority patent/US20240203569A1/en
Publication of WO2023272563A1 publication Critical patent/WO2023272563A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H40/00—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices
    • G16H40/20—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices for the management or administration of healthcare resources or facilities, e.g. managing hospital staff or surgery rooms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/40—Processing or translation of natural language
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • the present disclosure relates to the technical field of artificial intelligence, and in particular, relates to an intelligent triage method, an intelligent triage device, a computer-readable storage medium, and electronic equipment.
  • an intelligent recommendation system can recommend food, scenic spots, and transportation plans that meet user needs for users to choose from.
  • the intelligent recommendation system when using the intelligent recommendation system to recommend a doctor for a patient, it is usually based on the self-reported symptoms of the patient to match the highly relevant doctor label in the intelligent recommendation system, so as to recommend the corresponding doctor for the patient according to the retrieved doctor label.
  • the disclosure provides an intelligent triage method, an intelligent triage device, a computer-readable storage medium, and electronic equipment.
  • the present disclosure provides an intelligent triage method, including:
  • a matching degree calculation model is used to calculate the matching degree between the target patient-seeing patient vector and the plurality of doctor vectors, and to recommend a doctor for the target doctor-seeking patient according to the matching degree.
  • condition information includes text information; and obtaining the target patient vector according to the condition information includes:
  • the multiple word vectors are sequentially input into the pre-trained neural network to obtain the target patient visit vector.
  • the method further includes:
  • the doctor knowledge graph is constructed according to the relationships among the plurality of doctors.
  • the vectorization of each doctor in the doctor knowledge graph to obtain multiple doctor vectors includes:
  • Each doctor in the doctor knowledge map is vectorized through a graph neural network to obtain multiple doctor vectors.
  • the graph neural network is used to vectorize each doctor in the doctor knowledge map to obtain multiple doctor vectors, including:
  • the pre-trained graph neural network is used to iteratively update the respective node vectors to obtain the plurality of doctor vectors.
  • the iterative updating of the node vectors by using the pre-trained graph neural network to obtain the plurality of doctor vectors includes:
  • W p , W ph , W c , W ch are the parameters of the graph neural network
  • ⁇ is the activation function in the graph neural network
  • t is the number of network iterations
  • e i is the first doctor knowledge map
  • the node corresponding to i doctors, Np(e i ) is the parent node set of node e i
  • e k is the kth parent node of node e i
  • Nc(e i ) is the child node set of node e i
  • e j is the jth child node of node e i
  • h t (e i ) represents the vector of node e i after t network iterations.
  • the calculation of the matching degree between the target patient vector and the plurality of doctor vectors using a matching calculation model includes:
  • [doc i , pat j ] is the vector spliced by the i-th doctor vector doc i and the target patient vector pat j
  • W, v and b are matching parameters
  • ⁇ is the activation in the matching calculation model function.
  • the method further includes:
  • a matching degree score between the target patient vector and each doctor vector is calculated.
  • the training of the matching calculation model to obtain the matching parameters of the matching calculation model includes:
  • the training data set includes a positive training data set and a negative training data set;
  • the determining the matching parameters of the matching calculation model according to the objective function includes:
  • a stochastic gradient descent algorithm is used to update the matching parameters of the matching calculation model, and when the objective function converges, the training of the matching parameters is completed.
  • the objective function is:
  • (doc i , pat i ) is positive training data, indicating that doctor i is suitable for treating patient i
  • (doc j , pat i ) is negative training data, indicating that doctor j is not suitable for treating patient i
  • ⁇ represents the preset difference threshold between the matching score score(doc i , pat i ) of the positive training data and the matching score score(doc j , pat i ) of the negative training data.
  • the training of the graph neural network to obtain the neural network parameters of the graph neural network includes:
  • the neural network parameters of the graph neural network are iteratively updated using a backpropagation algorithm, and when the objective function converges, the training of the neural network parameters is completed.
  • the recommending a doctor for the target patient according to the degree of matching includes:
  • a doctor is recommended for the target patient based on the set of doctors to be recommended.
  • the recommending a doctor for the target patient based on the set of doctors to be recommended includes:
  • Each doctor in the set of doctors to be recommended is filtered according to the filter condition input by the target patient, so as to recommend a target doctor for the target patient.
  • the present disclosure provides an intelligent triage device, including:
  • the patient vector acquisition module is used to obtain the condition information of the target patient, and obtain the target patient vector according to the condition information;
  • the doctor vector acquisition module is used to vectorize each doctor in the doctor knowledge map to obtain multiple doctor vectors
  • the vector matching module is configured to calculate the matching degree between the target patient-seeing patient vector and the plurality of doctor vectors by using a matching degree calculation model, and recommend a doctor for the target doctor-seeking patient according to the matching degree.
  • the present disclosure provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, any one of the methods described above is implemented.
  • the present disclosure provides an electronic device, including: a processor; and a memory, configured to store executable instructions of the processor; wherein, the processor is configured to execute any one of the above-mentioned instructions by executing the executable instructions described method.
  • FIG. 1 shows a schematic diagram of an exemplary system architecture of an intelligent triage method and device that can be applied to an embodiment of the present disclosure
  • FIG. 2 shows a schematic structural diagram of a computer system suitable for implementing an electronic device according to an embodiment of the present disclosure
  • Fig. 3 schematically shows a flowchart of an intelligent triage method according to an embodiment of the present disclosure
  • Fig. 4 schematically shows a flow chart of obtaining target patient visit vectors according to an embodiment of the present disclosure
  • Fig. 5 schematically shows a flow chart of constructing a doctor's knowledge map according to an embodiment of the present disclosure
  • Fig. 6 schematically shows a schematic diagram of a doctor's knowledge graph according to an embodiment of the present disclosure
  • Fig. 7 schematically shows a flow chart of obtaining doctor vectors according to an embodiment of the present disclosure
  • Fig. 8 schematically shows a flow chart of doctor-patient matching according to an embodiment of the present disclosure
  • Fig. 9 schematically shows a block diagram of an intelligent triage device according to an embodiment of the present disclosure.
  • Example embodiments will now be described more fully with reference to the accompanying drawings.
  • Example embodiments may, however, be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concept of example embodiments to those skilled in the art.
  • the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
  • numerous specific details are provided in order to give a thorough understanding of embodiments of the present disclosure.
  • those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of the specific details being omitted, or other methods, components, devices, steps, etc. may be adopted.
  • well-known technical solutions have not been shown or described in detail to avoid obscuring aspects of the present disclosure.
  • Fig. 1 shows a schematic diagram of a system architecture of an exemplary application environment in which an intelligent triage method and device according to an embodiment of the present disclosure can be applied.
  • a system architecture 100 of an intelligent triage system may include one or more of terminal devices 101 , 102 , and 103 , a network 104 and a server 105 .
  • the network 104 is used as a medium for providing communication links between the terminal devices 101 , 102 , 103 and the server 105 .
  • Network 104 may include various connection types, such as wires, wireless communication links, or fiber optic cables, among others.
  • Terminal devices 101, 102, 103 may be various electronic devices, including but not limited to desktop computers, portable computers, smart phones, and tablet computers. It should be understood that the numbers of terminal devices, networks and servers in Figure 1 are only illustrative. According to the implementation needs, there can be any number of terminal devices, networks and servers.
  • the server 105 may be a server cluster composed of multiple servers.
  • the intelligent triage method provided by the embodiments of the present disclosure is generally executed by the server 105.
  • the intelligent triage device is generally set in the server 105. After the server matches the target patient with multiple symptomatic doctors, the matching result can be sent. to the terminal device, and the terminal device will display it to the patient for the patient to choose.
  • the intelligent triage method provided by the embodiment of the present disclosure can also be executed by one or more of the terminal devices 101, 102, 103, and correspondingly, the intelligent triage device can also be set in In the terminal devices 101, 102, 103, for example, after execution by the terminal device, the matching result can be directly displayed on the display screen of the terminal device, and the matching result can also be provided to the patient through a voice broadcast.
  • the matching result can be directly displayed on the display screen of the terminal device, and the matching result can also be provided to the patient through a voice broadcast.
  • This is not particularly limited.
  • FIG. 2 shows a schematic structural diagram of a computer system suitable for implementing the electronic device of the embodiment of the present disclosure.
  • a computer system 200 includes a central processing unit (CPU) 201 that can be programmed according to a program stored in a read-only memory (ROM) 202 or a program loaded from a storage section 208 into a random-access memory (RAM) 203 Instead, various appropriate actions and processes are performed.
  • ROM read-only memory
  • RAM random-access memory
  • various programs and data necessary for system operation are also stored.
  • the CPU 201, ROM 202, and RAM 203 are connected to each other via a bus 204.
  • An input/output (I/O) interface 205 is also connected to the bus 204 .
  • the following components are connected to the I/O interface 205: an input section 206 including a keyboard, a mouse, etc.; an output section 207 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 208 including a hard disk, etc. and a communication section 209 including a network interface card such as a LAN card, a modem, or the like.
  • the communication section 209 performs communication processing via a network such as the Internet.
  • a drive 210 is also connected to the I/O interface 205 as needed.
  • a removable medium 211 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 210 as necessary so that a computer program read therefrom is installed into the storage section 208 as necessary.
  • the intelligent triage method described in this disclosure may be executed by a processor of an electronic device.
  • the condition information of the target patient, the corpus information of multiple doctors, and the training data set used to build and train the matching calculation model can be input through the input part 206, for example, through the user interface of the electronic device Enter the condition information of the target patient, the corpus information of multiple doctors, etc.
  • the output part 207 can output the matching degree scores of the target patient and multiple doctors.
  • embodiments of the present disclosure include a computer program product, which includes a computer program carried on a computer-readable medium, where the computer program includes program codes for executing the methods shown in the flowcharts.
  • the computer program may be downloaded and installed from a network via communication portion 209 and/or installed from removable media 211 .
  • CPU central processing unit
  • various functions defined in the method and apparatus of the present application are performed.
  • the present application also provides a computer-readable medium.
  • the computer-readable medium may be included in the electronic device described in the above-mentioned embodiments; or it may exist independently without being assembled into the electronic device. middle.
  • the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by an electronic device, the electronic device is made to implement the methods described in the following embodiments. For example, the electronic device may implement the steps shown in FIG. 3 to FIG. 5 , and FIG. 7 and FIG. 8 .
  • the computer-readable medium shown in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two.
  • a computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
  • the intelligent recommendation system when using the intelligent recommendation system to recommend a doctor for a patient, it is usually based on the self-reported symptoms of the patient to match the highly relevant doctor label in the intelligent recommendation system, so as to recommend the corresponding doctor for the patient according to the retrieved doctor label.
  • the doctor label stored in the intelligent recommendation system is inconsistent with the doctor's actual professional ability, even if the doctor selected by the intelligent recommendation system has a high degree of matching with the patient, there may be cases where the recommended doctor is not good at treating the patient's disease situation, thereby reducing the patient experience.
  • the probability of the doctor being retrieved can be improved.
  • adding or blurring doctor labels may introduce noise and affect the accuracy of doctor-patient matching.
  • a label of this doctor in the database may be "good at oral diseases", at this time, the matching degree between this doctor and a patient who needs impacted tooth extraction will increase (because the impacted Tooth extraction is a dental disease). It can be seen that for patients who need extraction of impacted teeth, doctors who are good at endodontics are not symptomatic doctors, which may reduce the patient's experience during the patient's visit.
  • this example embodiment provides an intelligent triage method, which can be applied to the above-mentioned server 105, and can also be applied to one or more of the above-mentioned terminal devices 101, 102, 103.
  • the intelligent triage method may include the following steps S310 to S330:
  • Step S310 Obtain the condition information of the target patient, and obtain the vector of the target patient according to the condition information;
  • Step S320 Vectorize each doctor in the doctor knowledge map to obtain multiple doctor vectors
  • Step S330 Using the matching calculation model to calculate the matching degree between the target patient-seeing patient vector and the multiple doctor vectors, and recommending a doctor for the target patient-seeing patient according to the matching degree.
  • the intelligent triage method by obtaining the condition information of the target patient, and obtaining the vector of the target patient according to the condition information; vectorizing each doctor in the doctor knowledge graph to obtain multiple Doctor vector: using a matching calculation model to calculate the matching degree between the target patient-seeing patient vector and the multiple doctor vectors, and recommending a doctor for the target patient-seeing patient according to the matching degree.
  • a matching calculation model to calculate the matching degree between the target patient-seeing patient vector and the multiple doctor vectors, and recommending a doctor for the target patient-seeing patient according to the matching degree.
  • step S310 the condition information of the target patient is obtained, and the vector of the target patient is obtained according to the condition information.
  • the target patient may be a patient who needs to be triaged before the emergency department, or may be a patient who needs to be triaged in a normal outpatient clinic.
  • hospital staff can use the intelligent triage system to quickly triage the patient, so as to arrange the patient to see a doctor as soon as possible.
  • the hospital staff can use the intelligent triage system to triage the patient, or the patient can use the intelligent triage system to triage the patient, which is not specifically limited in this example.
  • patients can also use the intelligent triage system to determine the target hospital and symptomatic doctors before seeking medical treatment.
  • the condition information of the target patient can be obtained.
  • the condition information of the target patient may include the patient’s basic information such as name, age, etc., the patient’s self-reported symptoms such as cough, fever, etc., and the patient’s past medical records such as medical history, medication, etc. and other information, which is not specifically limited in this example.
  • the hospital staff may input the condition information of the target patient into the intelligent triage system, or the target patient may input his own condition information into the intelligent triage system.
  • the hospital staff or the target patient can input the disease information manually or by voice, which is not specifically limited in this example.
  • the patient After obtaining the condition information of the target patient, the patient can be vectorized according to the condition information of the patient to obtain the vector representation of the patient, so as to match the patient vector with the doctor vector, so as to determine the symptomatic doctor for the patient.
  • the vector representation of the target patient can be obtained according to step S410 and step S420.
  • step S410 each character in the character information is encoded to obtain multiple word vectors.
  • the patient's condition information can include text information, voice information, and image and video information.
  • the patient can input basic information, self-reported symptoms, past medical records and other information into the intelligent triage system, and the intelligent triage system can map the received voice information into corresponding text information, such as " Zhang San, 20 years old, had a toothache for three days.”
  • the patient can also manually input information such as basic information, self-reported symptoms, and past medical records into the intelligent triage system, and the intelligent triage system can directly obtain the patient's text information.
  • the text information of the patient can be encoded.
  • each text in “Zhang San, 20 years old, toothache for three days” can be encoded by Embedding (vector mapping), and each text can be represented by a low-dimensional vector, and multiple corresponding word vectors can be obtained.
  • the vector representation of "Zhang” is obtained.
  • Words can also be used as units to encode each character contained in a word to obtain multiple corresponding word vectors, such as obtaining the vector representation of "Zhang San”.
  • each text can be One-Hot (uniquely hot) encoded, and One-Hot encoding is also called one-bit effective encoding.
  • the method is to use N-bit status registers to encode N states, and each state There are independent register bits, and at any time, only one bit in the register is valid.
  • the vector dimension of One-Hot encoding will increase with the increase of the number of characters in the patient's medical condition information, which may increase the computational complexity.
  • dense vectors can also be used to represent each word.
  • the word2vec algorithm can be used to map each character in the acquired character disease information into a vector space, and each character can be represented by a word vector in the vector space.
  • Doc2vec algorithm, Glove algorithm, etc. can also be used to convert text into vectors.
  • Step S420 Input the multiple word vectors into the pre-trained neural network in sequence to obtain the target patient visit vector.
  • the word vectors corresponding to all the characters in the medical condition information can be regarded as a time series, and then the neural network (for example, recurrent neural network) operates on the word vector corresponding to each word.
  • the neural network For example, recurrent neural network
  • the eight word vectors can be input into the trained LSTM (Long Short-Term Memory, long-short-term memory) network to obtain the vector representation of the patient "Zhang San”.
  • the LSTM network is a time recurrent neural network, which is suitable for processing and predicting important events with relatively long intervals and delays in time series.
  • the "word vector 1" corresponding to "Zhang” can be input into the LSTM network first, and the hidden features of the "word vector 1" can be extracted through the LSTM network, and the hidden vector at this time, such as time t, can be output. Then, the hidden vector at time t can be spliced with the "word vector 2" corresponding to "three" at time t+1, and the spliced vector is input into the LSTM network, and the hidden features of the spliced vector are extracted and output The implicit vector at time t+1.
  • the word vector at the current moment can be sequentially spliced with the hidden vector delivered at the previous moment, and the feature extraction of the spliced vector is performed through the LSTM network, until finally the "word vector 8" corresponding to "day” is input into the LSTM network , and use the hidden vector output at the last moment to represent all the text, the hidden vector is the vector representation of the patient "Zhang San", for example, it can be recorded as pat j .
  • the GRU (Gated Recurrent Unit) network can also be used to operate on the word vector corresponding to each word.
  • the structure of the GRU network is simpler than that of the LSTM network, and the implementation effect is the same as that of the LSTM network.
  • the vector dimension of the patient vector can be the same as that of the doctor vector, for example, the doctor can be Mapping with patients is a 256-dimensional vector. It can be understood that the number of vector dimensions of the doctor vector and the patient vector is only illustrative. In other examples, both the doctor and the patient may be mapped to 128-dimensional vectors, which is not specifically limited in the present disclosure.
  • step S320 each doctor in the doctor knowledge graph is vectorized to obtain multiple doctor vectors.
  • the relationship between doctors can be used for modeling.
  • multiple doctors and the relationship between multiple doctors can be used to construct a doctor knowledge map.
  • mapping the doctors to be recommended in the doctor knowledge map into vectors and performing doctor-patient matching in the corresponding vector space it can be Patients recommend more symptomatic doctors, thereby improving the patient experience.
  • the knowledge graph is a graph-based data structure that can be composed of nodes and edges.
  • nodes can represent entities or concepts
  • edges can be composed of attributes or relationships.
  • a knowledge graph is a relational network that can connect different types of information together, and it is a way to effectively represent relationships. Therefore, using the knowledge map can analyze the problem from the perspective of "relationship".
  • the doctor knowledge graph may be a knowledge graph composed of multiple doctors (nodes) and relationships (edges) between multiple doctors. Referring to FIG. 5 , a doctor knowledge graph can be constructed according to steps S510 to S530.
  • Step S510 Acquire corpus information of multiple doctors.
  • a crawler can be used to crawl the corpus information of multiple doctors from the Internet, which can include information such as the doctor's name, age, graduate school, work location, field of expertise, published articles, etc.
  • Exemplary can utilize webpage crawler to realize the mining of doctor's information
  • webpage crawler refers to writing crawler script to obtain doctor's information
  • basic workflow can include: First, can select some URL (Uniform Resource Locator, Uniform Resource Locator) as seed Put the URL into the queue to be crawled, and then write a crawler script.
  • the script tool can be used to quickly and conveniently obtain the corpus information of multiple doctors.
  • the corpus information of multiple doctors may also be directly acquired through manual input or copy input.
  • Step S520 Perform semantic analysis on the corpus information through natural language processing, and extract the relationship among the multiple doctors.
  • the semantic analysis of the corpus information of multiple doctors can be performed by using natural language processing, and the relationship among the multiple doctors can be extracted.
  • Natural language processing means that the computer can accept the user's input in the form of natural language, and internally perform a series of operations such as processing and calculation through the algorithm defined by humans, so as to simulate human understanding of natural language and return the results expected by the user.
  • unsupervised learning can be used for clustering to realize relation extraction.
  • Semi-supervised learning can also be used to achieve relationship extraction. For example, a part of corpus information can be selected for labeling, and the labeled corpus information can be iterated. You can also use supervised learning to classify and perform a large number of labels to achieve relationship extraction, or you can also extract relationships by training an end-to-end labeling model based on deep learning end-to-end joint labeling.
  • five types of relationships among multiple doctors can be extracted, which can be teacher-student, colleague, colleague, alumni, and collaborator.
  • doctors A and Doctor B when they have the same teacher or the same colleagues, it can indicate that the diseases they are good at treating may be similar.
  • doctor A is good at treating pulp diseases, and there is a high probability that his students, seniors, or colleagues are also good at treating pulp diseases.
  • Step S530 Taking the plurality of doctors as entities, constructing the doctor knowledge graph according to the relationships among the plurality of doctors.
  • a doctor knowledge map can be constructed.
  • the entities of the doctor knowledge map can be various doctors, and the relationship between entities can include many kinds, for example, 5 kinds of relationships, namely teachers and students, peers, colleagues, alumni and collaborators .
  • a doctor knowledge graph of a directed graph model can be constructed.
  • the entities corresponding to the five nodes in the doctor knowledge graph are doctor A, doctor B, doctor C, doctor D and doctor E.
  • the associated edge includes an outgoing edge and an incoming edge, which can be associated with node doctor B through the outgoing edge, and can be associated with node doctor C through the incoming edge.
  • node doctor C can be called the parent node of node doctor A.
  • doctor C can be the teacher of doctor A
  • node doctor B can be called the child node of node doctor A
  • node doctor D can be the child node of node doctor B and node doctor C at the same time
  • node doctor D can be the parent node of node doctor E.
  • the constructed doctor knowledge graph can be stored for subsequent access and recall. Therefore, the doctor's knowledge map can be constructed in real time every time an intelligent triage is performed, or it can be pre-built and stored in the database. For example, the doctor's knowledge map can be stored in Neo4j (a high-performance NoSQL graph database) The database is called when recommending a symptomatic doctor for a patient.
  • Neo4j a high-performance NoSQL graph database
  • the doctors to be recommended in the doctor knowledge graph can be mapped into vectors, so as to facilitate doctor-patient matching in the corresponding vector space.
  • each doctor in the doctor knowledge graph can be vectorized by using a graph neural network (Graph Neural Networks, GNN) to obtain multiple doctor vectors.
  • the graph neural network can combine the doctor's knowledge map with the neural network, and perform end-to-end calculations on the doctor's knowledge map. The entire calculation process can be carried out along the structure of the doctor's knowledge map, which can not only retain the structure of the doctor's knowledge map, but also learn the structural information of the doctor's knowledge map.
  • a function mapping can be learned, by mapping the node e i in the doctor’s knowledge graph, the features of the node e i can be aggregated with the features of the neighbor nodes e j (such as the parent node) and e k (such as the child node) to generate the node e i
  • a new representation of that is, the hidden state h(e i ) of each node can be obtained. It can be seen that for each node's hidden state, information from neighboring nodes can be included.
  • the graph neural network can generate a new representation of node e i by iteratively updating the hidden states of all nodes.
  • graph neural network can be used to obtain representations of multiple nodes, that is, multiple doctor vectors can be obtained.
  • Step S710 Initialize the node vector corresponding to each doctor in the doctor knowledge graph.
  • Step S720 Using the pre-trained graph neural network to iteratively update each node vector to obtain the multiple doctor vectors.
  • Each node vector is updated to obtain the plurality of doctor vectors.
  • W p , W ph , W c , and W ch are the parameters in the graph neural network
  • ⁇ is the activation function in the graph neural network.
  • the activation function can be Leaky ReLU (Leaky Rectified Linear Unit, with a leaky linear rectification function)
  • t is the number of network iterations
  • e i is the node corresponding to the i-th doctor in the doctor knowledge map
  • Np(e i ) is the set of parent nodes of node e i , correspondingly, e k is the k-th parent of node e i Node
  • Nc(e i ) is the set of child nodes of node e i
  • e j is the jth child node of node e i
  • h t (e i ) represents the vector representation of node e i after t network iterations , can be recorded as doc i
  • the matching degree calculation model is used to calculate the matching degree between the target patient-seeing patient vector and the plurality of doctor vectors, and recommend a doctor for the target doctor-seeking patient according to the matching degree.
  • the vector representation corresponding to the patient can be obtained through the trained LSTM network, and then the matching degree between the patient and each doctor can be calculated through the trained matching calculation model , and sorted according to the size of the matching degree.
  • the matching degree calculation model it can be based on
  • [doc i , pat j ] is the vector spliced by the i-th doctor vector doc i and the target patient vector pat j
  • W, v and b are the matching parameters of the matching calculation model, for example, W can be 256* 512 parameter matrix, v, b can be a 256-dimensional column vector, or a 256-dimensional row vector. It should be noted that when both v and b are row vectors or column vectors, v or b can be transposed. And when v and b are not both row vectors or column vectors, transposition may not be performed.
  • ⁇ is the activation function in the matching degree calculation model, such as Leaky ReLU (Leaky Rectified Linear Unit, with a leaky linear rectification function).
  • the matching degree score range may be [0, 100], and a larger matching degree score may indicate that a doctor is better at handling a patient's condition.
  • the matching calculation model, graph neural network and LSTM network can be pre-trained.
  • the vector representation corresponding to each doctor and the vector representation of the target patient can be obtained, and each doctor can be The corresponding vector representation is stored in the database.
  • it can be stored in a Redis database or a MySQL database, and each doctor and the corresponding vector representation can be queried and obtained in real time.
  • Redis is a key-value storage system.
  • the Redis database can include: a key-value pair (key-value) formed by a doctor ID and a corresponding vector representation, wherein the key (key) is the doctor ID, and the value ( value) is the corresponding vector representation.
  • Redis can support more than 100K+ read and write frequencies per second, and has certain advantages in data reading and storage speed.
  • MySQL is a relational database management system.
  • the relational database stores data in different tables instead of storing all data in a unified manner, which increases storage speed and flexibility. It has stable advantages in data storage and can avoid data occurrence. lost.
  • the matching degree calculation model, the graph neural network and the LSTM network are trained according to steps S810 and S820 to obtain corresponding matching parameters and neural network parameters, so as to use the trained matching degree to calculate The model performs doctor-patient matching.
  • Step S810 Train the matching calculation model, the graph neural network and the neural network to obtain corresponding matching parameters and neural network parameters.
  • a training data set may be obtained, and the training data set may include a positive training data set and a negative training data set.
  • the negative training data set can be ⁇ (doc j , pat i ) ⁇ , which means that the jth doctor is not suitable for treating the ith patient's condition, that is, the jth doctor is not a symptomatic doctor for the ith patient.
  • the negative training data set ⁇ ( doc j , pat i ) ⁇ ⁇ ( doc j , pat i ) ⁇ .
  • 5 negative training data (doc j , pat i ) can be generated by random replacement. It can be understood that the number of negative training data generated by randomly replacing each positive training data is only illustrative, and any number of negative training data can be obtained, and combined with positive training data, the matching degree calculation model can be multiplied. training times to improve the performance of the matching calculation model.
  • the training data set can be input into the matching calculation model for training.
  • the matching parameter may be a parameter used to define a mapping relationship between doctor vectors and patient vectors in the matching calculation model.
  • the objective function can also be called the loss function, which is a performance function in the matching degree calculation model and can be used to measure the degree of inconsistency between the predicted value of the model and the real value.
  • the matching score of patient pat i and symptomatic doctor doc i can be the highest, and is at least ⁇ higher than the matching score of patient pat i and other doctors doc j , ⁇ >0 as the target to train the matching degree calculation model.
  • the corresponding objective function can be:
  • (doc i , pat i ) is the positive training data
  • (doc j , pat i ) is the negative training data
  • ⁇ indicates the matching degree score(doc i , pat i ) of the positive training data and the matching degree of the negative training data
  • score(doc j , pat i )+ ⁇ -score(doc i , pat i ) ⁇ 0 it can represent the relationship between training data (doc j , pat i ) and training data (doc i , pat i )
  • the difference of the matching score satisfies at least a high ⁇ , indicating that there is no loss in the matching calculation model at this time.
  • score(doc j , pat i )+ ⁇ -score(doc i , pat i )>0 it can represent the matching degree between training data (doc j , pat i ) and training data (doc i , pat i )
  • a stochastic gradient descent algorithm may be used to update the matching parameters of the matching calculation model.
  • the objective function is continuously calculated, and the parameters of the matching degree calculation model are updated according to the objective function.
  • the objective function may also be minimized by alternate least squares method, Adam optimization algorithm, etc., and the matching parameters are updated sequentially from the back to the front to optimize the matching parameters.
  • the training process of the graph neural network can be based on
  • the network parameters are updated iteratively in reverse.
  • the training of the neural network parameters is completed, and the final doctor vector corresponding to each node can be obtained.
  • the preset number of iterations t may be 20, and the graph neural network is constantly updating network parameters during the 20 reverse iterations.
  • the final network parameters and the final vector representation of each node can be obtained.
  • the training parameters can also be determined according to the objective function L. For example, when the objective function L converges to the minimum value, the network parameters in the graph neural network can be obtained, so as to determine the final vector representation of each node according to the network parameters.
  • the training process of the LSTM network can use the backpropagation algorithm to initialize parameters randomly, and continuously update the parameters as the training progresses.
  • the BP (error Back Propagation) algorithm can be used.
  • the original input can be calculated sequentially from front to back to obtain the output of the output layer.
  • the objective function can be calculated.
  • L the objective function
  • the parameters in the LSTM network can be obtained, so as to determine the final vector representation of the target patient according to the parameters.
  • the gradient descent algorithm, Adam optimization algorithm, etc. can also be used to minimize the objective function L, and the parameters in the LSTM network are sequentially updated from the back to the front.
  • Step S820 Based on the matching parameters and the neural network parameters, calculate the matching score between the target patient vector and each doctor vector.
  • the target patient vector and each doctor vector can be obtained correspondingly.
  • the matching parameter, the target patient vector and each doctor vector can be used to calculate the matching score between the target patient vector and each doctor vector.
  • the parameters in the matching degree calculation model, LSTM network and graph neural network can be trained at the same time.
  • L the objective function
  • the three models of the matching degree calculation model, LSTM network and graph neural network can be trained at the same time, ensuring higher precision and accuracy of each model, and improving training efficiency at the same time.
  • the matching scores of the target patient and multiple doctors may be sorted, for example, sorted in descending order.
  • the top-ranked doctors can be formed into a set of doctors to be recommended, for example, the set of doctors to be recommended can be obtained from the top five doctors. It is also possible to select doctors whose matching score is greater than the preset matching threshold to obtain a set of doctors to be recommended. For example, when the match score between the target patient and the doctor is greater than 80 points, the doctor can be regarded as the doctor to be recommended and added to the set of doctors to be recommended.
  • the set of doctors to be recommended can be recommended to the target patient.
  • the information of each doctor in the set of doctors to be recommended can be output to the terminal device for the patient to select a target doctor from the set of doctors to be recommended.
  • the patient can input specific filtering conditions according to their own needs to perform secondary filtering on each doctor in the set of doctors to be recommended.
  • doctors in the set of doctors to be recommended who exceed the distance required by the patient may be filtered out.
  • each doctor in the set of recommended doctors can also be filtered according to the cost, and the final doctor can be recommended for the patient.
  • the doctor in the doctor knowledge map can be mapped to a vector of the specified dimension through the graph neural network, and the patient's condition information can be mapped to a vector of the same dimension through the long short-term memory network, and the doctor can be calculated in the vector space.
  • the degree of matching with the patient so as to recommend the symptomatic doctor for the patient. It can avoid the inconsistency between the doctor's label stored in the database and the doctor's actual professional ability, and improve the accuracy of doctor-patient matching.
  • the intelligent triage method by obtaining the condition information of the target patient, and obtaining the vector of the target patient according to the condition information; vectorizing each doctor in the doctor knowledge graph to obtain multiple Doctor vector: using a matching calculation model to calculate the matching degree between the target patient-seeing patient vector and the multiple doctor vectors, and recommending a doctor for the target patient-seeing patient according to the matching degree.
  • a matching calculation model to calculate the matching degree between the target patient-seeing patient vector and the multiple doctor vectors, and recommending a doctor for the target patient-seeing patient according to the matching degree.
  • an intelligent triage device is also provided.
  • the device can be applied to a server or terminal equipment.
  • the intelligent triage device 900 may include a patient vector acquisition module 910, a doctor vector acquisition module 920 and a vector matching module 930, wherein:
  • the patient vector acquisition module 910 is used to acquire the condition information of the target patient, and obtain the target patient vector according to the condition information;
  • the doctor vector acquisition module 920 is used to vectorize each doctor in the doctor knowledge map to obtain multiple doctor vectors;
  • the vector matching module 930 is configured to use a matching calculation model to calculate the matching degree between the target patient-seeing patient vector and the plurality of doctor vectors, and recommend a doctor for the target patient-seeing patient according to the matching degree.
  • condition information includes text information
  • patient vector acquisition module 910 includes:
  • An information encoding module configured to encode each text in the text information to obtain multiple word vectors
  • the patient vector determination module is used to sequentially input the multiple word vectors into the pre-trained neural network to obtain the target patient vector.
  • the intelligent triage device 900 also includes:
  • the corpus information acquisition module is used to acquire the corpus information of multiple doctors
  • the doctor relationship extraction module is used to perform semantic analysis on the corpus information through natural language processing, and extract the relationship between the plurality of doctors;
  • the doctor knowledge map construction module is used to construct the doctor knowledge map according to the relationships among the multiple doctors by using the multiple doctors as entities.
  • the doctor vector acquisition module 920 includes:
  • the doctor vector acquisition sub-module is used to vectorize each doctor in the doctor knowledge map through a graph neural network to obtain multiple doctor vectors.
  • the doctor's vector acquisition submodule includes:
  • a vector initialization unit configured to initialize node vectors corresponding to each doctor in the doctor knowledge map
  • the doctor vector acquisition unit is configured to use the pre-trained graph neural network to iteratively update the respective node vectors to obtain the plurality of doctor vectors.
  • the doctor vector acquisition unit is configured to:
  • W p , W ph , W c , and W ch are the training parameters of the graph neural network
  • ⁇ is the activation function in the graph neural network
  • t is the number of network iterations
  • e i is the doctor's knowledge map.
  • Np(e i ) is the set of parent nodes of node e i
  • e k is the kth parent node of node e i
  • Nc(e i ) is the set of child nodes of node e i
  • e j is the jth child node of node e i
  • h t (e i ) represents the vector of node e i after t network iterations.
  • the vector matching module 930 is configured to
  • [doc i , pat j ] is the vector spliced by the i-th doctor vector doc i and the target patient vector pat j
  • W, v and b are matching parameters
  • ⁇ is the activation in the matching calculation model function.
  • the intelligent triage device 900 also includes:
  • the first model training module is used to train the matching calculation model, the graph neural network and the neural network to obtain corresponding matching parameters and neural network parameters;
  • a data calculation module configured to calculate a matching degree score between the target patient vector and each doctor vector based on the matching parameters and the neural network parameters.
  • the first model training module includes:
  • a data set acquiring unit configured to acquire a training data set, the training data set including a positive training data set and a negative training data set;
  • an objective function construction unit configured to input the training data set into the matching calculation model, and construct an objective function
  • a matching parameter determining unit configured to determine the matching parameters of the matching degree calculation model according to the objective function.
  • the matching parameter determination unit is configured to update the matching parameters of the matching calculation model by using a stochastic gradient descent algorithm, and when the objective function converges, complete the matching parameter training.
  • the objective function is configured as:
  • (doc i , pat i ) is positive training data, indicating that doctor i is suitable for treating patient i
  • (doc j , pat i ) is negative training data, indicating that doctor j is not suitable for treating patient i
  • ⁇ represents the preset difference threshold between the matching score score(doc i , pat i ) of the positive training data and the matching score score(doc j , pat i ) of the negative training data.
  • the first model training module is further configured to iteratively update the neural network parameters of the graph neural network using the backpropagation algorithm, and when the objective function converges, complete the The training of the neural network parameters.
  • the intelligent triage device 900 also includes:
  • a doctor to be recommended determination module is used to determine a set of doctors to be recommended according to the size of the matching degree
  • a target doctor recommendation module configured to recommend a doctor for the target patient based on the set of doctors to be recommended.
  • the target doctor recommendation module includes:
  • a doctor set recommendation module configured to recommend the set of doctors to be recommended to the target patient
  • a target doctor determination module configured to filter each doctor in the set of doctors to be recommended according to the filter conditions input by the target patient, so as to recommend a target doctor for the target patient.
  • Each module in the above-mentioned device can be a general-purpose processor, including: a central processing unit, a network processor, etc.; it can also be a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic devices, discrete gates or transistors Logic devices, discrete hardware components. Each module may also be implemented by software, firmware, and other forms. Each processor in the above device may be an independent processor, or may be integrated together.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Theoretical Computer Science (AREA)
  • General Business, Economics & Management (AREA)
  • Business, Economics & Management (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Pathology (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

一种智能分诊方法、装置、存储介质及电子设备，涉及人工智能技术领域。所述方法包括：获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量（S310）；将医生知识图谱中的各个医生向量化，得到多个医生向量（S320）；使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生（S330）。

Description

智能分诊方法、装置、存储介质及电子设备 技术领域
本公开涉及人工智能技术领域，具体而言，涉及一种智能分诊方法、智能分诊装置、计算机可读存储介质以及电子设备。
背景技术
随着互联网技术的发展，智能推荐系统已经深入人们的生活之中。例如，智能推荐系统可以为用户推荐符合用户需求的饮食、景点和交通规划等，以供用户进行选择。
目前，利用智能推荐系统为患者推荐诊疗医生时，通常是根据患者的自述症状在智能推荐系统中匹配相关度较高的医生标签，以根据检索到的医生标签为患者推荐对应的诊疗医生。
需要说明的是，在上述背景技术部分公开的信息仅用于加强对本公开的背景的理解，因此可以包括不构成对本领域普通技术人员已知的现有技术的信息。
发明内容
本公开提供一种智能分诊方法、智能分诊装置、计算机可读存储介质以及电子设备。
本公开提供一种智能分诊方法，包括：
获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；
将医生知识图谱中的各个医生向量化，得到多个医生向量；
使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
在本公开的一种示例性实施例中，所述病情信息包括文字信息；所述根据所述病情信息得到目标就诊患者向量，包括：
将所述文字信息中的每个文字进行编码，得到多个词向量；
将所述多个词向量依次输入预先训练好的神经网络中，得到所述目标就诊患者向量。
在本公开的一种示例性实施例中，所述方法还包括：
获取多个医生的语料信息；
对所述语料信息通过自然语言处理进行语义分析，提取所述多个医生之间的关系；
以所述多个医生为实体，根据所述多个医生之间的关系构建所述医生知识图谱。
在本公开的一种示例性实施例中，所述将医生知识图谱中的各个医生向量化，得到多个医生向量，包括：
通过图神经网络将所述医生知识图谱中的各个医生向量化，得到多个医生向量。
在本公开的一种示例性实施例中，所述通过图神经网络将所述医生知识图谱中的各个医生向量化，得到多个医生向量，包括：
初始化所述医生知识图谱中各个医生对应的节点向量；
利用预先训练好的所述图神经网络对所述各个节点向量进行迭代更新，以得到所述多个医生向量。
在本公开的一种示例性实施例中，所述利用预先训练好的所述图神经网络对所述节点向量进行迭代更新，以得到所述多个医生向量，包括：
根据
Figure PCTCN2021103444-appb-000001
更新所述各个节点向量，以得到所述多个医生向量；
其中，W p、W ph、W c、W ch为所述图神经网络的参数，σ为所述图神经网络中的激活函数，t为网络迭代次数，e i为所述医生知识图谱中第i个医生对应的节点，Np(e i)为节点e i的父节点集合，e k为节点e i的第k个父节点，Nc(e i)为节点e i的子节点集合，e j为节点e i的第j个子节点，h t(e i)表示在t次网络迭代后节点e i的向量。
在本公开的一种示例性实施例中，所述使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，包括：
根据
score(doc i，pat j)＝v Tσ(W[doc i，pat j]+b)
计算所述目标就诊患者向量与所述每个医生向量之间的匹配度得分；
其中，[doc i，pat j]为由第i个医生向量doc i和目标就诊患者向量pat j拼接后的向量，W、v和b为匹配参数，σ为所述匹配度计算模型中的激活函数。
在本公开的一种示例性实施例中，所述方法还包括：
对所述匹配度计算模型、所述图神经网络和所述神经网络进行训练，得到对应的匹配参数和神经网络参数；
基于所述匹配参数和所述神经网络参数，计算所述目标就诊患者向量与每个医生向量之间的匹配度得分。
在本公开的一种示例性实施例中，所述对所述匹配度计算模型进行训练，得到所述匹配度计算模型的匹配参数，包括：
获取训练数据集，所述训练数据集包括正训练数据集和负训练数据集；
将所述训练数据集输入所述匹配度计算模型中，并构建目标函数；
根据所述目标函数确定所述匹配度计算模型的匹配参数。
在本公开的一种示例性实施例中，所述根据所述目标函数确定所述匹配度计算模型的匹配参数，包括：
利用随机梯度下降算法对所述匹配度计算模型的匹配参数进行更新，当所述目标函数收敛时，完成对所述匹配参数的训练。
在本公开的一种示例性实施例中，所述目标函数为：
Figure PCTCN2021103444-appb-000002
其中，(doc i，pat i)为正训练数据，表示第i个医生适合处理第i个患者的病情，(doc j，pat i)为负训练数据，表示第j个医生不适合处理第i个患者的病情，γ表示正训练数据的匹配度得分score(doc i，pat i)与负训练数据的匹配度得分score(doc j，pat i)之间的预设差值阈值。
在本公开的一种示例性实施例中，所述对所述图神经网络进行训练，得到所述图神经网络的神经网络参数，包括：
利用反向传播算法对所述图神经网络的神经网络参数进行迭代更新，当所述目标函数收敛时，完成对所述神经网络参数的训练。
在本公开的一种示例性实施例中，所述根据所述匹配度的大小为所述目标就诊患者推荐医生，包括：
根据所述匹配度的大小确定待推荐医生集合；
基于所述待推荐医生集合为所述目标就诊患者推荐医生。
在本公开的一种示例性实施例中，所述基于所述待推荐医生集合为所述目标就诊患者推荐医生，包括：
向所述目标就诊患者推荐所述待推荐医生集合；
根据所述目标就诊患者输入的筛选条件对所述待推荐医生集合中的每个医生进行过滤，以为所述目标就诊患者推荐目标医生。
本公开提供一种智能分诊装置，包括：
患者向量获取模块，用于获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；
医生向量获取模块，用于将医生知识图谱中的各个医生向量化，得到多个医生向量；
向量匹配模块，用于使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
本公开提供一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现上述任意一项所述的方法。
本公开提供一种电子设备，包括：处理器；以及存储器，用于存储所述处理器的可执行指令；其中，所述处理器配置为经由执行所述可执行指令来执行上述任意一项所述的方法。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，并不能限制本公开。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本公开的实施例， 并与说明书一起用于解释本公开的原理。显而易见地，下面描述中的附图仅仅是本公开的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1示出了可以应用本公开实施例的一种智能分诊方法及装置的示例性系统架构的示意图；
图2示出了适于用来实现本公开实施例的电子设备的计算机系统的结构示意图；
图3示意性示出了根据本公开的一个实施例的智能分诊方法的流程图；
图4示意性示出了根据本公开的一个实施例的获取目标就诊患者向量的流程图；
图5示意性示出了根据本公开的一个实施例的构建医生知识图谱的流程图；
图6示意性示出了根据本公开的一个实施例的医生知识图谱的示意图；
图7示意性示出了根据本公开的一个实施例的获取医生向量的流程图；
图8示意性示出了根据本公开的一个实施例的进行医患匹配的流程图；
图9示意性示出了根据本公开的一个实施例的智能分诊装置的框图。
具体实施方式
现在将参考附图更全面地描述示例实施方式。然而，示例实施方式能够以多种形式实施，且不应被理解为限于在此阐述的范例；相反，提供这些实施方式使得本公开将更加全面和完整，并将示例实施方式的构思全面地传达给本领域的技术人员。所描述的特征、结构或特性可以以任何合适的方式结合在一个或更多实施方式中。在下面的描述中，提供许多具体细节从而给出对本公开的实施方式的充分理解。然而，本领域技术人员将意识到，可以实践本公开的技术方案而省略所述特定细节中的一个或更多，或者可以采用其它的方法、组元、装置、步骤等。在其它情况下，不详细示出或描述公知技术方案以避免喧宾夺主而使得本公开的各方面变得模糊。
此外，附图仅为本公开的示意性图解，并非一定是按比例绘制。图中相同的附图标记表示相同或类似的部分，因而将省略对它们的重复描述。附图中所示的一些方框图是功能实体，不一定必须与物理或逻辑上独立的实体相对应。可以采用软件形式来实现这些功能实体，或在一个或多个硬件模块或集成电路中实现这些功能实体，或在不同网络和/或处理器装置和/或微控制器装置中实现这些功能实体。
图1示出了可以应用本公开实施例的一种智能分诊方法及装置的示例性应用环境的系统架构的示意图。
如图1所示，智能分诊系统的系统架构100可以包括终端设备101、102、103中的一个或多个，网络104和服务器105。网络104用以在终端设备101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型，例如有线、无线通信链路或者光纤电缆等等。终端设备101、102、103可以是各种电子设备，包括但不限于台式计算机、便携式计算机、智能手机和平板电脑等。应该理解，图1中的终端设备、网络和服务 器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端设备、网络和服务器。比如服务器105可以是多个服务器组成的服务器集群等。
本公开实施例所提供的智能分诊方法一般由服务器105执行，相应地，智能分诊装置一般设置于服务器105中，服务器为目标就诊患者匹配到多个对症的诊疗医生后可以将匹配结果发送至终端设备，并由终端设备向患者进行展示，以供患者进行选择。但本领域技术人员容易理解的是，本公开实施例所提供的智能分诊方法也可以由终端设备101、102、103中的一个或多个执行，相应的，智能分诊装置也可以设置于终端设备101、102、103中，例如，由终端设备执行后可以将匹配结果直接显示在终端设备的显示屏上，也可以通过语音播报的方式将匹配结果提供给患者，本示例性实施例中对此不做特殊限定。
图2示出了适于用来实现本公开实施例的电子设备的计算机系统的结构示意图。
需要说明的是，图2示出的电子设备的计算机系统200仅是一个示例，不应对本公开实施例的功能和使用范围带来任何限制。
如图2所示，计算机系统200包括中央处理单元(CPU)201，其可以根据存储在只读存储器(ROM)202中的程序或者从存储部分208加载到随机访问存储器(RAM)203中的程序而执行各种适当的动作和处理。在RAM 203中，还存储有系统操作所需的各种程序和数据。CPU 201、ROM 202以及RAM 203通过总线204彼此相连。输入/输出(I/O)接口205也连接至总线204。
以下部件连接至I/O接口205：包括键盘、鼠标等的输入部分206；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分207；包括硬盘等的存储部分208；以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分209。通信部分209经由诸如因特网的网络执行通信处理。驱动器210也根据需要连接至I/O接口205。可拆卸介质211，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器210上，以便于从其上读出的计算机程序根据需要被安装入存储部分208。
在一些实施例中，可以由电子设备的处理器执行本公开中所述的智能分诊方法。在一些实施例中，可以通过输入部分206输入目标就诊患者的病情信息、多个医生的语料信息，以及用于构建和训练匹配度计算模型的训练数据集，例如，通过电子设备的用户交互界面输入目标就诊患者的病情信息、多个医生的语料信息等。在一些实施例中，可以通过输出部分207将目标就诊患者与多个医生的匹配度得分输出。
特别地，根据本公开的实施例，下文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信部分209从网络上被下载和安装，和/或从可拆卸介质211被安装。在该计算机程序被中央处理单元(CPU)201执行时，执行本申请的方法和装置中限定的各种功能。
作为另一方面，本申请还提供了一种计算机可读介质，该计算机可读介质可以是上述 实施例中描述的电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被一个该电子设备执行时，使得该电子设备实现如下述实施例中所述的方法。例如，所述的电子设备可以实现如图3至图5，以及图7和图8所示的各个步骤等。
需要说明的是，本公开所示的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开中，计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：无线、电线、光缆、RF等等，或者上述的任意合适的组合。
以下对本公开实施例的技术方案进行详细阐述：
目前，利用智能推荐系统为患者推荐诊疗医生时，通常是根据患者的自述症状在智能推荐系统中匹配相关度高的医生标签，以根据检索到的医生标签为患者推荐对应的诊疗医生。但是，当智能推荐系统中存储的医生标签与医生的实际专业能力不一致时，即使智能推荐系统甄选出的诊疗医生与患者的匹配度较高，也可能存在推荐的诊疗医生并不擅长治疗患者疾病的情况，进而降低患者的就诊体验。
示例性的，通过增加或者模糊化医生标签，可以提升医生被检索到的几率。但是，增加或者模糊化医生标签可能会引入噪声，影响医患匹配的准确性。例如，对于一个擅长牙髓治疗的医生，在数据库中该医生的一个标签可能为“擅长口腔疾病”，此时该医生和一个需要阻生牙拔除手术的患者的匹配度会升高(因为阻生牙拔除属于口腔科疾病)。可以看出，对于需要阻生牙拔除手术的患者来说，擅长牙髓治疗的医生并不是对症的诊疗医生，在该患者就诊过程中可能降低患者的就诊体验。
基于上述一个或多个问题，本示例实施方式提供了一种智能分诊方法，该方法可以应用于上述服务器105，也可以应用于上述终端设备101、102、103中的一个或多个，本示例性实施例中对此不做特殊限定。参考图3所示，该智能分诊方法可以包括以下步骤S310至步骤S330：
步骤S310.获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；
步骤S320.将医生知识图谱中的各个医生向量化，得到多个医生向量；
步骤S330.使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
在本公开示例实施方式所提供的智能分诊方法中，通过获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；将医生知识图谱中的各个医生向量化，得到多个医生向量；使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。本公开通过将医生知识图谱中的待推荐医生映射成向量，并在对应的向量空间进行医患匹配时，可以为患者推荐更加对症的诊疗医生，进而提高患者的就诊体验。
下面，对于本示例实施方式的上述步骤进行更加详细的说明。
在步骤S310中，获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量。
本示例实施方式中，目标就诊患者可以是急诊前需要分诊的患者，也可以是正常门诊需要分诊的患者。例如，对于急诊前需要分诊的患者，医院工作人员可以利用智能分诊系统为该患者进行快速分诊，以便于安排患者尽快就诊。对于正常门诊需要分诊的患者，可以是医院工作人员利用智能分诊系统为该患者进行分诊，也可以是患者自行利用智能分诊系统进行分诊，本示例中对此不作具体限定。其他示例中，还可以是患者在就医前自行利用智能分诊系统确定目标医院和对症的诊疗医生。
利用智能分诊系统进行分诊时，可以获取目标就诊患者的病情信息。其中，目标就诊患者的病情信息可以包括该患者的基本信息如姓名、年龄等信息，也可以包括该患者自述的病情症状如咳嗽、发烧等信息，还可以包括该患者的既往病历如病史、用药等信息，本示例中对此不作具体限定。示例性的，可以由医院工作人员将目标就诊患者的病情信息输入智能分诊系统，也可以由目标就诊患者自行将自身的病情信息输入智能分诊系统。其中，医院工作人员或目标就诊患者可以通过手动输入病情信息，也可以通过语音输入病情信息，本示例中对此不作具体限定。
获取目标就诊患者的病情信息后，可以根据该患者的病情信息将该患者向量化，得到该患者的向量表示，以将患者向量与医生向量进行匹配，从而为该患者确定对症的诊疗医生。参考图4所示，可以根据步骤S410和步骤S420得到目标就诊患者的向量表示。
在步骤S410中.将所述文字信息中的每个文字进行编码，得到多个词向量。
以目标就诊患者自行利用智能分诊系统进行分诊为例，该患者的病情信息可以包括文字信息，也可以包括语音信息，还可以包括图像和视频信息。一种示例中，该患者可以将其基本信息、自述的病情症状以及既往病历等信息语音输入智能分诊系统中，智能分诊系统可以将接收到的语音信息映射为对应的文字信息，如“张三，20岁，牙疼三天”。其他 示例中，该患者也可以将其基本信息、自述的病情症状以及既往病历等信息手动输入智能分诊系统中，智能分诊系统可以直接获取该患者的文字信息。
获取目标就诊患者的文字信息如“张三，20岁，牙疼三天”后，可以将该患者的文字信息进行编码。例如，可以将“张三，20岁，牙疼三天”中的每个文字进行Embedding(向量映射)编码，将每个文字分别用一个低维向量表示，可以得到对应的多个字向量，如得到“张”的向量表示。也可以以词为单位，对词中包含的每个文字进行编码，得到对应的多个词向量，如得到“张三”的向量表示。示例性的，可以将每个文字进行One-Hot(独热)编码，One-Hot编码又称作一位有效编码，其方法是使用N位状态寄存器来对N个状态进行编码，每个状态都有独立的寄存器位，并且在任意时候，寄存器中只有一位有效。需要说明的是，One-Hot编码的向量维度会随着患者文字病情信息中文字数量的增加而增加，可能会增加计算复杂度。其他示例中，也可以利用稠密向量来表示每个文字。例如，可以利用Word2vec算法将获取的文字病情信息中的每个文字映射到向量空间中，每个文字可以用该向量空间中的一个字向量表示。类似的，还可以利用Doc2vec算法、Glove算法等将文字转化为向量。
步骤S420.将所述多个词向量依次输入预先训练好的神经网络中，得到所述目标就诊患者向量。
可以理解的是，就诊患者的文字病情信息中的各个症状之间是关联的，因此，本示例中可以将文字病情信息中所有文字对应的字向量看作一个时序序列，进而可以利用神经网络(例如递归神经网络)对每个文字对应的字向量进行运算。示例性的，得到“张三，20岁，牙疼三天”中每个文字对应的字向量后，可以将该八个字向量依次输入训练好的LSTM(Long Short-Term Memory，长短时记忆网络)网络中，以得到患者“张三”的向量表示。其中，LSTM网络是一种时间递归神经网络，适合于处理和预测时间序列中间隔和延迟相对较长的重要事件。
具体的，可以先将“张”对应的“字向量1”输入LSTM网络中，可以通过LSTM网络对“字向量1”进行隐含特征的提取，输出该时刻如t时刻的隐含向量。然后，可以将t时刻的隐含向量和t+1时刻“三”对应的“字向量2”进行拼接，并将拼接向量输入LSTM网络中，并对该拼接向量进行隐含特征的提取，输出t+1时刻的隐含向量。类似的，可以依次将当前时刻的字向量与上一时刻传递下来的隐含向量进行拼接，通过LSTM网络对拼接向量进行特征提取，直到最后将“天”对应的“字向量8”输入LSTM网络中，并使用最后时刻输出的隐含向量来表示所有文字，该隐含向量即为患者“张三”的向量表示，如可以记为pat j。其他示例中，也可以利用GRU(Gated Recurrent Unit，门控制循环单元)网络对每个文字对应的字向量进行运算，GRU网络的结构比LSTM网络简单，且实现效果与LSTM网络相同。
需要说明的是，本公开实施例中为了进行医患匹配，即计算患者向量和医生向量之间的匹配度，优选的，患者向量的向量维度可以与医生向量的向量维度相同，如可以将医生 与患者均映射为256维的向量。可以理解的是，医生向量与患者向量的向量维度的数目仅仅是示意性的，其他示例中，也可以将医生与患者均映射为128维的向量，本公开对此不作具体限定。
在步骤S320中，将医生知识图谱中的各个医生向量化，得到多个医生向量。
为了避免智能推荐系统中存储的医生标签与医生的实际专业能力不一致的情形，可以利用医生之间的关系进行建模。示例性的，可以利用多个医生以及多个医生之间的关系构建医生知识图谱，通过将医生知识图谱中的待推荐医生映射成向量，并在对应的向量空间进行医患匹配时，可以为患者推荐更加对症的诊疗医生，进而提高患者的就诊体验。
其中，知识图谱是一种基于图的数据结构，可以由节点和边组成。在知识图谱中，节点可以表示实体或概念，边可以由属性或关系构成。知识图谱是一种可以将不同种类的信息连接在一起得到的关系网络，是一种可以将关系有效表示的方式。因此，利用知识图谱可以从“关系”的角度去分析问题。对应的，该示例中，医生知识图谱可以是一种由多个医生(节点)以及多个医生之间的关系(边)组成的知识图谱。参考图5所示，可以根据步骤S510至步骤S530构建医生知识图谱。
步骤S510.获取多个医生的语料信息。
一种示例实施方式中，可以利用爬虫从互联网爬取多个医生的语料信息，可以包括如医生的姓名、年龄、毕业院校、工作地点、擅长领域、发表文章等信息，本示例中对此不作具体限定。示例性的，可以利用网页爬虫实现医生信息的挖掘，网页爬虫是指编写爬虫脚本来获取医生信息，基本工作流程可以包括：首先，可以选取一些URL(Uniform Resource Locator，统一资源定位符)作为种子URL放入待爬取队列中，然后编写爬虫脚本，针对待爬取队列中的种子URL，可以模拟人工浏览的方式访问网站，将爬取到的包括医生信息的网页HTML(Hyper Text Markup Language，超文本标记语言)数据进行存储以及解析，并可以将解析得到的新链接作为下一层爬取的种子URL。该示例中，利用该脚本工具可以快速、便捷的获取多个医生的语料信息。在其他实施例中，也可以通过手动输入或者拷贝输入的方式直接获取多个医生的语料信息。
步骤S520.对所述语料信息通过自然语言处理进行语义分析，提取所述多个医生之间的关系。
获取多个医生的语料信息后，可以利用自然语言处理对多个医生的语料信息进行语义分析，并提取该多个医生之间的关系。自然语言处理是指计算机可以接受用户自然语言形式的输入，并在内部通过人类所定义的算法进行加工、计算等系列操作，以模拟人类对自然语言的理解，并返回用户所期望的结果。示例性的，可以利用非监督学习进行聚类来实现关系抽取。也可以利用半监督学习来实现关系抽取，例如，可以选取一部分语料信息进行标注，将标注后的语料信息进行迭代。还可以利用监督学习进行分类并进行大量标注来实现关系抽取，或者还可以基于深度学习端到端的联合标注，通过训练一个端到端标注模型来抽取关系。
一种示例中，可以提取多个医生之间的5种关系，可以分别是师生、同门、同事、校友和合作者。例如，对于医生A和医生B，当二人有相同的老师或者有相同的同事时，可以说明二人擅长治疗的疾病可能类似。示例性的，医生A擅长治疗牙髓疾病，其学生、师兄或者同事存在较大概率也擅长治疗牙髓疾病。
步骤S530.以所述多个医生为实体，根据所述多个医生之间的关系构建所述医生知识图谱。
在该示例中，可以构建医生知识图谱，该医生知识图谱的实体可以是各个医生，实体之间的关系可以包括多种，例如5种关系，分别是师生、同门、同事、校友和合作者。参考图6所示，为了更直观的表示各个医生之间的关系，可以构建有向图模型的医生知识图谱。该医生知识图谱中5个节点对应的实体分别是医生A、医生B、医生C、医生D和医生E。其中，对于节点医生A来说，关联的边包括一条出边和一入边，可以通过出边与节点医生B关联，可以通过入边与节点医生C关联。本示例中，可以将节点医生C称为节点医生A的父节点，如医生C可以是医生A的老师，可以将节点医生B称为节点医生A的子节点。类似的，可以看出，节点医生D同时可以为节点医生B和节点医生C的子节点，节点医生D又可以为节点医生E的父节点。
可以将构建好的医生知识图谱进行存储以备后续访问和调用。因此，医生知识图谱可以是每次进行智能分诊时实时构建的，也可以是预先构建并存储于数据库中的，例如，可以将医生知识图谱存储于Neo4j(一种高性能的NoSQL图形数据库)数据库，在为患者推荐对症的诊疗医生时进行调用。
一种示例实施方式中，可以将医生知识图谱中的待推荐医生映射成向量，以便于在对应的向量空间进行医患匹配。示例性的，可以通过图神经网络(Graph Neural Networks，GNN)将医生知识图谱中的各个医生向量化，得到多个医生向量。其中，图神经网络可以将医生知识图谱与神经网络进行结合，在医生知识图谱上进行端对端的计算。整个计算过程，可以沿着医生知识图谱的结构进行，既可以保留医生知识图谱的结构，还可以对医生知识图谱的结构信息进行学习。例如，可以学习一个函数映射，通过映射医生知识图谱中的节点e i可以聚合节点e i的特征与邻居节点e j(如父节点)、e k(如子节点)的特征以生成节点e i的新表示，也就是可以获取每个节点的隐藏状态h(e i)。可以看出，对于每个节点的隐藏状态，可以包含来自邻居节点的信息。
示例性的，图神经网络可以通过迭代式更新所有节点的隐藏状态来生成节点e i的新表示。具体的，参考图7所示，可以根据步骤S710和步骤S720利用图神经网络得到多个节点的表示，也就是得到多个医生向量。
步骤S710.初始化所述医生知识图谱中各个医生对应的节点向量。
例如，医生知识图谱中可以包括N个医生节点{e i，i＝1，…，N}，M条边{r j，j＝1，…，M}，该示例中，以M＝5为例进行说明。可以理解的是，根据实现要求，M的取值可以是任意的。可以随机初始化图神经网络中的参数和医生知识图谱中各个医生 对应的节点向量，得到各个节点的初始向量h 0(e i)，i＝1，…，N。
步骤S720.利用预先训练好的所述图神经网络对所述各个节点向量进行迭代更新，以得到所述多个医生向量。
一种示例中，可以根据
Figure PCTCN2021103444-appb-000003
更新各个节点向量，以得到所述多个医生向量。
其中，W p、W ph、W c、W ch为图神经网络中的参数，σ为图神经网络中的激活函数，如激活函数可以为Leaky ReLU(Leaky Rectified Linear Unit，带泄露线性整流函数)，t为网络迭代次数，e i为医生知识图谱中第i个医生对应的节点，Np(e i)为节点e i的父节点集合，对应的，e k为节点e i的第k个父节点，Nc(e i)为节点e i的子节点集合，对应的，e j为节点e i的第j个子节点，h t(e i)表示在t次网络迭代后节点e i的向量表示，可以记为doc i，向量维度可以与患者向量的向量维度相同，如可以为256维。示例性的，预设的迭代次数t可以为20，即图神经网络经过20次迭代后可以得到各个节点的最终向量表示。
在步骤S330中，使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
一种示例中，在获取到目标就诊患者的病情信息后，可以通过训练好的LSTM网络获得该患者对应的向量表示，然后通过训练好的匹配度计算模型计算该患者和每个医生的匹配度，并按照匹配度的大小进行排序。
具体的，在匹配度计算模型中，可以根据
score(doc i，pat j)＝v Tσ(W[doc i，pat j]+b)
计算目标就诊患者向量pat j与医生向量doc i之间的匹配度得分。其中，[doc i，pat j]为由第i个医生向量doc i和目标就诊患者向量pat j拼接后的向量，W、v和b为匹配度计算模型的匹配参数，如W可以为256*512的参数矩阵，v、b可以为256维的列向量，也可以为256维的行向量。需要说明的是，当v、b均为行向量或者列向量时，可以将v或b进行转置。而当v、b不均为行向量或者列向量时，可以不用进行转置。σ为匹配度计算模型中的激活函数，如可以为Leaky ReLU(Leaky Rectified Linear Unit，带泄露线性整流函数)。示例性的，匹配度得分区间可以为[0，100]，匹配度得分越大，可以表示医生更擅长处理患者的病情。
另一种示例中，可以对匹配度计算模型、图神经网络和LSTM网络预先进行训练，训练完成后，可以得到每个医生对应的向量表示和目标就诊患者的向量表示，并可以将每个医生对应的向量表示存储到数据库中。例如，可以存储在Redis数据库中，也可以存储在MySQL数据库中，可以实时查询并获取每个医生和对应的向量表示。其中，Redis是一个key-value存储系统，存储在Redis数据库中时，可以包括：医生标识和对应的向量表示形成的键值对(key-value)，其中键(key)为医生标识，值(value)为对应的向量表示。Redis作为 一个高效的缓存技术，Redis能支持超过100K+每秒的读写频率，在数据读取以及存储的速度上具备一定的优势。MySQL是一种关联数据库管理系统，关联数据库将数据保存在不同的表中，而不是将所有数据进行统一存储，增加存储速度并提高灵活性，在数据存储方面具有稳定的优势，可以避免数据发生丢失。
示例性的，可以参考图8所示，根据步骤S810和步骤S820对匹配度计算模型、图神经网络和LSTM网络进行训练，得到对应的匹配参数和神经网络参数，以使用训练好的匹配度计算模型进行医患匹配。
步骤S810.对所述匹配度计算模型、所述图神经网络和所述神经网络进行训练，得到对应的匹配参数和神经网络参数。
在训练匹配度计算模型时，可以获取训练数据集，该训练数据集可以包括正训练数据集和负训练数据集。其中，正训练数据集可以为{(doc i，pat i)}，i＝1，…，K，表示第i个医生适合处理第i个患者的病情，即第i个医生为第i个患者对症的诊疗医生。负训练数据集可以为{(doc j，pat i)}，表示第j个医生不适合处理第i个患者的病情，即第j个医生不是为第i个患者对症的诊疗医生。一种示例中，可以通过将(doc i，pat i)中的第i个医生向量doc i随机替换成其他医生向量doc j(j≠i)来获得负训练数据集{(doc j，pat i)}。示例性的，对于每个正训练数据(doc i，pat i)，可以通过随机替换的方法生成5个负训练数据(doc j，pat i)。可以理解的是，将每个正训练数据通过随机替换的方法生成负训练数据的数目仅仅是示意性的，可以获取任意数目的负训练数据，并结合正训练数据，对匹配度计算模型进行多次训练，以提高匹配度计算模型的性能。
获取训练数据集后，可以将该训练数据集输入匹配度计算模型中进行训练。在训练过程中，首先需要构建目标函数，以根据该目标函数确定匹配度计算模型的匹配参数。其中，匹配参数可以是该匹配度计算模型中用于定义医生向量与患者向量之间的映射关系的参数。目标函数也可以称作损失函数，是匹配度计算模型中的性能函数，可以用来估量模型的预测值与真实值的不一致程度。
一种示例中，可以以患者pat i与对症的诊疗医生doc i的匹配度得分为最高，且比患者pat i与其他医生doc j的匹配度得分至少高γ，γ＞0为目标训练该匹配度计算模型。示例性的，当γ＝1时，即score(doc i，pat i)＞score(doc j，pat i)+1，其中，i，j∈[1，K]，i≠j。
对应的目标函数可以为：
Figure PCTCN2021103444-appb-000004
其中，(doc i，pat i)为正训练数据，(doc j，pat i)为负训练数据，γ表示正训练数据的匹配度得分score(doc i，pat i)与负训练数据的匹配度得分score(doc j，pat i)之间的预设差值阈值。其中，当score(doc j，pat i)+γ-score(doc i，pat i)≤0时，可以表示训练数据(doc j，pat i)和训练数据(doc i，pat i)之间的匹配度得分的差值满足至少高γ，说明此时的匹配度计算模型没 有损失。当score(doc j，pat i)+γ-score(doc i，pat i)＞0时，可以表示训练数据(doc j，pat i)和训练数据(doc i，pat i)之间的匹配度得分的差值没有满足至少高γ，说明此时的匹配度计算模型存在损失，且score(doc j，pat i)+γ-score(doc i，pat i)得到的值越大，对应的损失越大。因此，当[score(doc j，pat i)+γ-score(doc i，pat i)] +＝0时，目标函数可以收敛至最小值。
示例性的，可以利用随机梯度下降算法对对匹配度计算模型的匹配参数进行更新。根据反向传播原理，不断计算目标函数，并根据目标函数更新匹配度计算模型的参数。当目标函数收敛至最小值时，完成对模型参数的训练，此时对应的匹配度计算模型参数即为匹配参数。其他示例中，也可以交替最小二乘法、Adam优化算法等最小化目标函数，从后向前依次更新匹配参数，对匹配参数进行优化。
示例性的，图神经网络的训练过程可以根据
Figure PCTCN2021103444-appb-000005
反向迭代式更新网络参数，当满足预设的迭代次数时，完成对神经网络参数的训练，可以得到最终各个节点对应的医生向量。示例性的，预设的迭代次数t可以为20，图神经网络在进行20次反向迭代的过程中，一直在不断的更新网络参数。迭代完成后，可以得到最终的网络参数和各个节点最终的向量表示。其他示例中，也可以根据目标函数L确定训练参数，如当目标函数L收敛至最小值时，可以得到图神经网络中的网络参数，以根据该网络参数确定各个节点最终的向量表示。
一种示例中，LSTM网络的训练过程可以采用反向传播算法，通过随机初始化参数，随着训练的加深，不断更新参数。例如，可以采用BP(error Back Propagation，误差反向传播)算法，具体的，可以根据原始输入从前向后依次计算，得到输出层的输出，通过计算当前输出与目标输出的差距，即计算目标函数L。当目标函数L收敛至最小时，可以得到LSTM网络中的参数，以根据该参数确定目标就诊患者最终的向量表示。其他示例中，也可以利用梯度下降算法、Adam优化算法等最小化目标函数L，从后向前依次更新LSTM网络中的参数。
步骤S820.基于所述匹配参数和所述神经网络参数，计算所述目标就诊患者向量与每个医生向量之间的匹配度得分。
得到神经网络参数后，对应的可以得到目标就诊患者向量和每个医生向量。可以根据匹配度计算模型，利用匹配参数、目标就诊患者向量和每个医生向量计算该目标就诊患者向量和每个医生向量之间的匹配度得分。
在以上训练过程中，可以对匹配度计算模型、LSTM网络和图神经网络三者中的参数同时进行训练。例如，以L为目标函数，首先可以调整匹配度计算模型中的各个参数，由于匹配度计算模型的计算过程中需要用到医生向量和目标就诊患者向量，则进一步反向传播到LSTM网络和图神经网络，调整LSTM网络和图神经网络中的参数。通过多次一层一层的反向传播，最终可以使各模型参数均趋于收敛，或者满足一定迭代次数后训练终止。 通过这样的训练方式，可以同时对匹配度计算模型、LSTM网络和图神经网络三个模型进行训练，保证各模型的精度和准确度更高，同时可以提升训练效率。
一种示例中，在得到目标就诊患者和多个医生的匹配度得分后，可以将匹配度得分进行排序，如降序排序。可以将排名靠前的医生组成待推荐医生集合，如可以由排名前五的医生得到待推荐医生集合。也可以选取匹配度得分大于预设匹配度阈值的医生，得到待推荐医生集合。如当目标就诊患者和医生的匹配度得分大于80分时，可以将该医生作为待推荐医生，并添加至待推荐医生集合中。
该示例中，可以向目标就诊患者推荐该待推荐医生集合，如可以将待推荐医生集合中各个医生的信息输出至终端设备，以供患者从待推荐医生集合中选择目标诊疗医生。患者接收到待推荐医生集合中各个医生的信息后，可以根据自身需求输入特定的筛选条件对待推荐医生集合中的各个医生进行二次过滤。示例性的，当患者选择一定距离内的医院就诊时，可以将待推荐医生集合中超过患者要求距离的医生过滤掉。类似的，当患者选择一定额度的就诊费用时，也可以根据费用对待推荐医生集合中各个医生进行过滤，为患者推荐最终的诊疗医生。
该方法中，可以通过图神经网络将医生知识图谱中的医生映射成指定维度的向量，以及可以通过长短时记忆网络将患者的病情信息映射成相同维度的向量，并在该向量空间中计算医生和患者的匹配度，从而为患者推荐对症的诊疗医生。可以避免数据库中存储的医生标签和医生实际专业能力不一致的情形，提高医患匹配的准确性。
在本公开示例实施方式所提供的智能分诊方法中，通过获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；将医生知识图谱中的各个医生向量化，得到多个医生向量；使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。本公开通过将医生知识图谱中的待推荐医生映射成向量，并在对应的向量空间进行医患匹配时，可以为患者推荐更加对症的诊疗医生，进而提高患者的就诊体验。
应当注意，尽管在附图中以特定顺序描述了本公开中方法的各个步骤，但是，这并非要求或者暗示必须按照该特定顺序来执行这些步骤，或是必须执行全部所示的步骤才能实现期望的结果。附加的或备选的，可以省略某些步骤，将多个步骤合并为一个步骤执行，以及/或者将一个步骤分解为多个步骤执行等。
进一步的，本示例实施方式中，还提供了一种智能分诊装置。该装置可以应用于一服务器或终端设备。参考图9所示，该智能分诊装置900可以包括患者向量获取模块910、医生向量获取模块920以及向量匹配模块930，其中：
患者向量获取模块910，用于获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；
医生向量获取模块920，用于将医生知识图谱中的各个医生向量化，得到多个医生向量；
向量匹配模块930，用于使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
在一种可选的实施方式中，所述病情信息包括文字信息；患者向量获取模块910包括：
信息编码模块，用于将所述文字信息中的每个文字进行编码，得到多个词向量；
患者向量确定模块，用于将所述多个词向量依次输入预先训练好的神经网络中，得到所述目标就诊患者向量。
在一种可选的实施方式中，智能分诊装置900还包括：
语料信息获取模块，用于获取多个医生的语料信息；
医生关系提取模块，用于对所述语料信息通过自然语言处理进行语义分析，提取所述多个医生之间的关系；
医生知识图谱构建模块，用于以所述多个医生为实体，根据所述多个医生之间的关系构建所述医生知识图谱。
在一种可选的实施方式中，医生向量获取模块920包括：
医生向量获取子模块，用于通过图神经网络将所述医生知识图谱中的各个医生向量化，得到多个医生向量。
在一种可选的实施方式中，医生向量获取子模块，包括：
向量初始化单元，用于初始化所述医生知识图谱中各个医生对应的节点向量；
医生向量获取单元，用于利用预先训练好的所述图神经网络对所述各个节点向量进行迭代更新，以得到所述多个医生向量。
在一种可选的实施方式中，医生向量获取单元被配置为：
根据
Figure PCTCN2021103444-appb-000006
更新所述各个节点向量，以得到所述多个医生向量；
其中，W p、W ph、W c、W ch为所述图神经网络的训练参数，σ为所述图神经网络中的激活函数，t为网络迭代次数，e i为所述医生知识图谱中第i个医生对应的节点，Np(e i)为节点e i的父节点集合，e k为节点e i的第k个父节点，Nc(e i)为节点e i的子节点集合，e j为节点e i的第j个子节点，h t(e i)表示在t次网络迭代后节点e i的向量。
在一种可选的实施方式中，向量匹配模块930被配置为用于根据
score(doc i，pat j)＝v Tσ(W[doc i，pat j]+b)
计算所述目标就诊患者向量与所述每个医生向量之间的匹配度得分；
其中，[doc i，pat j]为由第i个医生向量doc i和目标就诊患者向量pat j拼接后的向量，W、v和b为匹配参数，σ为所述匹配度计算模型中的激活函数。
在一种可选的实施方式中，智能分诊装置900还包括：
第一模型训练模块，用于对所述匹配度计算模型、所述图神经网络和所述神经网络进 行训练，得到对应的匹配参数和神经网络参数；
数据计算模块，用于基于所述匹配参数和所述神经网络参数，计算所述目标就诊患者向量与每个医生向量之间的匹配度得分。
在一种可选的实施方式中，第一模型训练模块包括：
数据集获取单元，用于获取训练数据集，所述训练数据集包括正训练数据集和负训练数据集；
目标函数构建单元，用于将所述训练数据集输入所述匹配度计算模型中，并构建目标函数；
匹配参数确定单元，用于根据所述目标函数确定所述匹配度计算模型的匹配参数。
在一种可选的实施方式中，匹配参数确定单元被配置为用于利用随机梯度下降算法对所述匹配度计算模型的匹配参数进行更新，当所述目标函数收敛时，完成对所述匹配参数的训练。
在一种可选的实施方式中，所述目标函数被配置为：
Figure PCTCN2021103444-appb-000007
其中，(doc i，pat i)为正训练数据，表示第i个医生适合处理第i个患者的病情，(doc j，pat i)为负训练数据，表示第j个医生不适合处理第i个患者的病情，γ表示正训练数据的匹配度得分score(doc i，pat i)与负训练数据的匹配度得分score(doc j，pat i)之间的预设差值阈值。
在一种可选的实施方式中，第一模型训练模块还被配置为用于利用反向传播算法对所述图神经网络的神经网络参数进行迭代更新，当所述目标函数收敛时，完成对所述神经网络参数的训练。
在一种可选的实施方式中，智能分诊装置900还包括：
待推荐医生确定模块，用于根据所述匹配度的大小确定待推荐医生集合；
目标医生推荐模块，用于基于所述待推荐医生集合为所述目标就诊患者推荐医生。
在一种可选的实施方式中，目标医生推荐模块包括：
医生集合推荐模块，用于向所述目标就诊患者推荐所述待推荐医生集合；
目标医生确定模块，用于根据所述目标就诊患者输入的筛选条件对所述待推荐医生集合中的每个医生进行过滤，以为所述目标就诊患者推荐目标医生。
上述智能分诊装置中各模块的具体细节已经在对应的智能分诊方法中进行了详细的描述，因此此处不再赘述。
上述装置中各模块可以是通用处理器，包括：中央处理器、网络处理器等；还可以是数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。各模块也可以由软件、固件等形式来实现。上述装置中的各处理器可以是独立的处理器，也可以集成在一起。
应当注意，尽管在上文详细描述中提及了用于动作执行的设备的若干模块或者单元，但是这种划分并非强制性的。实际上，根据本公开的实施方式，上文描述的两个或更多模块或者单元的特征和功能可以在一个模块或者单元中具体化。反之，上文描述的一个模块或者单元的特征和功能可以进一步划分为由多个模块或者单元来具体化。
应当理解的是，本公开并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本公开的范围仅由所附的权利要求来限制。

Claims (17)

  1. 一种智能分诊方法，其特征在于，包括：
    获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；
    将医生知识图谱中的各个医生向量化，得到多个医生向量；
    使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
  2. 根据权利要求1所述的智能分诊方法，其特征在于，所述病情信息包括文字信息；所述根据所述病情信息得到目标就诊患者向量，包括：
    将所述文字信息中的每个文字进行编码，得到多个词向量；
    将所述多个词向量依次输入预先训练好的神经网络中，得到所述目标就诊患者向量。
  3. 根据权利要求1所述的智能分诊方法，其特征在于，所述方法还包括：
    获取多个医生的语料信息；
    对所述语料信息通过自然语言处理进行语义分析，提取所述多个医生之间的关系；
    以所述多个医生为实体，根据所述多个医生之间的关系构建所述医生知识图谱。
  4. 根据权利要求1所述的智能分诊方法，其特征在于，所述将医生知识图谱中的各个医生向量化，得到多个医生向量，包括：
    通过图神经网络将所述医生知识图谱中的各个医生向量化，得到多个医生向量。
  5. 根据权利要求4所述的智能分诊方法，其特征在于，所述通过图神经网络将所述医生知识图谱中的各个医生向量化，得到多个医生向量，包括：
    初始化所述医生知识图谱中各个医生对应的节点向量；
    利用预先训练好的所述图神经网络对所述各个节点向量进行迭代更新，以得到所述多个医生向量。
  6. 根据权利要求5所述的智能分诊方法，其特征在于，所述利用预先训练好的所述图神经网络对所述节点向量进行迭代更新，以得到所述多个医生向量，包括：
    根据
    Figure PCTCN2021103444-appb-100001
    更新所述各个节点向量，以得到所述多个医生向量；
    其中，W p、W ph、W c、W ch为所述图神经网络的参数，σ为所述图神经网络 中的激活函数，t为网络迭代次数，e i为所述医生知识图谱中第i个医生对应的节点，Np(e i)为节点e i的父节点集合，e k为节点e i的第k个父节点，Nc(e i)为节点e i的子节点集合，e j为节点e i的第j个子节点，h t(e i)表示在t次网络迭代后节点e i的向量。
  7. 根据权利要求1所述的智能分诊方法，其特征在于，所述使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，包括：
    根据
    score(doc i，pat j)＝v Tσ(W[doc i，pat j]+b)
    计算所述目标就诊患者向量与所述每个医生向量之间的匹配度得分；
    其中，[doc i，pat j]为由第i个医生向量doc i和目标就诊患者向量pat j拼接后的向量，W、v和b为匹配参数，σ为所述匹配度计算模型中的激活函数。
  8. 根据权利要求1-7任一项所述的智能分诊方法，其特征在于，所述方法还包括：
    对所述匹配度计算模型、所述图神经网络和所述神经网络进行训练，得到对应的匹配参数和神经网络参数；
    基于所述匹配参数和所述神经网络参数，计算所述目标就诊患者向量与每个医生向量之间的匹配度得分。
  9. 根据权利要求8所述的智能分诊方法，其特征在于，所述对所述匹配度计算模型进行训练，得到所述匹配度计算模型的匹配参数，包括：
    获取训练数据集，所述训练数据集包括正训练数据集和负训练数据集；
    将所述训练数据集输入所述匹配度计算模型中，并构建目标函数；
    根据所述目标函数确定所述匹配度计算模型的匹配参数。
  10. 根据权利要求9所述的智能分诊方法，其特征在于，所述根据所述目标函数确定所述匹配度计算模型的匹配参数，包括：
    利用随机梯度下降算法对所述匹配度计算模型的匹配参数进行更新，当所述目标函数收敛时，完成对所述匹配参数的训练。
  11. 根据权利要求10所述的智能分诊方法，其特征在于，所述目标函数为：
    Figure PCTCN2021103444-appb-100002
    其中，(doc i，pat i)为正训练数据，表示第i个医生适合处理第i个患者的病情，(doc j，pat i)为负训练数据，表示第j个医生不适合处理第i个患者的病情，γ表示正训练数据的匹配度得分score(doc i，pat i)与负训练数据的匹配度得分score(doc j，pat i)之间的预设差值阈值。
  12. 根据权利要求8所述的智能分诊方法，其特征在于，所述对所述图神经网络进行训练，得到所述图神经网络的神经网络参数，包括：
    利用反向传播算法对所述图神经网络的神经网络参数进行迭代更新，当所述目标函数收敛时，完成对所述神经网络参数的训练。
  13. 根据权利要求1所述的智能分诊方法，其特征在于，所述根据所述匹配度的大小为所述目标就诊患者推荐医生，包括：
    根据所述匹配度的大小确定待推荐医生集合；
    基于所述待推荐医生集合为所述目标就诊患者推荐医生。
  14. 根据权利要求13所述的智能分诊方法，其特征在于，所述基于所述待推荐医生集合为所述目标就诊患者推荐医生，包括：
    向所述目标就诊患者推荐所述待推荐医生集合；
    根据所述目标就诊患者输入的筛选条件对所述待推荐医生集合中的每个医生进行过滤，以为所述目标就诊患者推荐目标医生。
  15. 一种智能分诊装置，其特征在于，包括：
    患者向量获取模块，用于获取目标就诊患者的病情信息，并根据所述病情信息得到目标就诊患者向量；
    医生向量获取模块，用于将医生知识图谱中的各个医生向量化，得到多个医生向量；
    向量匹配模块，用于使用匹配度计算模型计算所述目标就诊患者向量与所述多个医生向量的匹配度，并根据所述匹配度的大小为所述目标就诊患者推荐医生。
  16. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现权利要求1-14任一项所述方法。
  17. 一种电子设备，其特征在于，包括：
    处理器；以及
    存储器，用于存储所述处理器的可执行指令；
    其中，所述处理器配置为经由执行所述可执行指令来执行权利要求1-14任一项所述的方法。
PCT/CN2021/103444 2021-06-30 2021-06-30 智能分诊方法、装置、存储介质及电子设备 WO2023272563A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CN202180001717.6A CN115803821A (zh) 2021-06-30 2021-06-30 智能分诊方法、装置、存储介质及电子设备
PCT/CN2021/103444 WO2023272563A1 (zh) 2021-06-30 2021-06-30 智能分诊方法、装置、存储介质及电子设备
US17/907,775 US20240203569A1 (en) 2021-06-30 2021-06-30 Intelligent triage method and device, storage medium and electronic device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/103444 WO2023272563A1 (zh) 2021-06-30 2021-06-30 智能分诊方法、装置、存储介质及电子设备

Publications (1)

Publication Number Publication Date
WO2023272563A1 true WO2023272563A1 (zh) 2023-01-05

Family

ID=84690916

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/103444 WO2023272563A1 (zh) 2021-06-30 2021-06-30 智能分诊方法、装置、存储介质及电子设备

Country Status (3)

Country Link
US (1) US20240203569A1 (zh)
CN (1) CN115803821A (zh)
WO (1) WO2023272563A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117497158A (zh) * 2023-12-28 2024-02-02 智业软件股份有限公司 基于维度组合与权重计算推荐分诊等级的方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109635122A (zh) * 2018-11-28 2019-04-16 平安科技（深圳）有限公司 智能疾病问询方法、装置、设备及存储介质
CN110675944A (zh) * 2019-09-20 2020-01-10 京东方科技集团股份有限公司 分诊方法及装置、计算机设备及介质
CN110993078A (zh) * 2019-11-27 2020-04-10 华中科技大学同济医学院附属协和医院 一种医疗分诊方法、装置和存储介质
CN111785368A (zh) * 2020-06-30 2020-10-16 平安科技（深圳）有限公司 基于医疗知识图谱的分诊方法、装置、设备及存储介质
US20210109995A1 (en) * 2019-10-14 2021-04-15 International Business Machines Corporation Filtering spurious knowledge graph relationships between labeled entities

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109635122A (zh) * 2018-11-28 2019-04-16 平安科技（深圳）有限公司 智能疾病问询方法、装置、设备及存储介质
CN110675944A (zh) * 2019-09-20 2020-01-10 京东方科技集团股份有限公司 分诊方法及装置、计算机设备及介质
US20210109995A1 (en) * 2019-10-14 2021-04-15 International Business Machines Corporation Filtering spurious knowledge graph relationships between labeled entities
CN110993078A (zh) * 2019-11-27 2020-04-10 华中科技大学同济医学院附属协和医院 一种医疗分诊方法、装置和存储介质
CN111785368A (zh) * 2020-06-30 2020-10-16 平安科技（深圳）有限公司 基于医疗知识图谱的分诊方法、装置、设备及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117497158A (zh) * 2023-12-28 2024-02-02 智业软件股份有限公司 基于维度组合与权重计算推荐分诊等级的方法
CN117497158B (zh) * 2023-12-28 2024-04-16 智业软件股份有限公司 基于维度组合与权重计算推荐分诊等级的方法

Also Published As

Publication number Publication date
US20240203569A1 (en) 2024-06-20
CN115803821A (zh) 2023-03-14

Similar Documents

Publication Publication Date Title
CN111090987B (zh) 用于输出信息的方法和装置
US11288593B2 (en) Method, apparatus and device for extracting information
Jalal et al. An overview of R in health decision sciences
CN111737476B (zh) 文本处理方法、装置、计算机可读存储介质及电子设备
CN111078836B (zh) 基于外部知识增强的机器阅读理解方法、系统、装置
CN104714931B (zh) 用于选择用以代表表格式信息的方法和系统
CN113535984B (zh) 一种基于注意力机制的知识图谱关系预测方法及装置
US10599686B1 (en) Method and system for extracting information from graphs
WO2023029506A1 (zh) 病情分析方法、装置、电子设备及存储介质
US20210263971A1 (en) Automatic corpora annotation
Shaikh et al. Bloom’s learning outcomes’ automatic classification using lstm and pretrained word embeddings
CN113704460B (zh) 一种文本分类方法、装置、电子设备和存储介质
CN112836502B (zh) 一种金融领域事件隐式因果关系抽取方法
CN110322959A (zh) 一种基于知识的深度医疗问题路由方法及系统
Chen et al. DxFormer: a decoupled automatic diagnostic system based on decoder–encoder transformer with dense symptom representations
CN116821373A (zh) 基于图谱的prompt推荐方法、装置、设备及介质
WO2023272563A1 (zh) 智能分诊方法、装置、存储介质及电子设备
CN112069825A (zh) 面向警情笔录数据的实体关系联合抽取方法
Samant et al. An optimized crossover framework for social media sentiment analysis
CN116955638A (zh) 知识图谱构建方法、装置、计算机可读介质及电子设备
US11947503B2 (en) Autoregressive graph generation machine learning models
CN116543853A (zh) 药物相互作用预测方法、计算机设备及介质
Sosa et al. A Bayesian approach for de-duplication in the presence of relational data
Abedini et al. Epci: an embedding method for post-correction of inconsistency in the RDF knowledge bases
CN113704481A (zh) 一种文本处理方法、装置、设备及存储介质

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 17907775

Country of ref document: US

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21947517

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE