WO2021139231A1 - 基于神经网络模型的分诊方法、装置和计算机设备 - Google Patents

基于神经网络模型的分诊方法、装置和计算机设备 Download PDF

Info

Publication number
WO2021139231A1
WO2021139231A1 PCT/CN2020/118137 CN2020118137W WO2021139231A1 WO 2021139231 A1 WO2021139231 A1 WO 2021139231A1 CN 2020118137 W CN2020118137 W CN 2020118137W WO 2021139231 A1 WO2021139231 A1 WO 2021139231A1
Authority
WO
WIPO (PCT)
Prior art keywords
information
disease
triaged
person
triage
Prior art date
Application number
PCT/CN2020/118137
Other languages
English (en)
French (fr)
Inventor
林桂
黎旭东
Original Assignee
平安科技(深圳)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 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021139231A1 publication Critical patent/WO2021139231A1/zh

Links

Classifications

    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20ICT 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
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00Handling natural language data
    • G06F40/20Natural language analysis
    • G06F40/279Recognition of textual entities
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00Handling natural language data
    • G06F40/30Semantic analysis
    • G06F40/35Discourse or dialogue representation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/049Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • This application relates to the field of artificial intelligence, in particular to a method, device and computer equipment for triage based on a neural network model.
  • Triage products refer to artificial intelligence products that automatically perform triage based on disease information to remind users which department number should be assigned to see a doctor.
  • Existing products mainly rely on rules for symptom recognition and lack flexibility.
  • the inventor found that triage products are targeted at ordinary users, and the input is mainly colloquial symptoms, which is difficult to directly map to standard symptoms, and the accuracy of triage is poor.
  • the main purpose of this application is to provide a triage method, device and computer equipment based on a neural network model, aiming to solve the problem of low triage accuracy of triage products.
  • this application proposes a method for triage based on a neural network model, including:
  • the pre-training model BERT is used to semantically encode the disease information to obtain the disease code, wherein the pre-training model BERT vectorizes each character in the disease information to obtain the character vector of each character, and give each character A character mark position vector is obtained to obtain a character position vector, each character vector and its corresponding character mark position vector are combined to obtain a code vector of each character, and the code vector of each character is combined to obtain the disease coding;
  • the triage information is obtained according to the standard symptoms, and the triage information is fed back to the person to be triaged.
  • This application also provides a triage device based on a neural network model, including:
  • the receiving unit is used to receive the disease information input by the person to be triaged
  • the coding unit is used to semantically encode the disease information using a pre-training model BERT to obtain a disease code, wherein the pre-training model BERT vectorizes each character in the disease information to obtain the character of each character Vector, and mark a position vector for each character to obtain a character position vector, combine each character vector and its corresponding character mark position vector to obtain a coding vector for each character, and convert the coding vector for each character Combining to obtain the disease code;
  • the decoding calculation unit is used to input the disease code into a preset BiLSTM+CRF sequence labeling model for calculation to obtain the symptom entity corresponding to the disease information, and obtain the standard symptom corresponding to the disease information according to the symptom entity,
  • the BiLSTM performs an encode operation on the disease code, traverses the disease code back and forth to extract features, and inputs the features into the CRF, the CRF performs a decoding operation, calculates the label of each word, and obtains The symptom entity;
  • the obtaining feedback unit is configured to obtain triage information according to the standard symptoms, and feed back the triage information to the person to be triaged.
  • the present application also provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor implements a neural network model-based triage method when the processor executes the computer program;
  • the triage method based on neural network model includes:
  • the pre-training model BERT is used to semantically encode the disease information to obtain the disease code, wherein the pre-training model BERT vectorizes each character in the disease information to obtain the character vector of each character, and give each character A character mark position vector is obtained to obtain a character position vector, each character vector and its corresponding character mark position vector are combined to obtain a code vector of each character, and the code vector of each character is combined to obtain the disease coding;
  • the triage information is obtained according to the standard symptoms, and the triage information is fed back to the person to be triaged.
  • the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, a method for triage based on a neural network model is implemented;
  • the triage method based on neural network model includes:
  • the pre-training model BERT is used to semantically encode the disease information to obtain the disease code, wherein the pre-training model BERT vectorizes each character in the disease information to obtain the character vector of each character, and give each character A character mark position vector is obtained to obtain a character position vector, each character vector and its corresponding character mark position vector are combined to obtain a code vector of each character, and the code vector of each character is combined to obtain the disease coding;
  • the triage information is obtained according to the standard symptoms, and the triage information is fed back to the person to be triaged.
  • Fig. 1 is a flow chart of a method for triage based on a neural network model according to an embodiment of the application
  • FIG. 2 is a schematic diagram of the process of semantic coding of the training model BERT and sequence labeling of the BiLSTM+CRF sequence labeling model according to an embodiment of the application;
  • FIG. 3 is a schematic structural diagram of a triage device based on a neural network model according to an embodiment of the application
  • FIG. 4 is a schematic structural diagram of a computer device according to an embodiment of the application.
  • an embodiment of the present application provides a 1.
  • a method for triage based on a neural network model including the following steps:
  • the pre-training model BERT uses the pre-training model BERT to semantically encode the disease information to obtain a disease code, wherein the pre-training model BERT vectorizes each character in the disease information to obtain a character vector of each character, and Mark the position vector for each character to obtain the character position vector, combine each character vector and its corresponding character mark position vector to obtain the code vector of each character, and combine the code vector of each character to obtain the character position vector.
  • the disease code ;
  • the execution subject of the above method is a triage system, which relies on a server and other computer equipment with data processing capabilities.
  • the above-mentioned person to be triaged refers to a patient or a person who helps the patient log in to the triage system.
  • the aforementioned disease information refers to the text information or voice information input by the person to be triaged. If the person to be triaged inputs voice information, the system will now convert the voice information into text information.
  • BERT is Pre-training of Deep Bidirectional Transformers for Language Understanding.
  • Pre-training means that BERT is a pre-training model.
  • BERT is a pre-training model.
  • BERT is a deep two-way pre-trained language understanding model using Transformers as feature extractors.
  • BERT learned a wealth of linguistic information. Symptom recognition is named entity recognition, and its essence is a serialized annotation task.
  • the above-mentioned semantic encoding process is the process of vectorizing disease information, specifically: the pre-training model BERT vectorizes each character in the disease information to obtain the character vector of each character, and give each character Marking the position vector to obtain a character position vector, combining each character vector and its corresponding character marking position vector to obtain a code vector of each character, and combining the coding vector of each character to obtain the disease code.
  • the pre-training model BERT used in this application completely abandoned the RNN, and instead passed the processed disease information into a large Transformer model for processing, which must mark the position of each character in order to truly understand the context information.
  • the method of marking a position vector for each character is the method of Positional Encoding, which is a method of marking a position vector.
  • the so-called position vector is a vector obtained by performing vector training on the position where the character appears.
  • This application uses the pre-training model BERT to semantically encode the disease information, which can obtain more accurate disease information identification and improve the standard accuracy of the subsequent BiLSTM+CRF sequence labeling model.
  • the calculation process of the standard BiLSTM+CRF sequence labeling model As described in the above step S3, it is the calculation process of the standard BiLSTM+CRF sequence labeling model, the calculation process of decoding the above disease code, in this process, the disease entity corresponding to the disease information is marked, and then the entity link technology is used to obtain The standard symptom corresponding to the disease entity is the process of entity linking.
  • the BiLSTM+CRF sequence labeling model belongs to a two-way cyclic neural network, which can predict the probability of the label for the input word according to the context information. Specifically, the pre-trained BERT embedding is used to encode the sentence at the character level, and the word vectors are respectively formed to obtain the disease code as input, which is input into the BiLSTM+CRF sequence labeling model for calculation.
  • BiLSTM mainly encodes sentences. BiLSTM is more effective than one-way LSTM or GRU. Because the sentence is traversed before and after, it can better capture semantic features and play a role in feature extraction. Then the extracted features are input to the CRF layer for decoding operation, and the label of each word in the sequence is calculated.
  • the disease information is "What should I do if my head hurts", and the labels of the three words "Headache” are finally outputted as B, I, I, and the labels of "What to do” are all O, where B is Begin, the beginning of the noun phrase; I is the middle of the noun phrase; I is the middle of the noun phrase; the other word tags of the sentence are all O, that is, Other, a non-noun phrase. Therefore, "headache” is a noun phrase, and here it is a symptom entity.
  • step S4 that is, after obtaining the standard symptoms, search for the department corresponding to the standard symptom in the preset symptom-triage department relationship table. Further, according to the current time, the information of the doctor on duty in the department corresponding to the standard symptom will be collected and fed back to the person to be triaged.
  • the feedback to the person to be triaged is to generate a file in a preset format with the department information and the information of the doctor on duty and send it to the terminal installed with the triage system operated by the person to be triaged.
  • the above-mentioned files are files with preset buttons, and different buttons correspond to links of different physicians. For example, clicking the button corresponding to a physician can view the specific information of the physician, or make an appointment for the physician to see a doctor, etc. .
  • the step S1 of receiving the condition information input by the person to be triaged includes:
  • the information currently recorded in the input window is used as the disease information.
  • a symptom information input window will be displayed on the triage system used by the person to be triaged.
  • the person to be triaged can enter the symptom information in the symptom information input window.
  • click the trigger confirmation button which means that the input of the disease information is completed, and then enter the step S2 mentioned above.
  • the method further includes:
  • the disease information input window is empty, call a preset human medical knowledge graph, where the human medical knowledge graph is a knowledge graph with a multi-layer mapping relationship;
  • a human medical knowledge map with a multi-layer mapping relationship is given, such as the human body
  • the medical knowledge map is a color picture of the human body. Different areas correspond to different biological organs, such as head, stomach, heart, stomach, etc.
  • stomach is painful (the patient knows which specific location hurts, but does not know What is the human organ corresponding to this position), but you don’t know it’s stomach pain, you can click on the painful position (stomach) in the color picture of the human body.
  • the person to be triaged can determine the final symptom information corresponding to the patient according to their specific symptoms and the layer-by-layer prompts of the human medical knowledge map.
  • the disease information obtained through the human medical knowledge graph can be standard symptoms, so the above steps S2 and S3 can be skipped directly to reduce unnecessary calculations of the system.
  • the patient to be triaged when the patient to be triaged is unable to summarize and summarize the input of the patient’s illness, the patient is provided with a visualized human medical knowledge graph, so that the patient (the patient to be triaged) can click on the visualized and self-contained The area corresponding to the symptom, and then give specific prompts in detail, and finally determine the patient’s disease information, improve the accuracy and flexibility of triage, and can’t accurately describe it for people with lower education, the elderly, or children, etc. People with their own symptoms can provide you with an accurate description of their own symptoms.
  • the method further includes:
  • the pre-training model BERT and the BiLSTM+CRF sequence labeling model are updated and trained.
  • the person to be triaged uses the triage system for triage, consultation and treatment will be conducted according to the triage result.
  • Consultation includes online or offline, etc., which is not limited here.
  • the above-mentioned person to be triaged can determine whether the previous triage is correct, and log in to the triage system again after the consultation and treatment, and input feedback information, that is, the triage is correct or incorrect.
  • the triage system will record the feedback information, and then store it in association with the disease information to form training data.
  • the training data reaches the specified amount of data, the above-mentioned pre-training model BERT and BiLSTM+CRF sequence labeling model can be supervised and trained to improve the accuracy of each model.
  • the above-mentioned step of updating and training the pre-training model BERT and BiLSTM+CRF sequence labeling model after the training data is greater than the preset data amount includes:
  • update training is performed automatically, and the update training is not performed immediately after the amount of data reaches the above-mentioned preset data amount, but waits for a preset non-working time for training, so as to save the triage system from working.
  • the hardware computing resources of the time period are not limited to:
  • step S1 of receiving the condition information input by the person to be triaged it includes:
  • ID information search for historical disease information corresponding to the ID information
  • step S1 of receiving the disease information input by the person to be triaged it includes:
  • the historical triage information corresponding to the historical disease information is called and fed back to the person to be triaged.
  • the input disease information will be basically the same, so when the historical disease information
  • the above-mentioned similarity calculation can use existing similarity calculation methods, such as vectorizing the disease information through a vector dictionary, and then calculating the similarity.
  • the triage information can also call the corresponding historical case information according to the ID information, select the patients to be triaged together, and then directly call the feedback information of the selected historical disease information to the patients to be triaged.
  • the method further includes:
  • a preset weighted average calculation method is adopted to recommend to the preferred hospital for the person to be triaged.
  • the above triage information is mainly to let the user know which department to go to for medical treatment for the disease, but which hospital should he go to.
  • the triage system of this application will also give the preferred hospital for treatment, and the specific preference is
  • the method is to perform a weighted average calculation based on the department’s medical level score and distance. For example, the proportion of distance is 30%, and the department’s medical level score is 70. It can also be calculated in combination with travel convenience, such as choosing public transportation, Cycling, walking, driving and other travel modes are used to determine the specific preferred hospital.
  • the convenience of getting to the hospital within the range accounts for 20%, the distance accounts for 30%, and the department's medical level score accounts for 50%.
  • the medical level scores of the aforementioned departments are the average scores of historical users after scoring according to their treatment results, or the scores given by professional institutions after evaluating the hospital, which is not specifically limited here.
  • the method further includes:
  • the above-mentioned special department refers to a department of a specialized hospital, or a department whose medical level score is greater than a preset score.
  • the level of the standard symptoms is higher, it means that the condition of the person to be triaged is more serious and requires special treatment.
  • the department of the corresponding specialty hospital or the department with high medical level is preferred. It can prevent the patients to be triaged from going to hospitals with insufficient medical capabilities to treat their corresponding diseases, wasting time, or even missing treatment time.
  • This application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, small computers, large computers, including Distributed computing environment for any of the above systems or equipment, etc.
  • This application may be described in the general context of computer-executable instructions executed by a computer, such as a program module.
  • program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types.
  • This application can also be practiced in distributed computing environments. In these distributed computing environments, tasks are performed by remote processing devices connected through a communication network.
  • program modules can be located in local and remote computer storage media including storage devices.
  • the method before the step S7 of using the pre-training model BERT to semantically encode the disease information to obtain the disease code, the method includes:
  • the first symptom information is judged to be a standard symptom, and the execution of the “use the pre-training model BERT to semantically encode the symptom information to obtain the symptom code; input the symptom code into the preset BiLSTM+ Perform calculations in the CRF sequence annotation model to obtain the standard symptoms corresponding to the disease information" step, directly enter the step of "obtain triage information according to the standard symptoms, and feed the triage information back to the person to be triaged" .
  • the disease information obtained through the human medical knowledge graph can be directly used as a standard symptom, so the above steps S7 and S8 can be skipped directly to reduce unnecessary calculations of the system.
  • the patient to be triaged when the patient to be triaged is unable to summarize and summarize the input of the patient’s illness, the patient is provided with a visualized human medical knowledge graph, so that the patient (the patient to be triaged) can click on the visualized and self-contained The area corresponding to the symptom, and then give specific prompts in detail, and finally determine the patient’s disease information, improve the accuracy and flexibility of triage, and can’t accurately describe it for people with lower education, the elderly, or children, etc. People with their own symptoms can provide you with an accurate description of their own symptoms.
  • the neural network model-based triage method in the embodiment of the application uses the pre-training model BERT for semantic recognition, which improves the recognition accuracy of standard illnesses and at the same time improves the understanding of the colloquially input illness information. Furthermore, this application also provides a human medical knowledge graph for use by the elderly or children who cannot describe the condition of the disease, making the input of the person to be triaged more concise, avoiding the trouble of inputting text to a certain extent, and improving the usability.
  • the present application also provides a triage device based on a neural network model, including:
  • the receiving unit 10 is used for receiving disease information input by a person to be triaged
  • the encoding unit 20 is configured to use a pre-training model BERT to semantically encode the disease information to obtain a disease code, wherein the pre-training model BERT vectorizes each character in the disease information to obtain the information of each character Character vector, and mark position vector for each character to obtain a character position vector, combine each character vector and its corresponding character mark position vector to obtain the code vector of each character, and combine the code of each character Vector combination to obtain the disease code;
  • the decoding calculation unit 30 is configured to input the disease code into the preset BiLSTM+CRF sequence labeling model for calculation to obtain the symptom entity corresponding to the disease information, and obtain the standard symptom corresponding to the disease information according to the symptom entity , wherein the BiLSTM performs an encode operation on the disease code, traverses the disease code back and forth to extract features, and inputs the features into the CRF, and the CRF performs a decoding operation to calculate the label of each word, Obtain the symptom entity;
  • the obtaining feedback unit 40 is configured to obtain triage information according to the standard symptoms, and feed back the triage information to the person to be triaged.
  • the receiving unit 10 includes:
  • the presentation module is used to present the preset disease information input window and the confirmation button to end the input;
  • the first judgment module is used to judge whether the confirmation button is triggered
  • the second judgment module is used to detect whether the disease information input window is empty
  • the first determination module is configured to, if the disease information input window is not empty, use the information recorded in the input window as the disease information.
  • the second determination module is configured to call a preset human medical knowledge graph if the disease information input window is empty, wherein the human medical knowledge graph is a knowledge graph with a multi-layer mapping relationship;
  • the determining module is used to receive the click information of the person to be triaged clicking on the human medical knowledge map, and use the click information as the disease information, wherein the click information is the user based on the multi-layer mapping relationship of the human medical knowledge map. Symptom information obtained after layer screening.
  • the above-mentioned triage device based on the neural network model further includes:
  • a receiving feedback unit configured to receive feedback information of the person to be triaged after being triaged and treated, where the feedback information is information for determining whether the triage is correct after the person to be triaged has been triaged and treated;
  • An associative storage unit for associative storage of the feedback information and the disease information input by the person to be triaged as training data
  • the update unit is used to update and train the pre-training model BERT and BiLSTM+CRF sequence labeling model when the training data is larger than the preset data amount.
  • the above-mentioned triage device based on the neural network model further includes:
  • An acquiring ID unit used to acquire the ID information of the person to be triaged
  • the history search unit is used to search for historical disease information corresponding to the ID information according to the ID information;
  • the likeness calculation unit is used to calculate the expected similarity between the disease information and the historical disease information
  • the invoking feedback unit is configured to, if the similarity is greater than a preset value, invoking the historical triage information corresponding to the historical disease information to feed back to the person to be triaged.
  • the above-mentioned triage device based on the neural network model further includes:
  • a location acquiring unit configured to acquire location information of the person to be triaged
  • a collection unit configured to collect, according to the location information, the medical level scores of the hospitals within a designated area and the departments of each of the hospitals corresponding to the standard symptoms;
  • the first recommendation unit is configured to use a preset weighted average calculation method according to the distance between the location information and each of the hospitals and the medical level scores of each of the departments, and recommend to the person to be triaged the preferred Visit the hospital.
  • the above-mentioned triage device based on the neural network model further includes:
  • a grade determination unit for determining the grade of the standard symptom
  • Find a department unit which is used to find a special department corresponding to the standard symptom and grade if the level is greater than a preset threshold
  • the second recommendation unit is used to recommend the special department to the person to be triaged.
  • an embodiment of the present application also provides a computer device.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 4.
  • the computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor designed by the computer is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, a computer program, and a database.
  • the memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium.
  • the database of the computer equipment is used to store data such as standard symptoms and historical case information.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer program is executed by the processor to implement the neural network model-based triage method in any of the above embodiments.
  • FIG. 4 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied.
  • An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the neural network model-based triage method of any one of the above embodiments is implemented.
  • the computer storage medium may be non-volatile or volatile.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous Link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Medical Treatment And Welfare Office Work (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

一种基于神经网络模型的分诊方法、装置和计算机设备,涉及人工智能,其中方法包括:接收待分诊者输入的病症信息(S1);利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码(S2);将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应病症信息的标准症状(S3);依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者(S4)。在执行方法的时候,使用预训练模型BERT进行语义识别,提高标准病症的识别准确性,同时提高对口语化输入的病症信息的理解。上述预训练模型BERT和BiLSTM+CRF序列标注模型可以存储到区块链网络中。所述方法还提供人体医疗知识图谱,供老人或儿童等无法描述病情的人群使用,在一定程度上避免了输入文字的麻烦,提高可用度。

Description

基于神经网络模型的分诊方法、装置和计算机设备
本申请要求于2020年6月30日提交中国专利局、申请号为202010621759.7,申请名称为“基于神经网络模型的分诊方法、装置和计算机设备”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及到人工智能领域,特别是涉及到一种基于神经网络模型的分诊方法、装置和计算机设备。
背景技术
分诊产品是指根据病症信息自动进行分诊,以提醒用户应该挂哪一个科室的号进行看病的人工智能产品。现有产品对于症状识别主要依赖于规则,灵活性欠缺。发明人发现分诊产品定位于普通用户,输入以口语化症状为主,难以直接映射到标准症状,分诊准确性较差。
技术问题
本申请的主要目的为提供基于神经网络模型的分诊方法、装置和计算机设备,旨在解决分诊产品分诊准确性不高的问题。
技术解决方案
为了实现上述申请目的,本申请提出一种基于神经网络模型的分诊方法,包括:
接收待分诊者输入的病症信息;
利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
本申请还提供一种基于神经网络模型的分诊装置,包括:
接收单元,用于接收待分诊者输入的病症信息;
编码单元,用于利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
解码计算单元,用于将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
获取反馈单元,用于依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
本申请还提供一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,所述处理器执行所述计算机程序时实现一种基于神经网络模型的分诊方法;
其中,所述基于神经网络模型的分诊方法包括:
接收待分诊者输入的病症信息;
利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
本申请还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一种基于神经网络模型的分诊方法;
其中,所述基于神经网络模型的分诊方法包括:
接收待分诊者输入的病症信息;
利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
附图说明
图1 为本申请一实施例的基于神经网络模型的分诊方法的流程视图;
图2 为本申请一实施例的训练模型BERT进行语义编码和BiLSTM+CRF序列标注模型进行序列标注的流程示意图;
图3 为本申请一实施例的基于神经网络模型的分诊装置的结构示意图;
图4 为本申请一实施例的计算机设备的结构示意图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
本发明的最佳实施方式
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
参照图1,本申请实施例提供一种1、一种基于神经网络模型的分诊方法,包括以下步骤:
S1、接收待分诊者输入的病症信息;
S2、利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
S3、将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
S4、依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
上述方法的执行主体是分诊系统,该分诊系统依托于服务器等具有数据处理能力的计算机设备等。
如上述步骤S1所述,上述待分诊者是指患者或者帮助患者登陆分诊系统的人。上述病症信息是指待分诊者输入的文字信息或者语音信息,如果待分诊者输入的是语音信息,系统会现将语音信息转换成文字信息。
如上述步骤S2所述,上述预训练模型BERT全称是Pre-training of Deep Bidirectional Transformers for Language Understanding。Pre-training表示BERT是一个预训练模型,通过前期的大量语料的无监督训练,为下游任务学习大量的先验的语言、句法、词义等信息。Bidirectional 说明BERT采用的是双向语言模型的方式,能够更好的融合前后文的知识。简而言之,BERT是一个用Transformers作为特征抽取器的深度双向预训练语言理解模型。BERT在预训练过程中,学习到了丰富的语言学方面的信息。症状识别即命名实体识别,其本质属于序列化标注任务。上述语意编码的过程即为将病症信息向量化的过程,具体为:所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码。本申请中使用的预训练模型BERT,完全抛弃了RNN,取而代之的是将经过处理的病症信息传入大型的Transformer模型中进行处理,其必须对每一字符的位置进行标记,以便于真正理解上下文信息。在本申请实施例中,给每一个字符标记位置向量的方法是Positional Encoding的方法,该方法即为标注位置向量的方法,所谓位置向量,就是对字符出现的位置进行向量训练而得到的向量。本申请使用预训练模型BERT对所述病症信息进行语意编码,可以得到更加准确的识别病症信息,提高后续BiLSTM+CRF序列标注模型的标准准确性。
如上述步骤S3所述,即为标准的BiLSTM+CRF序列标注模型计算过程,将上述病症编码进行解码的计算过程,在此过程中,标注出病症信息对应的病症实体,然后利用实体链接技术得到与病症实体对应的标准症状,也就是实体链接的过程。BiLSTM+CRF序列标注模型属于双向的循环神经网络,能够根据上下文信息对输入的字给以预测标签的概率。具体地,使用预训练好的BERT embedding对该语句进行字符级别编码,分别形成字向量进而得到病症编码作为输入,输入到BiLSTM+CRF序列标注模型中进行计算。其中BiLSTM主要对语句进行encode操作,BiLSTM效果比单向LSTM或GRU更好,由于对语句进行前后遍历,更能捕获语义特征,起到特征提取的作用。然后将提取出的特征输入到CRF层进行解码操作,计算序列中每个字的标签。如图2所示,病症信息是“头很痛怎么办”,最后输出“头很痛”三个字的标签分别为B,I,I,“怎么办”的标签均为O,其中B为Begin,名词短语的开始;I为Intermediate即名词短语的中间;该句的其他字标签均为O,即Other,非名词短语。因此“头很痛”为名词短语,在此为症状实体。
如上述步骤S4所述,即为得到标准症状后,到预设的症状-分诊科室的关系表中查找与标准症状对应的科室。进一步地,会根据当前的时间,收集对应标准症状的科室的值班医师的信息等一起反馈给待分诊者。反馈给待分诊者即为将科室信息和值班医师信息生成预设格式的文件发送到待分诊者操作的安装有分诊系统的终端上。进一步地,上述文件是带有预设按钮的文件,不同的按钮对应不同的医师的链接等,比如点击对应某一医师的按钮,可以是查看该医师的具体信息,或者预约该医师进行看病等。
在一个实施例中,上述接收待分诊者输入的病症信息的步骤S1,包括:
呈现预设的病症信息输入窗口,以及结束输入的确认按钮;
判断所述确认按钮是否被触发;
若是,则检测所述病症信息输入窗口中是否为空;
若否,则将所述输入窗口中当前所记录的信息作为所述病症信息。
在本实施例中,会在待分诊者使用的分诊系统上显示一个病症信息输入窗口,待分诊者可以在该病症信息输入窗口中输入病症信息,当输入完成后点击触发确认按钮后,表示病症信息输入完成,则进入上述的步骤S2。
进一步地,上述检测所述病症信息输入窗口中是否为空的步骤之后,还包括:
若所述病症信息输入窗口中为空,则调用预设的人体医疗知识图谱,其中人体医疗知识图谱为多层映射关系的知识图谱;
接收待分诊者点击人体医疗知识图谱的点击信息,并将所述点击信息作为所述病症信息,其中,所述点击信息为用户依据人体医疗知识图谱的多层映射关系逐层筛选后得到的症状信息。
在本实施例中,当用户没有输入语音或者文字的病症信息时,说明待分诊者无法描述出具体的病症,此时,给出一个具有多层映射关系的人体医疗知识图谱,比如该人体医疗知识图谱是一个人体的彩图,不同的区域对应着不同的生物学器官,如头、肚子、心脏、胃等,当患者的胃部疼痛时(患者知道哪个具体的位置疼,但是不知道该位置对应的人体器官是什么),却不知道是胃痛时,可以点击人体的彩图中对应自己疼痛的位置(胃部),当点击胃部之后,会到弹出与胃部具有映射关系的下一层具体的病症或者更细致的部位等信息,然后待分诊者根据自身的具体病症以及人体医疗知识图谱的逐层提示,确定好最终的与患者对应的症状信息。在一个具体实施例中,通过人体医疗知识图谱获取到的病症信息可以为标准症状,所以可以直接跳过上述步骤S2和S3的步骤,减少系统无谓的计算。在本实施例中,当待分诊者无法概括总结的输入患者的病症的时候,给待分诊者提供可视化的人体医疗知识图谱,使患者(待分诊者)通过点击可视化的、与自身症状对应的区域,然后逐层细化的给出具体的提示,最终确定患者的病症信息,提高分诊的准确性和灵活性,为文化较低的人、老年人、或者儿童等无法准确描述自身病症的人群提供可以准确提供给你对应自身病症的表述。
在一个实施例中,上述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤S4之后,还包括:
接收待分诊者被分诊治疗后的反馈信息,其中,所述反馈信息为所述待分诊者分诊治疗后确定分诊是否正确的信息;
将所述反馈信息,以及所述待分诊者输入的病症信息进行关联存储作为训练数据;
当所述训练数据大于预设数据量后,对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练。
在本实施例中,当待分诊者使用分诊系统进行分诊后,会根据分诊结果进行咨询治疗等,咨询包括线上或线下等,在此不做限定。在咨询治疗的过程中,上述待分诊者可以判定之前的分诊是否正确,并且在咨询治疗之后再次登陆分诊系统,输入反馈信息,即分诊正确或错误的信息。分诊系统会记录反馈信息,然后与病症信息进行关联存储,形成训练数据。当训练数据达到指定的数据量之后,可以对上述的预训练模型BERT和BiLSTM+CRF序列标注模型进行监督训练,提高各模型的准确性。
上述当所述训练数据大于预设数据量后,对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练的步骤,包括:
当所述训练数据大于预设数据量时,判断当前时间是否为非工作时间;
若是,则对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练。在本实施例中,更新训练是自动进行的,而且并不是数据量达到上述预设数据量就立刻进行更新训练,而是等到一个预设的非工作时间进行训练,以节约分诊系统在工作时间段的硬件计算资源。
在一个实施例中,上述接收待分诊者输入的病症信息的步骤S1之前,包括:
获取所述待分诊者的ID信息;
依据所述ID信息,查找所述ID信息对应的历史病症信息;
所述接收待分诊者输入的病症信息的步骤S1之后,包括:
计算所述所述病症信息与所述历史病症信息的想相似度;
若所述相似度大于预设值,则调用所述对应所述历史病症信息的历史分诊信息反馈给所述待分诊者。
在本实施例中,在现实中,患者往往会因为会复发同一个病而需要再次医疗,所以当使用本申请中的分诊系统的时候,输入的病症信息会基本相同,所以当历史病症信息和当前输入的病症信息的相似度达到预设值的时候,可以跳过模型编码、解码等过程(步骤S2-S4),直接调用历史病症信息对应的反馈信息作为当前输入的病症信息的反馈信息,可以节约分诊系统的计算资源。上述相似度计算可以使用现有相似度计算方法,如先通过向量词典将病症信息向量化,然后再计算相似度等。进一步地,分诊信息还可以根据ID信息调用其对应的历史病例信息,共待分诊者选择,然后直接调用被选择的历史病症信息的反馈信息给所述待分诊者。
在一个实施例中,上述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤S4之后,还包括:
获取所述待分诊者的位置信息;
根据所述位置信息收集指定范围内的医院,以及各所述医院对应所述标准症状的科室的医疗水平分数;
依据所述位置信息与各所述医院的距离,以及各所述科室的医疗水平分数,采用预设的加权平均的计算方法,推荐给所述待分诊者优选的就诊医院。
在本实施例中,上述分诊信息主要是让用户知道病症需要到哪一个科室进行就医,但是应该去那一个医院呢,本申请的分诊系统还会给出优选的就诊医院,具体的优选方法是根据科室的医疗水平分数和距离等进行加权平均计算,比如距离的占比为30%,科室的医疗水平分数占比为70等,还可以结合出行便利性等进行计算,比如选择公交、骑行、步行、驾车等出行方式进行确定具体的优选医院。在一个具体实施例中,待分诊者选择驾车出行,其到范围内的医院的便利性的占比为20%,距离占比为30%,科室的医疗水平分数占比为50%等。上述科室的医疗水平分数是历史用户根据其治疗的结果进行打分后的平均得分,或者是专业机构对所述医院进行评估后给出的分数,在此不做具体限定。
在一个实施例中,上述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤S4之后,还包括:
确定所述标准症状的等级;
若所述等级大于预设阈值,则查找与所述标准症状和等级对应的特级科室;
将所述特级科室推荐给所述待分诊者。
在本实施例中,上述特级科室是指专科医院的科室,或者科室的医疗水平分数大于预设分数的科室。当标准症状的等级较高时,说明待分诊者的病情比较严重,需要特殊治疗,此时优选对应的专科医院的科室,或者医疗水平高的科室。可以防止待分诊者去到医疗实力不不足以治疗其对应的病症的医院,浪费时间,甚至错过救治时间等。
本申请可用于众多通用或专用的计算机系统环境或配置中。例如:个人计算机、服务器计算机、手持设备或便携式设备、平板型设备、多处理器系统、基于微处理器的系统、置顶盒、可编程的消费电子设备、网络PC、小型计算机、大型计算机、包括以上任何系统或设备的分布式计算环境等等。本申请可以在由计算机执行的计算机可执行指令的一般上下文中描述,例如程序模块。一般地,程序模块包括执行特定任务或实现特定抽象数据类型的例程、程序、对象、组件、数据结构等等。也可以在分布式计算环境中实践本申请,在这些分布式计算环境中,由通过通信网络而被连接的远程处理设备来执行任务。在分布式计算环境中,程序模块可以位于包括存储设备在内的本地和远程计算机存储介质中。
在一个实施例中,上述利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码的步骤S7之前,包括:
判断所述病症信息是否为基于所述人体医疗知识图谱获取的第一病症信息;
若是,则将所述第一病症信息判定为标准症状,停止执行所述“利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码;将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应病症信息的标准症状”的步骤,直接进入所述“依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者”的步骤。
在本实施例中,通过人体医疗知识图谱获取到的病症信息可以作为标准症状直接使用,所以可以直接跳过上述步骤S7和S8的步骤,减少系统无谓的计算。在本实施例中,当待分诊者无法概括总结的输入患者的病症的时候,给待分诊者提供可视化的人体医疗知识图谱,使患者(待分诊者)通过点击可视化的、与自身症状对应的区域,然后逐层细化的给出具体的提示,最终确定患者的病症信息,提高分诊的准确性和灵活性,为文化较低的人、老年人、或者儿童等无法准确描述自身病症的人群提供可以准确提供给你对应自身病症的表述。
本申请实施例的基于神经网络模型的分诊方法,使用预训练模型BERT进行语义识别,提高标准病症的识别准确性,同时提高对口语化输入的病症信息的理解。进一步地,本申请还提供人体医疗知识图谱,供老人或儿童等无法描述病情的人群使用,使所述待分诊者输入更简洁,在一定程度上避免了输入文字的麻烦,提高可用度。
参照图3,本申请还提供一种基于神经网络模型的分诊装置,包括:
接收单元10,用于接收待分诊者输入的病症信息;
编码单元20,用于利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
解码计算单元30,用于将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
获取反馈单元40,用于依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
在一个实施例中,所述接收单元10,包括:
呈现模块,用于呈现预设的病症信息输入窗口,以及结束输入的确认按钮;
第一判断模块,用于判断所述确认按钮是否被触发;
第二判断模块,用于检测所述病症信息输入窗口中是否为空;
第一判定模块,用于若所述病症信息输入窗口中不为空,则将所述输入窗口中所记录的信息作为所述病症信息。
第二判定模块,用于若所述病症信息输入窗口中为空,则调用预设的人体医疗知识图谱,其中,所述人体医疗知识图谱为多层映射关系的知识图谱;
确定模块,用于接收待分诊者点击人体医疗知识图谱的点击信息,并将所述点击信息作为所述病症信息,其中,所述点击信息为用户依据人体医疗知识图谱的多层映射关系逐层筛选后得到的症状信息。
在一个实施例中,上述基于神经网络模型的分诊装置,还包括:
接收反馈单元,用于接收所述待分诊者被分诊治疗后的反馈信息,其中,所述反馈信息为所述待分诊者分诊治疗后确定分诊是否正确的信息;
关联存储单元,用于将所述反馈信息,以及所述待分诊者输入的病症信息进行关联存储作为训练数据;
更新单元,用于当所述训练数据大于预设数据量后,对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练。
在一个实施例中,上述基于神经网络模型的分诊装置,还包括:
获取ID单元,用于获取所述待分诊者的ID信息;
历史查找单元,用于依据所述ID信息,查找所述ID信息对应的历史病症信息;
像似计算单元,用于计算所述所述病症信息与所述历史病症信息的想相似度;
调用反馈单元,用于若所述相似度大于预设值,则调用所述对应所述历史病症信息的历史分诊信息反馈给所述待分诊者。
在一个实施例中,上述基于神经网络模型的分诊装置,还包括:
位置获取单元,用于获取所述待分诊者的位置信息;
收集单元,用于根据所述位置信息收集指定范围内的医院,以及各所述医院对应所述标准症状的科室的医疗水平分数;
第一推荐单元,用于依据所述位置信息与各所述医院的距离,以及各所述科室的医疗水平分数,采用预设的加权平均的计算方法,推荐给所述待分诊者优选的就诊医院。
在一个实施例中,上述基于神经网络模型的分诊装置,还包括:
等级确定单元,用于确定所述标准症状的等级;
查找科室单元,用于若所述等级大于预设阈值,则查找与所述标准症状和等级对应的特级科室;
第二推荐单元,用于将所述特级科室推荐给所述待分诊者。
参照图4,本申请实施例中还提供一种计算机设备,该计算机设备可以是服务器,其内部结构可以如图4所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中,该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机程序和数据库。该内存器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的数据库用于存储标准症状、历史病例信息等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机程序被处理器执行时以实现上述任一实施例的基于神经网络模型的分诊方法。
本领域技术人员可以理解,图4中示出的结构,仅仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定。
本申请实施例还提供一种计算机可读存储介质,其上存储有计算机程序,计算机程序被处理器执行时实现上述任一实施例的基于神经网络模型的分诊方法。
所述计算机存储介质可以是非易失性,也可以是易失性。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机程序来指令相关的硬件来完成,所述的计算机程序可存储与一非易失性计算机可读取存储介质中,该计算机程序在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的和实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可以包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM一多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双速据率SDRAM(SSRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
需要说明的是,在本文中,术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、装置、物品或者方法不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、装置、物品或者方法所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括该要素的过程、装置、物品或者方法中还存在另外的相同要素。
以上所述仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种基于神经网络模型的分诊方法,其中,包括:
    接收待分诊者输入的病症信息;
    利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
    将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
    依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
  2. 根据权利要求1所述的基于神经网络模型的分诊方法,其中,所述接收待分诊者输入的病症信息的步骤,包括:
    呈现预设的病症信息输入窗口,以及结束输入的确认按钮;
    判断所述确认按钮是否被触发;
    若是,则检测所述病症信息输入窗口中是否为空;
    若所述病症信息输入窗口中不为空,则将所述输入窗口中所记录的信息作为所述病症信息。
  3. 根据权利要求2所述的基于神经网络模型的分诊方法,其中,所述检测所述病症信息输入窗口中是否为空的步骤之后,还包括:
    若所述病症信息输入窗口中为空,则调用预设的人体医疗知识图谱,其中,所述人体医疗知识图谱为多层映射关系的知识图谱;
    接收待分诊者点击人体医疗知识图谱的点击信息,并将所述点击信息作为所述病症信息,其中,所述点击信息为用户依据人体医疗知识图谱的多层映射关系逐层筛选后得到的症状信息。
  4. 根据权利要求1所述的基于神经网络模型的分诊方法,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    接收所述待分诊者被分诊治疗后的反馈信息,其中,所述反馈信息为所述待分诊者分诊治疗后确定分诊是否正确的信息;
    将所述反馈信息,以及所述待分诊者输入的病症信息进行关联存储作为训练数据;
    当所述训练数据大于预设数据量后,对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练。
  5. 根据权利要求1所述的基于神经网络模型的分诊方法,其中,所述接收待分诊者输入的病症信息的步骤之前,包括:
    获取所述待分诊者的ID信息;
    依据所述ID信息,查找所述ID信息对应的历史病症信息;
    所述接收待分诊者输入的病症信息的步骤之后,包括:
    计算所述所述病症信息与所述历史病症信息的想相似度;
    若所述相似度大于预设值,则调用所述对应所述历史病症信息的历史分诊信息反馈给所述待分诊者。
  6. 根据权利要求1所述的基于神经网络模型的分诊方法,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    获取所述待分诊者的位置信息;
    根据所述位置信息收集指定范围内的医院,以及各所述医院对应所述标准症状的科室的医疗水平分数;
    依据所述位置信息与各所述医院的距离,以及各所述科室的医疗水平分数,采用预设的加权平均的计算方法,推荐给所述待分诊者优选的就诊医院。
  7. 根据权利要求1所述的基于神经网络模型的分诊方法,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    确定所述标准症状的等级;
    若所述等级大于预设阈值,则查找与所述标准症状和等级对应的特级科室;
    将所述特级科室推荐给所述待分诊者。
  8. 一种基于神经网络模型的分诊装置,其中,包括:
    接收单元,用于接收待分诊者输入的病症信息;
    编码单元,用于利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
    解码计算单元,用于将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
    获取反馈单元,用于依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
  9. 一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,其中,所述处理器执行所述计算机程序时实现一种基于神经网络模型的分诊方法;
    其中,所述基于神经网络模型的分诊方法包括:
    接收待分诊者输入的病症信息;
    利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
    将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
    依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
  10. 根据权利要求9所述的计算机设备,其中,所述接收待分诊者输入的病症信息的步骤,包括:
    呈现预设的病症信息输入窗口,以及结束输入的确认按钮;
    判断所述确认按钮是否被触发;
    若是,则检测所述病症信息输入窗口中是否为空;
    若所述病症信息输入窗口中不为空,则将所述输入窗口中所记录的信息作为所述病症信息。
  11. 根据权利要求10所述的计算机设备,其中,所述检测所述病症信息输入窗口中是否为空的步骤之后,还包括:
    若所述病症信息输入窗口中为空,则调用预设的人体医疗知识图谱,其中,所述人体医疗知识图谱为多层映射关系的知识图谱;
    接收待分诊者点击人体医疗知识图谱的点击信息,并将所述点击信息作为所述病症信息,其中,所述点击信息为用户依据人体医疗知识图谱的多层映射关系逐层筛选后得到的症状信息。
  12. 根据权利要求9所述的计算机设备,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    接收所述待分诊者被分诊治疗后的反馈信息,其中,所述反馈信息为所述待分诊者分诊治疗后确定分诊是否正确的信息;
    将所述反馈信息,以及所述待分诊者输入的病症信息进行关联存储作为训练数据;
    当所述训练数据大于预设数据量后,对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练。
  13. 根据权利要求9所述的计算机设备,其中,所述接收待分诊者输入的病症信息的步骤之前,包括:
    获取所述待分诊者的ID信息;
    依据所述ID信息,查找所述ID信息对应的历史病症信息;
    所述接收待分诊者输入的病症信息的步骤之后,包括:
    计算所述所述病症信息与所述历史病症信息的想相似度;
    若所述相似度大于预设值,则调用所述对应所述历史病症信息的历史分诊信息反馈给所述待分诊者。
  14. 根据权利要求9所述的计算机设备,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    获取所述待分诊者的位置信息;
    根据所述位置信息收集指定范围内的医院,以及各所述医院对应所述标准症状的科室的医疗水平分数;
    依据所述位置信息与各所述医院的距离,以及各所述科室的医疗水平分数,采用预设的加权平均的计算方法,推荐给所述待分诊者优选的就诊医院。
  15. 一种计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现一种基于神经网络模型的分诊方法;
    其中,所述基于神经网络模型的分诊方法包括:
    接收待分诊者输入的病症信息;
    利用预训练模型BERT对所述病症信息进行语意编码,得到病症编码,其中,所述预训练模型BERT对所述病症信息中的每一个字符向量化,得到每一个字符的字符向量,以及给每一个字符标记位置向量,得到字符位置向量,将所述每一个字符向量及其对应的字符标记位置向量合并,得到每一个字符的编码向量,将所述每一个字符的编码向量组合得到所述病症编码;
    将所述病症编码输入到预设的BiLSTM+CRF序列标注模型中进行计算,得到对应所述病症信息的症状实体,并依据所述症状实体得到对应病症信息的标准症状,其中,所述BiLSTM对所述病症编码进行encode操作,对所述病症编码进行前后遍历提取特征,并将所述特征输入到CRF中,所述CRF进行解码操作,计算每个字的标签,得到所述症状实体;
    依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者。
  16. 根据权利要求15所述的计算机可读存储介质,其中,所述接收待分诊者输入的病症信息的步骤,包括:
    呈现预设的病症信息输入窗口,以及结束输入的确认按钮;
    判断所述确认按钮是否被触发;
    若是,则检测所述病症信息输入窗口中是否为空;
    若所述病症信息输入窗口中不为空,则将所述输入窗口中所记录的信息作为所述病症信息。
  17. 根据权利要求16所述的计算机可读存储介质,其中,所述检测所述病症信息输入窗口中是否为空的步骤之后,还包括:
    若所述病症信息输入窗口中为空,则调用预设的人体医疗知识图谱,其中,所述人体医疗知识图谱为多层映射关系的知识图谱;
    接收待分诊者点击人体医疗知识图谱的点击信息,并将所述点击信息作为所述病症信息,其中,所述点击信息为用户依据人体医疗知识图谱的多层映射关系逐层筛选后得到的症状信息。
  18. 根据权利要求15所述的计算机可读存储介质,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    接收所述待分诊者被分诊治疗后的反馈信息,其中,所述反馈信息为所述待分诊者分诊治疗后确定分诊是否正确的信息;
    将所述反馈信息,以及所述待分诊者输入的病症信息进行关联存储作为训练数据;
    当所述训练数据大于预设数据量后,对所述预训练模型BERT和BiLSTM+CRF序列标注模型进行更新训练。
  19. 根据权利要求15所述的计算机可读存储介质,其中,所述接收待分诊者输入的病症信息的步骤之前,包括:
    获取所述待分诊者的ID信息;
    依据所述ID信息,查找所述ID信息对应的历史病症信息;
    所述接收待分诊者输入的病症信息的步骤之后,包括:
    计算所述所述病症信息与所述历史病症信息的想相似度;
    若所述相似度大于预设值,则调用所述对应所述历史病症信息的历史分诊信息反馈给所述待分诊者。
  20. 根据权利要求15所述的计算机可读存储介质,其中,所述依据所述标准症状得到分诊信息,并将分诊信息反馈给所述待分诊者的步骤之后,还包括:
    获取所述待分诊者的位置信息;
    根据所述位置信息收集指定范围内的医院,以及各所述医院对应所述标准症状的科室的医疗水平分数;
    依据所述位置信息与各所述医院的距离,以及各所述科室的医疗水平分数,采用预设的加权平均的计算方法,推荐给所述待分诊者优选的就诊医院。
PCT/CN2020/118137 2020-06-30 2020-09-27 基于神经网络模型的分诊方法、装置和计算机设备 WO2021139231A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010621759.7A CN111785367A (zh) 2020-06-30 2020-06-30 基于神经网络模型的分诊方法、装置和计算机设备
CN202010621759.7 2020-06-30

Publications (1)

Publication Number Publication Date
WO2021139231A1 true WO2021139231A1 (zh) 2021-07-15

Family

ID=72761636

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/118137 WO2021139231A1 (zh) 2020-06-30 2020-09-27 基于神经网络模型的分诊方法、装置和计算机设备

Country Status (2)

Country Link
CN (1) CN111785367A (zh)
WO (1) WO2021139231A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113628726A (zh) * 2021-08-10 2021-11-09 海南榕树家信息科技有限公司 基于图神经网络的中医辩治推荐系统、方法和电子设备
CN115662593A (zh) * 2022-11-08 2023-01-31 北京健康在线技术开发有限公司 基于症状知识图谱的医患匹配方法、装置、设备及介质
CN117637092A (zh) * 2024-01-24 2024-03-01 创智和宇信息技术股份有限公司 一种基于人工智能模型的病历预编码方法及装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113268576B (zh) * 2021-06-02 2024-03-08 北京汇声汇语科技有限公司 一种基于深度学习的部门语义信息抽取的方法及装置
CN113793668A (zh) * 2021-09-17 2021-12-14 平安科技(深圳)有限公司 基于人工智能的症状标准化方法、装置、电子设备及介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103164616A (zh) * 2013-02-02 2013-06-19 杭州卓健信息科技有限公司 一种智能导诊系统和方法
CN110032648A (zh) * 2019-03-19 2019-07-19 微医云(杭州)控股有限公司 一种基于医学领域实体的病历结构化解析方法
US20200152226A1 (en) * 2018-11-13 2020-05-14 CurieAI, Inc. Design of Stimuli for Symptom Detection
CN111292821A (zh) * 2020-01-21 2020-06-16 上海联影智能医疗科技有限公司 一种医学诊疗系统

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110069631B (zh) * 2019-04-08 2022-11-29 腾讯科技(深圳)有限公司 一种文本处理方法、装置以及相关设备
CN110705293A (zh) * 2019-08-23 2020-01-17 中国科学院苏州生物医学工程技术研究所 基于预训练语言模型的电子病历文本命名实体识别方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103164616A (zh) * 2013-02-02 2013-06-19 杭州卓健信息科技有限公司 一种智能导诊系统和方法
US20200152226A1 (en) * 2018-11-13 2020-05-14 CurieAI, Inc. Design of Stimuli for Symptom Detection
CN110032648A (zh) * 2019-03-19 2019-07-19 微医云(杭州)控股有限公司 一种基于医学领域实体的病历结构化解析方法
CN111292821A (zh) * 2020-01-21 2020-06-16 上海联影智能医疗科技有限公司 一种医学诊疗系统

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113628726A (zh) * 2021-08-10 2021-11-09 海南榕树家信息科技有限公司 基于图神经网络的中医辩治推荐系统、方法和电子设备
CN113628726B (zh) * 2021-08-10 2023-12-26 海南榕树家信息科技有限公司 基于图神经网络的中医辨治推荐系统、方法和电子设备
CN115662593A (zh) * 2022-11-08 2023-01-31 北京健康在线技术开发有限公司 基于症状知识图谱的医患匹配方法、装置、设备及介质
CN117637092A (zh) * 2024-01-24 2024-03-01 创智和宇信息技术股份有限公司 一种基于人工智能模型的病历预编码方法及装置
CN117637092B (zh) * 2024-01-24 2024-04-23 创智和宇信息技术股份有限公司 一种基于人工智能模型的病历预编码方法及装置

Also Published As

Publication number Publication date
CN111785367A (zh) 2020-10-16

Similar Documents

Publication Publication Date Title
US11810671B2 (en) System and method for providing health information
WO2021139231A1 (zh) 基于神经网络模型的分诊方法、装置和计算机设备
WO2021139232A1 (zh) 基于医疗知识图谱的分诊方法、装置、设备及存储介质
US11610678B2 (en) Medical diagnostic aid and method
CN112507696B (zh) 基于全局注意力意图识别的人机交互导诊方法与系统
US20200027560A1 (en) Drawing conclusions from free form texts with deep reinforcement learning
CN111259111B (zh) 基于病历的辅助决策方法、装置、电子设备和存储介质
WO2021151356A1 (zh) 分诊数据处理方法、装置、计算机设备及存储介质
US20240170161A1 (en) Method for processing medical data, apparatus, and storage medium
CN113223735A (zh) 基于对话表征的分诊方法、装置、设备及存储介质
CN116975218A (zh) 文本处理方法、装置、计算机设备和存储介质
CN114999676A (zh) 用于自动回复医疗咨询的方法、系统、装置和介质
Dammavalam et al. AI based chatbot for hospital management system
CN117457162A (zh) 基于多编码器和多模态信息融合的急诊分诊方法及系统
CN116453674A (zh) 一种智慧医疗系统
US20220036180A1 (en) Reinforcement learning approach to approximate a mental map of formal logic
CN117112739A (zh) 增强意图理解的医疗对话系统
CN116595994A (zh) 基于提示学习的矛盾信息预测方法、装置、设备及介质
CN116151273A (zh) 基于Transformer和知识图谱的智能交互方法
Ma et al. Medical answer selection based on two attention mechanisms with birnn
CN113761899A (zh) 一种医疗文本生成方法、装置、设备及存储介质
CN113314236A (zh) 一种面向高血压的智能问答系统
CN110289065A (zh) 一种辅助生成医学电子报告的控制方法以及装置
CN113724882B (zh) 基于问诊会话构建用户画像的方法、装置、设备和介质
EP4362033A1 (en) Patient consent

Legal Events

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

Ref document number: 20912235

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20912235

Country of ref document: EP

Kind code of ref document: A1