WO2019136806A1 - 医疗模型训练方法、医疗识别方法、装置、设备及介质 - Google Patents

医疗模型训练方法、医疗识别方法、装置、设备及介质 Download PDF

Info

Publication number
WO2019136806A1
WO2019136806A1 PCT/CN2018/077475 CN2018077475W WO2019136806A1 WO 2019136806 A1 WO2019136806 A1 WO 2019136806A1 CN 2018077475 W CN2018077475 W CN 2018077475W WO 2019136806 A1 WO2019136806 A1 WO 2019136806A1
Authority
WO
WIPO (PCT)
Prior art keywords
medical
model
target medical
target
medical data
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2018/077475
Other languages
English (en)
French (fr)
Inventor
鲁宁
周凡
程诚
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
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 Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2019136806A1 publication Critical patent/WO2019136806A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/951—Indexing; Web crawling techniques
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques

Definitions

  • the present application relates to the field of data identification, and in particular, to a medical model training method, a medical identification method, a device, a device, and a medium.
  • the identification of the current disease is usually determined by the doctor according to the patient's symptoms, depends on the doctor's professional and experience, and can not enable the user to know his or her health according to his or her symptoms. Since the identification of current diseases is mainly diagnosed by doctors, the average user will be aware when the condition is serious, and then go to a medical institution for treatment, which may delay the condition.
  • the medical data such as symptoms, diseases and treatments provided in the current medical website do not form a complete system. The user does not have enough professional knowledge to inquire about their own health status according to their own symptoms.
  • the embodiment of the present application provides a medical model training method for solving the current disease identification, which is usually determined by a doctor according to the patient's symptoms, and depends on the doctor's professional and experience, and cannot enable the user to know his or her health condition according to his or her symptoms.
  • the problem is usually determined by a doctor according to the patient's symptoms, and depends on the doctor's professional and experience, and cannot enable the user to know his or her health condition according to his or her symptoms.
  • the embodiment of the present application provides a medical identification method to solve the problem that the medical data provided by the current medical website does not form a complete system, and the user does not have sufficient professional knowledge when querying, and cannot understand according to his own symptoms. The problem of your own health status.
  • an embodiment of the present application provides a medical model training method, including:
  • the target medical data is divided into preset proportions to obtain a training set and a test set;
  • the long-term and short-term memory network model is trained by using the target medical data in the training set to obtain the original medical model
  • the original medical model is tested using target medical data in the test set to obtain a target medical model.
  • an embodiment of the present application provides a medical model training apparatus, including:
  • a target medical data acquisition module for acquiring target medical data
  • a target medical data dividing module configured to divide the target medical data into a preset ratio to obtain a training set and a test set;
  • the original medical model acquisition module is configured to train the long and short time memory network model by using the target medical data in the training set to obtain the original medical model;
  • the target medical model obtaining module is configured to test the original medical model by using target medical data in the test set to obtain a target medical model.
  • the embodiment of the present application provides a medical identification method, including:
  • the medical data to be tested including at least one symptom characteristic
  • At least one of the symptom features is input to the target medical model for identification, and the target medical feature is acquired, and the target medical model is a model acquired by the medical model training method of the first aspect.
  • the embodiment of the present application provides a medical identification device, including:
  • the medical data acquisition module to be tested is configured to acquire medical data to be tested of the user, where the medical data to be tested includes at least one symptom feature;
  • a target medical feature obtaining module configured to input at least one symptom feature into the target medical model for obtaining a target medical feature, wherein the target medical model is a model obtained by using the medical model training method of the first aspect.
  • an embodiment of the present application provides a terminal device, including a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, where the processor executes the computer The following steps are implemented when reading the instruction:
  • the target medical data is divided into preset proportions to obtain a training set and a test set;
  • the long-term and short-term memory network model is trained by using the target medical data in the training set to obtain the original medical model
  • the original medical model is tested using target medical data in the test set to obtain a target medical model.
  • an embodiment of the present application provides a terminal device, including a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, where the processor executes the computer The following steps are implemented when reading the instruction:
  • the medical data to be tested including at least one symptom characteristic
  • At least one of the symptom features is input to the target medical model for identification, and the target medical feature is acquired, and the target medical model is a model acquired by the medical model training method of the first aspect.
  • the embodiment of the present application provides a computer readable medium storing computer readable instructions, where the computer readable instructions are executed by a processor to implement the following steps:
  • the target medical data is divided into preset proportions to obtain a training set and a test set;
  • the long-term and short-term memory network model is trained by using the target medical data in the training set to obtain the original medical model
  • the original medical model is tested using target medical data in the test set to obtain a target medical model.
  • an embodiment of the present application provides a computer readable medium storing computer readable instructions, where the computer readable instructions are executed by a processor to implement the following steps:
  • the medical data to be tested including at least one symptom characteristic
  • At least one of the symptom features is input to the target medical model for identification, and the target medical feature is acquired, and the target medical model is a model acquired by the medical model training method of the first aspect.
  • the target medical data is first acquired, and the target medical data is divided, and the training set and the test set are obtained, so as to be based on the target medical data in the training set.
  • the memory network model is trained to effectively update the weights of each layer in the long-short memory network model to ensure that the original medical model acquired by the training can identify the target medical features.
  • the original medical model is tested based on the target medical data in the test set to ensure that the acquired target medical features identify the target medical features with higher accuracy.
  • the medical data to be tested of the user is acquired to identify the medical data to be measured based on the target medical model to obtain the target medical feature.
  • the medical data to be measured is identified by the target medical model to ensure higher recognition accuracy and more accurate identification of the target medical features to assist the user in judging the risk of the disease so as to prevent it in time.
  • Embodiment 1 is a flow chart of a medical model training method provided in Embodiment 1.
  • FIG. 2 is a specific schematic diagram of step S11 of FIG. 1.
  • FIG. 3 is a specific schematic diagram of step S112 of FIG. 2.
  • FIG. 4 is a specific schematic diagram of step S13 of FIG. 1.
  • FIG. 5 is a specific schematic diagram of step S14 of FIG. 1.
  • FIG. 6 is a schematic block diagram of a medical model training device provided in Embodiment 2.
  • Embodiment 7 is a flow chart of a medical recognition method provided in Embodiment 3.
  • FIG. 8 is a specific schematic diagram of step S22 of FIG. 7.
  • Figure 9 is a schematic block diagram of the medical identification device provided in Embodiment 4.
  • FIG. 10 is a schematic diagram of a terminal device provided in Embodiment 6.
  • Fig. 1 is a flow chart showing a medical model training method in the present embodiment.
  • the medical model training method can quickly collect a large amount of medical data from the network to perform medical model training based on the collected medical data.
  • the medical model training method can be specifically applied to the medical knowledge base management system, which is used for identifying the symptom features input by the user, and recommending a list of suspected disease features for the user, which can effectively increase the accuracy of the recognition.
  • the medical model training method includes the following steps:
  • the target medical data is data for performing model training.
  • the target medical data includes, but is not limited to, medical data such as symptom characteristics, disease characteristics, and the like in the present embodiment.
  • the target medical data is acquired for subsequent model training.
  • S12 The target medical data is divided into preset proportions to obtain a training set and a test set.
  • the preset ratio is a preset ratio for classifying the target medical data.
  • the preset ratio may be a ratio obtained based on historical experience.
  • the training set is a learning sample data set, and the classifier is built by matching some parameters, that is, training the machine learning model by using the target medical data in the training set to determine the parameters of the machine learning model.
  • the test set is used to test the resolving power of a trained machine learning model, such as recognition rate.
  • the target medical data can be classified according to a ratio of 9:1, and 90% of the target medical data can be used as the training set, and the remaining 10% of the data is used as the test set.
  • the Long Short-Term Memory (LSTM) model is a time recurrent neural network model for training data with time series characteristics, and the data with time series characteristics in long and short time memory networks.
  • Model training enables the acquisition of a recognition model corresponding to the data.
  • the data with the time series characteristics is the target medical data
  • the model obtained through the target medical data in the training set is the original medical model.
  • the long and short time memory network model includes an input layer, an output layer and at least one hidden layer.
  • the weight of each layer in the long and short time memory network model refers to the weight of each layer connection in the neural network model, and the weight determines the final layer.
  • the information is output and the network has a memory function on the timing.
  • the weights of each layer in the long and short time memory network model can be effectively updated.
  • the long-and short-term memory network model can identify the target medical data more accurately by identifying the target medical data in the training set with time series characteristics.
  • the output layer of the LSTM model is subjected to regression processing using Softmax (Regression Model) for classifying the output weight matrix.
  • Softmax regression model
  • Softmax is a classification function commonly used in neural networks. It maps the output of multiple neurons to the interval [0,1], which can be understood as probability. It is simple and convenient to calculate, so as to carry out multi-classification. Output to make its output more accurate.
  • the target medical model is a model obtained by continuously testing and improving the original medical model.
  • each target medical data in the test set is input to the original medical model for testing, and the predicted medical features are obtained; and then the accuracy of the model is determined based on the predicted medical characteristics and the target medical data in the test set, so that the model-based Accuracy, get the target medical model.
  • the target medical data is first acquired, and the target medical data is divided, and the training set and the test set are acquired, so as to train the long and short time memory network model based on the target medical data in the training set, and the long and short time memory network can be effectively updated.
  • the weight of each layer in the model to ensure that the original medical model acquired by the training can identify the target medical features.
  • the original medical model is tested based on the target medical data in the test set to ensure that the target medical features are more accurate in identifying the target medical features.
  • step S11 the target medical data is acquired, which specifically includes the following steps:
  • the target webpage address is a webpage address corresponding to the target medical data to be obtained in advance. For example: 39 health website. Support for obtaining target medical data based on the target web page address by obtaining the target web page address to enable subsequent crawling of the target medical data using the web crawler technology.
  • the crawler tool is a tool for automatically crawling the webpage content corresponding to the webpage address according to certain rules, such as a Python crawler tool.
  • the crawler tool is used to crawl the webpage content corresponding to the target webpage address to obtain the original medical data, and each of the original medical data includes the actual disease characteristic and the corresponding at least one symptom characteristic, such as a cold (actual disease characteristic) corresponding flow. Symptoms such as snot, headache and fever.
  • the crawler tool is used to crawl the webpage corresponding to the target webpage address, and no manual search is needed, which is beneficial to improving the efficiency of data collection.
  • the crawler tool performs crawling data on the target webpage address in a periodic crawling manner, so that the original medical data is time-series, so that the target medical features acquired by the subsequent model training are also sequential.
  • S113 Perform data cleaning on the original medical data to obtain target medical data.
  • data cleaning refers to the method of processing raw medical data according to certain rules to obtain pure target medical data.
  • Target medical data refers to pure data processed in accordance with data cleaning rules.
  • the data cleaning rules include, but are not limited to, removing duplicate data. Since each raw medical data contains actual disease characteristics and corresponding at least one symptom characteristic, when cleaning according to the data cleaning rule, two or more original medical data having the same disease characteristics and symptom characteristics can be combined into one. Raw medical data.
  • the quality of the target medical data can be effectively improved, and the repeated original medical data is removed, so that when the target medical data is subsequently used for training, it is not necessary. Re-training the repeated raw medical data can effectively reduce the training time, save time and improve training efficiency.
  • the target webpage address is obtained first, and then the crawler tool is used to crawl the webpage corresponding to the target webpage address to obtain the original medical data, and no manual search is needed, which is beneficial to improving the efficiency of data collection. Finally, the original medical data is cleaned and the target medical data is obtained, so that when the target medical data is used for training, there is no need to re-train the repeated data, which can effectively reduce the training time, save time and improve training efficiency.
  • step S112 the crawler tool is used to crawl the webpage corresponding to the target webpage address to obtain the original medical data, which specifically includes the following steps:
  • S1121 Using a crawler tool, crawling at least one access address linked by the target webpage address according to a depth-first algorithm or a breadth-first algorithm, where each access address corresponds to a webpage.
  • the depth-first algorithm means that the web crawler starts from the start page, and a link follows a link to track down. After processing the line, it transfers to the next start page and continues to track the link.
  • the target webpage address corresponds to a target webpage, and the target webpage includes at least one accessing address linked to the starting page and the at least one starting page, and each accessing address corresponds to a visiting webpage.
  • the crawler tool includes a web page extraction tool and a web page download tool.
  • the webpage extraction tool is a tool for extracting an access address, and step S1121 specifically uses a webpage extraction tool to crawl at least one access address linked by the target webpage address.
  • the webpage downloading tool is a tool for downloading a webpage corresponding to an access address.
  • the breadth-first algorithm can also be used to continuously crawl a new webpage address from the current page into the message queue to be downloaded, and stop executing the crawler task until the preset stop condition is satisfied.
  • the breadth-first algorithm refers to inserting the link found by the newly downloaded webpage directly into the end of the queue to be crawled, that is, the web crawler first crawls all the webpages in the start page, and then selects one of the links. The webpage continues to crawl all pages linked in this page.
  • S1122 Store at least one access address in a message queue to be downloaded.
  • each access address crawled in step S121 is stored in the queue to be downloaded according to the chronological order of the crawling, so that when step S123 is performed, the crawling may be based on the webpage address in the message queue to be downloaded.
  • the message queue to be downloaded processes the access address according to the advanced first-in-first method, so that the crawl access address and the original medical data can be crawled asynchronously based on the access address, which is beneficial to improving the efficiency of obtaining the original medical data.
  • the webpage extraction tool in the crawler tool first grabs all the webpage content in the start page, and then selects at least one access address linked by the startpage, and continues to crawl the webpage of the visitor address link.
  • the start page is a target webpage address.
  • the crawler tool is used to extract data from the webpage corresponding to each access address in the download message queue to obtain original medical data.
  • the webpage downloading tool automatically downloads all the medical data in the webpage corresponding to the accessing address according to each access address of the message queue to be downloaded.
  • a plurality of webpage addresses including the original medical data are stored in the message queue to be downloaded, and the webpage downloading tool of the crawler tool sequentially obtains the accessing address from the to-be-downloaded message queue and downloads the webpage corresponding to the accessing address.
  • Raw medical data Specifically, the crawler tool obtains an access address from the head of the message queue to be downloaded and downloads the webpage corresponding to the access address, stores the downloaded original medical data in the database, and unregisters the corresponding webpage address in the message queue to be downloaded.
  • the webpage downloading tool in the crawler tool automatically crawls the webpage address including the original medical data from the Internet according to the crawler task set by the user, and does not need manual search, which is beneficial to improving data collection efficiency.
  • the crawler tool is used to crawl at least one access address linked by the target webpage address according to the depth-first algorithm or the breadth-first algorithm, and the obtained access address is stored in the message queue to be downloaded, and then the crawler tool is used based on The access address obtained in the message queue to be downloaded downloads the original medical data, so that the original medical data downloaded by the access address is processed asynchronously, which is beneficial to improving the efficiency of obtaining the original medical data.
  • obtaining the original medical data asynchronously by using the webpage extraction tool and the webpage downloading tool is beneficial to improving the efficiency of acquiring the original medical data.
  • step S13 the long-term and short-term memory network model is trained by using the target medical data in the training set to obtain the original medical model, which specifically includes the following steps:
  • the long and short time memory network model is initialized, wherein the long and short time memory neural network is a network interconnected in time, and the basic unit is called a neuron.
  • the long and short time memory network model includes an input layer, an output layer and at least one hidden layer, and the hidden layer includes an input gate, an forgetting gate, an output gate, a neuron state and a neuron output, and each layer in the long and short time memory network model. Multiple neurons can be included.
  • the Forgetting Gate determines the information to be discarded in the neuron state.
  • the input gate determines the information to be added in the neuron.
  • the output gate determines the information to be output in the neuron.
  • the state of the neuron determines the information that each gate discards, adds, and outputs, and is specifically expressed as the weight of the connection to each gate.
  • the neuron output determines the connection weight to the next layer.
  • the long and short time memory network model is initialized, that is, the weight of the connection between the layers of the long and short memory network model and the input gate, the forgetting gate, the output gate, the neuron state and the neuron output in the hidden layer.
  • the initial weight in this embodiment, the initial weight can be set to 1.
  • S132 Input target medical data in the training set in the long and short time memory network model, and calculate output values of each layer of the long and short time memory network model.
  • the target medical data in the training set acquired in a preset time period according to the unit time interval is input into the long and short time memory network model, and the output values of each layer are respectively calculated, including calculating the target medical data in the training set.
  • one neuron includes three activation functions f(sigmoid), g(tanh), and h(softmax).
  • the activation function can transform the weight result into a classification result, which can add some nonlinear factors to the neural network, so that the neural network can better solve the more complicated problems.
  • the data received and processed by a neuron includes: target medical data in the input training set: x, status data: s.
  • the parameters mentioned below also include that the input of the neuron is represented by a and the output is represented by b.
  • the subscripts ⁇ , ⁇ , and ⁇ represent the input gate, the forgetting gate, and the output gate, respectively.
  • the subscript c represents a neuron and t represents a moment.
  • the weights of the neurons connected to the input gate, the forgetting gate, and the output gate are recorded as w cl , w c ⁇ , and w c ⁇ , respectively .
  • the input gate receives the sample X t at the current moment, the output value b t-1 h at the previous moment, and the state data S t-1 c of the neuron at the previous moment, by connecting the target medical data of the input training set with the input gate
  • Get a scalar of 0-1 intervals This scalar controls the proportion of neurons that receive current information based on a combination of current state and past state.
  • the forgetting gate receives the sample X t of the current time, the output value b t-1 h of the previous moment, and the state data S t-1 c of the previous moment, by connecting the target medical data of the input training set and the weight of the forgetting gate.
  • right right w i ⁇ connected to the output value of a time value w h ⁇ forgetting gate and the gate is connected neurons forgetting value w c ⁇ , according to the formula Calculate the output of the forgotten gate Acting on the activation function f
  • a scalar of 0-1 interval is obtained, which controls the proportion of the past information received by the neuron based on the comprehensive judgment of the current state and the past state.
  • the neuron receives the sample X t of the current time, the output value b t-1 h of the previous time, and the state data S t-1 c of the previous time, the weight of the target medical data of the connected neuron and the input training set w Ic , the weight of the output value of the connected neuron and the previous moment, w hc, and the output scalar of the input gate and the forgetting gate, according to the formula Calculate the neuron status at the current time
  • the output gate receives the current time sample and the current time state data X t , the previous time output value b t-1 h and the current time state data.
  • the output values of the layers of the long and short time memory network model can be obtained by calculating the target medical data in the training set between the layers.
  • S133 Perform error back propagation update on each layer of the long and short time memory network model according to the output value, and obtain the updated weight of each layer, wherein the expression of the error back propagation update is Where T is the time, W is the weight, B is the output value, and ⁇ is the error term.
  • T is the time
  • W is the weight
  • B is the output value
  • ⁇ is the error term.
  • b t-1 h is the output value of the previous moment.
  • the error back-reversal update is performed on each layer of the long-short-time memory network model according to the output values of the layers of the long-short-time memory network model. Specifically, first based on the expression of the error term The error term of each layer can be found. Where ⁇ and ⁇ both represent error terms, in particular, An error term indicating the back-transition of the neuron output, An error term indicating the back-transmission of a neuron state, both of which indicate an error term, but the specific meaning is different.
  • the input of a neuron is represented by a
  • the output is represented by b.
  • the subscripts ⁇ , ⁇ , and ⁇ represent the input gate, the forgetting gate, and the output gate, respectively.
  • the subscript c represents a neuron and t represents a moment.
  • the weights of the neurons connected to the input gate, the forgetting gate, and the output gate are recorded as w cl , And w c ⁇ .
  • S c represents the neuron state
  • the activation function of the control gate is represented by f(sigmoid)
  • g(tanh) and h(softmax) represent the input activation function and the output activation function of the neuron, respectively.
  • K is the number of neurons in the output layer
  • H is the number of neurons in the hidden layer
  • the weight of each layer can be updated by calculating the weight gradient.
  • the expression of the weight update is Where T represents the time and W represents the weight, such as connection weights such as w cl , w c ⁇ and w c ⁇ .
  • B represents the output value, such as Wait for the output.
  • ⁇ represents the error term, such as Equal error term.
  • b t-1 h is the output value of the previous moment.
  • the parameters of the above expressions need to be corresponding. If the specific weight of the update is w cl , the output B is corresponding.
  • the error term ⁇ is the corresponding.
  • the required parameter value of the weight update expression can be obtained according to the expressions of step S132 and step S133. Then, according to the expression updated by the weight, the weight of each layer after the update is obtained.
  • the acquired original medical model can be obtained by applying the obtained weights of the updated layers to the long and short time memory network model. Further, the weights between the layers in the original medical model enable the original medical model to decide which old information to discard, which new information to add, and which information to output. At the output layer of the original medical model, a probability value is finally output, which indicates the closeness of the information to the original medical model after being processed by the original medical model, and can be widely applied to medical feature recognition to achieve accurate identification of medical features. .
  • step S14 the original medical model is tested by using the target medical data in the test set to obtain the target medical model, which specifically includes the following steps:
  • S141 Enter at least one symptom feature in each target medical data into the original medical model for testing, and obtain corresponding predicted medical features.
  • each target medical data contains at least one symptom characteristic and corresponding actual disease characteristics, such as a runny nose and/or a fever (symptom characteristic) corresponding to a cold (actual disease characteristic).
  • each target medical data in the test set is input to the original medical model for testing, that is, steps S31-S34 are performed to obtain corresponding predicted medical features.
  • S142 Acquire an identification quantity that matches an actual disease feature and a predicted medical feature in the same target medical data.
  • the predicted disease feature acquired in step S41 is matched with the actual disease feature in the same target medical data. If the matching is successful, the accurate number of the recognition result is increased by 1. If the matching is unsuccessful, the accurate result is recognized. The number that is constant, and finally the statistical recognition result is the number of identification.
  • S143 Obtain a recognition accuracy rate based on the total amount of data of the target quantity and the target medical data in the test set.
  • the number of identifications obtained in step S42 is divided by the total amount of data of all target medical data in the test set, to obtain the recognition accuracy rate.
  • the recognition accuracy can be formulated Perform calculation acquisition, where T is the exact number of recognition results, and N is the total amount of data of the target medical data.
  • the preset accuracy rate is a preset probability value for evaluating the quality of the model. Specifically, if the recognition accuracy rate is greater than the preset accuracy rate, determining that the original medical model is more accurate to use the original medical model as the target medical model; and if the recognition accuracy is not greater than the preset accuracy rate, determining the original The medical model is not accurate enough. It is necessary to retrain the original medical model with more target medical data. When the recognition accuracy of the original medical model obtained by the training is greater than the preset accuracy rate, the model training is completed to obtain the target medical model.
  • At least one symptom feature in each target medical data is input to the original medical model for testing, corresponding predicted medical features are acquired, and then the actual disease characteristics in the same target medical data are matched with the predicted medical features.
  • the amount and amount of data for the target medical data in the test set to obtain recognition accuracy. Finally based on the recognition accuracy. Obtain a target medical model to improve the accuracy of medical feature recognition.
  • the target medical data is first acquired, and the target medical data is divided, and the training set and the test set are acquired, so as to train the long and short time memory network model based on the target medical data in the training set, and the long and short time memory network can be effectively updated.
  • the weight of each layer in the model makes the original medical feature recognition obtained through the target medical data training more accurate.
  • the original medical model is tested based on the target medical data in the test set to ensure that the acquired target medical features identify the target medical features with higher accuracy.
  • Fig. 6 is a block diagram showing the principle of the medical model training device corresponding to the medical model training method in the first embodiment.
  • the medical model training device includes a target medical data acquisition module 11, a target medical data division module 12, an original medical model acquisition module 13, and a target medical model acquisition module 14.
  • the implementation functions of the target medical data acquisition module 11, the target medical data division module 12, the original medical model acquisition module 13, and the target medical model acquisition module 14 correspond one-to-one with the steps corresponding to the medical model training method in the embodiment, in order to avoid redundancy This embodiment is not described in detail.
  • the target medical data acquisition module 11 is configured to acquire target medical data.
  • the target medical data dividing module 12 is configured to divide the target medical data into a preset ratio to obtain a training set and a test set.
  • the original medical model acquisition module 13 is configured to train the long and short time memory network model by using the target medical data in the training set to obtain the original medical model.
  • the target medical model obtaining module 14 is configured to test the original medical model by using the target medical data in the test set to obtain the target medical model.
  • the target medical data acquisition module 11 includes a target webpage address acquisition unit 111, an original medical data acquisition unit 112, and a target medical data acquisition unit 113.
  • the target webpage address obtaining unit 111 is configured to acquire a target webpage address.
  • the original medical data obtaining unit 112 is configured to use the crawler tool to crawl the webpage corresponding to the target webpage address to obtain the original medical data.
  • the target medical data acquiring unit 113 is configured to perform data cleaning on the original medical data to acquire target medical data.
  • the original medical data acquisition unit 112 includes an access address acquisition sub-unit 1121, an access address storage sub-unit 1122, and an original medical data acquisition sub-unit 1123.
  • the access address obtaining sub-unit 1121 is configured to use a crawler tool to crawl at least one access address linked by the target webpage address according to a depth-first algorithm or a breadth-first algorithm, where each access address corresponds to a webpage.
  • the access address storage sub-unit 1122 stores at least one access address in the message queue to be downloaded.
  • the original medical data acquisition sub-unit 1123 uses the crawler tool to perform data extraction on the webpage corresponding to each access address in the download message queue to obtain the original medical data.
  • the original medical model acquisition module 13 includes a network model initialization unit 131, an output value calculation unit 132, a weight update unit 133, and an original medical model acquisition unit 134.
  • the network model initializing unit 131 is configured to initialize the long and short time memory network model.
  • the output value calculation unit 132 is configured to input positive and negative samples in the long and short time memory network model, and calculate output values of each layer of the long and short time memory network model.
  • the weight update unit 133 is configured to perform error back propagation update on each layer of the long and short time memory network model according to the output value, and obtain weights of the updated layers.
  • the original medical model obtaining unit 134 is configured to acquire the original medical model based on the weights of the updated layers.
  • the target medical model acquisition module 14 includes a predicted medical feature acquisition unit 141, an identification number acquisition unit 142, a recognition accuracy acquisition unit 143, and a target medical model acquisition unit 144.
  • the predicted medical feature acquiring unit 141 is configured to input at least one symptom feature in each target medical data to the original medical model for testing, and obtain a corresponding predicted medical feature.
  • the identification quantity obtaining unit 142 is configured to acquire the identification quantity that the actual disease feature in the same target medical data matches the predicted medical feature.
  • the recognition accuracy acquisition unit 143 is configured to acquire the recognition accuracy rate based on the identification quantity and the total amount of data of the target medical data in the test set.
  • the target medical model obtaining unit 144 is configured to acquire the target medical model when the recognition accuracy is greater than the preset accuracy.
  • Fig. 7 is a flow chart showing the medical recognition method in the embodiment.
  • the medical identification method can be applied to a terminal device of a medical institution or other institution to identify symptom characteristics and achieve the effect of intelligently identifying diseases. As shown in FIG. 7, the medical identification method includes the following steps:
  • S21 Obtain medical data of the user to be tested, and the medical data to be tested includes at least one symptom feature.
  • the medical data to be tested refers to medical data selected by the user for identifying the probability of the target medical feature.
  • the medical data to be tested includes at least one symptom characteristic, such as symptom characteristics such as runny nose, body temperature, and blood pressure.
  • the target medical feature is the diseased feature with the greatest probability obtained by using the target medical model to identify the symptom features.
  • at least one symptom feature is input to the target medical model for identification, that is, steps S131-S134 in Embodiment 1 are performed to acquire the target medical feature.
  • the medical data to be tested of the user is first acquired, and the medical data to be tested includes at least one symptom feature, so that at least one symptom feature is input to the target medical model for identification, and the target medical feature is obtained, thereby ensuring accurate acquisition of the recognition probability value. Sex to help users judge the risk of illness.
  • step S22 at least one symptom feature is input to the target medical model for identification, and the target medical feature is obtained, which specifically includes the following steps:
  • S221 Identify at least one symptom feature based on the target medical model, and acquire at least one identified medical feature and a corresponding recognition probability value.
  • the at least one symptom feature is input into the target medical model for identification, and the input at least one symptom feature is subjected to conversion processing based on the inter-layer weights, and at least one identification medical feature is outputted at the output layer.
  • Correspondence identification probability value is to identify a disease characteristic that it may have when identifying at least one symptom feature using the target medical model. Identifying the probability value is the possibility of identifying at least one symptom when the target medical model is used to identify the medical feature. Wherein, the recognition probability value may be a real number between 0-1.
  • the identified medical feature corresponding to the largest recognition probability value is selected as the target medical feature.
  • the database also associates the department corresponding to the target medical feature with at least one drug name.
  • the corresponding department and the at least one drug name are associated and outputted to enhance the practicability of the medical recognition model.
  • the at least one symptom feature is identified based on the target medical model, and at least one identified medical feature and a corresponding recognition probability value are acquired, so that the target medical feature is determined based on the recognition probability value, that is, the disease characteristic suspected of being sick is determined. Ensure that medical feature recognition results are more accurate and reliable.
  • the medical data to be tested of the user is acquired to identify the medical data to be measured based on the target medical model to obtain the target medical feature.
  • the medical data to be measured is identified by the target medical model to ensure higher recognition accuracy and more accurate identification of the target medical features to assist the user in judging the risk of the disease so as to prevent it in time.
  • Fig. 9 is a block diagram showing the principle of the medical recognition device corresponding to the medical recognition method in the first embodiment.
  • the medical identification device includes a medical data acquisition module 21 to be tested and a target medical feature acquisition module 22.
  • the implementation functions of the medical data acquisition module 21 and the target medical feature acquisition module 22 correspond to the steps corresponding to the medical identification method in the embodiment. To avoid redundancy, the present embodiment will not be described in detail.
  • the medical data acquisition module 21 to be tested is configured to acquire medical data to be tested of the user, and the medical data to be tested includes at least one symptom feature.
  • the target medical feature acquisition module 22 is configured to input at least one symptom feature into the target medical model for identification, and acquire the target medical feature.
  • the target medical feature acquisition module 22 includes a recognition probability value acquisition unit 221 and a target medical feature acquisition unit 222.
  • the recognition probability value obtaining unit 221 is configured to identify at least one symptom feature based on the target medical model, and acquire at least one identified medical feature and a corresponding recognition probability value.
  • the target medical feature acquiring unit 222 is configured to select the identified medical feature corresponding to the largest recognition probability value as the target medical feature.
  • the embodiment provides a computer readable storage medium having stored thereon computer readable instructions, which are implemented by a processor to implement the medical model training method of Embodiment 1, in order to avoid repetition, here No longer.
  • the computer readable instructions are executed by the processor, the functions of the modules/units in the medical model training device in Embodiment 2 are implemented. To avoid repetition, details are not described herein again.
  • the computer readable instructions are implemented by the processor to implement the functions of the steps in the medical identification method in the third embodiment. To avoid repetition, details are not described herein.
  • the functions of the modules/units in the medical identification device in Embodiment 4 are implemented. To avoid repetition, details are not described herein.
  • the computer readable storage medium can include any entity or device capable of carrying the computer readable instruction code, a recording medium, a USB flash drive, a removable hard drive, a magnetic disk, an optical disk, a computer memory, a read only memory (ROM, Read-Only) Memory), random access memory (RAM), electrical carrier signals, telecommunications signals, and software distribution media.
  • FIG. 10 is a schematic diagram of a terminal device according to an embodiment of the present application.
  • the terminal device 100 of this embodiment includes a processor 101, a memory 102, and computer readable instructions 103 stored in the memory 102 and executable on the processor 101.
  • the processor 101 implements the functions of the steps of the medical model training method in Embodiment 1 when the computer readable instructions 103 are executed, such as steps S11 to S14 shown in FIG.
  • the processor 101 implements the functions of the modules/units of the medical model training device in the embodiment 2 when the computer readable instructions 103 are executed, such as the functions of the modules 11 to 14 shown in FIG.
  • the processor 101 executes the computer readable instructions 103
  • the functions of the steps of the medical identification method in the third embodiment are implemented.
  • the computer readable instructions implement the functions of the various modules/units in the medical identification device of Embodiment 4 when the processor 101 executes the computer readable instructions 103. To avoid repetition, we will not go into details here.
  • computer readable instructions 103 may be partitioned into one or more modules/units, one or more modules/units being stored in memory 102 and executed by processor 101 to complete the application.
  • the one or more modules/units may be an instruction segment of a series of computer readable instructions 103 capable of performing a particular function for describing the execution of computer readable instructions 103 in the terminal device 100.
  • the computer readable instructions 103 may be divided into the target medical data acquisition module 11 , the target medical data division module 12 , the original medical model acquisition module 13 , and the target medical model acquisition module 14 in Embodiment 2, or in Embodiment 4
  • the specific functions of the modules are as described in the second embodiment or the fourth embodiment, and are not described here.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
  • the above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

一种医疗模型训练方法包括：获取目标医疗数据(S11)；将目标医疗数据按预设比例进行划分，获取训练集和测试集(S12)；采用训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型(S13)；采用测试集中的目标医疗数据对原始医疗模型进行测试，获取目标医疗模型(S14)。该医疗模型训练方法提高了疾病识别的准确率。还公开了一种医疗识别方法、装置、设备及介质。

Description

医疗模型训练方法、医疗识别方法、装置、设备及介质
本专利申请以2018年1月12日提交的申请号为201810030552.5，名称为“医疗模型训练方法、医疗识别方法、装置、设备及介质”的中国发明专利申请为基础，并要求其优先权。
技术领域
本申请涉及数据识别领域，尤其涉及一种医疗模型训练方法、医疗识别方法、装置、设备及介质。
背景技术
当前疾病的识别通常由医生根据病人的症状确定，需依赖于医生的专业和经验，无法使用户根据自身的症状情况及时了解自身的健康情况。由于当前疾病的识别主要由医生进行诊断，一般用户都是在病情比较严重时才会意识到，进而到医疗机构就诊，可能延误病情。当前医疗网站中提供的症状、疾病和治疗等医疗数据没有形成完整的系统，用户查询时没有足够的专业知识，无法根据自己的症状情况及时了解自身的健康状态。
发明内容
本申请实施例提供一种医疗模型训练方法，以解决当前疾病的识别通常由医生根据病人的症状确定，需依赖于医生的专业和经验，无法使用户根据自身的症状情况及时了解自身的健康情况的问题。
本申请实施例提供一种医疗识别方法，以解决当前医疗网站中提供的症状、疾病和治疗等医疗数据没有形成完整的系统，用户查询时没有足够的专业知识，无法根据自己的症状情况及时了解自身的健康状态的问题。
第一方面，本申请实施例提供一种医疗模型训练方法，包括：
获取目标医疗数据；
将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模 型；
采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
第二方面，本申请实施例提供一种医疗模型训练装置，包括：
目标医疗数据获取模块，用于获取目标医疗数据；
目标医疗数据划分模块，用于将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
原始医疗模型获取模块，用于采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
目标医疗模型获取模块，用于采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
第三方面，本申请实施例提供一种医疗识别方法，包括：
获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用第一方面所述医疗模型训练方法获取的模型。
第四方面，本申请实施例提供一种医疗识别装置，包括：
待测医疗数据获取模块，用于获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
目标医疗特征获取模块，用于将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用第一方面所述医疗模型训练方法获取的模型。
第五方面，本申请实施例提供一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
获取目标医疗数据；
将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
第六方面，本申请实施例提供一种终端设备，包括存储器、处理器以及存储在所述存 储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用第一方面所述医疗模型训练方法获取的模型。
第七方面，本申请实施例提供一种计算机可读介质，所述计算机可读介质存储有计算机可读指令，所述计算机可读指令被处理器执行时实现如下步骤：
获取目标医疗数据；
将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
第八方面，本申请实施例提供一种计算机可读介质，所述计算机可读介质存储有计算机可读指令，所述计算机可读指令被处理器执行时实现如下步骤：
获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用第一方面所述医疗模型训练方法获取的模型。
本申请实施例提供的医疗模型训练方法、装置、设备及介质中，先获取目标医疗数据，并对目标医疗数据进行划分，获取训练集和测试集，以便基于训练集中的目标医疗数据对长短时记忆网络模型进行训练，能够有效更新长短时记忆网络模型中各层的权值，以保证训练获取的原始医疗模型能够识别目标医疗特征。最后，再基于测试集中的目标医疗数据对原始医疗模型进行测试，以保证获取到的目标医疗特征识别目标医疗特征的准确率更高。
本申请实施例提供的医疗识别方法、装置、设备及介质中，通过获取用户的待测医疗数据，以便基于目标医疗模型对待测医疗数据进行识别，以获取目标医疗特征。通过目标医疗模型对待测医疗数据进行识别，以保证识别的准确率更高，较精准的识别目标医疗特征，以辅助用户判断患病风险，以便及时预防。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对本申请实施例的描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是实施例1中提供的医疗模型训练方法的一流程图。
图2是图1中步骤S11的一具体示意图。
图3是图2中步骤S112的一具体示意图。
图4是图1中步骤S13的一具体示意图。
图5是图1中步骤S14的一具体示意图。
图6是实施例2中提供的医疗模型训练装置的一原理框图。
图7是实施例3中提供的医疗识别方法的一流程图。
图8是图7中步骤S22的一具体示意图。
图9实施例4中提供的医疗识别装置的一原理框图。
图10是实施例6中提供的终端设备的一示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
实施例1
图1示出本实施例中医疗模型训练方法的流程图。该医疗模型训练方法可快速从网络中采集到大量的医疗数据，以便基于采集到的医疗数据进行医疗模型训练。该医疗模型训练方法可具体应用在医疗知识库管理系统这一数据管理系统中，用于对用户所输入的症状特征进行识别，为用户推荐疑似疾病特征列表，能够有效增加识别的准确率。如图1所示，该医疗模型训练方法包括如下步骤：
S11：获取目标医疗数据。
其中，目标医疗数据是用于进行模型训练的数据。该目标医疗数据包括但不限于本实施例中的症状特征、疾病特征等医疗数据。本实施例中，通过获取目标医疗数据，以便后续进行模型训练。
S12：将目标医疗数据按预设比例进行划分，获取训练集和测试集。
其中，预设比例是预先设定好的，用于对目标医疗数据进行分类的比例。该预设比例可以是根据历史经验获取的比例。其中，训练集(training set)是学习样本数据集，是通过匹配一些参数来建立分类器，即采用训练集中的目标医疗数据训练机器学习模型，以确定机器学习模型的参数。测试集(test set)是用于测试训练好的机器学习模型的分辨能力，如识别率。本实施例中，可按照9:1的比例对目标医疗数据进行分类，即可将90％的目标医疗数据作为训练集，剩余10％的数据作为测试集。
S13：采用训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型。
其中，长短时记忆网络(Long Short-Term Memory，简称LSTM)模型，是一种时间递归神经网络模型，用于训练具有时序性特点的数据，将该具有时序性特点的数据在长短时记忆网络模型训练，能够获取与该数据相对应的识别模型。本实施例中，该具有时序性特点的数据为目标医疗数据，通过训练集中的目标医疗数据获取的模型即为原始医疗模型。长短时记忆网络模型包括一输入层、一输出层和至少一隐藏层，长短时记忆网络模型中各层的权值是指神经网络模型中各层连接的权值，权值决定了各层最终输出的信息，并使得网络具有时序上的记忆功能。通过采用目标医疗数据对长短时记忆网络模型进行训练，能够有效更新长短时记忆网络模型中各层的权值。并且，长短时记忆网络模型通过对具有时序性特点的训练集中的目标医疗数据进行识别，可使目标医疗模型的识别结果更为准确。
本实施例中，LSTM模型的输出层采用Softmax(回归模型)进行回归处理，用于分类输出权重矩阵。Softmax(回归模型)是一种常用于神经网络的分类函数，它将多个神经元的输出，映射到[0,1]区间内，可以理解成概率，计算起来简单方便，从而来进行多分类输出，使其输出结果更准确。
S14：采用测试集中的目标医疗数据对原始医疗模型进行测试，获取目标医疗模型。
其中，目标医疗模型是经过不断对原始医疗模型进行测试、改进所得到的模型。本实施例中，将测试集中每一目标医疗数据输入到原始医疗模型进行测试，获取得到预测医疗特征；再基于预测医疗特征和测试集中的目标医疗数据，确定模型的准确率，以便基于模型的准确率，获取目标医疗模型。
本实施例中，先获取目标医疗数据，并对目标医疗数据进行划分，获取训练集和测试集，以便基于训练集中的目标医疗数据对长短时记忆网络模型进行训练，能够有效更新长短时记忆网络模型中各层的权值，以保证训练获取的原始医疗模型能够识别目标医疗特征。最后，再基于测试集中的目标医疗数据对原始医疗模型进行测试，以保证获取到的目 标医疗特征识别目标医疗特征的准确率更高。
在一具体实施方式中，如图2所示，步骤S11中，即获取目标医疗数据，具体包括如下步骤：
S111：获取目标网页地址。
其中，目标网页地址是预先定义好所要获取目标医疗数据对应的网页地址。例如：39健康网站。通过获取到目标网页地址，以使后续采用网络爬虫技术爬取目标医疗数据时，为基于目标网页地址获取目标医疗数据提供支持。
S112：采用爬虫工具爬取目标网页地址对应的网页，获取原始医疗数据。
其中，爬虫工具是按照一定的规则自动爬取网页地址所对应的网页内容的工具，例如Python爬虫工具。具体地，采用爬虫工具爬取目标网页地址所对应的网页内容，以获取原始医疗数据，每一原始医疗数据都包含实际疾病特征和对应的至少一个症状特征，例如感冒(实际疾病特征)对应流鼻涕、头疼和发烧等症状特征。本实施例中，采用爬虫工具爬取目标网页地址对应的网页，无需人工搜索，有利于提高数据采集的效率。
本实施例中，爬虫工具会采取周期性爬取的方式对目标网页地址进行爬取数据，以使原始医疗数据具有时序性，以使后续模型训练获取到的目标医疗特征也具有时序性。
S113：对原始医疗数据进行数据清洗，获取目标医疗数据。
其中，数据清洗是指对原始医疗数据按照一定规则进行处理，获取纯净的目标医疗数据的方法。目标医疗数据是指按照数据清洗规则进行处理得到的纯净的数据。该数据清洗规则包括但不限于去除重复的数据。由于每一原始医疗数据均包含实际疾病特征和对应的至少一个症状特征，依据数据清洗规则进行清洗时，可将实际疾病特征和症状特征均相同的两个或两个以上原始医疗数据合并为一原始医疗数据。本实施例中，通过对原始医疗数据进行数据清洗，以获取目标医疗数据，能够有效提升目标医疗数据的质量，并且，去除重复的原始医疗数据，以使后续采用目标医疗数据进行训练时，无需对重复的原始医疗数据进行再次训练，能够有效减少训练时长，节省时间，提高训练效率。
本实施例中，先获取目标网页地址，然后采用爬虫工具爬取目标网页地址对应的网页，以获取原始医疗数据，无需人工搜索，有利于提高数据采集的效率。最后对原始医疗数据进行数据清洗，获取目标医疗数据，以使后续采用目标医疗数据进行训练时，无需对重复的数据进行再次训练，能够有效减少训练时长，节省时间，提高训练效率。
在一具体实施方式中，如图3所示，步骤S112中，即采用爬虫工具爬取目标网页地址对应的网页，获取原始医疗数据，具体包括如下步骤：
S1121：采用爬虫工具，依据深度优先算法或广度优先算法爬取目标网页地址所链接的至少一个访问地址，每一访问地址对应一网页。
其中，深度优先算法是指网络爬虫会从起始页开始，一个链接一个链接跟踪下去，处理完这条线路之后再转入下一个起始页，继续追踪链接。目标网页地址对应一目标网页，该目标网页包括起始页和至少一个起始页所链接的至少一个访问地址，每一访问地址对应一访问网页。爬虫工具包括网页提取工具和网页下载工具。网页提取工具是用于提取访问地址的工具，步骤S1121具体是采用网页提取工具爬取目标网页地址所链接的至少一个访问地址。网页下载工具是用于下载访问地址对应的网页的工具。
本实施例中，还可采用广度优先算法不断从当前页面上爬取新的网页地址放入待下载消息队列中，直到预设停止条件满足时停止执行爬虫任务。其中，广度优先算法是指将新下载网页发现的链接直接插入到待抓取消息队列的末尾，也就是指网络爬虫会先抓取起始页中的所有网页，然后在选择其中的一个链接的网页，继续抓取在此网页中链接的所有网页。
S1122：将至少一个访问地址存储在待下载消息队列中。
具体地，将步骤S121中爬取到的每一访问地址依据爬取到的时间先后顺序存储在待下载消息队列中，以便在执行步骤S123时，可基于待下载消息队列中的网页地址进行爬取数据。待下载消息队列依据先进先入的方式对访问地址进行处理，可使爬取访问地址和基于访问地址爬取原始医疗数据异步处理，有利于提高获取原始医疗数据效率。
具体地，爬虫工具中的网页提取工具会先抓取起始页中的所有网页内容，然后再选择起始页所链接的至少一个访问地址，继续抓取此访问地址链接的网页。本实施例中，起始页为目标网页地址。
S1123：采用爬虫工具对待下载消息队列中的每一访问地址对应的网页进行数据提取，获取原始医疗数据。
具体地，采用网页下载工具根据待下载消息队列的每一访问地址自动下载该访问地址对应的网页中所有医疗数据。本实施例中，待下载消息队列中存储有多个包含原始医疗数据的网页地址，爬虫工具的网页下载工具依序从待下载消息队列中逐一获取访问地址并下载该访问地址对应的网页中的原始医疗数据。具体地，爬虫工具从待下载消息队列的队头获取到一访问地址并对该访问地址对应的网页进行下载，将下载的原始医疗数据存储在数据库后，注销待下载消息队列中相应的网页地址，重复上述步骤直至待下载消息队列中不存在访问地址，以获取爬虫工具爬取的所有网页中的原始医疗数据。本实施例中，爬虫工 具中的网页下载工具根据用户设置的爬虫任务自动从互联网上爬取包含原始医疗数据的网页地址，无需人工搜索，有利于提高数据采集效率。
本实施例中，采用爬虫工具，依据深度优先算法或广度优先算法爬取目标网页地址所链接的至少一个访问地址，并将获取到的访问地址存储在待下载消息队列中，再采用爬虫工具基于待下载消息队列中获取的访问地址下载原始医疗数据，使得访问地址下载的原始医疗数据异步处理，有利于提高原始医疗数据的获取效率。本实施例中，通过采用网页提取工具和网页下载工具异步处理获取原始医疗数据，有利于提高获取原始医疗数据的效率。
在一具体实施方式中，如图4所示，步骤S13中，即采用训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型，具体包括如下步骤：
S131：初始化长短时记忆网络模型。
本实施例中，对长短时记忆网络模型进行初始化操作，其中，长短时记忆神经网络是在时间上相互连接的网络，其基本单元称为神经元。长短时记忆网络模型包括一输入层、一输出层和至少一隐藏层，其隐藏层包括输入门、遗忘门、输出门、神经元状态和神经元输出，长短时记忆网络模型中的每一层可以包括多个神经元。遗忘门决定了在神经元状态中所要丢弃的信息。输入门决定了在神经元中所要增加的信息。输出门决定了在神经元中所要输出的信息。神经元状态决定了各个门丢弃、增加和输出的信息，具体表示为与各个门之间连接的权值。神经元输出决定了与下一层的连接权值。可以理解地，初始化长短时记忆网络模型，即为设置长短时记忆网络模型各层之间连接的权值以及隐藏层中输入门、遗忘门、输出门、神经元状态和神经元输出之间的初始权值，本实施例中初始权值可设为1。
S132：在长短时记忆网络模型中输入训练集中的目标医疗数据，计算长短时记忆网络模型各层的输出值。
本实施例中，采用按单元时间间隔在一预设时间段内获取的训练集中的目标医疗数据输入到长短时记忆网络模型中，分别计算各层的输出值，包括计算训练集中的目标医疗数据在输入门、遗忘门、输出门、神经元状态和神经元输出的输出。其中，一个神经元包括有三种激活函数f(sigmoid)、g(tanh)和h(softmax)。激活函数能够将权值结果转化成分类结果，其作用是能够给神经网络加入一些非线性因素，使得神经网络可以更好地解决较为复杂的问题。
一个神经元所接收和处理的数据包括：输入的训练集中的目标医疗数据：x，状态数 据：s。此外，以下提及的参数还包括：神经元的输入用a表示，输出用b表示。下标ι,φ和ω分别表示输入门、遗忘门和输出门。下标c表示神经元，t代表时刻。神经元跟输入门、遗忘门和输出门连接的权值分别记做w cl、w cφ和w cω。S c表示神经元状态。I表示输入层的神经元的个数，H是隐层神经元的个数，C是神经元状态的神经元个数，这里取C＝H。
输入门接收当前时刻的样本X t、上一时刻的输出值b t-1 h以及上一时刻神经元的状态数据S t-1 c，通过连接输入的训练集中的目标医疗数据与输入门的权值w il、连接上一时刻的输出值与输入门的权值w hl和连接神经元与输入门的权值w cl，根据公式
Figure PCTCN2018077475-appb-000001
计算得到输入门的输出
Figure PCTCN2018077475-appb-000002
将激活函数f作用于
Figure PCTCN2018077475-appb-000003
由公式
Figure PCTCN2018077475-appb-000004
得到一个0-1区间的标量。此标量控制了神经元根据当前状态和过去状态的综合判断所接收当前信息的比例。
遗忘门接收当前时刻的样本X t、上一时刻的输出值b t-1 h以及上一时刻的状态数据S t-1 c，通过连接输入的训练集中的目标医疗数据与遗忘门的权值w iφ、连接上一时刻的输出值与遗忘门的权值w hφ和连接神经元与遗忘门的权值w cφ，根据公式
Figure PCTCN2018077475-appb-000005
计算得到遗忘门的输出
Figure PCTCN2018077475-appb-000006
将激活函数f作用于
Figure PCTCN2018077475-appb-000007
由公式
Figure PCTCN2018077475-appb-000008
得到一个0-1区间的标量，此标量控制了神经元根据当前状态和过去状态的综合判断所接收过去信息的比例。
神经元接收当前时刻的样本X t、上一时刻的输出值b t-1 h以及上一时刻的状态数据S t-1 c、连接神经元与输入的训练集中的目标医疗数据的权值w ic、连接神经元与上一时刻的输出值的权值w hc以及输入门、遗忘门的输出标量，根据式
Figure PCTCN2018077475-appb-000009
Figure PCTCN2018077475-appb-000010
计算当前时刻的神经元状态
Figure PCTCN2018077475-appb-000011
输出门接收当前时刻的样本以及当前时刻的状态数据X t，上一时刻的输出值b t-1 h以及当前时刻的状态数据
Figure PCTCN2018077475-appb-000012
通过连接输入的训练集中的目标医疗数据与输出门的权值w iw、 连接上一时刻的输出值与输出门的权值w hw以及连接神经元与输出门的权值w cw，根据公式
Figure PCTCN2018077475-appb-000013
计算输出门的输出
Figure PCTCN2018077475-appb-000014
将激活函数f作用于
Figure PCTCN2018077475-appb-000015
上由公式
Figure PCTCN2018077475-appb-000016
得到一个0-1区间的标量。
神经元输出
Figure PCTCN2018077475-appb-000017
根据输出门输出的标量计算。具体地，神经元输出的输出根据公式
Figure PCTCN2018077475-appb-000018
计算得出。由上述对训练集中的目标医疗数据在各层间的计算可获取长短时记忆网络模型各层的输出值。
S133：根据输出值对长短时记忆网络模型各层进行误差反传更新，获取更新后的各层的权值，其中，误差反传更新的表达式为
Figure PCTCN2018077475-appb-000019
其中，T表示时刻，W表示权值，B表示输出值，δ表示误差项，
Figure PCTCN2018077475-appb-000020
为上一时刻神经元的状态数据，b t-1 h为上一时刻的输出值。
本实施例中，根据获取长短时记忆网络模型各层的输出值对长短时记忆网络模型各层进行误差反传更新。具体地，首先根据误差项的表达式
Figure PCTCN2018077475-appb-000021
可求出各层的误差项。其中，ε和δ均表示误差项，特别地，
Figure PCTCN2018077475-appb-000022
表示神经元输出反传的误差项，
Figure PCTCN2018077475-appb-000023
表示神经元状态反传的误差项，两者均表示误差项，但具体含义不同。在以下表达式中，神经元的输入用a表示，输出用b表示。下标ι,φ和ω分别表示输入门、遗忘门和输出门。下标c表示神经元，t代表时刻。神经元跟输入门、遗忘门和输出门连接的权值分别记做w cl、
Figure PCTCN2018077475-appb-000024
和w cω。S c表示神经元状态，控制门的激活函数用f(sigmoid)表示，g(tanh)和h(softmax)分别表示神经元的输入激活函数和输出激活函数。K是输出层神经元的个数，H是隐层神经元的个数，C是神经元状态的神经元个数，这里取C＝H。则输入门反传的误差项为
Figure PCTCN2018077475-appb-000025
遗忘门反传的误差项为
Figure PCTCN2018077475-appb-000026
神经元状态反传的误差项为
Figure PCTCN2018077475-appb-000027
其中，
Figure PCTCN2018077475-appb-000028
输出门反传的误差项为
Figure PCTCN2018077475-appb-000029
神经元输出反传的误差项为
Figure PCTCN2018077475-appb-000030
根据获得的各层误差项，再进行权值梯度的计算即可更新各层的权值，其中，权值更新的表达式为
Figure PCTCN2018077475-appb-000031
式中T表示时刻，W表示权值，如w cl、w cφ和w cω等连接权值。B表示输出值，如
Figure PCTCN2018077475-appb-000032
等输出。δ表示误差项，如
Figure PCTCN2018077475-appb-000033
等误差项。
Figure PCTCN2018077475-appb-000034
为上一时刻神经元的状态数据，b t-1 h为上一时刻的输出值。上述表达式各参数需相对应，如更新的具体权值为w cl时，则输出B为相对应的
Figure PCTCN2018077475-appb-000035
误差项δ为相对应的
Figure PCTCN2018077475-appb-000036
根据步骤S132和步骤S133的表达式可获得该权值更新表达式的所需参数值。则根据该权值更新的表达式进行运算即可获取更新后各层的权值。
S134：基于更新后的各层的权值，获取原始医疗模型。
本实施例中，将获取的更新后的各层的权值，应用到长短时记忆网络模型中即可获取原始医疗模型。进一步地，该原始医疗模型中各层之间的权值实现了原始医疗模型决定丢弃哪些旧信息、增加哪些新信息以及输出哪些信息的功能。在原始医疗模型的输出层最终会输出概率值，该概率值表示信息在通过原始医疗模型处理后与该原始医疗模型的贴近程度，可广泛应用于医疗特征识别，以达到准确识别医疗特征的目的。
在一具体实施方式中，如图5所示，步骤S14中，即采用测试集中的目标医疗数据对原始医疗模型进行测试，获取目标医疗模型，具体包括如下步骤：
S141：将每一目标医疗数据中的至少一个症状特征输入到原始医疗模型进行测试，获取对应的预测医疗特征。
其中，每一目标医疗数据包含至少一个症状特征和对应的实际疾病特征，如流鼻涕和/或发烧(症状特征)对应感冒(实际疾病特征)。具体地，将测试集中的每一目标医疗数据输入到原始医疗模型进行测试，即执行步骤S31-S34，以获取对应的预测医疗特征。
S142：获取同一目标医疗数据中实际疾病特征与预测医疗特征相匹配的识别数量。
具体地，将步骤S41获取到的预测疾病特征与同一目标医疗数据中的实际疾病特征相匹配，若匹配成功，则将识别结果准确的数量加1，若匹配不成功，则识别结果准确的数量不变，最后统计识别结果准确的数量即为识别数量。
S143：基于识别数量和测试集中的目标医疗数据的数据总量，获取识别准确率。
具体地，将步骤S42获取到的识别数量即识别结果准确的数量除以测试集中所有目标 医疗数据的数据总量，以获取识别准确率。可以理解地，识别准确率可以采用公式
Figure PCTCN2018077475-appb-000037
进行计算获取，其中，T为识别结果准确的数量，N为目标医疗数据的数据总量。
S144：若识别准确率大于预设准确率，则获取目标医疗模型。
其中，预设准确率是预先设置的用于评价模型好坏的概率值。具体地，若识别准确率大于预设准确率，则认定该原始医疗模型较准确，以将该原始医疗模型作为目标医疗模型；反之，若识别准确率不大于预设准确率，则认定该原始医疗模型不够准确，需采用更多的目标医疗数据重新训练原始医疗模型，在训练获取得的原始医疗模型的识别准确率大于预设准确率时，认为模型训练完成，以获取目标医疗模型。
本实施例中，将每一目标医疗数据中的至少一个症状特征输入到原始医疗模型进行测试，获取对应的预测医疗特征，然后基于同一目标医疗数据中实际疾病特征与预测医疗特征相匹配的识别数量和测试集中的目标医疗数据的数据总量，以获取识别准确率。最后基于识别准确率。获取目标医疗模型，以提高医疗特征识别的准确率。
本实施例中，先获取目标医疗数据，并对目标医疗数据进行划分，获取训练集和测试集，以便基于训练集中的目标医疗数据对长短时记忆网络模型进行训练，能够有效更新长短时记忆网络模型中各层的权值，使得通过目标医疗数据训练得到的原始医疗特征识别效果更精准。最后，再基于测试集中的目标医疗数据对原始医疗模型进行测试，以保证获取到的目标医疗特征识别目标医疗特征的准确率更高。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
实施例2
图6示出与实施例1中医疗模型训练方法一一对应的医疗模型训练装置的原理框图。如图6所示，该医疗模型训练装置包括目标医疗数据获取模块11、目标医疗数据划分模块12、原始医疗模型获取模块13和目标医疗模型获取模块14。其中，目标医疗数据获取模块11、目标医疗数据划分模块12、原始医疗模型获取模块13和目标医疗模型获取模块14的实现功能与实施例中医疗模型训练方法对应的步骤一一对应，为避免赘述，本实施例不一一详述。
目标医疗数据获取模块11，用于获取目标医疗数据.
目标医疗数据划分模块12，用于将目标医疗数据按预设比例进行划分，获取训练集和测试集。
原始医疗模型获取模块13，用于采用训练集中的目标医疗数据对长短时记忆网络模 型进行训练，获取原始医疗模型。
目标医疗模型获取模块14，用于采用测试集中的目标医疗数据对原始医疗模型进行测试，获取目标医疗模型。
优选地，目标医疗数据获取模块11包括目标网页地址获取单元111、原始医疗数据获取单元112和目标医疗数据获取单元113。
目标网页地址获取单元111，用于获取目标网页地址。
原始医疗数据获取单元112，用于采用爬虫工具爬取目标网页地址对应的网页，获取原始医疗数据。
目标医疗数据获取单元113，用于对原始医疗数据进行数据清洗，获取目标医疗数据。
优选地，原始医疗数据获取单元112包括访问地址获取子单元1121、访问地址存储子单元1122和原始医疗数据获取子单元1123。
访问地址获取子单元1121，用于采用爬虫工具，依据深度优先算法或广度优先算法爬取目标网页地址所链接的至少一个访问地址，每一访问地址对应一网页。
访问地址存储子单元1122，将至少一个访问地址存储在待下载消息队列中。
原始医疗数据获取子单元1123，采用爬虫工具对待下载消息队列中的每一访问地址对应的网页进行数据提取，获取原始医疗数据。
优选地，原始医疗模型获取模块13包括网络模型初始化单元131、输出值计算单元132、权值更新单元133和原始医疗模型获取单元134。
网络模型初始化单元131，用于初始化长短时记忆网络模型。
输出值计算单元132，用于在长短时记忆网络模型中输入正负样本，计算长短时记忆网络模型各层的输出值。
权值更新单元133，用于根据输出值对长短时记忆网络模型各层进行误差反传更新，获取更新后的各层的权值。
原始医疗模型获取单元134，用于基于更新后的各层的权值，获取原始医疗模型。
优选地，目标医疗模型获取模块14包括预测医疗特征获取单元141、识别数量获取单元142、识别准确率获取单元143和目标医疗模型获取单元144。
预测医疗特征获取单元141，用于将每一目标医疗数据中的至少一个症状特征输入到原始医疗模型进行测试，获取对应的预测医疗特征。
识别数量获取单元142，用于获取同一目标医疗数据中实际疾病特征与预测医疗特征相匹配的识别数量。
识别准确率获取单元143，用于基于识别数量和测试集中的目标医疗数据的数据总量，获取识别准确率。
目标医疗模型获取单元144，用于识别准确率大于预设准确率时，则获取目标医疗模型。
实施例3
图7示出本实施例中医疗识别方法的一流程图。该医疗识别方法可应用在医疗机构或者其他机构的终端设备上，以便对症状特征进行识别，达到智能识别疾病的效果。如图7所示，该医疗识别方法包括如下步骤：
S21：获取用户的待测医疗数据，待测医疗数据包括至少一个症状特征。
其中，待测医疗数据是指用户选择的用于识别目标医疗特征概率的医疗数据。本实施例中，待测医疗数据包括至少一个症状特征，如流鼻涕、体温和血压等症状特征。
S22：将至少一个症状特征输入到目标医疗模型进行识别，获取目标医疗特征，目标医疗模型是采用实施例1中医疗模型训练方法获取的模型。
其中，目标医疗特征是采用目标医疗模型对症状特征进行识别所得到的最大概率的患病特征。具体地，将至少一个症状特征输入到目标医疗模型进行识别，即执行实施例1中步骤S131-S134，以获取目标医疗特征。
本实施例中，先获取用户的待测医疗数据，待测医疗数据包括至少一个症状特征，以便至少一个症状特征输入到目标医疗模型进行识别，获取目标医疗特征，有利保障获取识别概率值的准确性，以辅助用户判断患病风险。
在一具体实施方式中，如图8所示，步骤S22，即将至少一个症状特征输入到目标医疗模型进行识别，获取目标医疗特征，具体包括如下步骤：
S221：基于目标医疗模型对至少一个症状特征进行识别，获取至少一个识别医疗特征和对应的识别概率值。
具体地，将至少一个症状特征输入到目标医疗模型中进行识别，在目标医疗模型中对输入的至少一个症状特征进行基于各层间权值的转换处理，在输出层输出至少一个识别医疗特征和对应识别概率值。该识别医疗特征是采用目标医疗模型对至少一个症状特征进行识别时，识别出其可能患有的疾病特征。识别概率值是采用目标医疗模型对至少一个症状进行识别时，识别其可能为识别医疗特征的可能性。其中，该识别概率值可以为0-1之间的实数。
S222：选取最大的识别概率值对应的识别医疗特征作为目标医疗特征。
具体地，选取最大的识别概率值对应的识别医疗特征作为目标医疗特征。其中，数据库还会关联存储有目标医疗特征所对应的科室和至少一个药品名称。本实施例中，当用户输入至少一个症状特征进行识别时，除了输出目标医疗特征，还会关联输出对应的科室和至少一个药品名称，以增强医疗识别模型的实用性。
本实施例中，基于目标医疗模型对至少一个症状特征进行识别，获取至少一个识别医疗特征和对应的识别概率值，以便基于识别概率值确定目标医疗特征，即确定疑似患病的疾病特征，以保证医疗特征识别结果更精确可靠。
本实施例中，通过获取用户的待测医疗数据，以便基于目标医疗模型对待测医疗数据进行识别，以获取目标医疗特征。通过目标医疗模型对待测医疗数据进行识别，以保证识别的准确率更高，较精准的识别目标医疗特征，以辅助用户判断患病风险，以便及时预防。
实施例4
图9示出与实施例1中医疗识别方法一一对应的医疗识别装置的原理框图。如图9所示，该医疗识别装置包括待测医疗数据获取模块21和目标医疗特征获取模块22。其中，待测医疗数据获取模块21和目标医疗特征获取模块22的实现功能与实施例中医疗识别方法对应的步骤一一对应，为避免赘述，本实施例不一一详述。
待测医疗数据获取模块21，用于获取用户的待测医疗数据，待测医疗数据包括至少一个症状特征。
目标医疗特征获取模块22，用于将至少一个症状特征输入到目标医疗模型进行识别，获取目标医疗特征。
优选地，目标医疗特征获取模块22包括识别概率值获取单元221和目标医疗特征获取单元222。
识别概率值获取单元221，用于基于目标医疗模型对至少一个症状特征进行识别，获取至少一个识别医疗特征和对应的识别概率值。
目标医疗特征获取单元222，用于选取最大的识别概率值对应的识别医疗特征作为目标医疗特征。
实施例5
本实施例提供一计算机可读存储介质，该计算机可读存储介质上存储有计算机可读指令，该计算机可读指令被处理器执行时实现实施例1中医疗模型训练方法，为避免重复，这里不再赘述。或者，该计算机可读指令被处理器执行时实现实施例2中医疗模型训练装置中各模块/单元的功能，为避免重复，这里不再赘述。或者，该计算机可读指令被处理 器执行时实现实施例3中医疗识别方法中各步骤的功能，为避免重复，此处不一一赘述。或者，该计算机可读指令被处理器执行时实现实施例4中医疗识别装置中各模块/单元的功能，为避免重复，此处不一一赘述。
该计算机可读存储介质可以包括：能够携带所述计算机可读指令代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、电载波信号、电信信号以及软件分发介质等。
实施例6
图10是本申请一实施例提供的终端设备的示意图。如图10所示，该实施例的终端设备100包括：处理器101、存储器102以及存储在存储器102中并可在处理器101上运行的计算机可读指令103。处理器101执行计算机可读指令103时实现实施例1中的医疗模型训练方法各步骤的功能，例如图1所示的步骤S11至S14。或者，处理器101执行计算机可读指令103时实现实施例2中的医疗模型训练装置各模块/单元的功能，例如图6所示模块11至14的功能。或者，处理器101执行计算机可读指令103时实现实施例3中医疗识别方法各步骤的功能，为避免重复，此处不一一赘述。或者，该计算机可读指令被处理器101执行计算机可读指令103时实现实施例4中医疗识别装置中各模块/单元的功能。为避免重复，此处不一一赘述。
示例性的，计算机可读指令103可以被分割成一个或多个模块/单元，一个或者多个模块/单元被存储在存储器102中，并由处理器101执行，以完成本申请。一个或多个模块/单元可以是能够完成特定功能的一系列计算机可读指令103的指令段，该指令段用于描述计算机可读指令103在终端设备100中的执行过程。例如，计算机可读指令103可以被分割成实施例2中的目标医疗数据获取模块11、目标医疗数据划分模块12、原始医疗模型获取模块13和目标医疗模型获取模块14，或者实施例4中的待测医疗数据获取模块21和目标医疗特征获取模块22，各模块的具体功能如实施例2或实施例4所述，在此不一一赘述。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施 例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种医疗模型训练方法，其特征在于，包括：
    获取目标医疗数据；
    将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
    采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
    采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
  2. 如权利要求1所述的医疗模型训练方法，其特征在于，所述获取目标医疗数据包括：
    获取目标网页地址；
    采用爬虫工具爬取所述目标网页地址对应的网页，获取原始医疗数据；
    对所述原始医疗数据进行数据清洗，获取目标医疗数据。
  3. 如权利要求1所述的医疗模型训练方法，其特征在于，所述采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型，包括：
    初始化长短时记忆网络模型；
    在所述长短时记忆网络模型中输入所述训练集中的目标医疗数据，计算所述长短时记忆网络模型各层的输出值；
    根据所述输出值对所述长短时记忆网络模型各层进行误差反传更新，获取更新后的所述各层的权值，其中，所述误差反传更新的表达式为
    Figure PCTCN2018077475-appb-100001
    其中，T表示时刻，W表示权值，B表示输出值，δ表示误差项，
    Figure PCTCN2018077475-appb-100002
    为上一时刻神经元的状态数据，b t-1 h为上一时刻的输出值；
    基于更新后的所述各层的权值，获取原始医疗模型。
  4. 如权利要求1所述的医疗模型训练方法，其特征在于，所述目标医疗数据包括一个实际疾病特征和对应的至少一个症状特征；
    所述采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型，包括：
    将每一所述目标医疗数据中的至少一个所述症状特征输入到所述原始医疗模型进行测试，获取对应的预测医疗特征；
    获取同一所述目标医疗数据中所述实际疾病特征与所述预测医疗特征相匹配的识别数量；
    基于所述识别数量和所述测试集中所述的目标医疗数据的数据总量，获取识别准确率；
    若所述识别准确率大于预设准确率，则获取所述目标医疗模型。
  5. 一种医疗识别方法，其特征在于，包括：
    获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
    将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用权利要求1-4任一项所述医疗模型训练方法获取的模型。
  6. 如权利要求5所述的医疗识别方法，其特征在于，将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，包括：
    基于所述目标医疗模型对至少一个所述症状特征进行识别，获取至少一个识别医疗特征和对应的识别概率值；
    选取最大的所述识别概率值对应的识别医疗特征作为所述目标医疗特征。
  7. 一种医疗模型训练装置，其特征在于，包括：
    目标医疗数据获取模块，用于获取目标医疗数据；
    目标医疗数据划分模块，用于将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
    原始医疗模型获取模块，用于采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
    目标医疗模型获取模块，用于采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
  8. 一种医疗识别装置，其特征在于，包括：
    待测医疗数据获取模块，用于获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
    目标医疗特征获取模块，用于将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用权利要求1-4任一项所述医疗模型训练方法获取的模型。
  9. 一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现如下 步骤：
    获取目标医疗数据；
    将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
    采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
    采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
  10. 如权利要求9所述的终端设备，其特征在于，所述获取目标医疗数据包括：
    获取目标网页地址；
    采用爬虫工具爬取所述目标网页地址对应的网页，获取原始医疗数据；
    对所述原始医疗数据进行数据清洗，获取目标医疗数据。
  11. 如权利要求9所述的终端设备，其特征在于，所述采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型，包括：
    初始化长短时记忆网络模型；
    在所述长短时记忆网络模型中输入所述训练集中的目标医疗数据，计算所述长短时记忆网络模型各层的输出值；
    根据所述输出值对所述长短时记忆网络模型各层进行误差反传更新，获取更新后的所述各层的权值，其中，所述误差反传更新的表达式为
    Figure PCTCN2018077475-appb-100003
    其中，T表示时刻，W表示权值，B表示输出值，δ表示误差项，
    Figure PCTCN2018077475-appb-100004
    为上一时刻神经元的状态数据，b t-1 h为上一时刻的输出值；
    基于更新后的所述各层的权值，获取原始医疗模型。
  12. 如权利要求9所述的终端设备，其特征在于，所述目标医疗数据包括一个实际疾病特征和对应的至少一个症状特征；
    所述采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型，包括：
    将每一所述目标医疗数据中的至少一个所述症状特征输入到所述原始医疗模型进行测试，获取对应的预测医疗特征；
    获取同一所述目标医疗数据中所述实际疾病特征与所述预测医疗特征相匹配的识别数量；
    基于所述识别数量和所述测试集中所述的目标医疗数据的数据总量，获取识别准确率；
    若所述识别准确率大于预设准确率，则获取所述目标医疗模型。
  13. 一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现如下步骤：
    获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
    将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用权利要求1-4任一项所述医疗模型训练方法获取的模型。
  14. 如权利要求13所述的终端设备，其特征在于，将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，包括：
    基于所述目标医疗模型对至少一个所述症状特征进行识别，获取至少一个识别医疗特征和对应的识别概率值；
    选取最大的所述识别概率值对应的识别医疗特征作为所述目标医疗特征。
  15. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现如下步骤：
    获取目标医疗数据；
    将所述目标医疗数据按预设比例进行划分，获取训练集和测试集；
    采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型；
    采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型。
  16. 如权利要求15所述的计算机可读存储介质，其特征在于，所述获取目标医疗数据包括：
    获取目标网页地址；
    采用爬虫工具爬取所述目标网页地址对应的网页，获取原始医疗数据；
    对所述原始医疗数据进行数据清洗，获取目标医疗数据。
  17. 如权利要求15所述的计算机可读存储介质，其特征在于，所述采用所述训练集中的目标医疗数据对长短时记忆网络模型进行训练，获取原始医疗模型，包括：
    初始化长短时记忆网络模型；
    在所述长短时记忆网络模型中输入所述训练集中的目标医疗数据，计算所述长短时记忆网络模型各层的输出值；
    根据所述输出值对所述长短时记忆网络模型各层进行误差反传更新，获取更新后的所述各层的权值，其中，所述误差反传更新的表达式为
    Figure PCTCN2018077475-appb-100005
    其中，T表示时刻，W表示权值，B表示输出值，δ表示误差项，
    Figure PCTCN2018077475-appb-100006
    为上一时刻神经元的状态数据，b t-1 h为上一时刻的输出值；
    基于更新后的所述各层的权值，获取原始医疗模型。
  18. 如权利要求15所述的计算机可读存储介质，其特征在于，所述目标医疗数据包括一个实际疾病特征和对应的至少一个症状特征；
    所述采用所述测试集中的目标医疗数据对所述原始医疗模型进行测试，获取目标医疗模型，包括：
    将每一所述目标医疗数据中的至少一个所述症状特征输入到所述原始医疗模型进行测试，获取对应的预测医疗特征；
    获取同一所述目标医疗数据中所述实际疾病特征与所述预测医疗特征相匹配的识别数量；
    基于所述识别数量和所述测试集中所述的目标医疗数据的数据总量，获取识别准确率；
    若所述识别准确率大于预设准确率，则获取所述目标医疗模型。
  19. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现如下步骤：
    获取用户的待测医疗数据，所述待测医疗数据包括至少一个症状特征；
    将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，所述目标医疗模型是采用权利要求1-4任一项所述医疗模型训练方法获取的模型。
  20. 如权利要求19所述的计算机可读存储介质，其特征在于，将至少一个所述症状特征输入到所述目标医疗模型进行识别，获取目标医疗特征，包括：
    基于所述目标医疗模型对至少一个所述症状特征进行识别，获取至少一个识别医疗特征和对应的识别概率值；
    选取最大的所述识别概率值对应的识别医疗特征作为所述目标医疗特征。
PCT/CN2018/077475 2018-01-12 2018-02-28 医疗模型训练方法、医疗识别方法、装置、设备及介质 Ceased WO2019136806A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810030552.5A CN108461152A (zh) 2018-01-12 2018-01-12 医疗模型训练方法、医疗识别方法、装置、设备及介质
CN201810030552.5 2018-01-12

Publications (1)

Publication Number Publication Date
WO2019136806A1 true WO2019136806A1 (zh) 2019-07-18

Family

ID=63220699

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/077475 Ceased WO2019136806A1 (zh) 2018-01-12 2018-02-28 医疗模型训练方法、医疗识别方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN108461152A (zh)
WO (1) WO2019136806A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111198893A (zh) * 2019-12-31 2020-05-26 南京医睿科技有限公司 一种数据更新方法、装置、可读介质及电子设备
CN111477321A (zh) * 2020-03-11 2020-07-31 北京大学第三医院（北京大学第三临床医学院） 具有自学习能力的治疗效果预测系统及治疗效果预测终端
CN112016702A (zh) * 2020-09-09 2020-12-01 平安科技（深圳）有限公司 基于迁移学习的医疗数据处理方法、装置、设备及介质
CN113626688A (zh) * 2021-07-21 2021-11-09 上海齐网网络科技有限公司 基于软件定义的智能化医疗数据采集方法及系统
CN114360732A (zh) * 2022-01-12 2022-04-15 平安科技（深圳）有限公司 医疗数据分析方法、装置、电子设备及存储介质
CN115331811A (zh) * 2022-09-19 2022-11-11 联仁健康医疗大数据科技股份有限公司 一种疾病类型的预测方法、装置、电子设备及存储介质

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109389972B (zh) * 2018-09-21 2020-11-03 四川长虹电器股份有限公司 语义云功能的质量测试方法、装置、存储介质和设备
CN109448858A (zh) * 2018-10-30 2019-03-08 中山大学肿瘤防治中心 结直肠癌医疗数据处理方法及装置、存储介质、电子设备
CN109378065A (zh) * 2018-10-30 2019-02-22 医渡云（北京）技术有限公司 医疗数据处理方法及装置、存储介质、电子设备
CN109615012A (zh) * 2018-12-13 2019-04-12 平安医疗健康管理股份有限公司 基于机器学习的就诊数据异常识别方法、设备及存储介质
CN109754852A (zh) * 2019-01-08 2019-05-14 中南大学 基于电子病历的心血管疾病风险预测方法
CN111859078B (zh) * 2019-04-29 2024-04-12 百度在线网络技术（北京）有限公司 一种信息识别方法、装置及电子设备
CN110399927B (zh) * 2019-07-26 2022-02-01 玖壹叁陆零医学科技南京有限公司 识别模型训练方法、目标识别方法及装置
CN110619423B (zh) * 2019-08-06 2023-04-07 平安科技（深圳）有限公司 多任务预测方法、装置、电子设备及存储介质
CN110755047B (zh) * 2019-10-30 2022-05-03 中国医学科学院生物医学工程研究所 一种心脑血管疾病预防方法、装置、终端及存储介质
CN113707328A (zh) * 2020-05-20 2021-11-26 阿里巴巴集团控股有限公司 数据处理方法、装置及计算设备
CN111816300A (zh) * 2020-06-30 2020-10-23 平安科技（深圳）有限公司 基于深度强化学习的数据分类方法、装置、设备及介质
CN112133423A (zh) * 2020-08-28 2020-12-25 北京世纪互联宽带数据中心有限公司 基于边缘计算的医疗数据处理方法、装置及网络设备
CN114840660B (zh) * 2022-05-25 2025-08-29 中国平安财产保险股份有限公司 业务推荐模型训练方法、装置、设备及存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040181512A1 (en) * 2003-03-11 2004-09-16 Lockheed Martin Corporation System for dynamically building extended dictionaries for a data cleansing application
WO2004088464A2 (en) * 2003-03-28 2004-10-14 Dun & Bradstreet, Inc. System and method for data cleansing
CN105786830A (zh) * 2014-12-19 2016-07-20 阿里巴巴集团控股有限公司 计算机系统中模型自适应调整方法、装置及系统
CN106250707A (zh) * 2016-08-12 2016-12-21 王双坤 一种基于深度学习算法处理头部结构像数据的方法
CN106407387A (zh) * 2016-09-18 2017-02-15 华南师范大学 一种针对医疗诊断文本的概念连接方法
CN106980811A (zh) * 2016-10-21 2017-07-25 商汤集团有限公司 人脸表情识别方法和人脸表情识别装置
CN106980899A (zh) * 2017-04-01 2017-07-25 北京昆仑医云科技有限公司 预测血管树血管路径上的血流特征的深度学习模型和系统

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10650305B2 (en) * 2016-07-08 2020-05-12 Baidu Usa Llc Systems and methods for relation inference
CN106934799B (zh) * 2017-02-24 2019-09-03 安翰科技(武汉)股份有限公司 胶囊内窥镜图像辅助阅片系统及方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040181512A1 (en) * 2003-03-11 2004-09-16 Lockheed Martin Corporation System for dynamically building extended dictionaries for a data cleansing application
WO2004088464A2 (en) * 2003-03-28 2004-10-14 Dun & Bradstreet, Inc. System and method for data cleansing
CN105786830A (zh) * 2014-12-19 2016-07-20 阿里巴巴集团控股有限公司 计算机系统中模型自适应调整方法、装置及系统
CN106250707A (zh) * 2016-08-12 2016-12-21 王双坤 一种基于深度学习算法处理头部结构像数据的方法
CN106407387A (zh) * 2016-09-18 2017-02-15 华南师范大学 一种针对医疗诊断文本的概念连接方法
CN106980811A (zh) * 2016-10-21 2017-07-25 商汤集团有限公司 人脸表情识别方法和人脸表情识别装置
CN106980899A (zh) * 2017-04-01 2017-07-25 北京昆仑医云科技有限公司 预测血管树血管路径上的血流特征的深度学习模型和系统

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111198893A (zh) * 2019-12-31 2020-05-26 南京医睿科技有限公司 一种数据更新方法、装置、可读介质及电子设备
CN111198893B (zh) * 2019-12-31 2023-05-02 医渡云（北京）技术有限公司 一种数据更新方法、装置、可读介质及电子设备
CN111477321A (zh) * 2020-03-11 2020-07-31 北京大学第三医院（北京大学第三临床医学院） 具有自学习能力的治疗效果预测系统及治疗效果预测终端
CN111477321B (zh) * 2020-03-11 2023-06-09 北京大学第三医院（北京大学第三临床医学院） 具有自学习能力的治疗效果预测系统及治疗效果预测终端
CN112016702A (zh) * 2020-09-09 2020-12-01 平安科技（深圳）有限公司 基于迁移学习的医疗数据处理方法、装置、设备及介质
CN112016702B (zh) * 2020-09-09 2023-07-28 平安科技（深圳）有限公司 基于迁移学习的医疗数据处理方法、装置、设备及介质
CN113626688A (zh) * 2021-07-21 2021-11-09 上海齐网网络科技有限公司 基于软件定义的智能化医疗数据采集方法及系统
CN113626688B (zh) * 2021-07-21 2023-09-01 上海齐网网络科技有限公司 基于软件定义的智能化医疗数据采集方法及系统
CN114360732A (zh) * 2022-01-12 2022-04-15 平安科技（深圳）有限公司 医疗数据分析方法、装置、电子设备及存储介质
CN114360732B (zh) * 2022-01-12 2024-04-09 平安科技（深圳）有限公司 医疗数据分析方法、装置、电子设备及存储介质
CN115331811A (zh) * 2022-09-19 2022-11-11 联仁健康医疗大数据科技股份有限公司 一种疾病类型的预测方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN108461152A (zh) 2018-08-28

Similar Documents

Publication Publication Date Title
WO2019136806A1 (zh) 医疗模型训练方法、医疗识别方法、装置、设备及介质
Li et al. CANet: cross-disease attention network for joint diabetic retinopathy and diabetic macular edema grading
CN109659033B (zh) 一种基于循环神经网络的慢性疾病病情变化事件预测装置
CN106874643B (zh) 基于词向量自动构建知识库实现辅助诊疗的方法和系统
Pise et al. Skin disease detection using machine learning
CN110675944A (zh) 分诊方法及装置、计算机设备及介质
CN111696661B (zh) 患者分群模型构建方法、患者分群方法及相关设备
WO2020224433A1 (zh) 基于机器学习的目标对象属性预测方法及相关设备
CN111863247B (zh) 基于结构磁共振影像的大脑年龄级联精炼预测方法及系统
WO2025112186A1 (zh) 状态预测方法、装置、计算机设备和存储介质
CN114446431A (zh) 专业数据的标注人员遴选方法、装置和电子设备
Kokala et al. Bone cancer prediction using regression algorithms and machine learning approaches
CN116350234A (zh) 基于gcnn-lstm模型的ecg心律失常分类方法及系统
CN116746947A (zh) 基于在线测试时间域适应的跨受试者脑电信号分类方法
CN117875319B (zh) 医疗领域标注数据的获取方法、装置、电子设备
CN117010971B (zh) 一种基于人像识别的智能健康险提供方法及系统
CN114238558B (zh) 一种电子病历的质检方法、装置、存储介质及设备
CN115409051A (zh) 压力识别方法、装置、设备和存储介质
CN119443178B (zh) 医学诊断模型训练方法、医学诊断方法和相关装置
CN120319506A (zh) 智能问诊对话方法、装置、计算机设备及存储介质
CN119811656A (zh) 多维医疗大数据融合与疾病预测识别引擎的构建和应用
CN115831356B (zh) 一种基于人工智能算法的辅助预测诊断方法
US20250087355A1 (en) A machine learning based framework using electroretinography for detecting early stage glaucoma
CN118176510A (zh) 医学图像分析的语境化
Bhalerao et al. Ecg classification using machine learning on wave samples for the indian population

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: 18899926

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205N DATED 02/09/2020)

122 Ep: pct application non-entry in european phase

Ref document number: 18899926

Country of ref document: EP

Kind code of ref document: A1