WO2020224433A1 - 基于机器学习的目标对象属性预测方法及相关设备 - Google Patents

基于机器学习的目标对象属性预测方法及相关设备 Download PDF

Info

Publication number
WO2020224433A1
WO2020224433A1 PCT/CN2020/086007 CN2020086007W WO2020224433A1 WO 2020224433 A1 WO2020224433 A1 WO 2020224433A1 CN 2020086007 W CN2020086007 W CN 2020086007W WO 2020224433 A1 WO2020224433 A1 WO 2020224433A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
target object
detection
neural network
regularity
Prior art date
Application number
PCT/CN2020/086007
Other languages
English (en)
French (fr)
Inventor
乔治
葛屾
晏阳天
王锴
吴贤
范伟
Original Assignee
腾讯科技（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 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Priority to KR1020217025407A priority Critical patent/KR20210113336A/ko
Priority to JP2021562367A priority patent/JP7191443B2/ja
Priority to EP20801439.9A priority patent/EP3968337A4/en
Publication of WO2020224433A1 publication Critical patent/WO2020224433A1/zh
Priority to US17/469,270 priority patent/US20210406687A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/40—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for data related to laboratory analysis, e.g. patient specimen analysis
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records

Definitions

  • This application relates to the field of data prediction technology, in particular to target object attribute prediction based on machine learning.
  • EHR (electronic health records) data can record every consultation record of the target object.
  • more and more clinical diagnosis estimation models can simulate the diagnosis process of doctors based on the EHR data of patients. To predict the future incidence of users.
  • the process of predicting the future morbidity of a user can be: taking the medical coding data in the EHR data as the attributes of the patient and inputting it into the clinical diagnosis estimation model.
  • the clinical diagnosis estimation model trains the medical coding data. Output the predicted diagnosis result.
  • the clinical diagnosis estimation model carries out the process of training the medical coding data, which can characterize the clinical diagnosis estimation model to simulate the diagnosis process of the doctor, so that the subsequent diagnosis results predicted by the clinical diagnosis estimation model can be used to evaluate the patient’s future incidence. prediction.
  • the embodiments of the present application provide a method and related equipment for predicting the attributes of a target object based on machine learning, which can solve the problem of low accuracy of the diagnosis result predicted by the clinical diagnosis estimation model.
  • the technical solution is as follows:
  • a method for predicting the attributes of a target object based on machine learning is provided, which is executed by a computer device, and the method includes:
  • the first neural network outputs a first regular feature and a second regular feature different from the first regular feature after two different time sequence calculations, where all The first regular feature represents the historical change law of the detection feature, and the second regular feature represents the future change law of the detection feature;
  • the second neural network extracts and outputs at least one local feature of the target object from the global feature
  • a device for predicting attributes of a target object based on machine learning includes:
  • An acquisition module configured to determine the detection characteristics of the target object according to the detection data of the target object and the attributes corresponding to the detection data
  • the calculation module is used to input the detection features into the first neural network; for the detection features in each time sequence of the detection features, the first neural network outputs the first regular feature and different features after two different time sequence calculations In the second regularity feature of the first regularity feature, wherein the first regularity feature represents the historical change law of the detection feature, and the second regularity feature represents the future change law of the detection feature;
  • the acquisition module is further configured to determine the global characteristics of the target object based on the first regular feature and the second regular feature;
  • An extraction module configured to input the global feature into a second neural network; the second neural network extracts and outputs at least one local feature of the target object from the global feature;
  • the prediction module is configured to predict the attributes of the target object based on at least one local feature of the target object.
  • a computer device includes: a processor; a memory for storing a computer program; wherein the processor is used for executing the computer program stored on the memory to realize the above-mentioned target object based on machine learning The operation performed by the attribute prediction method.
  • a computer-readable storage medium stores a computer program, which when executed by a processor, realizes the operations performed by the above-mentioned method for predicting the attributes of a target object based on machine learning.
  • a computer program product including instructions, which when run on a computer, causes the computer to perform the operations performed by the above-mentioned method for predicting the attributes of a target object based on machine learning.
  • the global feature of the target object based on the regular features that represent the history of the detection feature and the changing law in the future, and refine the global feature to obtain at least one local feature of the target object. Then, the refined local feature is more
  • the characteristics of the target object can be reflected, and the attributes of the target object can be predicted based on local features. Therefore, the accuracy of the predicted attributes can be improved.
  • the attributes of the target object are the predicted diagnosis results, the accuracy of the predicted diagnosis results can be improved.
  • FIG. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application.
  • FIG. 2 is a flowchart of a method for predicting attributes of a target object based on machine learning provided by an embodiment of the present application
  • FIG. 3 is a schematic diagram of a diagnosis estimation model provided by an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of a device for predicting attributes of a target object based on machine learning according to an embodiment of the present application
  • Fig. 5 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • the process of predicting the future morbidity of the user can be: taking the medical coding data in the EHR data as the attributes of the patient and inputting it into the clinical diagnosis estimation model, and the clinical diagnosis estimation model trains the medical coding data, The predicted diagnosis result can be output.
  • the process of clinical diagnosis estimation model training on medical coding data can represent the clinical diagnosis estimation model simulating the doctor's diagnosis process, so that the subsequent diagnosis results predicted by the trained clinical diagnosis estimation model can be used to determine the patient's future disease The situation is predicted.
  • the input of the clinical diagnosis estimation model is the medical code data. Since the medical code data includes data for thousands of diseases, it is very difficult for one patient. In other words, it is only possible to suffer from one or several diseases, and it is not possible to suffer from various diseases. Therefore, the useful data in the medical coding data is distributed relatively sparsely and discretely in the medical coding data. The coded data can only indicate the disease that the patient suffers, but not the overall physical state of the patient. Then, after the clinical diagnosis estimation model is trained with such medical coded data, the accuracy of the output predicted diagnosis result is low. The predicted diagnosis result determines the patient’s future morbidity is not accurate.
  • the present application provides a method for predicting the attributes of a target object based on machine learning.
  • the method specifically includes: determining the detection characteristics of the target object according to the detection data of the target object and the attributes corresponding to the detection data;
  • the detection features are input into the first neural network, and for the detection features in each time sequence in the detection features, the first neural network outputs the first regular feature and the second regular feature different from the first regular feature after two different sequence calculations.
  • the first law feature represents the historical change law of the detection feature
  • the second law feature represents the future change law of the detection feature
  • the global feature of the target object is determined; the global feature is input to the second neural network ,
  • the second neural network extracts and outputs at least one local feature of the target object from the global feature; and predicts the attribute of the target object based on the at least one local feature of the target object.
  • the global characteristics of the target object are determined based on regular features that represent the history of detection features and the law of future changes, and the global features are refined to Obtain at least one local feature of the target object, then the refined local features can better reflect the characteristics of the target object, and then predict the attributes of the target object based on the local features. Therefore, the accuracy of the predicted attributes can be improved. When it is a predicted diagnosis result, the accuracy of the predicted diagnosis result can be improved.
  • Machine Learning is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. Specializing in the study of how computers simulate or realize human learning behaviors in order to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve its own performance.
  • Machine learning is the core of artificial intelligence, the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence.
  • Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning techniques.
  • AI Artificial Intelligence
  • artificial intelligence is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.
  • artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a similar way to human intelligence.
  • Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.
  • Artificial intelligence technology is a comprehensive discipline, covering a wide range of fields, including both hardware-level technology and software-level technology.
  • Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, and mechatronics.
  • Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning.
  • the method for predicting the attributes of a target object based on machine learning relates to artificial intelligence, and particularly relates to machine learning in artificial intelligence.
  • the method for predicting the attributes of target objects based on machine learning can be applied to computer devices that can perform data processing, such as terminal devices, servers, etc.; wherein the terminal devices can specifically be smart phones, computers, and personal digital devices.
  • Assistant Personal Digital Assitant, PDA
  • the server can be an application server or a web server. In actual deployment, the server can be an independent server or a cluster server.
  • the terminal device can directly perform the attribute prediction of the target object according to the detection data of the target object input by the user and the attributes corresponding to the detection data. Forecast and display the forecast results for users to view.
  • the server first predicts the attributes of the target object according to the detection data of the target object uploaded by the terminal device and the attributes corresponding to the detection data. Obtain the prediction result; then send the prediction result to the terminal device so that the terminal device can display the received prediction result for the user to view.
  • the following is an example of applying the method for predicting the attributes of a target object based on machine learning provided by the embodiments of the present application to a terminal device.
  • the applicable application scenarios are exemplified.
  • the application scenario includes: a terminal device and a user; wherein the terminal device is used to execute the machine learning-based
  • the target object attribute prediction method is to predict the attribute of the target object to obtain the prediction result for the user to view.
  • the terminal device After the terminal device receives the attribute prediction instruction triggered by the user, the terminal device can determine the detection characteristics of the target object according to the detection data of the target object and the attributes corresponding to the detection data; input the detection characteristics into the first neural network, and target each of the detection characteristics. Time-series detection features, the first neural network outputs the first regular feature and the second regular feature different from the first regular feature after two different time-series calculations.
  • the first regular feature represents the historical change rule of the detection feature
  • the second The regular feature represents the future change law of the detection feature
  • the global feature of the target object is determined; the global feature is input to the second neural network, and the second neural network extracts and outputs the target object’s At least one local feature; based on at least one local feature of the target object, predict the attribute of the target object to obtain a prediction result, so that the terminal device can display the prediction result to the user.
  • the method for predicting attributes of a target object based on machine learning provided in the embodiments of the present application can also be applied to a server.
  • the application scenario includes: a server, a terminal device, and a user.
  • the terminal device After the terminal device receives the attribute prediction instruction triggered by the user, the terminal device generates an attribute prediction request according to the attribute prediction instruction, and sends the attribute prediction request to the server, so that the server can receive the attribute prediction request sent by the terminal device Then, according to the detection data of the target object and the corresponding attributes of the detection data, the detection feature of the target object is determined; the detection feature is input into the first neural network, and the first neural network adopts two different types of detection features for each time sequence in the detection feature. After calculating the time series, output the first law feature and the second law feature different from the first law feature.
  • the first law feature represents the historical change law of the detection feature
  • the second law feature represents the future change law of the detection feature
  • the feature and the second regular feature determine the global feature of the target object; the global feature is input to the second neural network, and the second neural network extracts and outputs at least one local feature of the target object from the global feature; based on the at least one local feature of the target object,
  • the prediction result is obtained by predicting the attributes of the target object, so that the server can feed back the obtained prediction result to the terminal device, so that the user can view the prediction result on the terminal device.
  • Fig. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application.
  • the environment includes a target object attribute prediction system 100 based on machine learning.
  • the target object attribute prediction system based on machine learning includes a preprocessing module 101, a detection feature extraction module 102, a regular feature extraction module 103, and a prediction module 104.
  • the preprocessing module 101 is used to process the detection data of the target object and the attributes corresponding to the detection data, and specifically convert the detection data of the user and the attributes corresponding to the detection data into data that can be calculated by the detection feature extraction module 102.
  • the detection feature extraction module 102 is configured to extract a mixed feature of a feature of the detection data and an attribute corresponding to the detection data, and the extracted mixed feature can be used as a target object detection feature. Specifically, the detection feature extraction module 102 may first extract the features of the attributes and the features of the detection data based on the data processed by the preprocessing module 101, then splice the features of the extracted attributes and the features of the detection data, and finally, the detection The feature extraction module 102 extracts detection features based on the splicing result.
  • the regular feature extraction module 103 is used to extract regular features and generate global features of the target object.
  • the regular features are used to represent the global change law of the detection feature.
  • the regular feature extraction module 103 may first extract based on the detection feature extraction module 102 extracting Then, the law feature extraction module 103 can obtain the global change law of the detection feature based on the historical change law and future change law of the detection feature, and finally according to the law feature representing the global change law , To determine the global characteristics of the target object.
  • the prediction module 104 is used to predict the attributes of the target object. Specifically, the prediction module 104 can refine the global features generated by the law feature extraction module 103 through a neural network to obtain the local features of the target object, and then the prediction module 104 uses The target local features are expressed in a concentrated expression of the acquired multiple local features. Finally, the prediction module 104 predicts the attributes of the target object based on the target local features.
  • each module in the target object attribute prediction system 100 based on machine learning can be implemented by one computer device or multiple computer devices.
  • the embodiments of the present application implement the functions of each module.
  • the number of computer equipment is not specifically limited.
  • Figure 1 introduces the respective functions of each module in the target object attribute prediction system based on machine learning.
  • Figure 2 which is an embodiment of the present application.
  • the computer device determines the detection feature of the target object according to the detection data of the target object and the attributes corresponding to the detection data.
  • the computer device can be any computer device.
  • the target object can be any user.
  • the detection data of the target object may include the detection data of each detection of the target object in the historical time period.
  • the detection data of each detection corresponds to a detection time. Therefore, the detection data of each detection may include the detection data related to the target object.
  • Various types of data taking the detection of the physical signs of the target object as an example, the data detected at one time, heartbeat data, blood pressure data, and other types of data, for any type of data, each detection may detect a lot of data. Multiple data detected at one time can form a timing sequence related to the detection time of this time. Therefore, a detection time may correspond to multiple timing sequences, and each type of timing sequence can be labeled to distinguish between a detection time. Many types of inspection data.
  • the embodiment of the present application does not specifically limit the detection time, and does not specifically limit the time interval between two detection times.
  • the embodiment of the present application does not limit the aforementioned historical time period, and the aforementioned historical time period may refer to any time period before the computer device uses the attribute prediction method to predict the attribute of the target object.
  • the detection data when the detection data is the physical sign data of the target object, the detection data can be the time series data stored in the EHR data.
  • the time series data includes the time of each consultation of the target object and the time of each consultation.
  • the consultation time is also the detection time.
  • the above attribute is used to indicate at least one state of the target object.
  • Each position of the attribute corresponds to a state.
  • the state identifier can be used to indicate whether the target object has a corresponding state.
  • the state identifier may include a first state identifier and a second state identifier. , Where the first state identifier is used to indicate that the target object has a corresponding state, and the second state identifier is used to indicate that the target object does not have a corresponding state, for example, when any position within the attribute has the first state identifier , It means that the target object has a state corresponding to the position. When any position in the attribute has a second state identifier, it means that the target object does not have a state corresponding to the position.
  • different character strings may be used to represent the first status identifier and the second status identifier, and the embodiment of the present application does not specifically limit the character string representing the first status identifier or the second status identifier.
  • the current detection data will be obtained. Based on the current detection data, the attributes of the target object can be determined, and the detection time corresponding to the determined target object’s attributes is the corresponding detection data of this time The inspection time. It can be seen that in the embodiment of the present application, one attribute corresponds to one detection time, and if the detection data of the target object includes at least one detection time, the detection data of the target object corresponds to at least one attribute.
  • the aforementioned at least one state may be the diseased state of the target object, or other states, and the embodiment of the present application does not specifically limit the at least one state.
  • the attribute corresponding to the detection data can be medical coding data.
  • a medical coding data can consist of 0 and 1. Each position in the medical coding data corresponds to a disease. When the data in a position is 0, it means that the target object does not have the disease corresponding to that position. When the data in any position is 1, it means that the target object has the disease corresponding to the position. It is understandable that here 0 is equivalent to the second state identifier, and 1 here is equivalent to the first state identifier.
  • the computer device can first preprocess the detection data of the target object and the attributes corresponding to the detection data, so that the detection data and the attributes corresponding to the detection data conform to the format required for subsequent calculations, and then perform the processing on the processed data Feature extraction to obtain the detection feature of the target object.
  • this step 201 can be implemented through the process shown in the following steps 2011-2014.
  • Step 2011 The computer device inputs the attribute corresponding to the detection data into the fully connected neural network, filters out the target state in the attribute through the fully connected neural network, performs weighting processing on the target state, and outputs the attribute corresponding to the detection data feature.
  • the target object Since multiple states are stored in the attributes of the target object, the target object has some states in the attributes, and the state of the target object is regarded as the target state.
  • the computer equipment preprocesses the attributes corresponding to the detection data.
  • the computer equipment can use the mulit-hot (multi-hot) vector to express the attributes corresponding to the detection data.
  • the mulit-hot vector is composed of 0 and 1. 0 represents that the target object does not have a corresponding state, and 1 represents that the target object has a corresponding state.
  • the computer device can input the mulit-hot vector into the fully connected neural network, and the fully connected network filters out the target state of the target object through the coding matrix, and weights the selected target state, and outputs the attributes corresponding to the detection data Characteristics.
  • this application performs weighting processing on the filtered target states, so that the processed results can concentrate the characteristics of the attributes corresponding to the mulit-hot vector.
  • W T is the coding feature matrix
  • the W T can be a pre-trained matrix, or a matrix trained in the process of calculating the attributes of the full neural network
  • j is the number of the jth detection time
  • x j Is the mulit-hot vector of the attribute corresponding to the j-th detection time, that is, x j represents the attribute mulit-hot vector corresponding to the j-th detection time
  • ⁇ j is the feature vector of the attribute corresponding to the j-th detection time, this feature
  • the vector represents the feature of the attribute corresponding to the jth detection, j is an integer greater than or equal to 1
  • b ⁇ represents the deviation parameter
  • b ⁇ can be a pre-trained parameter, or in the process of calculating the attribute
  • the computer device uses the mulit-hot vector of the attribute to represent the attribute, the dimensions of the mulit-hot vector may be more high.
  • the dimension of the feature ⁇ of the attribute can be reduced compared to the dimension of the mulit-hot vector, so that the process shown in step 2011 can also be regarded as a dimensionality reduction process. Conducive to subsequent calculations.
  • Step 2012 The computer device inputs the detection data into a time series analysis tool, extracts the characteristics of each type of data in the detection data at each time series through the time series analysis tool, and outputs a feature set.
  • the time series analysis tool can be the HCTSA (highly comparative time-series) code base.
  • the characteristics of each type of data in each time series can include the data distribution, entropy, and scaling attributes of this type of data. It can be seen that these features can represent the autocorrelation structure of this type of data, and because these features are based on actual detection Data is obtained, so these characteristics are interpretable.
  • the time series analysis tool can extract the characteristics of each type of data in the jth detection data based on the preset feature extraction rules among them, Represents the timing characteristics of the k-th data type during the j-th detection, z is a positive integer, k is the number of the data type, j is the number of the j-th detection, that is, the number of the j-th detection time, and j is a positive integer , 1 ⁇ j ⁇ J, J represents the total number of detections of the target object. It should be noted that the embodiment of the present application does not specifically limit the preset feature extraction rule.
  • the time series analysis tool stores the extracted features of each detection data in a feature set Finally, the time series analysis tool can output the feature set, so that the computer device can obtain the feature set.
  • the computer device Since the features in the feature set can only represent the autocorrelation structure of these various types of data, and cannot reflect the features of all detection data, the computer device also needs to perform the following step 2013 to obtain the features of the detection data.
  • step 2013, the computer device inputs the feature set into the deep cross network, and cross-processes each time sequence feature in the feature set through the deep cross neural network, and outputs the feature of the detection data.
  • DCN deep&cross network, deep cross network
  • the computing device can input the feature set into the cross network and the deep network respectively.
  • multiple timing features in the feature set can be crossed and output
  • the common features of all features in the feature set are extracted through the deep network.
  • the crossed features output by DCN are combined with the common features extracted by the deep network to obtain the features of the detection data.
  • the embodiment of the present application does not specifically limit the execution order of steps 2011-203.
  • the computing device may perform step 2011 first, and then perform steps 2012 and 2013; it may also perform steps 2012 and 2013 first, and then perform step 2011; it may also perform steps 2012 and 2013 while performing step 2011.
  • Step 2014 The computer device inputs the feature of the attribute and the feature of the detection data into a deep neural network, extracts the mixed feature of the detection data and the attribute corresponding to the detection data through the deep neural network, and outputs the detection feature.
  • the computer device can first splice the characteristics of the attribute and the characteristics of the detection data to obtain a spliced feature, and then input the merged feature into the deep nerve.
  • Each node in the deep neural network can calculate the data in ⁇ j according to the following second formula, so that The deep neural network can output the detection feature ⁇ j at the jth detection time, and the second formula can be expressed as:
  • W x is the first weight matrix
  • each weight of W x is used to represent the importance of each element in ⁇ j , through ⁇ j can be achieved in each element of the weighting process, further elements may be integrated in the ⁇ j, ReLU (rectified linear unit rectifying linear) function can be related to data mining preferably between features, thus, strong expression function RELU , Use the ReLU function to After processing, the processed result ⁇ j can express the characteristics of ⁇ j . Therefore, ⁇ j can be used as the detection feature of the jth detection.
  • ReLU rectifified linear unit rectifying linear
  • the computer device can then extract the detection characteristics at each detection time through the deep neural network.
  • each detection time The above detection features are called sub-detection features. Therefore, the detection feature finally output by the deep neural network includes at least one sub-detection feature.
  • the detection feature Since the characteristics of the detection data and the attributes of the attributes include the characteristics of various types of data in various time series, the detection feature has a multi-modality, and the detection feature can be regarded as a multi-modal feature.
  • the detection feature in the embodiment of the present application can better reflect the feature in the detection process of the target object; and,
  • the detection data is the actual detected data, which can be used as an objective basis to make the acquired detection features interpretable.
  • the attributes corresponding to the detection data are the result of subjective judgment. Therefore, based on the attributes and detection data, the acquired detection features are The accuracy is high.
  • FIG. 3 is a schematic diagram of a diagnostic estimation model provided by an embodiment of the present application, in the multi-modal feature extraction part.
  • the medical coding data of the computer equipment that is, the attributes corresponding to the detection data
  • the mulit-hot vector is input to the fully connected neural network
  • the fully connected neural network can output the medical coding data through calculation
  • the feature that is, the feature of the attribute
  • the feature process of the fully connected neural network that can output the medical code data through calculation is the process of embedding the medical code data.
  • the computer equipment performs feature extraction on the time series data (that is, the detection data) to obtain a feature set.
  • the computer equipment inputs the feature set to DCN, and DCN outputs cross-multiple time series feature mixture, that is, It is the feature of the detection data.
  • the computer equipment mixes the multi-time-series cross-mixing feature with the attribute feature, and obtains the multi-modal feature (that is, the detection feature) based on the mixed feature.
  • the computer device can also adopt other neural networks to obtain the characteristics of the detection data.
  • the computer device inputs the detection feature into the first neural network, and for the detection features in each time sequence in the detection feature, the first neural network outputs the first regular feature and the feature different from the first regular feature after two different sequence calculations
  • the second regularity feature of, the first regularity feature represents the historical change law of the detection feature
  • the second regularity feature represents the future change law of the detection feature.
  • the first neural network may be BiRNN (bidirectional recurrent neural networks) with an attention mechanism.
  • the BiRNN may be composed of a first sub-network and a second sub-network, where the first sub-network Used to obtain the first regular feature, and the second sub-network is used to obtain the second regular feature.
  • the computer device inputs the detection feature into the first sub-network of the first neural network according to the reverse sequence sequence, and performs reverse sequence calculation on the detection feature through the first sub-network, Obtain the first law feature; according to the forward sequence sequence, input the detection feature into the second sub-network of the first neural network, and perform forward sequence calculation on the detection feature through the first sub-network to obtain the second law feature.
  • the computer device can input the detection feature into the first sub-network in a forward time sequence, and the computer device can input the detection feature into the first sub-network in a reverse time sequence.
  • the detection feature is input into the second sub-network.
  • the first sub-network can calculate each sub-feature in the detection feature based on the calculation rules preset in the first sub-network, and finally, the first sub-network
  • the sub-network can output the first regular feature Among them, b is used to indicate the reverse direction, and the embodiment of the present application does not specifically limit the preset calculation rule in the first sub-network.
  • the input layer of the node input specifically, the computer device inputs ⁇ 1 into the first node of the input layer of the first sub-network, and inputs ⁇ 2 into the second node of the input layer of the first sub-network,...( And so on).
  • the second sub-network can calculate each sub-feature in the detection feature based on the calculation rules preset in the second sub-network, and finally, the second sub-network
  • the sub-network can output the second regular feature
  • f represents the forward direction
  • the embodiment of the application does not specifically limit the preset calculation rule in the second sub-network.
  • the computer device determines the global feature of the target object based on the first regularity feature and the second regularity feature.
  • the computer device can first obtain the law feature representing the global change law of the detection feature based on the first law feature and the second law feature, and then according to This regular feature obtains the global feature.
  • this step 203 can be implemented through the process shown in the following steps 2031-2033.
  • Step 2031 the computer device splices the first regularity feature and the second regularity feature to obtain a third regularity feature.
  • Step 2032 The computer device performs weighting processing on the third regularity feature to obtain a fourth regularity feature.
  • the fourth regularity feature is used to represent the global change law of the detected feature.
  • the computing device can perform weighting processing on the third regular feature through the attention mechanism in the first neural network.
  • this step 2032 can be implemented through the process shown in steps 11-13.
  • Step 11 The computer device performs weight learning based on the first attention mechanism and the third regularity feature to obtain at least one first weight, and the first weight is used to indicate that a piece of detection data corresponds to the piece of detection data The importance of the attribute.
  • the first attention mechanism is any attention mechanism in the first neural network
  • the computer device can perform weight learning based on the weight learning strategy in the first attention mechanism
  • the weight learning strategy It can be a location-based attention weight learning strategy
  • the location-based attention weight learning strategy can be expressed as: Among them, W ⁇ is the second weight vector, W ⁇ T is the transpose matrix of W ⁇ , b ⁇ is the second deviation parameter, It is the first weight corresponding to the jth detection.
  • weight learning strategy in the first attention mechanism may also be other attention weight learning strategies, and the embodiment of the application does not specifically limit the weight learning strategy in the first attention mechanism.
  • Step 12 The computer device performs normalization processing on the at least one first weight to obtain at least one second weight.
  • the at least one first weight value in step 11 is a value obtained through mathematical calculation, the at least one first weight value may be too large or too small.
  • the at least one first weight value may be normalized , So that the size of each second weight value obtained after processing is moderate, then, when the at least one first weight value is too large, the at least one first weight value can be reduced proportionally, when the at least one first weight value If the weight value is too small, the at least one first weight value may be proportionally enlarged to realize the normalization processing of the at least one first weight value.
  • each second weight value is only the result of normalizing a first weight value
  • the second weight value has the same function as the first weight value, and it is used to indicate that a detection data corresponds to the detection data. The importance of the attribute.
  • Step 13 The computer device performs weighting processing on the third regularity feature based on at least one second weight value to obtain the fourth regularity feature.
  • the computer device uses the at least one second weight Substituting the third formula into the third formula, the output of the third formula is used as the fourth rule feature to implement the weighting process for the third rule feature, the third formula can be expressed as Among them, c is the fourth law feature, It is the second weight corresponding to the jth detection, J is the total number of detections of the target object.
  • the third law feature is weighted by at least one second weight, and the third law feature is more concentrated expression. Therefore, the fourth regular feature can represent the global change regularity of the detection feature.
  • the first law feature and the second law feature are weighted, the first law feature and the second law feature can be integrated and represented by the fourth law feature, so the fourth law feature can represent the first law feature
  • the historical change law of can also represent the future change law represented by the first law feature. Therefore, the fourth feature law can represent the global change law of the detection feature.
  • Step 2033 The computer device determines the global feature based on the third regularity feature and the fourth regularity feature.
  • the computer device may determine the last detection time of the third regular feature
  • the corresponding regularity feature and the fourth regularity feature are substituted into the fourth formula, and the output of the fourth formula is taken as the global feature, where the fourth formula can be expressed as: among them, Is the global feature, h J is the regular feature corresponding to the last detection time of the target object in the third regular feature, [h J ,c] is the vector after h J and c are joined, W d is the third weight matrix, Is the transposed matrix of W d , b d is the third deviation parameter, and ReLU() refers to the linear rectification function.
  • the fourth law feature represents the global change law of the detection feature
  • the first law feature can represent the history change law of the detection feature
  • the second law feature can represent the future change law of the detection feature. Therefore, the three regular features are weighted. , The result obtained can represent the global characteristics of the target object.
  • the computer device inputs the global feature into a second neural network, and the second neural network extracts and outputs at least one local feature of the target object from the global feature.
  • the second neural network may be HMCN (hierarchical multi-label classification networks), and the second neural network may extract local features from the input global features. Since the global features cannot represent the details of the target object, the details of the target object can be extracted through the second neural network. Specifically, the second neural network can extract the details of the target object level by level, so that the final extracted details can satisfy the requirements. Demand for attribute forecasting.
  • HMCN hierarchical multi-label classification networks
  • Each layer of the second neural network can output one local feature.
  • the computing device inputs the global feature to the second neural network
  • the global feature can be input from the input layer to the output layer of the second neural network.
  • the second neural network can calculate the global feature layer by layer.
  • it can calculate the hierarchical characteristics of the first target layer and the local characteristics of the target object in the first target layer based on the output data of the second target layer.
  • the first target layer Is any layer of the second neural network
  • the second target layer is the upper layer of the first target layer in the second neural network
  • the level feature is used to indicate that the global feature is in the network layer of the second neural network
  • the hierarchical characteristics of the first target layer are determined by the global characteristics and the hierarchical characteristics of the second target layer.
  • the second target layer of the second neural network After the second target layer of the second neural network generates the hierarchical characteristics of the second target layer and the local characteristics of the target object in the second target layer, the second target layer can output the second target layer's Hierarchical features and the global features (output data of the second target layer), so that the first target layer can receive the hierarchical features of the second target layer and the global features. Then, following the global features output by each network layer of the second neural network, the global feature can be input to each network layer of the second neural network.
  • the first target layer A target layer may calculate the hierarchical characteristics of the first target layer based on the hierarchical characteristics of the second target layer and the global characteristics.
  • the hierarchical characteristics of the i-th layer of the second neural network It can be expressed as: Among them, G represents the overall situation, Is the fourth weight matrix, Is the hierarchical feature of layer i-1, Indicates the hierarchical characteristics of the i-th layer, b G is the fourth deviation parameter, and the i-th layer can be regarded as the first target layer.
  • the nodes on the first target layer may obtain the local characteristics of the target object on the first target layer based on the hierarchical characteristics of the first target layer and the global characteristics.
  • the local feature of the target object in the i-th layer It can be expressed as: Among them, L represents the network layer, Is the fifth weight matrix, b T is the fifth deviation parameter.
  • each layer of the second neural network is calculated based on the hierarchical features and global features of the previous layer, the local features of the target object on each layer of the second neural network are affected by the local features of the previous layer.
  • the influence of features since the level expression of each layer is determined by the level features of this layer, the local features generated by any network layer in the second neural network can be used as the parent of the local features generated by the next network layer.
  • the second neural network can extract the details of the target object level by level.
  • the computer device predicts the attribute of the target object based on at least one local feature of the target object.
  • a local feature can represent different levels of details of the target object, considering that there are more details, these can be processed intensively to obtain more detailed local features, and then attribute predictions are made based on this more detailed local feature.
  • this step 205 can be implemented through the process shown in the following steps 2051-2052.
  • Step 2051 The computing device performs weighting processing on at least one local feature of the target object to obtain the target local feature.
  • the target local feature is also a more detailed local feature.
  • the computing device can perform weighting processing on the local features of the target object through the attention mechanism in the second neural network. In a possible implementation manner, this step 2051 can be implemented through the process shown in steps 21-22.
  • Step 21 The computer device performs weight learning based on the second attention mechanism and the at least one local feature to obtain at least one third weight, and one third weight is used to indicate the importance of a local feature.
  • the second attention mechanism is any attention mechanism in the second neural network
  • the computer computing device performs weight learning based on the second attention mechanism and the at least one local feature
  • the second attention mechanism can be
  • the weight learning strategy in the second attention mechanism, the learning weight, the weight learning strategy in the second attention mechanism can be expressed as:
  • W ⁇ is the sixth weight matrix
  • b ⁇ is the sixth deviation parameter
  • M represents the number of local features.
  • weight learning strategy in the second attention mechanism may also be other attention weight learning strategies, and the embodiment of the application does not specifically limit the weight learning strategy in the second attention mechanism.
  • Step 22 The computer device performs weighting processing on at least one local feature of the target object based on the at least one third weight to obtain the target local feature.
  • the computer device substitutes at least one third weight and at least one local feature of the target object into the fifth formula, and uses the output of the fifth formula as the target local feature to implement weighting processing on at least one local feature of the target object.
  • the fifth formula can be expressed as:
  • a G is the attribute corresponding to the detection data corresponding to the current time series
  • N is the number of layers of the second neural network.
  • At least one local feature is weighted by at least one third weight, so that the obtained target local feature is more detailed.
  • Step 2052 Predict the attributes of the target object based on the local features of the target.
  • the computer device can substitute the local feature of the target into a sixth formula to predict the attribute of the target object.
  • the sixth formula is used to predict the attribute of the target object.
  • the sixth formula can be expressed as: among them, It is the attribute corresponding to the predicted target object in the J+1th detection data, W G is the seventh weight matrix, and b G is the seventh deviation parameter.
  • the second neural network may determine whether to output the currently predicted attribute according to the global loss and the local loss in the second neural network.
  • the second neural network after the global feature is input to the second neural network, if the global loss and the local loss in the second neural network meet preset conditions, the second neural network outputs the current prediction Attribute, otherwise, the second neural network adjusts the weight matrix in the second neural network until the global loss and local loss in the second neural network meet the preset conditions, the local loss is the second neural network in each layer The difference between the expected output data and the actual output data, and the global loss is the difference between the expected final output data of the second neural network and the actual final output data.
  • the second neural network can predict the local features of any layer at the next detection time (referred to as predicted local features), then, the predicted local features of the i-th layer It can be expressed as
  • the second neural network can predict the attributes of at least one target object.
  • the second neural network can use a cross-entropy strategy to calculate the The local loss of any layer, then, the i-th local loss L li can be expressed as:
  • Q is the number of target objects, Based on the global characteristics of the Q-th target object, the actual output data of the i-th layer, To predict the output data of the i-th layer based on the global characteristics of the Q-th target object.
  • the second neural network can calculate the attributes of at least one target object in the next detection. Then, the The second neural network can predict the attributes of at least one target object at the next detection, and the cross-entropy strategy is used to calculate the global loss L G , where L G can be expressed as:
  • the global loss and local loss in the second neural network meet the preset conditions, it means that the difference between the local features generated by each layer of the second neural network and the expected local features reaches the preset accuracy, which can ensure The accuracy of the local features of each layer of the second neural network is relatively high, so that the accuracy of the predicted attributes can be improved.
  • the computer device since the second neural network is calculated based on numerical values, and each state in the attributes of the target object is actually represented by a state identifier, the computer device also needs to actualize the second neural network.
  • the output data is converted into attributes composed of state identifiers.
  • the actual output data of the second neural network may include at least one probability value. Each probability value corresponds to a state in the attribute of the target object.
  • the computer device When any probability value is greater than the target value , Indicating that the target object has the target state corresponding to any probability, the computer device stores the first state identifier in the position of the target state in the attribute; when any probability value is less than or equal to the target value, the target If the object does not have the target state corresponding to any one of the probabilities, the computer device stores the second state identifier in the position of the target state in the attribute. Then, by judging each probability value, the actual expression of the attribute can be obtained.
  • the embodiment of the application does not specifically limit the target value.
  • the attention loop network is equivalent to the first neural network.
  • the computing device inputs global features to each layer of the second neural network in the neural-level multi-label modeling part, and the first layer generates the first-level hierarchical features based on the global features
  • the computer equipment is calculating the local loss L l1 of the first layer, and the first layer will Output to the second layer, so that the second layer can perform a calculation process similar to the first layer, and finally all the layers of the second neural network can get a Computer equipment will be M Output to the attention set (attentional ensembel), in the attention set, based on the second attention mechanism, generate predicted output data And then based on the predicted output data Generate a global loss L G. Then, when the loss of the global and local loss L G L i satisfies a preset condition, the second neural network output can be
  • the method provided by the embodiment of the present application determines the global feature of the target object based on the regular feature representing the history of the detection feature and the future change rule, and refines the global feature to obtain at least one local feature of the target object, then ,
  • the refined local features can better reflect the features of the target object, and then predict the attributes of the target object based on the local features. Therefore, the accuracy of the predicted attributes can be improved.
  • the prediction can be improved The accuracy of the diagnosis results.
  • the detection feature in the embodiment of the present application can better reflect the feature in the detection process of the target object, compared to the feature of the target object obtained only based on the detection data.
  • the detection data is the actual detected data, which can be used as an objective basis to make the acquired detection features interpretable, and the attributes corresponding to the detection data are the result of subjective judgment. Therefore, based on the attributes and detection data, the acquired detection The accuracy of features is high.
  • the global loss and local loss in the second neural network meet the preset conditions, it means that the local features generated by each layer of the second neural network have reached the expected value, so that the output layer of the second neural network can be guaranteed The accuracy of the output local features is high.
  • FIG. 4 is a schematic structural diagram of a device for predicting attributes of a target object based on machine learning according to an embodiment of the present application, and the device includes:
  • the acquiring module 401 is configured to determine the detection characteristics of the target object according to the detection data of the target object and the attributes corresponding to the detection data;
  • the calculation module 402 is configured to input the detection features into a first neural network; for the detection features in each time sequence of the detection features, the first neural network outputs the first regular feature after two different time sequence calculations And a second regularity feature that is different from the first regularity feature, wherein the first regularity feature represents a historical change law of the detection feature, and the second regularity feature represents a future change law of the detection feature;
  • the acquiring module 401 is further configured to determine the global characteristics of the target object based on the first regular feature and the second regular feature;
  • the extraction module 403 is configured to input the global feature into a second neural network; the second neural network extracts and outputs at least one local feature of the target object from the global feature;
  • the prediction module 404 is configured to predict the attributes of the target object based on at least one local feature of the target object.
  • the obtaining module 401 is used to:
  • the feature of the attribute and the feature of the detection data are input into a deep neural network, and a mixed feature of the detection data and the attribute corresponding to the detection data is extracted through the deep neural network, and the detection feature is output.
  • the obtaining module 401 is used to:
  • the global feature is determined.
  • the obtaining module 401 is specifically configured to:
  • weight learning is performed to obtain at least one first weight, and one of the first weights is used to represent the attribute of a piece of detection data corresponding to the piece of detection data Importance;
  • weighting is performed on the third regularity feature to obtain the fourth regularity feature.
  • the prediction module 404 includes:
  • a processing unit configured to perform weighting processing on at least one local feature of the target object to obtain the target local feature
  • the prediction unit is configured to predict the attributes of the target object based on the local features of the target.
  • the processing unit is used for;
  • weighting is performed on the at least one local feature to obtain the target local feature.
  • each layer of the second neural network outputs one local feature.
  • the device further includes an output module, configured to, after the global feature is input to the second neural network, if the global loss and the local loss in the second neural network meet a preset condition, then The second neural network outputs the current predicted attributes, the local loss is the difference between the expected output data of the second neural network in each layer and the actual output data, and the global loss is the second neural network The difference between the final output data expected by the network and the actual final data.
  • an output module configured to, after the global feature is input to the second neural network, if the global loss and the local loss in the second neural network meet a preset condition, then The second neural network outputs the current predicted attributes, the local loss is the difference between the expected output data of the second neural network in each layer and the actual output data, and the global loss is the second neural network The difference between the final output data expected by the network and the actual final data.
  • the device further includes a generating module for generating the local features output by the first target layer based on the hierarchical features of the first target layer in the second neural network and the local features generated by the second target layer,
  • the hierarchical feature of the first target layer is used to indicate the state of the global feature in the first target layer
  • the second target layer is the upper layer of the first target layer in the second neural network .
  • the hierarchical characteristics of the first target layer are determined by the global characteristics and the hierarchical characteristics of the second target layer.
  • FIG. 5 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • the computer device 500 may have relatively large differences due to different configurations or performance, and may include one or more CPUs (central processing units, processors) 501 And one or more memories 502, wherein at least one instruction is stored in the memory 502, and the at least one instruction is loaded and executed by the processor 501 to achieve the machine learning-based goals provided by the foregoing method embodiments Object attribute prediction method.
  • the computer device 500 may also have components such as a wired or wireless network interface, a keyboard, and an input/output interface for input and output.
  • the computer device 500 may also include other components for implementing device functions, which will not be repeated here.
  • a computer-readable storage medium such as a memory including instructions, which can be executed by a processor in a terminal to complete the method for predicting attributes of a target object based on machine learning in the foregoing embodiment.
  • the computer-readable storage medium may be ROM (read-only memory, read-only memory), RAM (random access memory, random access memory), CD-ROM (compact disc read-only memory, CD-ROM) , Magnetic tapes, floppy disks and optical data storage devices.
  • a computer program product including instructions, which when run on a computer, cause the computer to execute the method for predicting the attributes of a target object based on machine learning in the foregoing embodiments.
  • the device for predicting attributes of the target object based on machine learning provided in the above embodiment, only the division of the above functional modules is used as an example for illustration. In actual applications, the above function can be assigned differently according to needs.
  • the function module is completed, that is, the internal structure of the device is divided into different function modules to complete all or part of the functions described above.
  • the device for predicting the attributes of a target object based on machine learning provided in the above embodiments belongs to the same concept as the embodiments of the method for predicting attributes of a target object based on machine learning. The specific implementation process is detailed in the method embodiments, and will not be repeated here.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于机器学习的目标对象属性预测方法及相关设备，属于数据预测技术领域。方法通过基于表示检测特征历史和未来的变化规律的规律特征，确定出目标对象的全局特征，并对全局特征进行细化，以得到目标对象的至少一个局部特征，那么细化后的局部特征更能体现目标对象的特点，进而根据局部特征预测目标对象的属性，可以提高预测的属性的精度，当目标对象的属性为预测的诊断结果时，则可以提高预测的诊断结果的精度。

Description

基于机器学习的目标对象属性预测方法及相关设备
本申请要求于2019年05月09日提交中国专利局、申请号为201910386448.4、申请名称为“属性预测方法、装置、计算机设备及计算机可读存储介质”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
本申请涉及数据预测技术领域，特别涉及基于机器学习的目标对象属性预测。
背景技术
EHR(electronic health records，电子健康记录)数据可以记录目标对象每一次的问诊记录，随着技术的发展，越来越多的临床诊断估计模型可以基于患者的EHR数据，模拟医生的诊断过程，以对用户未来的发病情况进行预测。
在相关技术中，预测用户未来的发病情况的过程可以为：将EHR数据中的医疗编码数据作为患者的属性，且输入至临床诊断估计模型中，临床诊断估计模型对医疗编码数据进行训练，可以输出预测的诊断结果。其中，临床诊断估计模型进行对医疗编码数据进行训练的过程，可以表征临床诊断估计模型模拟医生的诊断过程，使得后续可以根据由临床诊断估计模型预测的诊断结果，来对患者未来的发病情况进行预测。
然而，因上述诊断预测过程中由临床诊断估计模型预测的诊断结果精确度较低，导致基于该预测的诊断结果确定的患者未来的发病情况不准确。
发明内容
本申请实施例提供了一种基于机器学习的目标对象属性预测方法及相关设备，能够解决临床诊断估计模型预测的诊断结果精确度低的问题。所述技术方案如下：
一方面，提供了一种基于机器学习的目标对象属性预测方法，由计算机设备执行，所述方法包括：
根据目标对象的检测数据以及所述检测数据对应的属性，确定所述目标对 象的检测特征；
将所述检测特征输入第一神经网络；
针对所述检测特征中各个时序上的检测特征，所述第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于所述第一规律特征的第二规律特征，其中，所述第一规律特征表示所述检测特征的历史变化规律，所述第二规律特征表示所述检测特征的未来变化规律；
基于所述第一规律特征和第二规律特征，确定所述目标对象的全局特征；
将所述全局特征输入第二神经网络；
所述第二神经网络从所述全局特征提取并输出所述目标对象的至少一个局部特征；
基于所述目标对象的至少一个局部特征，对所述目标对象的属性进行预测。
另一方面，提供了一种基于机器学习的目标对象属性预测装置，所述装置包括：
获取模块，用于根据目标对象的检测数据以及所述检测数据对应的属性，确定所述目标对象的检测特征；
计算模块，用于将所述检测特征输入第一神经网络；针对所述检测特征中各个时序上的检测特征，所述第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于所述第一规律特征的第二规律特征，其中，所述第一规律特征表示所述检测特征的历史变化规律，所述第二规律特征表示所述检测特征的未来变化规律；
所述获取模块，还用于基于所述第一规律特征和第二规律特征，确定所述目标对象的全局特征；
提取模块，用于将所述全局特征输入第二神经网络；所述第二神经网络从所述全局特征提取并输出所述目标对象的至少一个局部特征；
预测模块，用于基于所述目标对象的至少一个局部特征，对所述目标对象的属性进行预测。
另一方面，提供了一种计算机设备，该计算机设备包括：处理器；用于存放计算机程序的存储器；其中，该处理器用于执行存储器上所存放的计算机程序，实现上述基于机器学习的目标对象属性预测方法所执行的操作。
另一方面，提供了一种计算机可读存储介质，该计算机可读存储介质内存储有计算机程序，该计算机程序被处理器执行时实现上述基于机器学习的目标对象属性预测方法所执行的操作。
另一方面，提供了一种包括指令的计算机程序产品，当其在计算机上运行时，使得所述计算机执行上述基于机器学习的目标对象属性预测方法所执行的操作。
本申请实施例提供的技术方案带来的有益效果是：
通过基于表示检测特征历史和未来的变化规律的规律特征，确定出目标对象的全局特征，并对全局特征进行细化，以得到目标对象的至少一个局部特征，那么，细化后的局部特征更能体现目标对象的特点，进而根据局部特征预测目标对象的属性，因此，可以提高预测的属性的精度，当目标对象的属性为预测的诊断结果时，则可以提高预测的诊断结果的精度。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例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的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的一种实施环境的示意图；
图2是本申请实施例提供的一种基于机器学习的目标对象属性预测方法的流程图；
图3是本申请实施例提供的一种诊断估计模型的示意图；
图4是本申请实施例提供的一种基于机器学习的目标对象属性预测装置的结构示意图；
图5是本申请实施例提供的一种计算机设备的结构示意图。
具体实施方式
为了便于理解本申请提供的属性预测方法，下面先介绍属性预测的相关技术以及发明人的研究发现。
在相关技术中，预测用户未来的发病情况的过程可以为：将EHR数据中的医疗编码数据作为患者的属性，且输入至临床诊断估计模型中，临床诊断估计模型进行对医疗编码数据进行训练，可以输出预测的诊断结果。其中，临床诊断估计模型对医疗编码数据进行训练的过程，可以表征临床诊断估计模型模拟医生的诊断过程，使得后续可以根据由训练好的临床诊断估计模型预测的诊断结果，来对患者未来的发病情况进行预测。
然而，发明人在对上述相关技术的研究中发现：在上述诊断预测过程中，临床诊断估计模型输入的是医疗编码数据，由于医疗编码数据包括成千上种疾病的数据，对于一位患者而言，仅有可能患有其中的一种或者几种疾病，并不可能患有各种各样的疾病，因此，导致医疗编码数据内有用数据在医疗编码数据中分布比较稀疏且离散，并且医疗编码数据仅能表示患者患有的疾病，并不能表示患者整体的身体状态，那么，在临床诊断估计模型利用这样的医疗编码数据进行训练后，输出的预测的诊断结果精确度低，导致基于该预测的诊断结果确定的患者未来的发病情况不准确。
为了解决上述相关技术存在的技术问题，本申请提供一种基于机器学习的目标对象属性预测方法，该方法具体包括：根据目标对象的检测数据以及检测数据对应的属性，确定目标对象的检测特征；将检测特征输入第一神经网络，针对检测特征中各个时序上的检测特征，第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于第一规律特征的第二规律特征，第一规律特征表示检测特征的历史变化规律，第二规律特征表示检测特征的未来变化规律；基于第一规律特征和第二规律特征，确定目标对象的全局特征；将全局特征输入第二神经网络，第二神经网络从全局特征提取并输出目标对象的至少一个局部特征；基于目标对象的至少一个局部特征，对目标对象的属性进行预测。
可见，在本申请提供的基于机器学习的目标对象属性预测方法中，通过基于表示检测特征历史和未来的变化规律的规律特征，确定出目标对象的全局特征，并对全局特征进行细化，以得到目标对象的至少一个局部特征，那么，细化后的局部特征更能体现目标对象的特点，进而根据局部特征预测目标对象的属性，因此，可以提高预测的属性的精度，当目标对象的属性为预测的诊断结果时，则可以提高预测的诊断结果的精度。
本申请实施例提供的基于机器学习的目标对象属性预测方法是基于机器学习技术实现的。其中，机器学习(Machine Learning,ML)是一门多领域交叉学科，涉及概率论、统计学、逼近论、凸分析、算法复杂度理论等多门学科。专门研究计算机怎样模拟或实现人类的学习行为，以获取新的知识或技能，重新组织已有的知识结构使之不断改善自身的性能。机器学习是人工智能的核心，是使计算机具有智能的根本途径，其应用遍及人工智能的各个领域。机器学习和深度学习通常包括人工神经网络、置信网络、强化学习、迁移学习、归纳学习、式教学习等技术。
另外，机器学习是人工智能(Artificial Intelligence,AI)的一个重要方向。其中，人工智能是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。换句话说，人工智能是计算机科学的一个综合技术，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式做出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。
人工智能技术是一门综合学科，涉及领域广泛，既有硬件层面的技术也有软件层面的技术。人工智能基础技术一般包括如传感器、专用人工智能芯片、云计算、分布式存储、大数据处理技术、操作/交互系统、机电一体化等技术。人工智能软件技术主要包括计算机视觉技术、语音处理技术、自然语言处理技术以及机器学习/深度学习等几大方向。
基于上述内容可知，本申请实施例提供的基于机器学习的目标对象属性预测方法涉及人工智能，尤其涉及人工智能中的机器学习。
应理解，本申请实施例提供的基于机器学习的目标对象属性预测方法可以应用于可以进行数据处理的计算机设备，如终端设备、服务器等；其中，终端设备具体可以为智能手机、计算机、个人数字助理(Personal Digital Assitant，PDA)、平板电脑等；服务器具体可以为应用服务器，也可以为Web服务器，在实际部署时，该服务器可以为独立服务器，也可以为集群服务器。
若本申请实施例提供的基于机器学习的目标对象属性预测方法由终端设备执行时，则终端设备可以直接根据用户输入的目标对象的检测数据以及该检 测数据对应的属性，对目标对象的属性进行预测，并显示预测结果，以供用户查看。若本申请实施例提供的基于机器学习的目标对象属性预测方法由服务器执行时，则服务器先根据终端设备上传的目标对象的检测数据以及该检测数据对应的属性，对目标对象的属性进行预测，得到预测结果；再将该预测结果发送给终端设备，以便终端设备将接收的预测结果进行显示，以供用户查看。
为了便于理解本申请实施例提供的技术方案，下面对本申请实施例提供的基于机器学习的目标对象属性预测方法应用于终端设备为例，对本申请实施例提供的基于机器学习的目标对象属性预测方法适用的应用场景进行示例性介绍。
作为本申请实施例提供的基于机器学习的目标对象属性预测方法的一种可能的应用场景，该应用场景包括：终端设备和用户；其中，终端设备用于执行本申请实施例提供的基于机器学习的目标对象属性预测方法，对目标对象的属性进行预测得到预测结果，以供用户查看。
终端设备在接收到用户触发的属性预测指令之后，终端设备可以根据目标对象的检测数据以及检测数据对应的属性，确定目标对象的检测特征；将检测特征输入第一神经网络，针对检测特征中各个时序上的检测特征，第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于第一规律特征的第二规律特征，第一规律特征表示检测特征的历史变化规律，第二规律特征表示检测特征的未来变化规律；基于第一规律特征和第二规律特征，确定目标对象的全局特征；将全局特征输入第二神经网络，第二神经网络从全局特征提取并输出目标对象的至少一个局部特征；基于目标对象的至少一个局部特征，对目标对象的属性进行预测，得到预测结果，以便终端设备将该预测结果显示给用户。
应理解，在实际应用中，也可以将本申请实施例提供的基于机器学习的目标对象属性预测方法应用于服务器。基于此，作为本申请实施例提供的基于机器学习的目标对象属性预测方法的另一种可能的应用场景，该应用场景包括：服务器、终端设备和用户。其中，终端设备在接收到用户触发的属性预测指令之后，终端设备将根据属性预测指令生成属性预测请求，并将属性预测请求发送给服务器，以使服务器能够在接收到终端设备发送的属性预测请求之后，根 据目标对象的检测数据以及检测数据对应的属性，确定目标对象的检测特征；将检测特征输入第一神经网络，针对检测特征中各个时序上的检测特征，第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于第一规律特征的第二规律特征，第一规律特征表示检测特征的历史变化规律，第二规律特征表示检测特征的未来变化规律；基于第一规律特征和第二规律特征，确定目标对象的全局特征；将全局特征输入第二神经网络，第二神经网络从全局特征提取并输出目标对象的至少一个局部特征；基于目标对象的至少一个局部特征，对目标对象的属性进行预测得到预测结果，以便服务器可以将得到的预测结果反馈给终端设备，使得用户能够在终端设备上查看该预测结果。
应理解，上述应用场景仅为示例，在实际应用中，本申请实施例提供的基于机器学习的目标对象属性预测方法还可以应用于其他进行属性预测的应用场景，在此不对本申请实施例提供的基于机器学习的目标对象属性预测方法做任何限定。
为使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请实施方式作进一步地详细描述。
图1是本申请实施例提供的一种实施环境的示意图。参见图1，该环境包括基于机器学习的目标对象属性预测系统100，该基于机器学习的目标对象属性预测系统包括预处理模块101，检测特征提取模块102、规律特征提取模块103以及预测模块104。
其中，预处理模块101，用于处理目标对象的检测数据以及检测数据对应的属性，具体将用户的检测数据以及检测数据对应的属性转化为检测特征提取模块102可以进行计算的数据。
检测特征提取模块102，用于提取检测数据的特征和检测数据对应的属性的混合特征，提取出的混合特征可以作为目标对象检测特征。具体地，检测特征提取模块102可以先基于预处理模块101处理后的数据，提取属性的特征以及检测数据的特征，然后，将提取到的属性的特征以及检测数据的特征进行拼接，最后，检测特征提取模块102基于拼接结果，提取检测特征。
规律特征提取模块103，用于提取规律特征以及生成目标对象的全局特征，该规律特征用于表示检测特征的全局变化规律，具体地，规律特征提取模块 103可以先提取基于检测特征提取模块102提取到的检测特征的历史变化规律和未来变化规律，然后，规律特征提取模块103可以基于检测特征的历史变化规律和未来变化规律，获取检测特征的全局变化规律，最后根据代表全局变化规律的规律特征，确定目标对象的全局特征。
预测模块104，用于预测目标对象的属性，具体地，预测模块104可以通过神经网络对规律特征提取模块103生成的全局特征进行细化进，得到目标对象的局部特征，然后，预测模块104使用目标局部特征集中表达获取的多个局部特征进行集中，最后，预测模块104基于目标局部特征对目标对象的属性进行预测。
需要说明的是，该基于机器学习的目标对象属性预测系统100中的各个模块的功能可以用一个计算机设备来实现，也可以用多个计算机设备来实现，本申请实施例实现各个模块的功能的计算机设备的数量不做具体限定。
图1介绍了基于机器学习的目标对象属性预测系统中的各个模块的各自功能，为了体现基于机器学习的目标对象属性预测系统进行属性预测的具体过程，参见图2，图2是本申请实施例提供的一种基于机器学习的目标对象属性预测方法的流程图。如图2所示，该基于机器学习的目标对象属性预测方法包括：
201、计算机设备根据目标对象的检测数据以及该检测数据对应的属性，确定该目标对象的检测特征。
该计算机设备可以是任一计算机设备。目标对象可以是任一个用户。
目标对象的检测数据可以包括目标对象在历史时间段内的每次检测的检测数据，每次检测的检测数据均对应一个检测时间，因此，每次检测的检测数据可以包括与该目标对象有关的多种类型的数据，以检测目标对象的体征为例，一次检测到的数据心跳数据、血压数据等类型的数据，对于任一类型的数据而言，每一次检测，可能检测出很多个数据，一次检测出的多个数据可以组成一个和这一次检测时间有关的时序序列，因此，一个检测时间可能对应多个时序序列，可以为每种类型的时序序列进行标号，以区分一个检测时间上的多中类型的检测数据。需要说明的是，本申请实施例对检测时间不做具体限定，对两个检测时间之间的时间间隔不做具体限定。另外，本申请实施例不限定上述历 史时间段，上述历史时间段可以是指在计算机设备利用属性预测方法对目标对象的属性进行预测之前的任一时间段。
那么，当该检测数据为目标对象的体征数据时，该检测数据可以是EHR数据内存储的时序数据，该时序数据包括目标对象每次问诊的问诊时间，以及在每个问诊时间上检测的目标对象的体征数据，可以理解的是，该问诊时间也即是检测时间。
上述属性用于指示目标对象的至少一个状态，该属性的每个位置对应一个状态，可以用状态标识表示该目对象是否具有对应的状态，该状态标识可以包括第一状态标识和第二状态标识，其中，该第一状态标识用于表示目标对象具有对应的状态，该第二状态标识用于表示目标对象不具有对应的状态，例如，当该属性内的任一位置上具有第一状态标识时，则说明目标对象具有该位置对应的状态，当该属性内的任一位置上具有第二状态标识时，则说明目标对象不具有该位置对应的状态。另外，可以使用不同的字符串来表示第一状态标识和第二状态标识，本申请实施例对表示该第一状态标识或第二状态标识的字符串不做具体限定。
当每次对目标对象进行检测后，会得到本次检测数据，基于本次的检测数据，可以确定目标对象的属性，而且该确定的目标对象的属性对应的检测时间就是本次的检测数据对应的检验时间。可见，在本申请实施例中，一个属性对应一个检测时间，而且若该目标对象的检测数据包括至少一个检测时间，则该目标对象的检测数据对应至少一个属性。
上述至少一个状态可以为目标对象的患病状态，还可以为其他状态，本申请实施例对该至少一个状态不做具体限定。当该目标对象的状况为患病状态时，该检测数据对应的属性可以是医疗编码数据，一个医疗编码数据可以由0和1组成，在医疗编码数据中的每一个位置对应一个疾病，当任一位置上的数据为0时，代表目标对象不患有该位置对应的疾病，当任一位置上的数据为1时，代表目标对象患有该位置对应的疾病，可以理解的是，此处的0相当于第二状态标识，此处的1相当于第一状态标识。
为了便于计算，该计算机设备可以先对目标对象的检测数据以及该检测数据对应的属性进行预处理，使得检测数据以及检测数据对应的属性符合后续计 算要求的格式，然后在对处理后的数据进行特征提取，得到该目标对象的检测特征。在一种可能的实现方式中，本步骤201可以通过下述步骤2011-2014所示的过程来实现。
步骤2011、该计算机设备将该检测数据对应的属性输入全连接神经网络，通过该全连接神经网络筛选出该属性中的目标状态，对该目标状态进行加权处理，输出该检测数据对应的属性的特征。
由于目标对象的属性内存储有多个状态，该目标对象具有属性内的一些状态，将目标对象所具有的状态作为目标状态。
为了便于计算，计算机设备对该检测数据对应的属性进行预处理，在一种可能的实现方式中，该计算机设备可以将该检测数据对应的属性用mulit-hot(多热)向量来表示，以达到对该检测数据对应的属性进行预处理的目的。该mulit-hot向量由0和1组成，0代表目标对象不具有对应的状态，1代表目标对象具有对应的状态。
那么，该计算机设备可以将mulit-hot向量输入全连接神经网络，由全连接网络通过编码矩阵筛选出目标对象的目标状态，并对筛选出的目标状态进行加权处理，输出该检测数据对应的属性的特征。其中，本申请通过对筛选出的目标状态进行加权处理，使得处理后的结果可以集中mulit-hot向量对应的属性的特征。
具体地，该全连接网络中的每个网络节点可以通过第一公式对该mulit-hot向量内的数据进行计算，该第一公式可以表示为：π j＝ReLU(W Tx j+bπ)，其中，W T为编码特征矩阵，该W T可以是预先训练出的矩阵，也可是全神经网络计算属性的特征过程中，训练出的矩阵；j为第j次检测时间的编号；x j为第j个检测时间对应的属性的mulit-hot向量，也就是x j代表第j次检测时所对应的属性mulit-hot向量；π j为第j个检测时间对应属性的特征向量，该特征向量代表第j次检测时所对应的属性的特征，j为大于等于1的整数；b π表示偏差参数，而且bπ可以是预先训练出的参数，也可是全神经网络计算属性的特征过程中，训练出的参数。
当J表示目标对象总共的检测次数时，若该计算机设备通过全神经网络进行对该检测数据对应的所有属性进行计算后，可以得到检测数据所对应属性的 特征π＝[π 1,...,π J]，需要说明的是，该计算机设备除了通过全连接网络计算属性的特征外，还可以通过其他神经网络计算属性的特征。
由于一个属性用于指示目标对象的至少一个状态，那么，当该属性指示的状态较多时，若该计算机设备利用该属性的mulit-hot向量来表示该属性，则mulit-hot向量的维度可能比较高。此时，可以通过筛选目标状态的方式，使得属性的特征π的维度相比mulit-hot向量的维度有所降低，从而使得步骤2011所示的过程也可以视为一个降维的过程，如此有利于后续的计算。
步骤2012、该计算机设备将该检测数据输入时间序列分析工具，通过时间序列分析工具提取该检测数据中每一类数据在各个时序的特征，输出特征集合。
时间序列分析工具可以是HCTSA(highly comparative time-series，高度比较的时间序列分析)代码库。每一类数据在各个时序的特征可以包括表示这一类数据的数据分布、熵、缩放属性等特征，可见这些特征可以表示这一类数据的自相关结构，并且由于这些特征是基于实际的检测数据得到的，因此这些特征具有可解释性。
当该计算机设备将第j次的检测数据输入到时间序列分析工具后，时间序列分析工具可以基于预设特征提取规则，提取第j次的检测数据中每一类数据的特征
Figure PCTCN2020086007-appb-000001
其中，
Figure PCTCN2020086007-appb-000002
表示第j次检测时第k个数据类型的时序特征，z为正整数，k为数据类型的编号，j表示第j次检测的编号，也就是第j次检测时间的编号，j为正整数，1≤j≤J，J表示目标对象总共的检测次数。需要说明的是，本申请实施例对该预设特征提取规则不做具体限定。
若J表示目标对象总共的检测次数，则当将这J次的检测数据都通过时间序列分析工具处理后，该时间序列分析工具将提取到的每一次检测数据的特征存储在一个特征集合
Figure PCTCN2020086007-appb-000003
最后该时间序列分析工具可以输出该特征集合，从而该计算机设备可以得到该特征集合。
由于特征集合中的特征仅可以表示这各类数据的自相关结构，并不能体现所有检测数据的特征，因此，该计算机设备还需要通过执行下述步骤2013来获取检测数据的特征。
步骤2013、该计算机设备将特征集合输入深交叉网络，通过该深交叉神 经网络对该特征集合内的各个时序特征进行交叉处理，输出该检测数据的特征。
DCN(deep&cross network，深交叉网络)由交叉网络和一个深度网络，该计算设备可以将特征集合分别输入至交叉网络和深度网络，通过交叉网络可以将特征集合内的多条时序特征进行交叉，输出交叉后的特征，通过深度网络提取特征集合内所有特征的普通特征。最后，将DCN输出的交叉后的特征与深度网络提取的普通特征相结合，得到该检测数据的特征。
需要说明的是，本申请实施例对步骤2011-203的执行顺序不做具体限定。例如，该计算设备可以先执行步骤2011，再执行步骤2012和2013；也可以先执行步骤2012和2013，再执行步骤2011；还可以在执行步骤2011的同时执行步骤2012和2013。
步骤2014、该计算机设备将该属性的特征以及该检测数据的特征输入深度神经网络，通过该深度神经网络提取该检测数据和该检测数据对应的属性的混合特征，输出该检测特征。
该计算机设备可以先将该属性的特征以及检测数据的特征进行拼接，得到一个拼接后的特征，然后，再将并接后的特征输入至该深度神经中。
在一种可能的实现方式，采用连接数组的函数concat[]，将第j次检测时所对应属性的特征π j和第j次检测时的检测数据的特征τ j拼接为χ j，则χ j＝concat[τ j,τ j]，然后，将χ j作为该深度神经网络的输入，该深度神经网中的每个节点可以根据如下第二公式对χ j中的数据进行计算，从而使得该深度神经网络可以输出第j个检测时间上的检测特征ξ j，该第二公式可以表示为：
Figure PCTCN2020086007-appb-000004
其中，W x为第一权值矩阵，
Figure PCTCN2020086007-appb-000005
是W x的转置矩阵；b x第一偏差参数。
由于W x每个权值用于表示χ j中的每个元素的重要程度，通过
Figure PCTCN2020086007-appb-000006
可以实现对χ j中的每个元素进行加权处理，进而可以集成χ j中的元素，ReLU(rectified linear unit线性整流)函数可以较好的挖掘数据间相关特征，因此，ReLU函数表达能力比较强，使用ReLU函数对
Figure PCTCN2020086007-appb-000007
进行处理，处理后结果ξ j可以表达χ j所具有的特征，因此，ξ j可以作为第j次检测的检测特征。
当将每个检测时间所对应的属性的特征以及检测数据的特征进行拼接处理，该计算机设备再通过该深度神经网络可以提取每个检测时间上的检测特征， 为了便于描述，将每个检测时间上的检测特征称为子检测特征，因此，深度神经网络最终输出的检测特征包括至少一个子检测特征，该计算机设备可以按照时序将该至少一个子检测特征存储在该检测特征内，得到检测特征ξ＝[ξ 1,ξ 2,...ξ J]，其中，ξ j为第j个子检测特征，也即是第j个检测时间所对应的检测特征，j为正整数，1≤j≤J，J表示目标对象总共的检测次数。
由于检测数据的特征与属性的特征包括有各个时序的各种类型数据的特征，因此该检测特征具有多模态，则该检测特征可以视为多模态特征。
由于检测特征是基于检测数据和检测数据对应的属性获得的，相对于仅基于检测数据获得的目标对象的特征，本申请实施例中的检测特征更能体现目标对象检测过程中的特征；并且，检测数据为实际的检测到的数据，可以作为客观依据，使得获取的检测特征具有可解释性，检测数据对应的属性为主观判断的结果，因此，基于属性和检测数据，所获取的检测特征的精度较高。
为了便于理解本步骤2011-2014所示的过程，参见图3中的多模态特征提取部分，图3是本申请实施例提供的一种诊断估计模型的示意图，在多模态特征提取部分内，可见，首先将计算机设备医疗编码数据(即是检测数据对应的属性)转换为至mulit-hot向量，将mulit-hot向量输入全连接神经网络，全连接神经网络通过计算就可以输出医疗编码数据的特征(即是属性的特征)，需要说明的是，全连接神经网络通过计算就可以输出医疗编码数据的特征过程为对医疗编码数据进行嵌入的过程。
其次，计算机设备对时序数据(即是检测数据)进行特征提取，得到特征集合，再次，计算机设备将特征集合输入至DCN，DCN输出多时序交叉混合特征(cross multiple time series feature mixture)，也即是检测数据的特征，最后，计算机设备将多时序交叉混合特征与属性特征进行混合，并基于混合后的特征，获取多模态特征(即是检测特征)。
需要说明的是，该计算机设备除了使用深度神经网络，获取检测数据的特征外，还可以采取其他神经网络，来获取检测数据的特征。
202、该计算机设备将该检测特征输入第一神经网络，针对检测特征中各个时序上的检测特征，第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于第一规律特征的第二规律特征，该第一规律特征表示该检测特征 的历史变化规律，该第二规律特征表示该检测特征的未来变化规律。
第一神经网络可以是具有注意力(attention)机制的BiRNN(bidirectional recurrent neural networks，双向回归神经网络)，该BiRNN可以由一个第一子网络和一个第二子网络组成，其中，第一子网络用于获取第一规律特征，第二子网络用于获取第二规律特征。
在一种可能的实现方式中，该计算机设备按照反向时序顺序，将该检测特征输入该第一神经网络的第一子网络，通过该第一子网络对该检测特征进行反向时序计算，得到该第一规律特征；按照正向时序顺序，将该检测特征输入该第一神经网络的第二子网络，通过该第一子网络对该检测特征进行正向时序计算，得到该第二规律特征。
由于检测特征内的至少一个子检测特征是按照时间顺序排序的，则该计算机设备可以按照正向时序的方式，将检测特征输入第一子网络，该计算机设备可以按照反向时序的方式，将检测特征输入至第二子网络内。
在一种可能的实现方式中，该计算机设备按照从后往前的顺序，将检测特征ξ＝[ξ 1,ξ 2,...ξ J]中的每个子检测特征依次输入第一子网络内的输入层的节点，具体地，该计算机设备将ξ J输入第一子网络的输入层的第一个节点，将ξ J-1输入第一子网络的输入层的第二个节点，……(以此类推)。当该计算机设备将该检测特征输入至第一子网络后，该第一子网络可以基于第一子网络内预设的计算规则，对检测特征内的各个子特征进行计算，最终，该第一子网络可以输出该第一规律特征
Figure PCTCN2020086007-appb-000008
其中，b用于表示反向，本申请实施例对该第一子网络内预设的计算规则不做具体限定。
在一种可能的实现方式中，该计算机设备按照从前往后的顺序，将检测特征ξ＝[ξ 1,ξ 2,...ξ J]中的每个子检测特征依次输入第二子网络内的输入层的节点输入，具体地，该计算机设备将ξ 1输入第一子网络的输入层的第一个节点，将ξ 2输入第一子网络的输入层的第二个节点，……(以此类推)。当该计算机设备将该检测特征输入至第二子网络后，该第二子网络可以基于第二子网络内预设的计算规则，对检测特征内的各个子特征进行计算，最终，该第二子网络可以输出该第二规律特征
Figure PCTCN2020086007-appb-000009
其中，f代表正向，本申请实施例对该第二子网络内预设的计算规则不做具体限定。
203、该计算机设备基于该第一规律特征和第二规律特征，确定该目标对象的全局特征。
由于检测特征的历史变化规律是由第一规律特征来表示，检测特征的未来变化规律是由第二规律特征来表示的，那么第一规律特征和第二规律特征中的任一规律特征均不能表示检测特征的全局变化规律，为了得到更为精确的目标对象的全局特征，该计算机设备可以先基于第一规律特征以及第二规律特征，获取表示检测特征的全局变化规律的规律特征，再根据这个规律特征，获取全局特征。
在一种可能的实现方式中，本步骤203可以通过下述步骤2031-2033所示的过程来实现。
步骤2031、该计算机设备将该第一规律特征和第二规律特征进行拼接，得到第三规律特征。
当第一子网络输出该第一规律特征
Figure PCTCN2020086007-appb-000010
且第二子网络输出该第二规律特征
Figure PCTCN2020086007-appb-000011
后，该计算机设备可以通过第一神经网络对第一规律特征进行拼接，得到第三规律特征h＝[h 1,...,h J]，其中，
Figure PCTCN2020086007-appb-000012
j为正整数，1≤j≤J，J表示目标对象总共的检测次数。
步骤2032、该计算机设备对该第三规律特征进行加权处理，得到第四规律特征，该第四规律特征用于表示该检测特征的全局变化规律。
该计算设备可以通过第一神经网络中的注意力机制，对第三规律特征进行加权处理。在一种可能的实现方式中，本步骤2032可以通过步骤11-13所示的过程来实现。
步骤11、该计算机设备基于第一注意力机制以及该第三规律特征，进行权值学习，得到至少一个第一权值，该第一权值用于表示一个检测数据与该一个检测数据对应的属性的重要程度。
其中，该第一注意力机制为该第一神经网络内的任一注意力机制，该计算机设备可以基于该第一注意力机制中的权值学习策略，进行权值学习，该权值学习策略可以是基于位置的注意力权值学习策略，那么基于位置的注意力权值学习策略可以表示为：
Figure PCTCN2020086007-appb-000013
其中，W τ为第二权值向量，W τ T为W τ的转置矩阵，b τ为第二偏差参数，
Figure PCTCN2020086007-appb-000014
为第j次检测时对应的第一权值。
那么，该计算设备可以基于第三规律特征h＝[h 1,...,h J]中的各个规律特征以及上述位置的注意力权值学习策略，进行权值学习，可以得到J个第一权值，该J个第一权值也即是至少一个第一权值。
需要说明的是，该第一注意力机制中的权值学习策略还可以是其他注意力权值学习策，本申请实施例对该第一注意力机制中的权值学习策略不做具体限定。
步骤12、该计算机设备对该至少一个第一权值进行归一化处理，得到至少一个第二权值。
由于步骤11第一权值为通过数学计算得到的值，那么，该至少一个第一权值的可能过大或者过小，为了方便计算，可以对该至少一个第一权值进行归一化处理，使得处理后得到的各个第二权值的大小适中，那么，当该至少一个第一权值的过大时，可以将该至少一个第一权值成比例的缩小，当该至少一个第一权值的过小时，可以将该至少一个第一权值成比例的放大，以实现对该至少一个第一权值的归一化处理。
由于每个第二权值仅是对一个第一权值归一化的结果，因此，第二权值和第一权值的作用一样，均用于指示一个检测数据与该一个检测数据对应的属性的重要程度。
步骤13、该计算机设备基于至少一个第二权值，对该第三规律特征进行加权处理，得到该第四规律特征。
该计算机设备将该至少一个第二权值
Figure PCTCN2020086007-appb-000015
代入第三公式，将第三公式的输出作为该第四规律特征，以实现对第三规律特征加权处理，该第三公式可以表示为
Figure PCTCN2020086007-appb-000016
其中，c为第四规律特征，
Figure PCTCN2020086007-appb-000017
为与第j次检测时对应的第二权值，J该目标对象总共的检测次数。
由于一个第二权值用于表示一个检测数据与该一个检测数据对应的属性的重要程度，通过至少一个第二权值对该第三规律特征进行加权处理，第三规律特征更加集中的表达，因此，第四规律特征可以表示检测特征的全局变化规律。
由于将对第一规律特征和第二规律特征进行了加权处理，使得第一规律特 征和第二规律特征可以由第四规律特征集成表示，因此第四规律特征既能表示第一规律特征所表示的历史变化规律，也能表示第一规律特征所表示的未来变化规律，因此，第四特征规律可以表示检测特征的全局变化规律。
步骤2033、该计算机设备基于该第三规律特征和该第四规律特征，确定该全局特征。
由于相邻检测数据之间的相关性最高以及相邻的属性之间的相关性最高，为了进一步预测目标对象下一次检测时的属性，该计算机设备可以将该第三规律特征中最后一次检测时间对应的规律特征和该第四规律特征代入至第四公式中，将第四公式的输出作为该全局特征，其中，第四公式可以表示为：
Figure PCTCN2020086007-appb-000018
Figure PCTCN2020086007-appb-000019
其中，
Figure PCTCN2020086007-appb-000020
为该全局特征，h J为第三规律特征中目标对象最后一次检测时间对应的规律特征，[h J,c]为h J和c拼接后的向量，W d为第三权值矩阵，
Figure PCTCN2020086007-appb-000021
为W d的转置矩阵，b d第三偏差参数，ReLU()是指线性整流函数。
由于第四规律特征表示检测特征的全局变化规律，第一规律特征可以表示检测特征的历史变化规律，第二规律特征可以表示检测特征的未来变化规律，因此通过对这个三个规律特征进行加权处理，得到的结果可以表示目标对象的全局特征。
204、该计算机设备将该全局特征输入第二神经网络，第二神经网络从全局特征提取并输出目标对象的至少一个局部特征。
该第二神经网络可以是HMCN(hierarchical multi-label classification networks，分层多标签分类网络)，而且第二神经网络可以从输入的全局特征中进行局部特征的提取。由于全局特征不能表示目标对象的细节，因此，可以通过第二神经网络的提取目标对象的细节，具体地，第二神经网络可以逐级提取该目标对象的细节，使得最终提取的细节可以满足用于属性预测的需求。
该第二神经网络的每一层可以输出一个该局部特征。当该计算设备将该全局特征输入该第二神经网络后，该全局特征可以从输入层一直输入至第二神经 网络的输出层，该第二神经网络可以对全局特征进行逐层计算，该第二神经网络的第一目标层在计算时，可以基于第二目标层的输出数据，计算第一目标层的层级特征以及计算该目标对象在第一目标层的局部特征，其中，第一目标层为该第二神经网络的任一层，第二目标层为该第二神经网络中该第一目标层的上一层，该层级特征用于表示该全局特征在该第二神经网络的网络层中的状态，该第一目标层的层级特征由该全局特征以及该第二目标层的层级特征来决定来确定。
当该第二神经网络的第二目标层生成第二目标层的层级特征以及该目标对象在第二目标层的局部特征后，该第二目标层可以向第一目标层输出第二目标层的层级特征以及该全局特征(第二目标层的输出数据)，以便第一目标层可以接收第二目标层的层级特征以及该全局特征。那么，随着第二神经网络各个网络层输出的全局特征使得该全局特征可以输入至该第二神经网络的各个网络层。
由于第一目标层的层级特征由该全局特征以及该第二目标层的层级特征来决定来确定，那么，当第一目标层接收到第二目标层的层级特征以及该全局特征后，该第一目标层可以基于第二目标层的层级特征以及和该全局特征，计算第一目标层的层级特征。在一种可能的实现方式中，该第二神经网络第i层的层级特征
Figure PCTCN2020086007-appb-000022
可以表示为：
Figure PCTCN2020086007-appb-000023
其中，G代表全局，
Figure PCTCN2020086007-appb-000024
为第四权重矩阵，
Figure PCTCN2020086007-appb-000025
为第i-1层的层级特征，
Figure PCTCN2020086007-appb-000026
表示第i层的层级特征，b G为第四偏差参数，第i层可以视为第一目标层。
该第一目标层上的节点可以基于第一目标层的层级特征以及该全局特征，获取该目标对象在第一目标层上的局部特征。在一种可能的实现方式中，该目标对象在第i层的局部特征
Figure PCTCN2020086007-appb-000027
可以表示为：
Figure PCTCN2020086007-appb-000028
其中，L代表网络层，
Figure PCTCN2020086007-appb-000029
为第五权重矩阵，b T为第五偏差参数。
由于该第二神经网络的每一层都是基于上一层的层级特征以及全局特征来进计算的，那么目标对象在第二神经网络的每一层上的局部特征都受到上一层的局部特征的影响，由于每一层的层级表达由本层的层级特征决定，所以，该第二神经网络中任一个网络层生成的局部特征可以作为下一个网络层所生成的局部特征的父级，因此，第二神经网络可以实现逐级提取该目标对象的细 节。
205、该计算机设备基于该目标对象的至少一个局部特征，对该目标对象的属性进行预测。
由于一个局部特征可以代表目标对象不同级别的细节，考虑到细节较多，可以对这些进行集中处理，获取更加细节的局部特征，然后在根据这个更加细节的局部特征，进行属性预测。
在一种可能的实现方式中，本步骤205可以通过下述步骤2051-2052所示的过程来实现。
步骤2051、该计算设备对该目标对象的至少一个局部特征进行加权处理，得到目标局部特征。
该目标局部特征也即是更加细节的局部特征。计算设备可以通过第二神经网络中的注意力机制，对目标对象的局部特征进行加权处理。在一种可能的实现方式中，本步骤2051可以通过步骤21-22所示的过程来实现。
步骤21、该计算机设备基于第二注意力机制以及该至少一个局部特征，进行权值学习，得到至少一个第三权值，一个第三权值用于表示一个局部特征的重要程度。
其中，该第二注意力机制为该第二神经网络内的任一注意力机制，机算计设备基于第二注意力机制以及该至少一个局部特征，进行权值学习，可以该第二注意力机制中的权值学习策略，学习权值，该第二注意力机制中的权值学习策略可以表示为：
Figure PCTCN2020086007-appb-000030
其中，
Figure PCTCN2020086007-appb-000031
为第i个第三权值，W α为第六权值矩阵，b α为第六偏差参数，
Figure PCTCN2020086007-appb-000032
为第j个检测时间对应的参量权值，M表示局部特征的个数。
需要说明的是，该第二注意力机制中的权值学习策略还可以是其他注意力权值学习策，本申请实施例对该第二注意力机制中的权值学习策略不做具体限定。
步骤22、该计算机设备基于该至少一个第三权值，对该目标对象的至少 一个局部特征进行加权处理，得到该目标局部特征。
该计算机设备将至少一个第三权值和该目标对象的至少一个局部特征代入第五公式，将第五公式的输出作为目标局部特征，以实现对目标对象的至少一个局部特征的加权处理，该第五公式可以表示为：
Figure PCTCN2020086007-appb-000033
其中，A G为当前时序对应的检测数据所对应的属性，N为第二神经网络的层数。
由于一个第三权值一个局部特征的重要程度，通过至少一个第三权值对至少一个局部特征进行加权处理，使得到的目标局部特征更加细节化。
步骤2052、基于该目标局部特征，对该目标对象的属性进行预测。
该计算机设备可以将该目标局部特征，代入第六公式，以对目标对象的属性进行预测，该第六公式用于预测该目标对象的属性，该第六公式可以表示为：
Figure PCTCN2020086007-appb-000034
其中，
Figure PCTCN2020086007-appb-000035
为预测的目标对象在第J+1次检测数据所对应的属性，W G为第七权值矩阵，b G为第七偏差参数。
在一些实施例中，第二神经网络可以根据第二神经网络中的全局损失以及局部损失，确定是否输出当前预测的属性。
在一些可能的实现方式中，当将该全局特征输入至该第二神经网络后，若该第二神经网络中的全局损失以及局部损失满足预设条件，则该第二神经网络输出当前预测的属性，否则，该第二神经网络调整该第二神经网络内的权重矩阵，直至第二神经网络中的全局损失以及局部损失满足预设条件，该局部损失为该第二神经网络在每一层期望的输出数据与实际的输出数据的差值，该全局损失为该第二神经网络期望的最终输出数据与实际的最终输出数据的差值。
当第二神经网络的任一层计算结束后，该第二神经网络可以预测该任一层在下一个检测时间时的局部特征(简称预测局部特征)，那么，第i层的预测局部特征
Figure PCTCN2020086007-appb-000036
可以表示为
Figure PCTCN2020086007-appb-000037
其中，
Figure PCTCN2020086007-appb-000038
为第i层的第八权值矩阵，b L第八偏差参数。
第二神经网络可以预测至少一个目标对象的属性，当基于第二神经网络预测至少一个目标对象的属性时，基于该任一层的预测局部特征，该第二神经网络可以采用交叉熵策略计算该任一层的局部损失，那么，第i的局部损失L li可以表示为：
Figure PCTCN2020086007-appb-000039
其中，Q为目标对象的数目，
Figure PCTCN2020086007-appb-000040
为基于第Q个目标对象的全局特征，第i层实际的输出数据，
Figure PCTCN2020086007-appb-000041
为基于第Q个目标对象的全局特征，预测的第i层的输出数据。
当基于第二神经网络预测至少一个目标对象的属性时，若该第二神经网络每一层都计算结束后，该第二神经网络可以计算至少一个目标对象下次检测时的属性，然后，该第二神经网络可以预测的至少一个目标对象下次检测时的属性，采用交叉熵策略计算全局损失L G，其中，L G可以表示为：
Figure PCTCN2020086007-appb-000042
其中，
Figure PCTCN2020086007-appb-000043
为实际输出的第Q个目标对象下一次检测时的属性，
Figure PCTCN2020086007-appb-000044
为预测的第Q个目标对象下一次检测时的属性。
该预设条件可以表示为Loss＝L G+γ(L l1+L l2...L lp)，其中，P为大于2的整数，例如，p＝3，Loss为预设的收敛值，γ为预定义的参数，用于平衡全局损失以及局部损失，当该计算机设备将Loss、γ、L l1、L l2、...、L lp输入至上述预设条件的公式后，若上述公式成立，则该第二神经网络中的全局损失以及局部损失满足预设条件，否则该第二神经网络中的全局损失以及局部损失不满足预设条件。
当第二神经网络中的全局损失以及局部损失满足预设条件时，说明第二神经网络的每一层的生成的局部特征与期望的局部特征之间的差值达到预设精度，从而可以保证该第二神经网络的每一层的局部特征的精度较高，进而可以提高预测的属性精度。
需要说明的是，由于第二神经网络都是基于数值进行计算的，而目标对象的属性内的每个状态实际上是由状态标识表示的，所以，该计算机设备还需将 第二神经网络实际输出的数据转换为由状态标识组成的属性，第二神经网络实际输出的数据可以包括至少一个概率值，每个概率值对应目标对象的属性内的一个状态，当任一概率值大于目标值时，说明该目标对象具有该任一概率对应的目标状态，则该计算机设备将第一状态标识存储在目标状态在属性内的位置；当任一概率值小于或等于该目标值时，说明该目标对象不具有该任一概率对应的目标状态，则该计算机设备将第二状态标识存储在目标状态在属性内的位置。那么，通过对每个概率值进行判断，可以得到属性的实际表达方式。本申请实施例对该目标值不做具体限定。
为了进一步表明步骤203-204所示的过程，参见图3中神经层次多标签建模部分，从该部分可知，注意力循环网络(attention recurrent networks)的输出数据(即是全局特征)至神经层次多标签建模部分，注意力循环网络相当于第一神经网络。具体地，该计算设备将全局特征输入至神经层次多标签建模部分中的第二神经网络的每一层，第一层根据全局特征生成第一层的层级特征
Figure PCTCN2020086007-appb-000045
再根据
Figure PCTCN2020086007-appb-000046
生成第一层的局部特征
Figure PCTCN2020086007-appb-000047
基于
Figure PCTCN2020086007-appb-000048
可以进行数据预测，得到第一层预测的输出数据
Figure PCTCN2020086007-appb-000049
计算机设备在计算第一层的局部损失L l1，第一层将
Figure PCTCN2020086007-appb-000050
输出给第二层，使得第二层可以进行类似第一层的计算过程，最终第二神经网络的所有的层都可以得到一个
Figure PCTCN2020086007-appb-000051
计算机设备将M个
Figure PCTCN2020086007-appb-000052
输出至注意力集合(attentional ensembel)，在注意力集合中，基于第二注意机制，生成预测的输出数据
Figure PCTCN2020086007-appb-000053
进而根据预测的输出数据
Figure PCTCN2020086007-appb-000054
生成全局损失L G。那么，当全局损失L G和局部损失L i都满足预设条件时，该第二神经网络可以输出
Figure PCTCN2020086007-appb-000055
本申请实施例提供的方法，通过基于表示检测特征历史和未来的变化规律的规律特征，确定出目标对象的全局特征，并对全局特征进行细化，以得到目标对象的至少一个局部特征，那么，细化后的局部特征更能体现目标对象的特征，进而根据局部特征目标对象预测属性，因此，可以提高预测的属性的精度，当目标对象的属性为预测的诊断结果时，则可以提高预测的诊断结果的精度。 并且，由于检测特征是基于检测数据和检测数据对应的属性获得的，相对于仅基于检测数据获得的目标对象的特征，本申请实施例中的检测特征更能体现目标对象检测过程中的特征。并且，检测数据为实际的检测到的数据，可以作为客观依据，使得获取的检测特征具有可解释性，检测数据对应的属性为主观判断的结果，因此，基于属性和检测数据，所获取的检测特征的精度较高。并且，当第二神经网络中的全局损失以及局部损失满足预设条件时，说明第二神经网络的每一层的生成的局部特征均达到了期望值，从而可以保证该第二神经网络的输出层输出的局部特征的精度较高。
图4是本申请实施例提供的一种基于机器学习的目标对象属性预测装置的结构示意图，该装置包括：
获取模块401，用于根据目标对象的检测数据以及所述检测数据对应的属性，确定所述目标对象的检测特征；
计算模块402，用于将所述检测特征输入第一神经网络，；针对所述检测特征中各个时序上的检测特征，所述第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于所述第一规律特征的第二规律特征，其中，所述第一规律特征表示所述检测特征的历史变化规律，所述第二规律特征表示所述检测特征的未来变化规律；
所述获取模块401，还用于基于所述第一规律特征和第二规律特征，确定所述目标对象的全局特征；
提取模块403，用于将所述全局特征输入第二神经网络；所述第二神经网络从所述全局特征提取并输出所述目标对象的至少一个局部特征；
预测模块404，用于基于所述目标对象的至少一个局部特征，对所述目标对象的属性进行预测。
可选地，所述获取模块401用于：
将所述检测数据对应的属性输入全连接神经网络，通过所述全连接神经网络删除对所述检测数据对应的属性中的非必要因素，得到所述检测数据对应的属性的特征；
将所述检测数据输入时间序列分析工具，通过时间序列分析工具提取所述检测数据中每一类数据在各个时序的特征，输出特征集合；
将特征集合输入深交叉神经网络，通过深交叉神经网络对所述特征集合内的各个时序特征进行交叉处理，得到所述检测数据的特征；
将所述属性的特征以及所述检测数据的特征输入深度神经网络，通过所述深度神经网络提取所述检测数据和所述检测数据对应的属性的混合特征，输出所述检测特征。
可选地，所述获取模块401用于：
将所述第一规律特征和第二规律特征进行拼接，得到第三规律特征；
对所述第三规律特征进行加权处理，得到第四规律特征，所述第四规律特征用于表示所述检测特征的全局变化规律；
基于所述第三规律特征和所述第四规律特征，确定所述全局特征。
可选地，所述获取模块401具体用于：
基于第一注意力机制以及所述第三规律特征，进行权值学习，得到至少一个第一权值，一个所述第一权值用于表示一个检测数据与所述一个检测数据对应的属性的重要程度；
对所述至少一个第一权值进行归一化处理，得到至少一个第二权值；
基于所述至少一个第二权值，对所述第三规律特征进行加权处理，得到所述第四规律特征。
可选地，所述预测模块404包括：
处理单元，用于对所述目标对象的至少一个局部特征进行加权处理，得到目标局部特征；
预测单元，用于基于所述目标局部特征，对所述目标对象的属性进行预测。
可选地，所述处理单元用于；
基于第二注意力机制以及所述至少一个局部特征，进行权值学习，得到至少一个第三权值，一个第三权值用于表示一个局部特征的重要程度；
基于所述至少一个第三权值，对所述至少一个局部特征进行加权处理，得到所述目标局部特征。
可选地，所述第二神经网络的每一层输出一个所述局部特征。
可选地，所述装置还包括输出模块，用于当将所述全局特征输入至所述第二神经网络后，若所述第二神经网络中的全局损失以及局部损失满足预设条件， 则所述第二神经网络输出当前预测的属性，所述局部损失为所述第二神经网络在每一层期望的输出数据与实际的输出数据的差值，所述全局损失为所述第二神经网络期望的最终输出数据与实际的最终数据的差值。
可选地，该装置还包括生成模块，用于基于所述第二神经网络中第一目标层的层级特征以及第二目标层生成的局部特征，生成所述第一目标层输出的局部特征，第一目标层的所述层级特征用于表示所述全局特征在第一目标层的中的状态，所述第二目标层为所述第二神经网络中所述第一目标层的上一层。
可选地，所述第一目标层的层级特征由所述全局特征以及所述第二目标层的层级特征来决定。
图5是本申请实施例提供的一种计算机设备的结构示意图，该计算机设备500可因配置或性能不同而产生比较大的差异，可以包括一个或一个以上CPU(central processing units，处理器)501和一个或一个以上的存储器502，其中，所述存储器502中存储有至少一条指令，所述至少一条指令由所述处理器501加载并执行以实现上述各个方法实施例提供的基于机器学习的目标对象属性预测方法。当然，该计算机设备500还可以具有有线或无线网络接口、键盘以及输入输出接口等部件，以便进行输入输出，该计算机设备500还可以包括其他用于实现设备功能的部件，在此不做赘述。
在示例性实施例中，还提供了一种计算机可读存储介质，例如包括指令的存储器，上述指令可由终端中的处理器执行以完成上述实施例中基于机器学习的目标对象属性预测方法。例如，所述计算机可读存储介质可以是ROM(read-only memory，只读存储器)、RAM(random access memory，随机存取存储器)、CD-ROM(compact disc read-only memory，只读光盘)、磁带、软盘和光数据存储设备等。
在示例性实施例中，还提供了一种包括指令的计算机程序产品，当其在计算机上运行时，使得所述计算机执行上述实施例中基于机器学习的目标对象属性预测方法。
上述所有可选技术方案，可以采用任意结合形成本公开的可选实施例，在此不再一一赘述。
需要说明的是：上述实施例提供的基于机器学习的目标对象属性预测装置 在预测属性时，仅以上述各功能模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能模块完成，即将装置的内部结构划分成不同的功能模块，以完成以上描述的全部或者部分功能。另外，上述实施例提供的基于机器学习的目标对象属性预测装置与基于机器学习的目标对象属性预测方法实施例属于同一构思，其具体实现过程详见方法实施例，这里不再赘述。
本领域普通技术人员可以理解实现上述实施例的全部或部分步骤可以通过硬件来完成，也可以通过程序来指令相关的硬件完成，所述的程序可以存储于一种计算机可读存储介质中，上述提到的存储介质可以是只读存储器，磁盘或光盘等。
以上所述仅为本申请的较佳实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。

Claims (16)

  1. 一种基于机器学习的目标对象属性预测方法，由计算机设备执行，所述方法包括：
    根据目标对象的检测数据以及所述检测数据对应的属性，确定所述目标对象的检测特征；
    将所述检测特征输入第一神经网络；
    针对所述检测特征中各个时序上的检测特征，所述第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于所述第一规律特征的第二规律特征，其中，所述第一规律特征表示所述检测特征的历史变化规律，所述第二规律特征表示所述检测特征的未来变化规律；
    基于所述第一规律特征和第二规律特征，确定所述目标对象的全局特征；
    将所述全局特征输入第二神经网络；
    所述第二神经网络从所述全局特征提取并输出所述目标对象的至少一个局部特征；
    基于所述目标对象的至少一个局部特征，对所述目标对象的属性进行预测。
  2. 根据权利要求1所述的方法，所述根据目标对象的检测数据以及所述检测数据对应的属性，确定所述目标对象的检测特征，包括：
    将所述检测数据对应的属性输入全连接神经网络，通过所述全连接神经网络筛选出所述属性中的目标状态，对所述目标状态进行加权处理，输出所述属性的特征；
    将所述检测数据输入时间序列分析工具，通过时间序列分析工具提取所述检测数据中每一类数据在各个时序的特征，输出特征集合；
    将特征集合输入深交叉神经网络，通过深交叉神经网络对所述特征集合内的各个时序特征进行交叉处理，得到所述检测数据的特征；
    将所述属性的特征以及所述检测数据的特征输入深度神经网络，通过所述深度神经网络提取所述检测数据和所述检测数据对应的属性的混合特征，输出所述检测特征。
  3. 根据权利要求1所述的方法，所述基于所述第一规律特征和第二规律 特征，确定所述目标对象的全局特征，包括：
    将所述第一规律特征和第二规律特征进行拼接，得到第三规律特征；
    对所述第三规律特征进行加权处理，得到第四规律特征，所述第四规律特征用于表示所述检测特征的全局变化规律；
    基于所述第三规律特征和所述第四规律特征，确定所述全局特征。
  4. 根据权利要求3所述的方法，所述对所述第三规律特征进行加权处理，得到第四规律特征，包括：
    基于第一注意力机制以及所述第三规律特征，进行权值学习，得到至少一个第一权值，一个所述第一权值用于表示一个检测数据与所述一个检测数据对应的属性的重要程度；
    对所述至少一个第一权值进行归一化处理，得到至少一个第二权值；
    基于所述至少一个第二权值，对所述第三规律特征进行加权处理，得到所述第四规律特征。
  5. 根据权利要求1所述的方法，所述基于所述目标对象的至少一个局部特征，对所述目标对象的属性进行预测，包括：
    对所述目标对象的至少一个局部特征进行加权处理，得到目标局部特征；
    基于所述目标局部特征，对所述目标对象的属性进行预测。
  6. 根据权利要求5所述的方法，所述对所述目标对象的至少一个局部特征进行加权处理，得到目标局部特征，包括：
    基于第二注意力机制以及所述至少一个局部特征，进行权值学习，得到至少一个第三权值，一个第三权值用于表示一个局部特征的重要程度；
    基于所述至少一个第三权值，对所述至少一个局部特征进行加权处理，得到所述目标局部特征。
  7. 根据权利要求1所述的方法，所述第二神经网络的每一层输出一个所述局部特征。
  8. 根据权利要求1所述的方法，所述方法还包括：
    当将所述全局特征输入至所述第二神经网络后，若所述第二神经网络中的全局损失以及局部损失满足预设条件，则所述第二神经网络输出当前预测的属 性，所述局部损失为所述第二神经网络在每一层期望的输出数据与实际的输出数据的差值，所述全局损失为所述第二神经网络期望的最终输出数据与实际的最终数据的差值。
  9. 根据权利要求1所述的方法，所述方法还包括：
    基于所述第二神经网络中第一目标层的层级特征以及第二目标层生成的局部特征，生成所述第一目标层输出的局部特征，第一目标层的所述层级特征用于表示所述全局特征在第一目标层的中的状态，所述第二目标层为所述第二神经网络中所述第一目标层的上一层。
  10. 根据权利要求9所述的方法，所述第一目标层的层级特征由所述全局特征以及所述第二目标层的层级特征来决定。
  11. 一种基于机器学习的目标对象属性预测装置，所述装置包括：
    获取模块，用于根据目标对象的检测数据以及所述检测数据对应的属性，确定所述目标对象的检测特征；
    计算模块，用于将所述检测特征输入第一神经网络；针对所述检测特征中各个时序上的检测特征，所述第一神经网络通过两种不同的时序计算后输出第一规律特征及不同于所述第一规律特征的第二规律特征，其中，所述第一规律特征表示所述检测特征的历史变化规律，所述第二规律特征表示所述检测特征的未来变化规律；
    所述获取模块，还用于基于所述第一规律特征和第二规律特征，确定所述目标对象的全局特征；
    提取模块，用于将所述全局特征输入第二神经网络；所述第二神经网络从所述全局特征提取并输出所述目标对象的至少一个局部特征；
    预测模块，用于基于所述目标对象的至少一个局部特征，对所述目标对象的属性进行预测。
  12. 根据权利要求11所述的装置，所述获取模块用于：
    将所述检测数据对应的属性输入全连接神经网络，通过所述全连接神经网络删除对所述检测数据对应的属性中的非必要因素，得到所述检测数据对应的属性的特征；
    将所述检测数据输入时间序列分析工具，通过时间序列分析工具提取所述检测数据中每一类数据在各个时序的特征，输出特征集合；
    将特征集合输入深交叉神经网络，通过深交叉神经网络对所述特征集合内的各个时序特征进行交叉处理，得到所述检测数据的特征；
    将所述属性的特征以及所述检测数据的特征输入深度神经网络，通过所述深度神经网络提取所述检测数据和所述检测数据对应的属性的混合特征，输出所述检测特征。
  13. 根据权利要求11所述的装置，所述获取模块用于：
    将所述第一规律特征和第二规律特征进行拼接，得到第三规律特征；
    对所述第三规律特征进行加权处理，得到第四规律特征，所述第四规律特征用于表示所述检测特征的全局变化规律；
    基于所述第三规律特征和所述第四规律特征，确定所述全局特征。
  14. 一种计算机设备，其特征在于，所述计算机设备包括一个或多个处理器和一个或多个存储器，所述一个或多个存储器中存储有至少一条指令，所述指令由所述一个或多个处理器加载并执行以实现如权利要求1至权利要求10任一项所述的基于机器学习的目标对象属性预测方法所执行的操作。
  15. 一种计算机可读存储介质，其特征在于，所述存储介质中存储有至少一条指令，所述指令由处理器加载并执行以实现如权利要求1至权利要求10任一项所述的基于机器学习的目标对象属性预测方法所执行的操作。
  16. 一种包括指令的计算机程序产品，当其在计算机上运行时，使得所述计算机执行权利要求1至10任意一项所述的基于机器学习的目标对象属性预测方法所执行的操作。
PCT/CN2020/086007 2019-05-09 2020-04-22 基于机器学习的目标对象属性预测方法及相关设备 WO2020224433A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
KR1020217025407A KR20210113336A (ko) 2019-05-09 2020-04-22 머신 러닝에 기반한 타깃 객체 속성 예측 방법 및 관련 디바이스
JP2021562367A JP7191443B2 (ja) 2019-05-09 2020-04-22 機械学習に基づくターゲットオブジェクト属性予測方法、関連機器及びコンピュータプログラム
EP20801439.9A EP3968337A4 (en) 2019-05-09 2020-04-22 Target object attribute prediction method based on machine learning and related device
US17/469,270 US20210406687A1 (en) 2019-05-09 2021-09-08 Method for predicting attribute of target object based on machine learning and related device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910386448.4 2019-05-09
CN201910386448.4A CN110111885B (zh) 2019-05-09 2019-05-09 属性预测方法、装置、计算机设备及计算机可读存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/469,270 Continuation US20210406687A1 (en) 2019-05-09 2021-09-08 Method for predicting attribute of target object based on machine learning and related device

Publications (1)

Publication Number Publication Date
WO2020224433A1 true WO2020224433A1 (zh) 2020-11-12

Family

ID=67489143

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/086007 WO2020224433A1 (zh) 2019-05-09 2020-04-22 基于机器学习的目标对象属性预测方法及相关设备

Country Status (6)

Country Link
US (1) US20210406687A1 (zh)
EP (1) EP3968337A4 (zh)
JP (1) JP7191443B2 (zh)
KR (1) KR20210113336A (zh)
CN (1) CN110111885B (zh)
WO (1) WO2020224433A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114925806A (zh) * 2022-03-30 2022-08-19 北京达佳互联信息技术有限公司 信息处理方法、信息处理模型训练方法及装置

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110111885B (zh) * 2019-05-09 2023-09-19 腾讯科技（深圳）有限公司 属性预测方法、装置、计算机设备及计算机可读存储介质
CN112133441B (zh) * 2020-08-21 2024-05-03 广东省人民医院 一种mh术后裂孔状态预测模型的建立方法和终端
CN115994632A (zh) * 2023-03-24 2023-04-21 北京搜狐新动力信息技术有限公司 一种点击率预测方法、装置、设备及可读存储介质
CN116072298B (zh) * 2023-04-06 2023-08-15 之江实验室 一种基于层级标记分布学习的疾病预测系统
CN116824281B (zh) * 2023-08-30 2023-11-14 浙江大学 一种隐私保护的图像分类方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080010024A1 (en) * 2003-09-23 2008-01-10 Prediction Sciences Llp Cellular fibronectin as a diagnostic marker in cardiovascular disease and methods of use thereof
CN106778014A (zh) * 2016-12-29 2017-05-31 浙江大学 一种基于循环神经网络的患病风险预测方法
CN108648829A (zh) * 2018-04-11 2018-10-12 平安科技（深圳）有限公司 疾病预测方法及装置、计算机装置及可读存储介质
CN109698017A (zh) * 2018-12-12 2019-04-30 中电健康云科技有限公司 医疗病历数据生成方法及装置
CN110111885A (zh) * 2019-05-09 2019-08-09 腾讯科技（深圳）有限公司 属性预测方法、装置、计算机设备及计算机可读存储介质

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3302081B2 (ja) * 1993-03-23 2002-07-15 北陸電力株式会社 電力需要量予測装置
JP3767954B2 (ja) * 1996-11-07 2006-04-19 富士通株式会社 需要予測装置
US6171256B1 (en) * 1998-04-30 2001-01-09 Physio-Control Manufacturing Corporation Method and apparatus for detecting a condition associated with acute cardiac ischemia
US7634360B2 (en) * 2003-09-23 2009-12-15 Prediction Sciences, LL Cellular fibronectin as a diagnostic marker in stroke and methods of use thereof
WO2014055718A1 (en) * 2012-10-04 2014-04-10 Aptima, Inc. Clinical support systems and methods
JP6276681B2 (ja) 2014-11-14 2018-02-07 アオイ電子株式会社 試料固定装置および試料分析装置
US10699151B2 (en) * 2016-06-03 2020-06-30 Miovision Technologies Incorporated System and method for performing saliency detection using deep active contours
CN106777874A (zh) * 2016-11-18 2017-05-31 中国科学院自动化研究所 基于循环神经网络构建预测模型的方法
CN110532571B (zh) * 2017-09-12 2022-11-18 腾讯科技（深圳）有限公司 文本处理方法及相关装置
US11101029B2 (en) * 2018-07-17 2021-08-24 Petuum Inc. Systems and methods for predicting medications to prescribe to a patient based on machine learning
CN109599177B (zh) * 2018-11-27 2023-04-11 华侨大学 一种基于病历的深度学习预测医疗轨迹的方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080010024A1 (en) * 2003-09-23 2008-01-10 Prediction Sciences Llp Cellular fibronectin as a diagnostic marker in cardiovascular disease and methods of use thereof
CN106778014A (zh) * 2016-12-29 2017-05-31 浙江大学 一种基于循环神经网络的患病风险预测方法
CN108648829A (zh) * 2018-04-11 2018-10-12 平安科技（深圳）有限公司 疾病预测方法及装置、计算机装置及可读存储介质
CN109698017A (zh) * 2018-12-12 2019-04-30 中电健康云科技有限公司 医疗病历数据生成方法及装置
CN110111885A (zh) * 2019-05-09 2019-08-09 腾讯科技（深圳）有限公司 属性预测方法、装置、计算机设备及计算机可读存储介质

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114925806A (zh) * 2022-03-30 2022-08-19 北京达佳互联信息技术有限公司 信息处理方法、信息处理模型训练方法及装置

Also Published As

Publication number Publication date
EP3968337A1 (en) 2022-03-16
CN110111885B (zh) 2023-09-19
JP2022530868A (ja) 2022-07-04
EP3968337A4 (en) 2022-06-29
CN110111885A (zh) 2019-08-09
JP7191443B2 (ja) 2022-12-19
KR20210113336A (ko) 2021-09-15
US20210406687A1 (en) 2021-12-30

Similar Documents

Publication Publication Date Title
WO2020224433A1 (zh) 基于机器学习的目标对象属性预测方法及相关设备
CN109559300A (zh) 图像处理方法、电子设备及计算机可读存储介质
CN106295186A (zh) 一种基于智能推理的辅助疾病诊断的方法与系统
CN116994709B (zh) 一种个性化的饮食与运动推荐方法、系统及电子设备
Sharma et al. Application of fuzzy logic and genetic algorithm in heart disease risk level prediction
CN110363090A (zh) 智能心脏疾病检测方法、装置及计算机可读存储介质
CN113707323B (zh) 基于机器学习的疾病预测方法、装置、设备及介质
CN113673244B (zh) 医疗文本处理方法、装置、计算机设备和存储介质
Xu et al. Intelligent emotion detection method based on deep learning in medical and health data
CN113722474A (zh) 文本分类方法、装置、设备及存储介质
CN109935337A (zh) 一种基于相似性度量的病案查找方法及系统
CN116434976A (zh) 一种融合多源知识图谱的药物重定位方法和系统
CN116994695A (zh) 报告生成模型的训练方法、装置、设备及存储介质
El-Magd et al. An interpretable deep learning based approach for chronic obstructive pulmonary disease using explainable artificial intelligence
CN112216379A (zh) 一种基于智能联合学习的疾病诊断系统
CN114191665A (zh) 机械通气过程中人机异步现象的分类方法和分类装置
CN116612843B (zh) 一种心理测评掩饰性行为识别方法及系统
Hasan et al. DEVELOPMENT OF HEART ATTACK PREDICTION MODEL BASED ON ENSEMBLE LEARNING.
CN109361539A (zh) 一种基于发育网络的性能告警方法和装置
Alazwari et al. Improved Coyote Optimization Algorithm and Deep Learning Driven Activity Recognition in Healthcare
Ritul et al. Identifying Alzheimer’s Dementia Patients Using MRI Clinical Data
Singh et al. Malaria parasite recognition in thin blood smear images using squeeze and excitation networks
WO2024216763A1 (zh) 图像处理方法、装置、电子设备及存储介质
Haritha et al. Parkinson Disease Identification Using Hand Written Images
Li et al. Artificial Intelligence-Based Assessment of Physical Education and Training Effectiveness

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: 20801439

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 20217025407

Country of ref document: KR

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 2021562367

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2020801439

Country of ref document: EP

Effective date: 20211209