WO2023221739A1 - 一种基于因果校验数据生成的全科多疾病预测系统 - Google Patents

一种基于因果校验数据生成的全科多疾病预测系统 Download PDF

Info

Publication number
WO2023221739A1
WO2023221739A1 PCT/CN2023/089993 CN2023089993W WO2023221739A1 WO 2023221739 A1 WO2023221739 A1 WO 2023221739A1 CN 2023089993 W CN2023089993 W CN 2023089993W WO 2023221739 A1 WO2023221739 A1 WO 2023221739A1
Authority
WO
WIPO (PCT)
Prior art keywords
general
event
causal
disease
data
Prior art date
Application number
PCT/CN2023/089993
Other languages
English (en)
French (fr)
Inventor
李劲松
王丰
张航
池胜强
田雨
周天舒
Original Assignee
之江实验室
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 之江实验室 filed Critical 之江实验室
Publication of WO2023221739A1 publication Critical patent/WO2023221739A1/zh

Links

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
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • 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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • the invention belongs to the field of medical and health information technology, and specifically relates to a multi-disease prediction system for general practitioners based on causality verification data generation.
  • Generating simulation data through data generation methods is a common method to solve the problem of insufficient training samples for machine learning models.
  • Existing data generation methods are mainly based on generative adversarial networks.
  • Generative adversarial networks perform well when generating image data.
  • data and complex structures there are many types of data and complex structures.
  • structured medical data which contains multiple types of patient-centered feature data, is heterogeneous in time and space, and the data distribution is relatively complex.
  • Traditional generative adversarial networks are difficult to handle structured data with complex distribution.
  • training with few sample data is prone to problems such as training instability, gradient disappearance, and model collapse.
  • the calculation method of causal effect size based on propensity score is the most common method to measure the causal relationship between variables. Most of the existing propensity score calculation methods are based on logistic regression. However, due to the wide variety of data, complex structures, and often linear inseparability in general practice scenarios, propensity score calculation methods based on logistic regression do not perform well in general practice scenarios. .
  • the present invention proposes a propensity score calculation method based on the general propensity score network from the perspective of causality, and on this basis, proposes a medical treatment method based on a generative adversarial network based on causal verification.
  • the data generation method solves the problem that the data generated by the generative adversarial network based on correlation analysis is difficult to understand, and builds a comprehensive set of
  • the multi-disease prediction system solves the problem of poor model performance and low robustness in general practice scenarios due to the small number of training samples.
  • a general practice multi-disease prediction system based on causality verification data generation including:
  • Disease statistics module used to count the number of samples of various general diseases, and obtain a small sample of general diseases based on the sample ratio of various general diseases;
  • the first event variable set is formed according to the general patient characteristic variable set, the disease label variable set forms the second event variable set, and any first event variable and any second event variable form an event pairing;
  • Data generation module For a few samples of general diseases, a data generation model is built based on the generative adversarial network of causal verification, and the trained data generation model is used to generate simulation data;
  • the data generation model includes a generator and a discriminator, and the generator and the discriminator are trained iteratively and alternately;
  • the training process of the generator includes: generating random noise for each small-sample general disease, inputting the random noise and the corresponding disease label into the generator to obtain the generated sample; calculating the causal effect value of all event pairs of the generated sample; The sample and the corresponding disease label are input into the discriminator to obtain the discrimination result; the total loss of the generator includes the discriminator adversarial loss, causal loss and regular term loss; the causal loss is the generated sample corrected for the frequency of general diseases in a small sample The KL divergence loss between the causal effect values of all event pairs and the causal effect values of all event pairs of the original data;
  • the training process of the discriminator includes: randomly extracting positive samples from the original data, and extracting the same number of negative samples with different disease labels; generating the same number of random noises, and using the generator to obtain generated samples; combining the positive samples, negative samples, The generated samples are input into the discriminator respectively to obtain the discrimination results;
  • Model prediction module Obtain the characteristic data and disease label data of general practice patients to be trained, use the data generation model to generate general practice disease data for a small number of samples of general practice diseases; jointly train the training samples and the generated general practice disease data based on The general practice multi-disease prediction model of general practice causal graph convolutional neural network uses the trained general practice multi-disease prediction model to predict general practice diseases.
  • the general practice patient's binary variable data is used to train the general practice tendency score network; the general practice patient's characteristic variable data and label variable data are converted into binary classification variables.
  • the categorical variables Convert to binary categorical variables through one-hot encoding.
  • For continuous variables convert to categorical variables through binning and then convert into binary variables through one-hot encoding.
  • the general propensity score network includes an input layer, a local connection layer, a sigmoid activation layer and an output layer;
  • the number of input layer nodes and the number of output layer nodes are both the number M of first event variables in the first event variable set;
  • Both the local connection layer and the sigmoid activation layer contain ⁇ M nodes, ⁇ 2; the u-th node of the input layer is connected to all nodes except the ⁇ (u-1)+1 to ⁇ u-th nodes of the local connection layer;
  • the ⁇ (u-1)+1 to ⁇ u local connection layer nodes are connected in a one-to-one correspondence with the ⁇ (u-1)+1 to ⁇ u sigmoid activation layer nodes; the ⁇ (u-1)+1 to ⁇ u
  • the sigmoid activation layer node is only connected to the u-th output layer node.
  • first event variable a For each first event variable a, input the covariate data corresponding to the training sample into the local connection layer to obtain the first characteristic representation of the tendency, and input the first characteristic representation of the tendency into the sigmoid activation layer to obtain the second characteristic representation of the tendency, Input the second feature representation of the tendency into the output layer to obtain the predicted value of the first event variable a; use the predicted values of all first event variables and the true values of all first event variables to calculate the tendency loss.
  • the trained general propensity score network is used to calculate the general propensity score of general patient i for the first event variable a. Use the general propensity score to calculate the causal effect value ATE a, b of the first event variable a and the second event variable b .
  • the calculation formula is as follows:
  • n the total number of patients to be studied
  • Ti the true value of the first event variable of the i-th patient
  • Y i the true value of the second event variable of the i-th patient.
  • the generator is composed of a multi-layer generator module.
  • the generator module includes a normalization layer, a fully connected layer and an activation layer.
  • the last layer of the generator module of the generator is The activation layer is a sigmoid activation layer; during the training process, the random noise and the corresponding disease label are input into the normalization layer of the first generator module, and the normalized data is input into the fully connected layer of the first generator module to obtain the input data.
  • a feature representation input the first feature representation into the activation layer of the first generator module to obtain the second feature representation of the input data, use the second feature representation as the input data of the next layer of generator module, and finally pass the last layer of generator
  • the sigmoid activation layer of the module is used to generate samples.
  • the causal loss L causal calculation formula is as follows:
  • a r indicates the set of first event variables paired with the second event variable r;
  • R indicates the small-sample general disease set obtained by the disease statistics module ;
  • q r represents the frequency of general disease r in a small sample.
  • the discriminator adversarial loss L ⁇ is calculated as follows:
  • N is the amount of random noise data, is the probability that the i-th generated sample is judged by the discriminator to be the real data of the corresponding disease
  • the total loss L d of the discriminator is calculated as follows:
  • m d is the number of positive samples
  • y k is the disease label corresponding to the positive sample
  • y k is the disease label corresponding to the positive sample
  • They are positive samples x k and negative samples respectively.
  • the probability that the generated sample d k is judged by the discriminator to be the real data of the disease y k .
  • model prediction module includes:
  • Each first event variable constitutes a first event node in the event relationship graph
  • each second event variable constitutes a second event node in the event relationship graph
  • an edge is constructed for each event pairing
  • the general practice causal graph convolutional neural network includes multiple causal graph convolution modules.
  • the causal graph convolution module includes a causal graph convolution layer. and activation layer;
  • the node embedding representation is input into the causal graph convolution layer of the first causal graph convolution module to obtain the first graph feature representation h (0) :
  • H (0) represents the node embedding representation
  • W (0) represents the causal graph convolution layer weight
  • I represents the identity matrix
  • * represents the multiplication of each element of the matrix
  • the output of the previous causal graph convolution module is input into the next causal graph convolution module until the final disease prediction result is obtained.
  • the present invention While amplifying the data, the present invention considers the causal logic between features, making the generated data more consistent with the real situation. Model training on this part of the data can improve model performance.
  • the present invention proposes a generative adversarial network based on causal verification, which makes the generated data more consistent with real causal logic and has certain causal interpretability.
  • the present invention proposes a general causal graph convolutional neural network to improve the robustness of general practice multi-disease prediction models.
  • Figure 1 is a structural block diagram of a general practice multi-disease prediction system generated based on causality verification data provided by an embodiment of the present invention
  • Figure 2 is a flow chart of the implementation of the causality verification module provided by the embodiment of the present invention.
  • Figure 3 is a structural diagram of the general subject propensity score network provided by the embodiment of the present invention.
  • Figure 4 is a structural diagram of a generative adversarial network based on causality verification provided by an embodiment of the present invention
  • Figure 5 is a flow chart of the implementation of the model prediction module provided by the embodiment of the present invention.
  • the present invention provides a medical data generation method based on a generative adversarial network based on causal verification, and based on this method, a set of methods is constructed to solve the problem of poor prediction of small-sample diseases in general multi-disease prediction models due to fewer training samples.
  • Multi-disease prediction system for general practice problem As shown in Figure 1, the multi-disease prediction system for general practitioners based on causality verification data generation provided by the present invention includes a disease statistics module, a causality verification module, a data generation module and a model prediction module.
  • the sample ratio is the ratio of the number of samples for the disease with the largest number of samples to the number of samples for various diseases. For example, for the four general diseases of cold, gastritis, diarrhea, and fever, the number of samples corresponds to 10, 20, 30, and 40, respectively. Ratio 4, 2, 4/3, 1.
  • Obtain the patient's characteristic variable data and label variable data Convert the feature variable data and label variable data into binary variables as follows. For categorical variables, one-hot encoding is used to convert them into binary variables. For continuous variables, they are converted to categorical variables through binning and then converted into binary variables through one-hot encoding.
  • the feature variable set constitutes the first event variable set
  • the label variable set constitutes the second event variable set.
  • the first event variable set is a clinical manifestation set, such as ⁇ hypertension, fever, chest tightness ⁇
  • the second event variable set is a general disease set, such as ⁇ cold, gastritis, cardiovascular disease ⁇ .
  • the causal effect value calculation method is as follows.
  • first event variable a and the second event variable b constitute event pairing ⁇ ; define the covariates corresponding to event pairing ⁇ as variables in the first event variable set except the first event variable a, and use event pairing hypertension-cold
  • the covariates are variables other than the hypertension variable in the first event variable set ⁇ high blood pressure, fever, chest tightness ⁇ , that is, ⁇ fever, chest tightness ⁇ . Due to the diversity and complexity of general practice scenario data, the traditional propensity score calculation method based on logistic regression has limited ability to handle nonlinearly separable data. Therefore, the present invention constructs a general practice tendency score network for general practice scenarios, uses the binary variable data of general practice patients to train the general practice tendency score network, and uses the trained general practice tendency score network to calculate the general practice tendency Score.
  • the general propensity score represents the patient's probability of experiencing the first event given covariates. Take ⁇ hypertension, fever, chest tightness ⁇ as an example, that is, the probability of high blood pressure in patients who have fever and chest tightness.
  • the general propensity score network includes an input layer, a local connection layer, a sigmoid activation layer and an output layer.
  • the number of input layer nodes and the number of output layer nodes are both the number M of first event variables in the first event variable set.
  • Both the local connection layer and the sigmoid activation layer contain ⁇ M nodes, ⁇ is an adjustable parameter, ⁇ 2, the u-th node of the input layer is the same as the local connection layer except ⁇ (u-1)+1 to ⁇ u All nodes except layer nodes are connected.
  • the ⁇ (u-1)+1 to ⁇ u local connection layer nodes are connected to the ⁇ (u-1)+1 to ⁇ u sigmoid activation layer nodes in a one-to-one correspondence.
  • the ⁇ (u-1)+1 to ⁇ u sigmoid activation layer nodes are only connected to the u-th output layer node.
  • the beneficial effect of the local connection layer is that the local connection layer ensures that the input layer is locally connected to the output layer.
  • One event variable, the covariate feature node of the input layer, the local connection layer, the sigmoid activation layer and the first event variable node of the output layer form a local network.
  • the local connection layer ensures that the local networks are independent of each other, so that the predicted first An event variable will not be used for prediction.
  • Figure 3 is an example of a general subject propensity score network.
  • input layer node 1 it is connected to all nodes in the local connection layer except nodes 1 and 2.
  • the local connection layer nodes 1 is connected to sigmoid activation layer node 1
  • local connection layer node 2 is connected to sigmoid activation layer node 2
  • local connection layer nodes 1 and 2 are only connected to output layer node 1.
  • the propensity loss is calculated using the predicted values of all first event variables and the true values of all first event variables.
  • the propensity loss function L p is as follows:
  • m p represents the total number of training samples
  • ⁇ f a represents the true value of the first event variable a of the training sample f
  • the general propensity score is used to calculate the causal effect value ATE between the first event variable and the second event variable.
  • the causal effect value ATE a and b between the first event variable a and the second event variable b are calculated as follows:
  • n the total number of patients to be studied
  • Ti represents the true value of the first event variable of the i-th patient
  • Y i represents the true value of the second event variable of the i-th patient
  • a data generation model is built using a generative adversarial network based on causality verification, and the trained data generation model is used to generate simulation data.
  • the data generation model includes a generator and a discriminator.
  • the generator G(z, c) consists of a multi-layer generator module, where z represents random noise and c represents the disease label of the sample to be generated.
  • the generator module includes a normalization layer, a fully connected layer and an activation layer.
  • the activation layer of the last generator module of the generator is the sigmoid activation layer, and the activation layers of the remaining generator modules can be the relu activation layer, sigmoid activation layer, or tanh activation layer.
  • the discriminator D is composed of a multi-layer discriminator module, which includes a fully connected layer, a dropout layer and an activation layer.
  • Figure 4 is a structural diagram of a generative adversarial network based on causality checking. According to the generator training process and the discriminator training process, the generator and the discriminator are trained iteratively and alternately, and finally the trained data generation model is obtained. The training process is explained in detail below.
  • S2 Input the random noise z and the corresponding disease label c into the normalization layer of the first generator module.
  • the normalization layer is used to normalize the input data, including batch standardization, sample standardization, etc., and input the normalized data into the first generator module.
  • the fully connected layer of the generator module is used to obtain the first feature representation of the input data.
  • the first feature representation is input into the activation layer of the first generator module to obtain the second feature representation of the input data.
  • the second feature representation is generated as the next layer.
  • the input data of the generator module is input and output layer by layer, and finally the generated sample is obtained through the sigmoid activation layer of the last layer of the generator module.
  • S3 Use the causality check module to calculate the causal effect values of all event pairs of the generated sample.
  • S4 Input the generated sample and disease label into the discriminator, and obtain the probability y * that the discriminator will distinguish the generated sample as the real data corresponding to the disease.
  • the adversarial loss of the discriminator measures the degree to which the generated samples of the generator are judged to be true by the discriminator. The smaller the adversarial loss of the discriminator, the easier it is for the generated samples to be judged to be true.
  • the discriminator adversarial loss L ⁇ is calculated as follows:
  • the causal loss measures the degree of causal consistency between the generator's generated samples and the original data. The smaller the causal loss, the more consistent the internal causal relationship between the generated samples and the original data. Specifically, the causal loss is the KL divergence loss between the causal effect values of all event pairs of the generated sample corrected by the small-sample general disease frequency q r and the causal effect values of all event pairs of the original data. For diseases with very few samples, the variance of the causal effect values corresponding to the calculated original data is larger, and smaller weights are assigned to improve the stability of training.
  • the calculation formula of causal loss L causal is as follows:
  • a r represents the set of first event variables paired with the second event variable r;
  • q r represents the frequency of general disease r in a small sample.
  • the total generator loss is as follows:
  • S1 Randomly select m d patient samples from the original data, that is, the general practice data set As a positive sample, x k and y k respectively represent the feature data and disease label of the extracted kth positive sample.
  • Model prediction module the implementation process is shown in Figure 5.
  • First build an event relationship diagram including:
  • Each first event variable in the first event variable set constitutes a first event node in the event relationship graph
  • each second event variable in the second event variable set constitutes a second event node in the event relationship graph.
  • an edge is constructed to complete the construction of the event relationship graph.
  • first event variable set ⁇ fever, chest tightness ⁇ and second event variable set ⁇ acute respiratory infection ⁇ as an example. There is an edge between fever and acute respiratory tract infection, and there is an edge between chest tightness and acute respiratory infection.
  • a graph representation learning algorithm is used to generate embedding representations of the first event node and the second event node.
  • the corresponding degree matrix ⁇ and adjacency matrix A are constructed based on the event relationship graph.
  • the causal effect matrix ⁇ is constructed using the causal effect values of the original data.
  • the number of rows and columns of the causal effect matrix ⁇ is the same, which is the number of first event nodes plus the number of second event nodes.
  • the general practice causal graph convolutional neural network contains multiple causal graph convolution modules.
  • the causal graph convolution module includes a causal graph convolution layer and an activation layer.
  • the causal graph convolution layer is a graph convolution layer modified by the causal effect matrix. The robustness of the model is improved by adding causal effect modification.
  • H (0) represents the node embedding representation
  • W (0) represents the weight of the causal graph convolution layer of the first causal graph convolution module, which can be trained
  • I represents the identity matrix
  • * represents the multiplication of each element of the matrix.
  • the output of the previous causal graph convolution module is input into the next causal graph convolution module until the final disease prediction result is obtained.
  • Calculate the loss of the general causal graph convolutional neural network, and the loss function is the cross-entropy loss function.
  • the present invention proposes a general subject propensity score network suitable for calculating the general subject propensity score;
  • the causal effect calculation method is used to perform causal verification on the general subject data generated by the generative adversarial network, so that the generated data is more consistent with Real causal logic;
  • the generator training process generates the same number of noise points from the binomial distribution for each few-sample disease and serves as input to the generator;
  • the discriminator training process extracts positive samples from the original data and The same number of samples with different labels are used as negative samples, together with the negative samples generated by the generator, to train the discriminator;
  • a generative adversarial network based on causality verification is used to amplify the general data and improve
  • a general practice multi-disease prediction model based on the general practice causal graph convolutional neural network is proposed, and the causal effect value is integrated to improve the prediction performance of the general practice multi-disease prediction
  • first, second, third, etc. may use the terms first, second, third, etc. to describe various information, the information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other.
  • first information may also be called second information, and similarly, the second information may also be called first information.
  • word “if” as used herein may be interpreted as "when” or “when” or “in response to determining.”

Landscapes

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

Abstract

本发明公开了一种基于因果校验数据生成的全科多疾病预测系统，本发明针对全科场景，从因果性的角度出发，提出了基于全科倾向性得分网络的倾向性得分计算方法；相较于传统生成式对抗网络可解释性差的问题，本发明提出了基于因果校验的生成式对抗网络，使得生成的数据更加符合真实的因果逻辑；针对现有图卷积神经网络仅从相关性角度建模的问题，本发明提出了基于全科因果图卷积神经网络的全科多疾病预测模型，融入因果效应值以提升全科多疾病预测系统对疾病的预测性能，解决了全科场景因训练样本少导致模型表现差以及鲁棒性不高的问题。

Description

一种基于因果校验数据生成的全科多疾病预测系统 技术领域
本发明属于医疗健康信息技术领域，具体涉及一种基于因果校验数据生成的全科多疾病预测系统。
背景技术
随着信息技术的发展，机器学习已经成为推动医疗发展的重要力量。全科医学作为医疗领域受众面最广的医学学科，是机器学习模型在医疗场景应用的重点领域之一。然而，由于全科疾病多而繁杂，样本获取成本高等问题，部分疾病常常难以获取大量的训练数据，导致现有的全科多疾病预测系统对少样本疾病的预测效果不佳。目前迫切需要一套针对少样本的全科多疾病预测系统。
通过数据生成方法生成模拟数据是解决机器学习模型训练样本不足的常见方法。现有的数据生成方式主要是基于生成式对抗网络。生成式对抗网络在生成图像数据的时候表现良好。然而全科场景，数据种类繁多且结构复杂，尤其是结构化的医疗数据，包含以患者为中心的多种类特征数据，在时间、空间上存在异质性，数据分布较为复杂。传统生成式对抗网络难以处理分布复杂的结构化数据。一方面，使用少样本数据训练容易出现训练不稳定、梯度消失、模式崩溃的问题。另一方面，仅考虑了变量之间的相关性，没有考虑变量之间的因果关系，将导致其生成的数据常常难以理解，不符合常识，使用这些数据用于模型训练，可能不能提高甚至会削弱模型的训练效果。例如，感冒可分为病毒性感冒和细菌性感冒，也会用两种药。如果基于相关性模型生成发烧患者的数据，可能会生成同时使用病毒性感冒药和细菌性感冒药的情况，这会对后续构建模型产生干扰。
基于倾向性得分的因果效应值计算方法是最为常见衡量变量之间因果关系的方法。现有的倾向性得分计算方法大多是基于逻辑斯特回归，然而全科场景由于数据种类繁多，结构复杂且常常线性不可分，基于逻辑斯特回归的倾向性得分计算方法在全科场景表现不佳。
发明内容
本发明针对现有技术的不足，从因果性的角度出发，提出了基于全科倾向性得分网络的倾向性得分计算方法，并在此基础上提出了基于因果校验的生成式对抗网络的医疗数据生成方法，解决了基于相关性分析的生成式对抗网络生成数据难以理解的问题，构建了一套全科 多疾病预测系统，解决了全科场景因训练样本少导致模型表现差以及鲁棒性不高的问题。
本发明的目的是通过以下技术方案实现的：一种基于因果校验数据生成的全科多疾病预测系统，包括：
(1)疾病统计模块：用于统计各种全科疾病样本数，根据各种全科疾病样本比率得到少样本全科疾病；
(2)因果校验模块：根据全科病人的特征变量集构成第一事件变量集合，疾病标签变量集构成第二事件变量集合，任意第一事件变量同任意第二事件变量构成一个事件配对；
构建并训练全科倾向性得分网络，使用训练完成的全科倾向性得分网络计算全科倾向性得分，所述全科倾向性得分表示全科病人在协变量条件下发生第一事件的概率；使用全科倾向性得分计算所有事件配对的因果效应值；
(3)数据生成模块：对于少样本全科疾病，基于因果校验的生成式对抗网络构建数据生成模型，使用训练完成的数据生成模型生成模拟数据；
所述数据生成模型包括生成器和判别器，所述生成器和所述判别器迭代交替训练；
所述生成器的训练过程包括：对于每种少样本全科疾病生成随机噪声，将随机噪声以及对应的疾病标签输入生成器得到生成样本；计算生成样本的所有事件配对的因果效应值；将生成样本以及对应的疾病标签输入判别器，得到判别结果；所述生成器的总损失包括判别器对抗损失、因果损失和正则项损失；所述因果损失为经过少样本全科疾病频率矫正的生成样本的所有事件配对的因果效应值与原始数据的所有事件配对的因果效应值的KL散度损失；
所述判别器的训练过程包括：从原始数据中随机抽取正样本，并抽取相同数量但疾病标签不同的负样本；生成相同数量随机噪声，使用生成器得到生成样本；将正样本、负样本、生成样本分别输入判别器，得到判别结果；
(4)模型预测模块：获取待训练全科病人的特征数据和疾病标签数据，对少样本全科疾病使用数据生成模型生成全科疾病数据；将训练样本以及生成的全科疾病数据共同训练基于全科因果图卷积神经网络的全科多疾病预测模型，使用训练完成的全科多疾病预测模型对全科疾病进行预测。
进一步地，所述因果校验模块中，使用全科病人的二分类变量数据训练全科倾向性得分网络；将全科病人的特征变量数据和标签变量数据转换成二分类变量，对于类别变量，通过独热编码转换成二分类变量，对于连续变量，通过分箱转换至类别变量之后通过独热编码转换成二分类变量。
进一步地，所述全科倾向性得分网络包括输入层、局部连接层、sigmoid激活层和输出层；
输入层节点个数和输出层节点个数均为第一事件变量集合中的第一事件变量个数M；局 部连接层和sigmoid激活层均包含τM个节点，τ≥2；输入层的第u个节点同局部连接层的除第τ(u-1)+1到τu个节点之外的所有节点相连；第τ(u-1)+1到τu个局部连接层节点同第τ(u-1)+1到τu个sigmoid激活层节点一一对应连接；第τ(u-1)+1到τu个sigmoid激活层节点仅同第u个输出层节点相连。
进一步地，所述全科倾向性得分网络的训练过程如下：
对于每个第一事件变量a，将训练样本对应的协变量数据输入局部连接层得到倾向性第一特征表示，将所述倾向性第一特征表示输入sigmoid激活层得到倾向性第二特征表示，将所述倾向性第二特征表示输入输出层得到第一事件变量a的预测值；使用所有第一事件变量的预测值同所有第一事件变量的真实值计算倾向性损失。
进一步地，所述因果校验模块中，使用训练完成的全科倾向性得分网络计算全科病人i对于第一事件变量a的全科倾向性得分使用全科倾向性得分计算第一事件变量a同第二事件变量b的因果效应值ATEa，b，计算公式如下：
其中n表示待研究病人总数，Ti表示第i个病人第一事件变量真实值；Yi表示第i个病人第二事件变量真实值。
进一步地，所述数据生成模块中，所述生成器由多层生成器模块构成，所述生成器模块包括规范化层、全连接层和激活层，所述生成器的最后一层生成器模块的激活层为sigmoid激活层；在训练过程中，将随机噪声以及对应的疾病标签输入第一生成器模块的规范化层，将规范化后的数据输入第一生成器模块的全连接层得到输入数据的第一特征表示，将第一特征表示输入第一生成器模块的激活层得到输入数据的第二特征表示，将第二特征表示作为下一层生成器模块的输入数据，最终通过最后一层生成器模块的sigmoid激活层得到生成样本。
进一步地，所述数据生成模块中，因果损失Lcausal计算公式如下：
其中表示原始数据的第一事件变量a同第二事件变量r的因果效应值，表示生成样本的第一事件变量a同第二事件变量r的因果效应值；Ar表示与第二事件变量r配对的第一事件变量集合；R表示疾病统计模块得到的少样本全科疾病集合；qr表示少样本全科疾病r的频率。
进一步地，所述数据生成模块中，判别器对抗损失Lζ计算公式如下：
其中N为随机噪声数据量，为第i个生成样本经判别器判别为对应疾病真实数据的概率；
正则项损失Lregular计算公式如下：
Lregular＝||w||
其中||·||表示L1范数，w表示生成器模型参数。
进一步地，所述数据生成模块中，所述判别器的总损失Ld计算公式如下：
其中md为正样本数量，yk为正样本对应的疾病标签，分别为抽取的第k个正样本、抽取的第k个负样本、使用生成器得到的第k个生成样本，分别为正样本xk、负样本生成样本dk经判别器判定为疾病yk真实数据的概率。
进一步地，所述模型预测模块包括：
构建事件关系图：每个第一事件变量构成事件关系图中的一个第一事件节点，每个第二事件变量构成事件关系图中的一个第二事件节点，对于每个事件配对构建一条边；
生成第一事件节点和第二事件节点的节点嵌入表示；基于事件关系图构建度数矩阵Φ和邻接矩阵A；使用原始数据的因果效应值构建因果效应矩阵Ψ；
构建基于全科因果图卷积神经网络的全科多疾病预测模型，所述全科因果图卷积神经网络包含多个因果图卷积模块，所述因果图卷积模块包括因果图卷积层和激活层；
将节点嵌入表示输入第一因果图卷积模块的因果图卷积层，得到第一图特征表示h(0)：
其中H(0)表示节点嵌入表示，W(0)表示因果图卷积层权值，I表示单位矩阵，*表示矩阵各元素相乘；
将h(0)输入第一因果图卷积模块的激活层得到第一因果图卷积模块的输出H(1)；
将上一因果图卷积模块的输出输入下一因果图卷积模块，直到得到最终疾病预测结果。
本发明的有益效果是：
1.本发明对数据进行扩增的同时，考虑了特征之间的因果逻辑，使得产生的数据更加符合真实情况，这部分数据进行模型训练能够提升模型性能。
2.相较于传统生成式对抗网络可解释性差的问题，本发明提出了基于因果校验的生成式对抗网络，使得生成的数据更加符合真实的因果逻辑，具有一定的因果可解释性。
3.针对现有图卷积神经网络仅从相关性角度建模的问题，本发明提出了全科因果图卷积神经网络，提升全科多疾病预测模型的鲁棒性。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例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发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例提供的基于因果校验数据生成的全科多疾病预测系统结构框图；
图2为本发明实施例提供的因果校验模块实现流程图；
图3为本发明实施例提供的全科倾向性得分网络结构图；
图4为本发明实施例提供的基于因果校验的生成式对抗网络结构图；
图5为本发明实施例提供的模型预测模块实现流程图。
具体实施方式
为了更好的理解本申请的技术方案，下面结合附图对本申请实施例进行详细描述。
应当明确，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其它实施例，都属于本申请保护的范围。
在本申请实施例中使用的术语是仅仅出于描述特定实施例的目的，而非旨在限制本申请。在本申请实施例和所附权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式，除非上下文清楚地表示其他含义。
本发明提供一种基于因果校验的生成式对抗网络的医疗数据生成方法，并基于该方法构建了一套用以解决全科多疾病预测模型中因训练样本较少导致模型对少样本疾病预测较差问题的全科多疾病预测系统。如图1所示，本发明提供的基于因果校验数据生成的全科多疾病预测系统包括疾病统计模块、因果校验模块、数据生成模块和模型预测模块。
下述说明进一步给出了符合本申请要求的基于因果校验数据生成的全科多疾病预测系统各模块实现的部分实施例。
一、疾病统计模块
对所有种类全科疾病，统计各种疾病的样本数，计算各种疾病的样本比率。样本比率为样本数最多的疾病的样本数同各种疾病样本数的比率，例如对于感冒、胃炎、腹泻、发烧四种全科疾病，分别对应样本数10、20、30、40，分别对应样本比率4、2、4/3、1。
对于疾病样本比率大于设定阈值(可调参数，根据实际情况设定)的疾病，将其加入少样本全科疾病集合R，计算第r种少样本全科疾病的频率其中countr为第r种疾病的样本数。
二、因果校验模块，实现流程如图2所示。
获取病人的特征变量数据以及标签变量数据。将特征变量数据以及标签变量数据按以下方法转换成二分类变量。对于类别变量，通过独热编码转换成二分类变量。对于连续变量，通过分箱转换至类别变量之后再通过独热编码转换成二分类变量。
特征变量集构成第一事件变量集合，标签变量集构成第二事件变量集合。第一事件变量集合为临床表现集合，例如{高血压，发烧，胸闷}，第二事件变量集合为全科疾病集合，例如{感冒，胃炎，心血管疾病}。
对于第一事件变量集合中的任意一个第一事件变量同第二事件变量集合中的任意一个第二事件变量构成一个事件配对，计算所有事件配对的因果效应值，因果效应值计算方法如下。
记第一事件变量a和第二事件变量b构成事件配对δ；定义事件配对δ对应的协变量为第一事件变量集合中除第一事件变量a之外的变量，以事件配对高血压-感冒为例，协变量即第一事件变量集合{高血压，发烧，胸闷}中除高血压变量之外的变量，即{发烧，胸闷}。由于全科场景数据多样且复杂，传统的基于逻辑斯特回归的倾向性得分计算方法在处理非线性可分的数据能力有限。因此本发明构建了针对全科场景的全科倾向性得分网络，使用全科病人的二分类变量数据训练全科倾向性得分网络，并使用训练完成的全科倾向性得分网络计算全科倾向性得分。
全科倾向性得分表示病人在协变量条件下发生第一事件的概率。以{高血压，发烧，胸闷}为例，即发生发烧，胸闷的病人，其发生高血压的概率。
全科倾向性得分网络包括输入层、局部连接层、sigmoid激活层和输出层。
具体地，输入层节点个数以及输出层节点个数均为第一事件变量集合中的第一事件变量个数M。局部连接层以及sigmoid激活层均包含τM个节点，τ为可调节参数，τ≥2，输入层的第u个节点同局部连接层的除第τ(u-1)+1到τu个局部连接层节点之外的所有节点相连。第τ(u-1)+1到τu个局部连接层节点同第τ(u-1)+1到τu个sigmoid激活层节点一一对应进行连接。第τ(u-1)+1到τu个sigmoid激活层节点仅同第u个输出层节点相连。局部连接层的有益效果为，局部连接层保证了输入层同输出层局部连接，对于每一个待预测的第 一事件变量，输入层的协变量特征节点同局部连接层、sigmoid激活层以及输出层的第一事件变量节点构成一个局部网络，局部连接层保证了局部网络之间相互独立，使得被预测的第一事件变量不会用于预测。
图3为一全科倾向性得分网络示例，该示例中M＝3，τ＝2，对于输入层节点1，其同局部连接层除节点1，2之外的所有节点相连，局部连接层节点1连接sigmoid激活层节点1，局部连接层节点2连接sigmoid激活层节点2，局部连接层节点1，2仅同输出层节点1相连。
全科倾向性得分网络的训练流程如下：
对于每一个第一事件变量a，将训练样本对应的协变量数据输入局部连接层得到倾向性第一特征表示，将倾向性第一特征表示输入sigmoid激活层，得到倾向性第二特征表示，将倾向性第二特征表示输入输出层得到第一事件变量a的预测值。使用所有第一事件变量的预测值同所有第一事件变量的真实值计算倾向性损失，倾向性损失函数Lp如下：
其中mp表示训练样本总数，γf，a表示训练样本f的第一事件变量a的真实值，表示训练样本f的第一事件变量a的预测值。
使用训练完成的全科倾向性得分网络计算全科病人i的对于第一事件变量a的全科倾向性得分使用全科倾向性得分计算第一事件变量同第二事件变量的因果效应值ATE，第一事件变量a同第二事件变量b的因果效应值ATEa，b公式如下：
其中n表示待研究病人总数，Ti表示第i个病人第一事件变量真实值；Yi表示第i个病人第二事件变量真实值，Yi＝1表示第i个病人发生了第二事件，Yi＝0表示第i个病人未发生第二事件。
三、数据生成模块
对于少样本全科疾病集合R，基于因果校验的生成式对抗网络构建数据生成模型，使用训练完成的数据生成模型生成模拟数据。
具体地，数据生成模型包括生成器和判别器。生成器G(z，c)由多层生成器模块构成，其中z表示随机噪声，c表示待生成样本的疾病标签，生成器模块包括规范化层、全连接层和激活层。生成器的最后一层生成器模块的激活层为sigmoid激活层，其余生成器模块的激活层可以为relu激活层、sigmoid激活层、tanh激活层。判别器D由多层判别器模块构成，判别器模块包括全连接层、Dropout层和激活层。
图4为基于因果校验的生成式对抗网络结构图。按照生成器训练流程以及判别器训练流程，迭代交替训练生成器以及判别器，最终得到训练完成的数据生成模型，下面详细阐述训练流程。
(1)生成器训练流程
S1：对于少样本全科疾病集合R中的每种疾病r，从二项分布中随机生成mg个噪声点对应的疾病标签cr＝{r，r，...，r}。对于所有v种疾病，生成N＝mg×v个随机噪声数据和疾病标签数据，随机噪声数据z＝{z1，z2，...，zv}，疾病标签数据c＝{c1，c2，...，cv}。
S2：将随机噪声z以及对应的疾病标签c输入第一生成器模块的规范化层，规范化层用于对输入数据进行规范化操作，包括批标准化、样本标准化等，将规范化后的数据输入第一生成器模块的全连接层，得到输入数据的第一特征表示，将第一特征表示输入第一生成器模块的激活层，得到输入数据的第二特征表示，将第二特征表示作为下一层生成器模块的输入数据，逐层输入输出，最终通过最后一层生成器模块的sigmoid激活层得到生成样本。
S3：使用因果校验模块计算生成样本的所有事件配对的因果效应值。
S4：将生成样本以及疾病标签输入判别器，得到判别器将生成样本判别为对应疾病真实数据的概率y*。
S5：计算生成器总损失L，包括判别器对抗损失Lζ、因果损失Lcausal以及正则项损失Lregular。
判别器对抗损失衡量了生成器的生成样本被判别器判别为真的程度，判别器对抗损失越小，生成样本越易被判别为真。判别器对抗损失Lζ计算公式如下：
其中为第i个生成样本经判别器判别为对应疾病真实数据的概率。
因果损失衡量了生成器的生成样本同原始数据因果符合程度，因果损失越小，生成样本的内在因果关系同原始数据越一致。具体地，因果损失为经过少样本全科疾病频率qr矫正的生成样本的所有事件配对的因果效应值与原始数据的所有事件配对的因果效应值的KL散度损失。对于样本特少的疾病，计算的原始数据对应的因果效应值方差较大，赋予较小的权重以提升训练的稳定性。因果损失Lcausal计算公式如下：
其中表示原始数据的第一事件变量a同第二事件变量r的因果效应值，表示生成样本的第一事件变量a同第二事件变量r的因果效应值；Ar表示与第二事件变量r配对的第一事件变量集合；qr表示少样本全科疾病r的频率。
正则项损失Lregular计算公式如下：
Lregular＝||w||
其中||·||表示L1范数，w表示生成器模型参数。
生成器总损失如下：
L＝Lζ+Lcausal+Lregular
(2)判别器训练流程
S1：从原始数据即全科数据集中随机抽取md个病人样本作为正样本，xk，yk分别表示抽取的第k个正样本的特征数据和疾病标签。
S2：从原始数据中随机抽取md个病人样本作为负样本，分别表示抽取的第k个负样本的特征数据和疾病标签。抽样时需保证第k个正样本同第k个负样本对应的疾病标签不一样，即
S3：从二项分布中随机采样出md个噪声点并使用生成器得到生成样本，第k个生成样本dk表示如下：
S4：将抽取的正负样本，以及生成样本分别输入判别器D，得到预测的疾病标签。
S5：计算判别器总损失Ld，计算公式如下：
其中分别为正样本、负样本、生成样本经判别器D判别为疾病yk真实数据的概率。
四、模型预测模块，实现流程如图5所示。
获取待训练全科病人的特征数据以及疾病标签数据。对训练样本不足的疾病使用数据生 成模块中训练完成的数据生成模型生成全科疾病数据。将训练样本连同生成的全科疾病数据一同用于训练全科多疾病预测模型，具体流程如下：
首先构建事件关系图，包括：
对于第一事件变量集合的每一个第一事件变量构成事件关系图中的一个第一事件节点，对于第二事件变量集合中的每一个第二事件变量构成事件关系图中的一个第二事件节点。对于每一个病人的每一对第一事件变量同第二事件变量构建一条边，从而完成事件关系图构建。
以一个病人的第一事件变量集合{发烧，胸闷}，第二事件变量集合{急性呼吸道感染}为例。发烧同急性呼吸道感染之间构建一条边，胸闷同急性呼吸道感染构建一条边。
使用图表示学习算法生成第一事件节点以及第二事件节点的嵌入表示。基于事件关系图构建对应的度数矩阵Φ以及邻接矩阵A。使用原始数据的因果效应值构建因果效应矩阵Ψ，因果效应矩阵Ψ的行数和列数相同，为第一事件节点个数加上第二事件节点个数。因果效应矩阵Ψ的第α行第β列元素记为ψα，β，如果第α行为第一事件节点，第β列为第二事件节点，则否则ψα，β＝0。
构建基于全科因果图卷积神经网络的全科多疾病预测模型，全科因果图卷积神经网络包含多个因果图卷积模块，因果图卷积模块包括因果图卷积层和激活层。因果图卷积层为经过因果效应矩阵修正的图卷积层，通过加入因果效应修正来提高模型鲁棒性。将节点嵌入表示输入第一因果图卷积模块的因果图卷积层，得到第一图特征表示h(0)：

A∧＝A+I
其中H(0)表示节点嵌入表示，W(0)表示第一因果图卷积模块的因果图卷积层的权值，可训练得到，I表示单位矩阵，*表示矩阵各元素相乘。
将第一图特征表示h(0)输入第一因果图卷积模块的激活层得到第一因果图卷积模块的输出H(1)：
H(1)＝σ(h(0))
其中σ(·)表示激活函数
将上一因果图卷积模块的输出输入下一因果图卷积模块，直到得到最终疾病预测结果。计算全科因果图卷积神经网络损失，损失函数为交叉熵损失函数。
迭代训练全科因果图卷积神经网络，得到训练完成的全科多疾病预测模型，并使用训练完成的全科多疾病预测模型对全科疾病进行预测。
本发明针对全科场景，提出适用于计算全科倾向性得分的全科倾向性得分网络；利用因果效应计算方法对生成式对抗网络生成的全科数据进行因果校验，使得生成的数据更符合真实的因果逻辑；生成器训练过程，对每一个少样本疾病从二项分布中生成相同数量的噪声点，并一同作为生成器的输入；判别器训练过程，从原始数据抽取正样本，并抽取相同数量但是标签不同的样本作为负样本，连同生成器生成的负样本一起用于训练判别器；针对少样本全科疾病，使用基于因果校验的生成式对抗网络对全科数据扩增，提升全科多疾病预测系统对少样本疾病的预测性能；提出基于全科因果图卷积神经网络的全科多疾病预测模型，融入因果效应值以提升全科多疾病预测系统对疾病的预测性能。
还需要说明的是，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、商品或者设备中还存在另外的相同要素。
上述对本说明书特定实施例进行了描述。其它实施例在所附权利要求书的范围内。在一些情况下，在权利要求书中记载的动作或步骤可以按照不同于实施例中的顺序来执行并且仍然可以实现期望的结果。另外，在附图中描绘的过程不一定要求示出的特定顺序或者连续顺序才能实现期望的结果。在某些实施方式中，多任务处理和并行处理也是可以的或者可能是有利的。
在本说明书一个或多个实施例使用的术语是仅仅出于描述特定实施例的目的，而非旨在限制本说明书一个或多个实施例。在本说明书一个或多个实施例和所附权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式，除非上下文清楚地表示其他含义。还应当理解，本文中使用的术语“和/或”是指并包含一个或多个相关联的列出项目的任何或所有可能组合。
应当理解，尽管在本说明书一个或多个实施例可能采用术语第一、第二、第三等来描述各种信息，但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开。例如，在不脱离本说明书一个或多个实施例范围的情况下，第一信息也可以被称为第二信息，类似地，第二信息也可以被称为第一信息。取决于语境，如在此所使用的词语“如果”可以被解释成为“在……时”或“当……时”或“响应于确定”。
以上所述仅为本说明书一个或多个实施例的较佳实施例而已，并不用以限制本说明书一个或多个实施例，凡在本说明书一个或多个实施例的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本说明书一个或多个实施例保护的范围之内。

Claims (10)

  1. 一种基于因果校验数据生成的全科多疾病预测系统，其特征在于，包括：
    (1)疾病统计模块：用于统计各种全科疾病样本数，根据各种全科疾病样本比率得到少样本全科疾病；所述样本比率为样本数最多的疾病的样本数同各种疾病样本数的比率，对于全科疾病样本比率大于设定阈值的全科疾病，将其加入少样本全科疾病集合R，计算第r种少样本全科疾病的频率其中countr为第r种全科疾病样本数；
    (2)因果校验模块：根据全科病人的特征变量集构成第一事件变量集合，根据全科病人的疾病标签变量集构成第二事件变量集合，任意第一事件变量同任意第二事件变量构成一个事件配对；
    构建并训练全科倾向性得分网络，使用训练完成的全科倾向性得分网络计算全科倾向性得分，所述全科倾向性得分表示全科病人在协变量条件下发生第一事件的概率；使用全科倾向性得分计算所有事件配对的因果效应值；
    (3)数据生成模块：对于少样本全科疾病，基于因果校验的生成式对抗网络构建数据生成模型，使用训练完成的数据生成模型生成模拟数据；
    所述数据生成模型包括生成器和判别器，所述生成器和所述判别器迭代交替训练；
    所述生成器的训练过程包括：对于每种少样本全科疾病生成随机噪声，将随机噪声以及对应的疾病标签输入生成器得到生成样本；计算生成样本的所有事件配对的因果效应值；将生成样本以及对应的疾病标签输入判别器，得到判别结果；所述生成器的总损失包括判别器对抗损失、因果损失和正则项损失；所述因果损失为经过少样本全科疾病频率矫正的生成样本的所有事件配对的因果效应值与原始数据的所有事件配对的因果效应值的KL散度损失；
    所述判别器的训练过程包括：从原始数据中随机抽取正样本，并抽取相同数量但与正样本疾病标签不同的负样本；生成相同数量随机噪声，使用生成器得到生成样本；将正样本、负样本、生成样本分别输入判别器，得到判别结果；
    (4)模型预测模块：获取待训练全科病人的特征数据和疾病标签数据，对少样本全科疾病使用数据生成模型生成全科疾病数据；将训练样本以及生成的全科疾病数据共同训练基于全科因果图卷积神经网络的全科多疾病预测模型，使用训练完成的全科多疾病预测模型对全科疾病进行预测。
  2. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述因果校验模块中，使用全科病人的二分类变量数据训练全科倾向性得分网络；将全科病 人的特征变量数据和标签变量数据转换成二分类变量，对于类别变量，通过独热编码转换成二分类变量，对于连续变量，通过分箱转换至类别变量之后通过独热编码转换成二分类变量。
  3. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述全科倾向性得分网络包括输入层、局部连接层、sigmoid激活层和输出层；
    输入层节点个数和输出层节点个数均为第一事件变量集合中的第一事件变量个数M；局部连接层和sigmoid激活层均包含τM个节点，τ≥2；输入层的第u个节点同局部连接层的除第τ(u-1)+1到τu个节点之外的所有节点相连；第τ(u-1)+1到τu个局部连接层节点同第τ(u-1)+1到τu个sigmoid激活层节点一一对应连接；第τ(u-1)+1到τu个sigmoid激活层节点仅同第u个输出层节点相连。
  4. 根据权利要求3所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述全科倾向性得分网络的训练过程如下：
    对于每个第一事件变量a，将训练样本对应的协变量数据输入局部连接层得到倾向性第一特征表示，将所述倾向性第一特征表示输入sigmoid激活层得到倾向性第二特征表示，将所述倾向性第二特征表示输入输出层得到第一事件变量a的预测值；使用所有第一事件变量的预测值同所有第一事件变量的真实值计算倾向性损失。
  5. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述因果校验模块中，使用训练完成的全科倾向性得分网络计算全科病人i对于第一事件变量a的全科倾向性得分使用全科倾向性得分计算第一事件变量a同第二事件变量b的因果效应值ATEa，b，计算公式如下：
    其中n表示待研究病人总数，Ti表示第i个病人第一事件变量真实值；Yi表示第i个病人第二事件变量真实值。
  6. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述数据生成模块中，所述生成器由多层生成器模块构成，所述生成器模块包括规范化层、全连接层和激活层，所述生成器的最后一层生成器模块的激活层为sigmoid激活层；在训练过程中，将随机噪声以及对应的疾病标签输入第一生成器模块的规范化层，将规范化后的数据输入第一生成器模块的全连接层得到输入数据的第一特征表示，将第一特征表示输入第一生成器模块的激活层得到输入数据的第二特征表示，将第二特征表示作为下一层生成器模块的输入数据，最终通过最后一层生成器模块的sigmoid激活层得到生成样本。
  7. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于， 所述数据生成模块中，因果损失Lcausal计算公式如下：
    其中表示原始数据的第一事件变量a同第二事件变量r的因果效应值，表示生成样本的第一事件变量a同第二事件变量r的因果效应值；Ar表示与第二事件变量r配对的第一事件变量集合；所述第二事件变量集合为全科疾病集合，所述第二事件变量r对应少样本全科疾病集合R中的少样本全科疾病r。
  8. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述数据生成模块中，判别器对抗损失Lζ计算公式如下：
    其中N为随机噪声数据量，为第i个生成样本经判别器判别为对应疾病真实数据的概率；
    正则项损失Lregular计算公式如下：
    Lregular＝||w||
    其中||·||表示L1范数，w表示生成器模型参数。
  9. 根据权利要求1所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述数据生成模块中，所述判别器的总损失Ld计算公式如下：
    其中md为正样本数量，yk为正样本对应的疾病标签，分别为抽取的第k个正样本、抽取的第k个负样本、使用生成器得到的第k个生成样本，分别为正样本xk、负样本生成样本dk经判别器判定为疾病yk真实数据的概率。
  10. 根据权利要求1-9中任一项所述的基于因果校验数据生成的全科多疾病预测系统，其特征在于，所述模型预测模块包括：
    构建事件关系图：每个第一事件变量构成事件关系图中的一个第一事件节点，每个第二事件变量构成事件关系图中的一个第二事件节点，对于每个事件配对构建一条边；
    生成第一事件节点和第二事件节点的节点嵌入表示；基于事件关系图构建度数矩阵Φ和邻接矩阵A；使用原始数据的因果效应值构建因果效应矩阵Ψ；
    构建基于全科因果图卷积神经网络的全科多疾病预测模型，所述全科因果图卷积神经网络包含多个因果图卷积模块，所述因果图卷积模块包括因果图卷积层和激活层；
    将节点嵌入表示输入第一因果图卷积模块的因果图卷积层，得到第一图特征表示h(0)：
    其中H(0)表示节点嵌入表示，W(0)表示因果图卷积层权值，I表示单位矩阵，*表示矩阵各元素相乘；
    将h(0)输入第一因果图卷积模块的激活层得到第一因果图卷积模块的输出H(1)；
    将上一因果图卷积模块的输出输入下一因果图卷积模块，直到得到最终疾病预测结果。
PCT/CN2023/089993 2022-05-20 2023-04-23 一种基于因果校验数据生成的全科多疾病预测系统 WO2023221739A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210547826.4 2022-05-20
CN202210547826.4A CN114664452B (zh) 2022-05-20 2022-05-20 一种基于因果校验数据生成的全科多疾病预测系统

Publications (1)

Publication Number Publication Date
WO2023221739A1 true WO2023221739A1 (zh) 2023-11-23

Family

ID=82037230

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/089993 WO2023221739A1 (zh) 2022-05-20 2023-04-23 一种基于因果校验数据生成的全科多疾病预测系统

Country Status (2)

Country Link
CN (1) CN114664452B (zh)
WO (1) WO2023221739A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117577333A (zh) * 2024-01-17 2024-02-20 浙江大学 基于因果特征学习的多中心临床预后预测系统

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114664452B (zh) * 2022-05-20 2022-09-23 之江实验室 一种基于因果校验数据生成的全科多疾病预测系统
CN114864099B (zh) * 2022-07-05 2022-11-01 浙江大学 一种基于因果关系挖掘的临床数据自动化生成方法及系统
CN115862869B (zh) * 2022-12-15 2023-06-09 山东大学 一种基于因果网络不确定性推理的疾病预测预警系统
CN116434969B (zh) * 2023-06-14 2023-09-12 之江实验室 一种基于因果结构不变性的多中心慢性病预测装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210104330A1 (en) * 2019-10-02 2021-04-08 Kenneth Neumann Systems and methods for generating a genotypic causal model of a disease state
CN113298230A (zh) * 2021-05-14 2021-08-24 西安理工大学 一种基于生成对抗网络的不平衡数据集的预测方法
WO2021174723A1 (zh) * 2020-03-02 2021-09-10 平安科技（深圳）有限公司 训练样本扩充方法、装置、电子设备及存储介质
CN113990495A (zh) * 2021-12-27 2022-01-28 之江实验室 一种基于图神经网络的疾病诊断预测系统
CN114220549A (zh) * 2021-12-16 2022-03-22 无锡中盾科技有限公司 一种基于可解释机器学习的有效生理学特征选择和医学因果推理方法
CN114664452A (zh) * 2022-05-20 2022-06-24 之江实验室 一种基于因果校验数据生成的全科多疾病预测系统

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2541450A1 (en) * 2011-07-01 2013-01-02 Stichting Katholieke Universiteit Patient mobile computing system and method for exacerbation prediction
JP6354192B2 (ja) * 2014-02-14 2018-07-11 オムロン株式会社 因果ネットワーク生成システム
WO2019089949A1 (en) * 2017-11-01 2019-05-09 Praeveni, Sbc Methods and systems for a medical screening system
CN112151130B (zh) * 2019-01-15 2022-11-04 合肥工业大学 一种基于文献检索的决策支持系统和构建方法
CN111563533B (zh) * 2020-04-08 2023-05-02 华南理工大学 基于图卷积神经网络融合多种人脑图谱的受试者分类方法
CN113052316B (zh) * 2021-03-01 2022-01-11 浙江师范大学 基于因果推理的知识追踪方法、系统、设备及存储介质
CN113128552B (zh) * 2021-03-02 2024-02-02 杭州电子科技大学 一种基于深度可分离因果图卷积网络的脑电情绪识别方法
CN113112311B (zh) * 2021-05-12 2023-07-25 北京百度网讯科技有限公司 训练因果推断模型的方法、信息提示方法以装置
CN114022725A (zh) * 2021-10-09 2022-02-08 北京鹰瞳科技发展股份有限公司 一种训练多病种转诊系统的方法、多病种转诊系统以及方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210104330A1 (en) * 2019-10-02 2021-04-08 Kenneth Neumann Systems and methods for generating a genotypic causal model of a disease state
WO2021174723A1 (zh) * 2020-03-02 2021-09-10 平安科技（深圳）有限公司 训练样本扩充方法、装置、电子设备及存储介质
CN113298230A (zh) * 2021-05-14 2021-08-24 西安理工大学 一种基于生成对抗网络的不平衡数据集的预测方法
CN114220549A (zh) * 2021-12-16 2022-03-22 无锡中盾科技有限公司 一种基于可解释机器学习的有效生理学特征选择和医学因果推理方法
CN113990495A (zh) * 2021-12-27 2022-01-28 之江实验室 一种基于图神经网络的疾病诊断预测系统
CN114664452A (zh) * 2022-05-20 2022-06-24 之江实验室 一种基于因果校验数据生成的全科多疾病预测系统

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117577333A (zh) * 2024-01-17 2024-02-20 浙江大学 基于因果特征学习的多中心临床预后预测系统
CN117577333B (zh) * 2024-01-17 2024-04-09 浙江大学 基于因果特征学习的多中心临床预后预测系统

Also Published As

Publication number Publication date
CN114664452A (zh) 2022-06-24
CN114664452B (zh) 2022-09-23

Similar Documents

Publication Publication Date Title
WO2023221739A1 (zh) 一种基于因果校验数据生成的全科多疾病预测系统
Das et al. Group decision making in medical system: An intuitionistic fuzzy soft set approach
CN106778014A (zh) 一种基于循环神经网络的患病风险预测方法
Norgaard et al. Synthetic sensor data generation for health applications: A supervised deep learning approach
Salih et al. Modified alexnet convolution neural network for covid-19 detection using chest x-ray images
CN109119156A (zh) 一种基于bp神经网络的医疗诊断系统
WO2024008043A1 (zh) 一种基于因果关系挖掘的临床数据自动化生成方法及系统
Hussain et al. Novel deep learning architecture for predicting heart disease using CNN
CN111882066A (zh) 基于深度表征学习的反事实推理设备
KR20110098286A (ko) 퍼지 추론기법을 이용한 한방 자가 진단방법
US20230181120A1 (en) System and method for testing for sars-cov-2/covid-19 based on wearable medical sensors and neural networks
Irungu et al. A CNN Transfer Learning-Electrocardiogram (ECG) Signal Approach to Predict COVID-19
Sridhar et al. Mobile application development for disease diagnosis based on symptoms using machine learning techniques
CN110598789A (zh) 一种基于模糊感知器的人体疲劳状态预测方法及系统
JP7247292B2 (ja) 加齢による黄斑変性症用の分類モデルを学習させる電子機器及び方法
CN113379779B (zh) 堆叠宽度学习系统的边缘计算方法、装置、介质和设备
Azam et al. Classification of COVID-19 Symptoms Using Multilayer Perceptron
CN113593698B (zh) 一种基于图注意网络的中医证型识别方法
Ladić et al. Face mask classification using MLP Classifier
CN116936108B (zh) 一种面向不平衡数据的疾病预测系统
TWI768951B (zh) 訓練老年性黃斑部退化的分類模型的電子裝置和方法
BELHIA et al. Automatic Detection of COVID-19 based in Artificial Intelligence Tools
Sun et al. Select informative symptoms combination for diagnosing syndrome
Ligi et al. COVID-19 Detection from Pulmonary CT Images using Neural Networks based on Dropout-Driven Hidden Layers
Afreen et al. Deep Convlutional models on medical data for detection of Covid Lung X-Rays through Transfer Learning

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: 23806692

Country of ref document: EP

Kind code of ref document: A1