CN117034142B - Unbalanced medical data missing value filling method and system - Google Patents

Abstract

The invention discloses a method and system for filling missing values in unbalanced medical data. The invention uses bulldozer distance to construct the loss of the generator and the discriminator, which can solve the problem of possible vanishing gradients in the generator during the training process; the patient label is used as The supervision signal is added to the generator to increase the diversity of patient data generated by the generator; an auxiliary classifier is added to predict the patient data after the padding unit is filled, and the prediction results are fed back to the generator to improve the generation effect of the generator. ; Use random numbers to fill in the missing parts of the patient data, use the filled patient data as the input of the generator, and learn the relationship between the missing values and other data through the generator, avoiding the problem of collecting enough complete samples during the training process ;The generator loss consists of three parts. By constructing different losses, the generator can consider the effect of filling from different angles, thereby improving the accuracy of the filling results.

Description

A method and system for filling missing values in unbalanced medical data

Technical field

The invention belongs to the field of medical information technology, and in particular relates to a method and system for filling missing values in unbalanced medical data.

Background technique

Electronic Health Records (EHR, Electronic Health Records) store information related to patient treatment, including basic patient information, diagnosis information, examination information, medication information, etc. This information provides the basis for medical data mining. However, due to factors such as collection equipment failure and unstable transmission, there will be a large amount of missing data in electronic health records. These missing data will not only increase the complexity and difficulty of statistical analysis, but also lead to inaccurate analysis results. Therefore, solving the missing value filling problem in electronic health records is of great significance to improving the quality of data mining.

Generative Adversarial Networks (GAN) is a neural network that captures the distribution of training data and creates new data through the learned data distribution. It is currently commonly used in image generation, text generation and other fields. In recent years, some experts and scholars have also applied the GAN method to the field of filling missing data. However, in real life, because the electronic medical record data of hospital patients are often unbalanced, the number of patients with different types of diseases varies greatly. If you directly use There are some problems when the GAN method is used to fill missing values in unbalanced medical data. On the one hand, the filling effect lacks diversity. On unbalanced samples, the generator deceives the discriminator by only focusing on the filling quality of types with a large number of samples and ignoring the filling quality of those types with a small number of data, resulting in final filling. The data only belongs to a certain type of disease. On the other hand, GAN methods are trained on imbalanced data, and the generator is more prone to the vanishing gradient problem. The "Wasserstein GAN" article points out that under the optimal discriminator, minimizing the loss of the generator is equivalent to minimizing the JS divergence (JSD, Jensen-Shannon Divergence) between the real distribution and the generated distribution. When the real distribution and the generated distribution When the distributions do not overlap or the overlapping part can be ignored, the JS divergence is a fixed constant log2. At this time, the gradient of the generator disappears, making it difficult to train the network.

Contents of the invention

The purpose of the present invention is to provide a method and system for filling missing values of unbalanced medical data based on a generative adversarial network to improve the quality of filling missing values of medical data in view of the shortcomings of the existing technology.

The purpose of the present invention is achieved through the following technical solutions:

In a first aspect, the present invention provides a method for filling missing values in unbalanced medical data, which method includes:

Use the hospital's information system to obtain patient data;

Use data filling models to fill in missing values in patient data;

The data filling model includes a data processing unit, a generator, a filling unit, a discriminator and an auxiliary classifier; the generator and the discriminator constitute a generative adversarial network;

In the data processing unit, a mask matrix is used to record the locations of missing values in the patient's original data, 0 is used to pre-fill the missing values in the patient's original data, random numbers are used to fill in the missing values in the patient's original data, and the data is input into the generator;

The generator is used to learn the distribution of input patient data, generate new patient data, and input the filling unit, and the input of the generator includes patient data and patient labels;

The filling unit is used to fill in missing values in the original patient data using the new patient data generated by the generator;

The discriminator is used to identify each input patient data and determine whether it is an observation value. The input of the discriminator includes the patient data after filling in the filling unit and the patient after using 0 to pre-fill the missing values in the patient's original data. Data, the output is the probability that each patient's data is an observation;

The auxiliary classifier is used to predict the patient data after filling in the filling unit, and feed the prediction results back to the generator;

The training process includes pre-training the auxiliary classifier and filling the model with formal training data. During the pre-training process, the auxiliary classifier is trained using non-missing patient data and the network parameters of the auxiliary classifier are determined. The network parameters of the auxiliary classifier are not involved in the formal training process. Update: During the formal training process, the discriminator is trained first and then the generator is trained. The discriminator and generator are continuously trained against each other until the data filling model converges;

Input the patient data and patient labels that need to fill in missing values into the trained data filling model. After passing through the data processing unit, generator and filling unit, the filled patient data is output.

Further, the acquired patient data is preprocessed before being input into the data filling model, specifically: performing one-hot encoding operations on discrete data, and performing maximum and minimum value normalization operations on continuous data.

Further, the patient’s original data is recorded as , of which/> Represents the original data of the i-th patient, n is the number of patients, k is the number of features; the mask matrix is recorded as/> , of which/> Used to mark the observed values and missing values in the i-th patient's original data. The observed values are 1 and the missing values are 0; 0 is used to pre-fill the missing values in the patient's original data. The filled data matrix is marked as/> , of which/> Represents the patient data after pre-filling the missing values in the original data of the i-th patient with 0; creating a random matrix is recorded as/> , of which/> is a randomly generated random number vector that conforms to the standard normal distribution, used to fill in the missing values in the i-th patient's original data; use the random numbers in the random matrix to fill in the missing values in the patient's original data, and the filled data matrix records for/> , of which/> Represents the patient data obtained after using random numbers to fill the missing values in the original data of the i-th patient,/> ,/> Represents Hadama product.

Further, the loss function of the generator consists of three parts. The first part is to calculate the difference between the observation value generated by the generator and the actual observation value, using the mean square error as the loss function; the second part is to generate the adversarial network. The generator loss uses Wasserstein distance as the loss function; the third part of the loss is to calculate the gap between the predicted label of the patient data filled by the auxiliary classifier on the padded unit and the patient's true label, using the cross-entropy function as the loss function.

Furthermore, the loss function of the generator is ;

The first part of the loss function ;

The second part of the loss function ;

The third part of the loss function ;

in Represents the output value of the generator when the i-th patient data is used as input, /> , G() represents the patient data obtained after passing through the generator, _yi represents the real label of the i-th patient, D() represents the result of the patient data passing through the discriminator, t _i represents the filled-in original data of the i-th patient Patient data after cell filling,/> Represents the predicted label of the i-th patient by the auxiliary classifier, /> and/> is a hyperparameter, · represents the vector inner product.

Further, in the filling unit, the patient data generated by the generator is used Fill in the missing values in the patient's original data X, and the filled data matrix is marked as/> , where t _i represents the patient data after the original data of the i-th patient has been filled by the padding unit,/> , of which/> Represents the output value of the generator when the i-th patient data is used as input.

Further, the calculation formula of the loss function _LD of the discriminator is as follows:

;

where D() represents the result obtained after the patient data passes through the discriminator, represents the patient data after pre-filling the missing values in the i-th patient's original data with 0, t _i represents the patient data after filling the i-th patient's original data with the padding unit, and · represents the vector inner product.

Further, the calculation formula of the loss function _LD of the discriminator is as follows:

;

where D() represents the result obtained after the patient data passes through the discriminator, represents the patient data after pre-filling the missing values in the i-th patient's original data with 0, t _i represents the patient data after filling the i-th patient's original data with the padding unit, and · represents the vector inner product.

Furthermore, in the process of formally training the data to fill the model, first input the patient data containing missing values, the discriminator calculates the loss, and the gradient back propagation updates the discriminator network parameters; then the generator calculates the loss, and the gradient back propagation updates the generator network Parameters; the discriminator and generator are continuously trained against each other until the data-filled model converges.

In a second aspect, the present invention provides a system for filling missing values in unbalanced medical data. The system includes a data acquisition module, a data filling model building module and a data filling module; the data acquisition module is used to obtain patients using the hospital's information system. data;

The data filling model building module is used to build and train a data filling model; the data filling model includes a data processing unit, a generator, a filling unit, a discriminator and an auxiliary classifier, and the generator and the discriminator constitute a generative adversarial network;

In the data processing unit, a mask matrix is used to record the locations of missing values in the patient's original data, 0 is used to pre-fill the missing values in the patient's original data, random numbers are used to fill in the missing values in the patient's original data, and the data is input into the generator;

The generator is used to learn the distribution of input patient data, generate new patient data, and input the filling unit, and the input of the generator includes patient data and patient labels;

The filling unit is used to fill in missing values in the original patient data using the new patient data generated by the generator;

The discriminator is used to identify each input patient data and determine whether it is an observation value. The input of the discriminator includes the patient data after filling in the filling unit and the patient after using 0 to pre-fill the missing values in the patient's original data. Data, the output is the probability that each patient's data is an observation;

The auxiliary classifier is used to predict the patient data after filling in the filling unit, and feed the prediction results back to the generator;

The training process includes pre-training the auxiliary classifier and filling the model with formal training data. During the pre-training process, the auxiliary classifier is trained using non-missing patient data and the network parameters of the auxiliary classifier are determined. The network parameters of the auxiliary classifier are not involved in the formal training process. Update: During the formal training process, the discriminator is trained first and then the generator is trained. The discriminator and generator are continuously trained against each other until the data filling model converges;

The data filling module is used to input the patient data and patient labels that need to fill missing values into the trained data filling model, and output the filled patient data after passing through the data processing unit, generator and filling unit.

In a third aspect, the present invention provides a device for filling missing values in unbalanced medical data, which includes a memory and one or more processors. The memory stores executable code. When the processor executes the executable code, Implement the missing value filling method for imbalanced medical data as described in the first aspect.

In a fourth aspect, the present invention provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the method for filling missing values in unbalanced medical data as described in the first aspect is implemented.

The beneficial effects of the present invention are:

1. This invention uses bulldozer distance (Wasserstein distance) instead of JS divergence to construct the loss of the generator and discriminator. Wasserstein distance has superior smoothing properties compared to JS divergence. Even if the two distributions do not overlap, Wasserstein distance can still reflect their Far and near, it can solve the vanishing gradient problem that may occur in the generator during the training process.

2. The present invention adds patient labels to the generator as supervision signals, helping the generator to identify different patient data in unbalanced medical electronic medical records, and increasing the diversity of patient data generated by the generator.

3. The present invention adds an auxiliary classifier to classify and predict the patient data after filling in the filling units, and feeds the prediction results back to the generator to improve the generation effect of the generator.

4. The present invention uses random numbers to fill in the missing parts of patient data, uses the filled patient data as the input of the generator, and learns the relationship between the missing values and other data through the generator, avoiding the need to collect enough complete data during the training process. Sample question.

5. The generator loss proposed by this invention consists of three parts, which are the loss between the patient observation values generated by the generator and the patient's actual observation values, and the loss between the discriminator's prediction of the patient's missing values generated by the generator and the true value. The loss, the loss between the predicted label of the patient data filled by the auxiliary classifier and the real label of the patient after filling in the filling unit, by constructing different losses, allows the generator to consider the effect of filling from different angles, thereby improving the accuracy of the filling results .

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the drawings of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of a method for filling missing values in unbalanced medical data provided by an exemplary embodiment;

Figure 2 is a table form of patient original data provided by an exemplary embodiment;

Figure 3 is a schematic diagram of the data filling model architecture provided by an exemplary embodiment;

Figure 4 is a schematic diagram of the processing process of the data processing unit provided by an exemplary embodiment;

Figure 5 is a structural diagram of a system for filling missing values in unbalanced medical data provided by an exemplary embodiment;

Figure 6 is a structural diagram of a device for filling missing values in unbalanced medical data provided by an exemplary embodiment.

Detailed ways

In order to better understand the technical solution of the present application, the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

It should be clear that the described embodiments are only some of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of this application.

The terminology used in the embodiments of the present application is only for the purpose of describing specific embodiments and is not intended to limit the present application. As used in the embodiments and the appended claims, the singular forms "a," "the" and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise.

The present invention provides a method for filling missing values in unbalanced medical data, as shown in Figure 1. The method includes two parts: data acquisition and data filling. The data acquisition part uses the hospital's information system to extract patient structured data, and the data filling part Use the data filling model to fill in the missing values in the patient data. The specific implementation process of each part is explained in detail below.

1. Data acquisition

First, the hospital's information system is used to extract patient structured data, which includes basic patient information, diagnosis results, examination information, medication information, etc.

Then perform data preprocessing on the extracted patient data, specifically:

Perform a one-hot encoding operation on the extracted discrete data. The discrete data includes patient characteristics such as diagnosis and medication;

Perform maximum and minimum normalization operations on the extracted continuous data. The continuous data includes the patient's weight, age, blood pressure and other characteristics. The maximum and minimum normalization formula is: , where x _ij represents the value of the j-th feature of the i-th patient, /> Represents the minimum value in the j-th feature,/> Represents the maximum value in the jth feature.

Finally, the patient diagnosis result is used as the patient label, recorded as , where n represents the number of patients,/> is the label of the i-th patient, u is the disease type of the diagnosis result in the extracted patient data, which is 10 in this embodiment. The patient's original data is represented by , of which/> represents the original data of the i-th patient, and k is the number of extracted patient features. As shown in Figure 2, the patient _'s original data

2. Data filling

Use the data imputation model to fill in missing values in patient data. As shown in Figure 3, the data filling model includes a data processing unit, a generator G, a filling unit, a discriminator D, and an auxiliary classifier C. The generator G and the discriminator D constitute a generative adversarial network.

Specifically, the role of the data processing unit is to use random numbers to fill in the missing values of the patient's original data. Since the patient's original data obtained by the data acquisition part contains missing values, normal numerical operations cannot be performed. The data processing unit needs to be used to fill in the missing values containing Patient raw data with missing values were processed. The role of the generator is to learn the distribution of input patient data and generate new patient data. The function of the imputation unit is to use the new patient data generated by the generator to fill in the missing values of the original patient data. The function of the discriminator is to identify each input data and determine whether the input data is an observation value (no missing value). The role of the auxiliary classifier is to predict the filled patient data and feed the prediction results back to the generator. Next, each part of the data filling model is introduced separately.

2.1 Data processing unit

Since the original patient data obtained by the data acquisition part contains missing values, normal numerical operations cannot be performed. In order to enable normal numerical operations on these data, a data processing unit is required to process these original patient data. The processing process includes two parts. .

The first part is to record the location of missing values in the patient's raw data and prefill the missing values in the patient's raw data with 0. First use the mask matrix Record the location of missing values in the patient's original data, the mask vector in the mask matrix M/> Used to mark which positions in the original data of the i-th patient are observed values and which positions are missing values, and use 1 to represent observed values and 0 to represent missing values. For example, [1,1,0,1] indicates that the third feature of the patient is a missing value, and other features have observed values. Then fill the missing values in the patient's original data with 0 and use the data matrix /> Represents the populated patient data, where/> Represents the patient data after pre-filling the missing values in the original data of the i-th patient with 0.

The second part is to use random numbers to fill in the missing values in the patient's raw data. First create a random matrix , of which/> is a randomly generated random number vector that conforms to the standard normal distribution and is used to fill in the missing values in the original data of the i-th patient,/> Represents the k-th feature used to populate the original data of the i-th patient. Then use/> Represents the patient data obtained after using random numbers to fill the missing values in the original data of the i-th patient,/> by/> Calculated, where/> Represents Hadama product. and use data matrix Represents the patient data after filling all missing values in the patient's original data with random numbers.

Figure 4 is an example of the processing process of the data processing unit, in which the patient's original data X obtained through the data acquisition part contains missing values; data matrix It is the patient data after pre-filling the missing values in the patient's original data s position.

2.2 Generator

The generator is used to learn the distribution of input patient data and generate new patient data. Its inputs include patient data and patient labels, where the patient labels are used as supervision signals to allow the generator to learn the labels of each patient. The generator consists of three layers of fully connected networks, with the number of nodes in each layer being k, k, k, where k represents the feature dimension of the patient's original data X. The first two layers are hidden layers, the last layer is the output layer, the activation function of the first two layers is ReLU, the activation function of the last layer is Tanh, and the root mean square backpropagation function (RMSprop function) is used as the optimization of the generator function. Working with data matrices

represents the output of the generator, where /> Represents the output value of the generator when the i-th patient data is used as input, /> Represents the patient data after using random numbers to fill the missing values in the original data of the i-th patient, y _i represents the label of the i-th patient, and G() represents the patient data obtained after passing through the generator G.

The loss of the entire generator consists of three parts:

The first part of the loss L ₁ is to calculate the difference between the patient observation value generated by the generator and the patient's actual observation value. Here, the mean square error is used as this part of the loss function. When the difference between the patient observation value generated by the generator and the patient's actual observation value The smaller the gap, the closer the patient observation values generated by the generator are to the actual patient observation values.

The second part of the loss L ₂ is the generator loss of the traditional generative adversarial network, where Wasserstein distance is used instead of cross entropy as the loss function. The Wasserstein distance can be passed Approximately calculated, where D(T) represents the result of the patient data T filled by the padding unit passing through the discriminator D,/> represents the Hadamard product, M represents the mask matrix marking the position of missing values in the patient's original data, and E[] represents the mathematical expectation. When the patient missing values generated by the generator are closer to the real values, the patient missing values generated by the generator are more likely to be identified as observed values by the discriminator, and the L ₂ value will be smaller at this time, and vice versa.

The third part of the loss _L3 is to calculate the gap between the predicted label of the patient data after the auxiliary classifier has filled the padding unit and the patient's true label. Here, the cross entropy function is used as the loss function. When the auxiliary classifier fills the padding unit The smaller the gap between the predicted label of the patient data and the patient's true label, the better the filled data is.

The loss function L _G of the entire generator is as follows:

;

in Represents the patient data after pre-filling the missing values in the original data of the i-th patient with 0, /> Represents the output value of the generator when the i-th patient data is used as input, /> represents the Hadamard product, · represents the vector inner product, M is the mask matrix, /> is a mask vector, which records which positions in the original data of the i-th patient are observed values (no missing values) and which positions are missing values, /> is to calculate the mean square error between the observation value generated by the generator of the i-th patient and the actual observation value,/> Represents the patient data after using random numbers to fill the missing values in the original data of the i-th patient, G() represents the patient data obtained after passing through the generator G, y _i represents the true label of the i-th patient, /> represents the predicted label of the i-th patient by the auxiliary classifier; T represents the patient data after the padding unit is filled, t _i represents the patient data after the i-th patient data is filled by the padding unit, and D() represents the patient data after passing through the discriminator. The result obtained;/> and/> are hyperparameters, which are 0.3 and 0.2 respectively in this embodiment.

2.3 Filling units

The padding unit is the patient data generated using the generator Fill in the missing values in the original patient data X and output the filled patient data, use /> Represents the patient data after the padding unit is filled, where t _i represents the patient data after the i-th patient’s original data is filled by the padding unit, and t _i is given by/> Calculated, where/> Represents the output value of the generator when the i-th patient data is used as input, /> represents the patient data after pre-filling the missing values in the original data of the i-th patient with 0, m _i represents the mask vector of the i-th patient, /> Represents Hadama product.

2.4 Discriminator

The purpose of the discriminator is to determine whether each input data is an observation value. Its inputs are the patient data T after filling in the padding unit and the patient data after using 0 to pre-fill the missing values in the patient's original data. , whose outputs are for T and/> Each data in is the probability of an observation. The discriminator consists of three layers of fully connected networks. The number of nodes in each layer is k, k, k, where k represents the feature dimension of the patient's original data X. The first two layers are hidden layers, the activation function is ReLU, and the last layer is the output layer. , no activation function, and use the RMSprop function as the optimization function of the discriminator. The loss of the discriminator is to calculate the distribution difference between the original patient data and the patient data after padding unit padding, using Wasserstein distance instead of JS divergence to measure the distribution difference between the patient original data and the patient data after padding unit padding. The loss function _LD of the discriminator is calculated as follows:

;

where D() represents the result obtained after the patient data passes through the discriminator, represents the patient data after pre-filling the missing values in the i-th patient's original data with 0, m _i represents the mask vector of the i-th patient, and t _i represents the patient data after the i-th patient's original data has been filled by the padding unit. During the training process, when _LD is smaller, the Wasserstein distance indicating the true distribution and the generated distribution is smaller, and the data filling training is better.

2.5 Auxiliary classifier

The auxiliary classifier predicts the patient data after filling the filled units. During the pre-training process of the data filling model, the auxiliary classifier is first trained using non-missing patient data, and the network parameters of the auxiliary classifier are determined. During the formal training process of the data filling model, the network parameters of the auxiliary classifier do not participate in the update. The entire auxiliary classifier consists of three layers of fully connected networks, the first two layers are hidden layers, and the last layer is the output layer. The number of nodes in the first two layers is artificially set. In this embodiment, it is set to 128 and 64 respectively. The number of nodes in the last layer is u. u is the disease type of the diagnosis result in the extracted patient data. In this embodiment, it is 10. . The activation function of the first two layers is ReLU, and the activation function of the last layer is Softmax. The loss of the auxiliary classifier is calculated by calculating the difference between the predicted label of the auxiliary classifier and the patient's true label. When the gap is smaller, it means that the prediction effect of the auxiliary classifier is better. Here, the cross entropy function is used as the loss function of the auxiliary classifier to assist classification. The calculation formula of the loss function L _C of the device is as follows:

;

in is the label of the i-th patient, which is obtained from the diagnosis result in the patient data through a one-hot encoding operation, u is the disease type of the diagnosis result in the obtained patient data,/> Represents the predicted label of the i-th patient by the auxiliary classifier, which is a vector of length u. Each value in the vector represents the probability that the auxiliary classifier predicts the patient to suffer from the corresponding disease. , t _i represents the patient data after the i-th patient's original data has been filled by the padding unit, C() represents the result obtained after the patient data passes through the auxiliary classifier, · represents the vector inner product.

The training of the entire data filling model is divided into the following two stages:

The first stage is the pre-training process. The purpose of pre-training is to train the auxiliary classifier and determine the network parameters of the auxiliary classifier. When training the auxiliary classifier, first initialize the auxiliary classifier network parameters, then use non-missing patient data as training data to calculate the loss function of the auxiliary classifier, and gradient backpropagation updates the auxiliary classifier network parameters until the auxiliary classifier converges. After the auxiliary classifier network parameters are determined, the auxiliary classifier network parameters will not be updated during the formal training process of the second-stage data filling model.

The second stage is the process of training the data filling model. During the data filling model training process, the training strategy is to first train the discriminator and then the generator. When training data to fill the model, first input patient data containing missing values, the discriminator calculates the loss, and gradient backpropagation updates the discriminator network parameters; then the generator calculates the loss, gradient backpropagation updates the generator network parameters, the discriminator and The generator continues adversarial training until the data filling model converges; in the early stage of training, the discriminator can easily distinguish which data are observation values and which data are filling values. As the training deepens, the generator learns the distribution of patient data and generates The data is very close to the patient's observation values, and the discriminator cannot determine which data are observation values and which data are filling values. The generator and discriminator reach Nash equilibrium, and the trained data filling model converges at this time.

After the data filling model is trained, the model can be used to fill in missing values in patient data. During the filling process, first perform data preprocessing on the patient data that needs to be filled, perform a one-hot encoding operation on the discrete data in the patient data that needs to be filled, and perform a maximum and minimum value normalization operation on the continuous data. Then select the patient diagnosis result as the patient label, use the patient data and patient labels containing missing values as the input of the data filling model, and output the filled patient data through the data processing unit, generator and filling unit.

On the other hand, the present invention also provides an unbalanced medical data missing value filling system, as shown in Figure 5. The system includes a data acquisition module, a data filling model building module and a data filling module; the data acquisition module is used to Use the hospital's information system to obtain patient data;

The data filling model building module is used to build and train a data filling model; the data filling model includes a data processing unit, a generator, a filling unit, a discriminator and an auxiliary classifier, and the generator and the discriminator constitute a generative adversarial network;

In the data processing unit, a mask matrix is used to record the locations of missing values in the patient's original data, 0 is used to pre-fill the missing values in the patient's original data, random numbers are used to fill in the missing values in the patient's original data, and the data is input into the generator;

The generator is used to learn the distribution of input patient data, generate new patient data, and input the filling unit, and the input of the generator includes patient data and patient labels;

The filling unit is used to fill in missing values in the original patient data using the new patient data generated by the generator;

The discriminator is used to identify each input patient data and determine whether it is an observation value. The input of the discriminator includes the patient data after filling in the filling unit and the patient after using 0 to pre-fill the missing values in the patient's original data. Data, the output is the probability that each patient's data is an observation;

The auxiliary classifier is used to predict the patient data after filling in the filling unit, and feed the prediction results back to the generator;

The training process includes pre-training the auxiliary classifier and filling the model with formal training data. During the pre-training process, the auxiliary classifier is trained using non-missing patient data and the network parameters of the auxiliary classifier are determined. The network parameters of the auxiliary classifier are not involved in the formal training process. Update: During the formal training process, the discriminator is trained first and then the generator is trained. The discriminator and generator are continuously trained against each other until the data filling model converges;

The data filling module is used to input the patient data and patient labels that need to fill missing values into the trained data filling model, and output the filled patient data after passing through the data processing unit, generator and filling unit.

Corresponding to the foregoing embodiment of a method for filling missing values in unbalanced medical data, the present invention also provides an embodiment of a device for filling missing values in unbalanced medical data.

Referring to Figure 6, an embodiment of the present invention provides a device for filling missing values in unbalanced medical data, including a memory and one or more processors. The memory stores executable code, and the processor executes the executable code. The code is used to implement a missing value filling method for unbalanced medical data in the above embodiment.

The embodiment of a device for filling missing values in unbalanced medical data provided by the present invention can be applied to any device with data processing capabilities, and any device with data processing capabilities can be a device or device such as a computer. The device embodiment may be implemented by software, or may be implemented by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running them through the processor of any device with data processing capabilities. From the hardware level, as shown in Figure 6, it is a hardware structure diagram of any device with data processing capabilities where the unbalanced medical data missing value filling device provided by the present invention is located. In addition to the processor shown in Figure 6 , memory, network interface, and non-volatile memory, any device with data processing capabilities where the device is located in the embodiment may also include other hardware according to the actual functions of any device with data processing capabilities. In this regard No longer.

The specific implementation process of the functions and roles of each unit in the above equipment can be found in the implementation process of the corresponding steps in the above method, and will not be described again here.

As for the equipment embodiment, since it basically corresponds to the method embodiment, please refer to the partial description of the method embodiment for relevant details. The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

An embodiment of the present invention also provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the method for filling missing values of unbalanced medical data in the above embodiment is implemented.

The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities as described in any of the foregoing embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart memory card (SMC), an SD card, or a flash memory card equipped on the device. (Flash Card) etc. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

The above embodiments are used to illustrate the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modifications and changes made to the present invention fall within the protection scope of the present invention.

Claims (10)

1. A method of filling an unbalanced medical data loss value, comprising:

acquiring patient data by using an informatization system of a hospital;

filling the missing values in the patient data by using a data filling model;

the data filling model comprises a data processing unit, a generator, a filling unit, a discriminator and an auxiliary classifier; the generator and the discriminator form a generating countermeasure network;

the data processing unit records the position of the missing value in the original data of the patient by using a mask matrix, pre-fills the missing value in the original data of the patient by using 0, fills the missing value in the original data of the patient by using a random number, and inputs the missing value into the generator;

the generator is used for learning the distribution of the input patient data, generating new patient data and inputting a filling unit, and the input of the generator comprises the patient data and a patient label;

the filling unit is used for filling the missing value in the original patient data by utilizing the new patient data generated by the generator;

the input of the discriminator comprises patient data filled by the filling unit and patient data filled with the missing value in the original patient data by 0, and the probability of each patient data being an observed value is output;

the auxiliary classifier is used for predicting the patient data filled by the filling unit and feeding back a prediction result to the generator;

the training process comprises pre-training the auxiliary classifier and formally training a data filling model, wherein in the pre-training process, the auxiliary classifier is trained by using undelayed patient data, and the network parameters of the auxiliary classifier are determined, and in the formally training process, the network parameters of the auxiliary classifier do not participate in updating; training the discriminator and then training the generator in the formal training process, wherein the discriminator and the generator are used for continuously performing countermeasure training until the data filling model converges;

patient data and patient labels which need to be filled with missing values are input into a trained data filling model, and the filled patient data is output after passing through a data processing unit, a generator and a filling unit.

2. The method for filling unbalanced medical data loss values according to claim 1, wherein the data filling model is input after the acquired patient data is subjected to data preprocessing, specifically: and performing single-heat encoding operation on the discrete data, and performing maximum and minimum normalization operation on the continuous data.

3. The method of claim 1, wherein the patient raw data is recorded asWherein->Raw data representing the ith patient, n being the number of patients, k being the number of features; mask matrix is marked as->Wherein->The method is used for marking the observed value and the missing value in the original data of the ith patient, wherein the observed value is 1, and the missing value is 0; pre-filling the missing values in the original data of the patient with 0, the filled data matrix is marked +.>Wherein->Representing patient data after the i-th patient raw data is prefilled with 0; creating a random matrix mark as +.>Wherein->The random number vector which is randomly generated and accords with standard normal distribution is used for filling the missing value in the original data of the ith patient; filling the missing values in the original data of the patient by using random numbers in a random matrix, and marking the filled data matrix as

Wherein->Representing patient data obtained after filling the missing values in the ith patient raw data with random numbers, a->，/>Representing the hadamard product.

4. A method of filling unbalanced medical data loss values according to claim 3, wherein the generator's loss function is composed of three parts, the first part being the calculation of the difference between the observed value generated by the generator and the actual observed value, using the mean square error as the loss function; the second part is to generate the generator loss against the network, using the wasperstein distance as a loss function; the third partial loss is the difference between the prediction label of the patient data filled by the filling unit and the real label of the patient by the auxiliary classifier, and a cross entropy function is used as a loss function.

5. The method of claim 4, wherein the generator has a loss function；

First partial loss function；

Second partial loss function；

Third partial loss functionNumber of digits；

Wherein the method comprises the steps ofOutput value of generator representing ith patient data as input, +.>G () represents patient data obtained after passing through the generator, y _i Representing the actual label of the ith patient, D () represents the result of patient data after passing through the arbiter, t _i Representing patient data filled with the ith patient raw data via the filling unit,/patient data filled with the ith patient raw data via the filling unit>Predictive tag for the i patient representing the auxiliary classifier,>and->Is a hyper-parameter, representing the vector inner product.

6. A method of filling an unbalanced medical data loss value according to claim 3, wherein the shim cells are filled with patient data generated by a generatorFilling up the missing value in the original data X of the patient, and marking the filled up data matrix as +.>Wherein t is _i Representing patient data filled with the ith patient raw data via the filling unit,/patient data filled with the ith patient raw data via the filling unit>Wherein->Represents the output value of the generator when the ith patient data is input.

7. A method of filling an unbalanced medical data loss value according to claim 3, wherein the loss function L of the arbiter _D The calculation formula is as follows:

；

where D () represents the result of patient data after passing through the arbiter,representing patient data after prefilling the missing values in the ith patient raw data with 0, t _i Representing the patient data filled in by the filling unit with the ith patient raw data, representing the vector inner product.

8. The method for filling unbalanced medical data loss values according to claim 1, wherein the auxiliary classifier has a loss function L _C The calculation formula is as follows:

；

wherein the method comprises the steps ofThe label of the ith patient is obtained from the diagnosis result in the patient data through a single-heat coding operation, u is the disease type of the diagnosis result in the patient data,/->The predictive label of the auxiliary classifier for the ith patient is a vector with length u, and the vector isAnd each value of (2) represents the probability that the auxiliary classifier predicts that the patient suffers from the corresponding disease, n is the number of patients,，t _i representing the patient data after the i-th patient raw data is padded by the padding unit, C () represents the result obtained after the patient data is passed through the auxiliary classifier, and C represents the vector inner product.

9. The method for filling an unbalanced medical data loss value according to any one of claims 1 to 8, wherein in the process of formally training a data filling model, patient data containing the loss value is firstly input, loss is calculated by the discriminator, and network parameters of the discriminator are updated by gradient back propagation; then the generator calculates the loss, and the gradient back propagation updates the generator network parameters; the arbiter and generator continue the countermeasure training until the data population model converges.

10. The unbalanced medical data missing value filling system is characterized by comprising a data acquisition module, a data filling model construction module and a data filling module; the data acquisition module is used for acquiring patient data by using an informatization system of a hospital;

the data filling model construction module is used for constructing and training a data filling model; the data filling model comprises a data processing unit, a generator, a filling unit, a discriminator and an auxiliary classifier, wherein the generator and the discriminator form a generating countermeasure network;

the data processing unit records the position of the missing value in the original data of the patient by using a mask matrix, pre-fills the missing value in the original data of the patient by using 0, fills the missing value in the original data of the patient by using a random number, and inputs the missing value into the generator;

the generator is used for learning the distribution of the input patient data, generating new patient data and inputting a filling unit, and the input of the generator comprises the patient data and a patient label;

the filling unit is used for filling the missing value in the original patient data by utilizing the new patient data generated by the generator;

the input of the discriminator comprises patient data filled by the filling unit and patient data filled with the missing value in the original patient data by 0, and the probability of each patient data being an observed value is output;

the auxiliary classifier is used for predicting the patient data filled by the filling unit and feeding back a prediction result to the generator;

the training process comprises pre-training the auxiliary classifier and formally training a data filling model, wherein in the pre-training process, the auxiliary classifier is trained by using undelayed patient data, and the network parameters of the auxiliary classifier are determined, and in the formally training process, the network parameters of the auxiliary classifier do not participate in updating; training the discriminator and then training the generator in the formal training process, wherein the discriminator and the generator are used for continuously performing countermeasure training until the data filling model converges;

the data filling module is used for inputting patient data and patient labels which need to be filled with missing values into a trained data filling model, and outputting the filled patient data after passing through the data processing unit, the generator and the filling unit.