CN109658996B - Physical examination data completion method and device based on side information and application - Google Patents

Abstract

The invention discloses a medical examination data completion method based on side information, including (1) constructing and complementing a medical examination-disease matrix, a pathogenic factor-disease matrix, and a pathogenic factor-physical examination matrix according to the side information; (2) respectively Establish encoding and decoding networks D2F Net, D2C Net and F2C Net between any two matrices; (3) Jointly train D2F Net, D2C Net and F2C Net, after training, the pathogenic factor-disease matrix and the pathogenic factor-physical examination matrix (4) Input the medical examination-disease matrix to be completed into D2F Net, D2C Net, and use the completed pathogenic factor-disease matrix, pathogenic factor-physical examination matrix and F2C Net, and complete the calculation after calculation. General Physical Exam-Disease Matrix. A medical examination data complementing device based on side information is also disclosed, which can complement medical examination data and disease results according to existing information.

Description

A method, device and application for medical examination data completion based on side information

technical field

The invention belongs to the fields of data statistics and artificial intelligence, and in particular relates to a method, device and application for completing physical examination data based on side information.

Background technique

The traditional physical examination plan is to go through a series of physical examinations for disease screening: according to the needs of different diseases, the physical examination of the relevant physiological characteristics items is carried out under the arrangement and recommendation of the doctor or medical manual, and then the doctor will pass the relevant physiological characteristics examination results to the patient. Diagnose possible diseases. Due to the variety of physical examination items, different hospitals, doctors and eras have different examination methods, resulting in a variety of physical examination items that cannot be unified, resulting in a waste of relevant medical resources and making patients fearless.

With the continuous development of science and technology, the research on medical knowledge such as the correlation of physiological characteristics implied by different physical examination items and the degree of influence between physiological characteristics and diseases tends to be perfected, and the problems of matrix completion and side information have also been developed. Matrix Completion (abbreviated as: MC) is the process of estimating unknown elements based on known elements, thereby restoring the matrix to a complete process. It is a key difficulty in artificial intelligence research projects. Complete matrix for completion. This task has important applications in data mining, e-commerce marketing, engineering control, image and video processing.

In medical projects, the unification of different medical physical examination items depends on the matrix completion algorithm, which infers the effect of unknown physical examination items through related physical examination items. However, because the matrix completion technology often uses methods such as linear transformation and local information interpolation, there are few studies on nonlinear transformation using background knowledge, and the results are not perfect.

Side information refers to using the existing information Y to assist in encoding the information X, which can make the encoding length of the information X shorter. For side information, see Multi-User Source Coding. A popular example is: Suppose you go to the racecourse to bet on horses, and you can get an optimal investment plan according to the odds of each horse. But if you know some historical data on horse betting, such as the results of the last few games, you can come up with a better investment plan. Historical data in horse betting is side information.

The side information algorithm is an algorithm based on the missing information in the side information completion matrix, that is, to find the relevant and irrelevant data points in the information flow, and the improvement of constraints and auxiliary matrix completion technology is applied to various fields that require matrix completion. . Side information method is still a branch of traditional machine learning, and there is not enough attempt to combine it with artificial neural network and deep learning.

In the medical field, data missing is so serious and labeled data is sparse, but matrix completion methods are rarely applied.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a medical examination data complementing method and device based on side information, which can complement medical examination data and disease results according to the existing information.

Another object of the present invention is to provide an application of an apparatus for complementing medical examination data based on side information, which is used to reconstruct a disease.

In order to realize the above-mentioned purpose of the invention, the following technical solutions are provided:

In a first aspect, a method for completing medical examination data based on side information includes the following steps:

(1) Construct a column to represent physiological characteristics and disease subtypes, rows to represent patients, and element values to be a physical examination-disease matrix of the patient's physiological characteristics detection values and disease types; columns to represent disease subtypes, rows to represent causative factors, and element values The causative factor-disease matrix of the probability that the causative factor causes the disease; and the causative factor-physical examination matrix in which the column represents the physiological feature, the row represents the causative factor, and the element value is the correlation between the causative factor and the physiological feature;

(2) For the physical examination-disease matrix, supplement the detected values of physiological characteristics according to the data of the physical examination items, and supplement the disease types according to the subjective diagnosis results of the doctor; for the pathogenic factor-disease matrix and the pathogenic factor-physical examination matrix, according to medical knowledge, supplement the The probability of known causative factors causing known disease subtypes, supplementing the correlation between known causative factors and physiological characteristics;

(3) Establish an encoding and decoding network D2F Net in the physical examination-disease matrix and the pathogenic factor-disease matrix, the physical examination-disease matrix and the pathogenic factor-physical examination matrix, and the pathogenic factor-disease matrix and the pathogenic factor-physical examination matrix, respectively, D2CNet and F2C Net;

(4) Jointly train the encoding and decoding networks D2F Net, D2C Net and F2C Net. After the training, the pathogenic factor-disease matrix and the pathogenic factor-physical examination matrix have been completed;

(5) Input the physical examination-disease matrix to be completed into D2F Net and D2C Net, and use the completed causative factor-disease matrix, causative factor-physical examination matrix and F2C Net to complete the physical examination-disease matrix after calculation .

The medical examination data completion method can complete the unknown information by encoding and decoding according to the existing data information, which greatly reduces the heavy workload and reduces the economic and physical burden of the patient. Help different hospitals and doctors apply different physical examination results uniformly to ensure that medical resources are not wasted. In a second aspect, an apparatus for completing medical examination data based on side information, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor,

The computer memory stores the causative factor-disease matrix, causative factor-physical examination matrix and D2F Net, parameters of D2C Net and F2C Net that are completed by the side information-based medical examination data completion method described in the first aspect ;

The computer processor implements the following steps when executing the computer program:

Receive the input medical examination-disease matrix to be completed, use the completed pathogenic factor-disease matrix, pathogenic factor-physical examination matrix, D2F Net, D2C Net and F2C Net to calculate the medical examination-disease matrix, and output the completed Physical Exam - Disease Matrix.

The medical examination data completion device can complete unknown information by encoding and decoding according to the existing data information and the determined pathogenic factor-disease matrix and pathogenic factor-physical examination matrix, which greatly reduces the burden of sound. In addition, it can help different hospitals and doctors apply different physical examination results uniformly, so as to ensure that medical resources are not wasted. A third aspect provides an application for obtaining disease results by using the apparatus for completing medical examination data based on side information as described in the second aspect, and searching and obtaining disease results according to the output complemented medical examination-disease matrix.

According to the predicted disease subtype obtained by complementing the medical examination-disease matrix output by the medical examination data complementing device, the accuracy rate can reach more than 95%, which can assist doctors in disease diagnosis.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts.

Fig. 1 is a schematic form of the physical examination-disease matrix provided by the embodiment;

Fig. 2 is a schematic form of the causative factor-disease matrix provided by the embodiment;

Fig. 3 is a schematic form of the pathogenic factor-physical examination matrix provided by the embodiment;

FIG. 4 is a schematic diagram of an encoding-decoding network constructed between a physical examination-disease matrix, a causative factor-disease matrix, and a causative factor-physical examination matrix provided by the embodiment.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and do not limit the protection scope of the present invention.

In order to solve the problems of expensive and laborious physical examination and heavy physical examination work for doctors, the method for completing physical examination data based on side information provided in this embodiment specifically includes the following steps:

S101 , constructing a physical examination-disease matrix, a pathogenic factor-disease matrix, and a pathogenic factor-physical examination matrix.

For the physical examination-disease matrix, the columns represent physiological characteristics and disease subtypes, the rows represent patients, and the element values are the patient's physiological characteristics detection values and disease types. Among them, physiological characteristics refer to some physiological information of the human body, which are generally items of physical examination, including height, weight, heart rate, and blood routine. Figure 1 presents a schematic physical examination-disease matrix, which does not contain any real information and is only used to describe the structure of the physical examination-disease matrix. As shown in Figure 1, the rows represent different patients, the columns represent different physical examination items, such as globulin, Hongxibiao, alanine aminotransferase, etc., and the columns also represent the physical examination results of patients, such as A, B, C, D, E, F , G, etc.

In the physical examination-disease matrix, for the physiological features represented by yin and yang, the detected value corresponding to the physiological feature represented by yang is represented by 1, and the detected value corresponding to the physiological feature represented by yin is represented by 0.

For the causative factor-disease matrix, the columns represent disease subtypes, which are divided into dominant and recessive. For known diseases, it is a dominant disease subtype, and for an unknown disease, it is a recessive disease subtype. Disease factors, pathogenic factors are also divided into dominant and recessive factors, known pathogenic factors are dominant pathogenic factors, unknown pathogenic factors are recessive pathogenic factors, and the element value is the pathogenic factor that causes the disease. probability. Suppose the causative factor-disease matrix is an M×N matrix, and its row M represents M causative factors, of which only m (<M) are dominant, and its column N represents N disease subtypes, of which only There are n (<N) species that are dominant. Figure 2 is an exemplary causative factor-disease matrix, in which subtypes A, B, C, D, E, F, and G are known types of diseases, and the remaining unknown types 1, unknown types 2, unknown types 3. Unknown type 4 is the unknown disease subtype, a, b, and c are known types of pathogenic factors, and the remaining 6 are unknown pathogenic factors. If as shown in Figure 2, M=9, m=3; N=11, n=7. Among them, M and N must be larger than m and n; and as for how much larger, appropriate estimation is made through experience.

For an m×n matrix composed of known disease subtypes and known causative factors, the element values, that is, the probability that the causative factor causes the disease (ie, the probability of occurrence between the causative factor and the disease), are based on medical knowledge. Or the medical knowledge proof is completed, that is, the numbers 0.4, 0.1 and other values in Figure 2 are filled according to the medical knowledge or medical knowledge proof, that is, the establishment of the edge information of the causative factor-disease matrix, which is unknown in the M×N matrix The element values corresponding to disease subtypes and unknown causative factors cannot be filled, so they are left blank.

For the causative factor-physical examination matrix, the columns represent physiological characteristics (that is, physical examination data), the rows represent causative factors, and the element value is the correlation between causative factors and physiological characteristics, which is constructed based on medical knowledge and medical statistical results. , according to the degree of correlation, it can be expressed as high, medium and low, as shown in Figure 3; it can also use positive weights to express positive correlations, negative weights to express negative correlations, and 0 to express irrelevance, that is, the matching is achieved. The edge information of the disease factor-physical examination matrix is established.

Step 102, establish an encoding and decoding network D2F Net in the physical examination-disease matrix and the pathogenic factor-disease matrix, the physical examination-disease matrix and the pathogenic factor-physical examination matrix, and the pathogenic factor-disease matrix and the pathogenic factor-physical examination matrix, respectively, D2C Net and F2C Net, as shown in Figure 4.

Among them, the network structures of D2F Net, D2C Net and F2C Net are all self-encoders built by convolutional layers and self-decoders built by deconvolution. The convolution layer and the deconvolution layer are generally 3 to 4 layers, and a reconstruction objective function is established on each layer. In the self-decoder, the reconstruction difference corresponding to each layer is required to be as small as possible.

If the size of the physical examination-disease matrix, the causative factor-disease matrix and the causative factor-physical examination matrix are large, use a large-capacity neural network such as ResNeXt for encoding, and use the inverse corresponding to the convolutional layer in the neural network. The convolutional layer builds a self-decoder, in which the neural network cannot contain a pooling layer that will cause information loss, and the pooling layer and dropout layer need to be removed.

S103 , jointly train the encoding and decoding networks D2F Net, D2C Net and F2C Net. After the training, the causative factor-disease matrix and the causative factor-physical examination matrix have been completed.

When completing the causative factor-disease matrix, D2F Net and F2C Net are used to complete the causative factor-disease matrix, specifically,

For D2F Net, the physical examination-disease matrix is used as the input variable, the autoencoder is used to encode the physical examination-disease matrix to generate the reconstructed pathogenic factor-disease matrix, and the autodecoder is used to decode the reconstructed pathogenic factor-disease matrix, Generate a reconstructed medical examination-disease matrix, taking the sum of the squared deviation loss function of the medical examination-disease matrix and the reconstructed medical examination-disease matrix, and the sum of the squared deviation loss function of the factor-disease matrix and the reconstructed factor-disease matrix as the Loss function L ₁ of D2F Net;

For F2C Net, the pathogenic factor-physical examination matrix is used as the input variable, and the autoencoder is used to encode the pathogenic factor-physical examination matrix to generate the reconstructed pathogenic factor-disease matrix, and the autodecoder is used to reconstruct the pathogenic factor-disease matrix. The matrix is decoded to generate a reconstructed causative factor-physical examination matrix, and the squared dispersion loss function of the causative factor-physical examination matrix and the reconstructed causative factor-physical examination matrix, and the causative factor-disease matrix and the reconstructed pathogenic The sum of the squared deviation and loss function of the factor-disease matrix is used as the loss function L ₂ of F2C Net;

The sum of the loss function L ₁ and the loss function L ₂ L ¹ is used as the total loss function to complete the causative factor-disease matrix.

When completing the pathogenic factor-physical examination matrix, F2C Net and D2C Net are used to complete the pathogenic factor-physical examination matrix. Specifically,

For F2C Net, the pathogenic factor-disease matrix is used as the input variable, and the autoencoder is used to encode the pathogenic factor-disease matrix to generate a reconstructed pathogenic factor-physical examination matrix. The matrix is decoded to generate a reconstructed causative factor-disease matrix, with the squared variance loss function of the causative factor-disease matrix and the reconstructed causative factor-disease matrix, and the causative factor-physical examination matrix and the reconstructed pathogenic The sum of the squared deviation and loss function of the factor-physical examination matrix is used as the loss function L ₃ of F2C Net;

For D2C Net, using the physical examination-disease matrix as the input variable, the autoencoder is used to encode the physical examination-disease matrix to generate the reconstructed pathogenic factor-physical examination matrix, and the autodecoder is used to decode the reconstructed pathogenic factor-physical examination matrix. Generates a reconstructed medical-disease matrix with a sum of squared dispersion loss function of the medical-disease matrix and the reconstructed medical-disease matrix, and a causative factor-disease matrix and a reconstructed causative factor-disease matrix with a squared deviation loss The sum of functions is used as the loss function L ₄ of D2C Net;

The sum of the loss function L ₃ and the loss function L ₄ L ² is used as the total loss function to complete the causative factor-physical examination matrix.

When completing the physical examination-disease matrix, D2C Net and D2F Net are used to complete the physical examination-disease matrix. Specifically,

For D2C Net, using the pathogenic factor-physical examination matrix as the input variable, the autoencoder is used to encode the pathogenic factor-physical examination matrix to generate the reconstructed physical examination-disease matrix, and the autodecoder is used to decode the reconstructed physical examination-disease matrix, Generate a reconstructed causative factor-physical examination matrix, with the squared dispersion loss function of the causative factor-physical examination matrix and the reconstructed causative factor-physical examination matrix, and the squared deviation of the physical examination-disease matrix and the reconstructed physical examination-disease matrix. The sum of the loss function is used as the loss function L ₅ of D2C Net;

For D2F Net, the pathogenic factor-disease matrix is used as the input variable, the autoencoder is used to encode the pathogenic factor-disease matrix to generate the reconstructed medical examination-disease matrix, and the autodecoder is used to decode the reconstructed medical examination-disease matrix, Generate a reconstructed causative factor-disease matrix, with the squared dispersion loss function of the causative factor-disease matrix and the reconstructed causative factor-disease matrix, and the squared deviation of the medical examination-disease matrix and the reconstructed medical examination-disease matrix The sum of the loss function is used as the loss function L ₆ of D2F Net;

_The sum L3 ^of the loss function L5 and the loss function L6 is used as the total loss function to complete the medical examination _- disease matrix.

During joint training, the sum of L ¹ , L ² and L ³ is used as the total loss function, and the reverse transfer is performed to update the network parameters of D2F Net, D2C Net and F2C Net and complete the pathogenic factor-disease matrix, pathogenicity Factor-Physical Exam Matrix.

The above physical examination-disease matrix is a matrix with complete element values, and the causative factor-disease matrix and the causative factor-physical examination matrix are only incomplete matrices established by information, that is, they do not include the correspondence of unknown causative factors and unknown disease subtypes. Through the joint training of S103, the corresponding causative factor-disease matrix and causative factor-physical examination matrix are completed by using the autoencoding and decoding functions of the physical examination-disease matrix and the three networks of D2F Net, D2C Net and F2C Net. , thus finding the probability of occurrence between unknown causative factors and unknown disease subtypes, as well as the correlation between unknown causative factors and physiological characteristics.

S104, input the physical examination-disease matrix to be completed into D2F Net and D2C Net, and use the completed pathogenic factor-disease matrix, pathogenic factor-health examination matrix and F2C Net to complete the physical examination-disease matrix after calculation.

This embodiment also provides a side information-based medical examination data complementing device, including a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. There are causative factor-disease matrix, causative factor-physical examination matrix and parameters of D2F Net, D2C Net and F2C Net completed by the above-mentioned physical examination data completion method;

The computer processor implements the following steps when executing the computer program:

Receive the input medical examination-disease matrix to be completed, use the completed pathogenic factor-disease matrix, pathogenic factor-physical examination matrix, D2F Net, D2C Net and F2C Net to calculate the medical examination-disease matrix, and output the completed Physical Exam - Disease Matrix.

The above-mentioned physical examination data completion method and device can complete unknown information by encoding and decoding according to the existing data information and the determined pathogenic factor-disease matrix and pathogenic factor-physical examination matrix, which greatly alleviates the problem. A heavy workload can reduce the economic and physical burden of patients. In addition, it can also help different hospitals and doctors to apply different physical examination results uniformly, so as to ensure that medical resources are not wasted. After the above-mentioned physical examination data completion device outputs the completed physical examination-disease matrix, the physical examination-disease matrix contains the completed disease types, and the doctor can search and obtain the disease result according to the completed physical examination-disease matrix. The accuracy rate can reach more than 95%, which can assist doctors in diagnosing diseases.

The above-mentioned specific embodiments describe in detail the technical solutions and beneficial effects of the present invention. It should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, additions and equivalent substitutions made within the scope shall be included within the protection scope of the present invention.

Claims (4)

1. A physical examination data completion method based on side information comprises the following steps:

(1) constructing a physical examination-disease matrix with columns representing physiological characteristics and disease subtypes, rows representing patients and element values being physiological characteristic detection values and disease types of the patients; the columns represent disease subtypes, the rows represent disease-causing agents, and the element values are disease-causing agent-disease matrices of probability of disease-causing agents; and columns represent physiological characteristics, rows represent pathogenic factors, and element values are pathogenic factor-physical examination matrixes of the correlation between the pathogenic factors and the physiological characteristics;

(2) supplementing physiological characteristic detection values according to physical examination item data and supplementing disease types according to subjective diagnosis results of doctors aiming at the physical examination-disease matrix; supplementing the probability that the known pathogenic factors cause known disease subtypes and the correlation between the known pathogenic factors and physiological characteristics according to medical knowledge for the pathogenic factor-disease matrix and the pathogenic factor-physical examination matrix;

(3) establishing a code-decoding network D2F Net, a code-decoding network D2C Net and a code-decoding network F2C Net in a physical examination-disease matrix and a pathogenic factor-disease matrix, a physical examination-disease matrix and a pathogenic factor-physical examination matrix, and a pathogenic factor-disease matrix and a pathogenic factor-physical examination matrix respectively; the network structures of the coding and decoding network D2F Net, the coding and decoding network D2C Net and the coding and decoding network F2C Net are self-encoders built by convolutional layers and self-decoders built by deconvolution;

(4) jointly training the coding and decoding network D2F Net, the coding and decoding network D2C Net and the coding and decoding network F2CNet, and completing a pathogenic factor-disease matrix and a pathogenic factor-physical examination matrix after training is finished;

in joint training, to complement the overall loss function L of the disease-causative-agent matrix¹And (3) complementing the overall loss function L of the disease-causing factor-physical examination matrix²And the overall loss function L of the completion physical examination-disease matrix³The sum of the three is used as a total loss function, reverse transfer is carried out, and the network parameters, the completion pathogenic factor-disease matrix and the completion pathogenic factor-physical examination matrix of the coding and decoding network D2F Net, the coding and decoding network D2C Net and the coding and decoding network F2C Net are updated;

when the pathogenic factor-disease matrix is complemented, the pathogenic factor-disease matrix is complemented by adopting a coding and decoding network D2F Net and a coding and decoding network F2C Net, for the coding and decoding network D2F Net, a physical examination-disease matrix is used as an input variable, a self-encoder is used for coding the physical examination-disease matrix to generate a reconstructed pathogenic factor-disease matrix, a self-decoder is used for decoding the reconstructed pathogenic factor-disease matrix to generate a reconstructed physical examination-disease matrix, the sum of the squares of deviations of the physical examination-disease matrix and the reconstructed physical examination-disease matrix and the sum of the squares of deviations of the factor-disease matrix and the reconstructed factor-disease matrix are used as a loss function L of the coding and decoding network D2F Net₁；

For the encoding and decoding network F2C Net, a pathogenic factor-physical matrix is used as an input variable, the pathogenic factor-physical matrix is encoded by using an auto-encoder to generate a reconstructed pathogenic factor-disease matrix, the reconstructed pathogenic factor-disease matrix is decoded by using an auto-decoder to generate a reconstructed pathogenic factor-physical matrix, and the sum of the dispersion square sum loss function of the pathogenic factor-physical matrix and the reconstructed pathogenic factor-physical matrix and the dispersion square sum loss function of the pathogenic factor-disease matrix and the reconstructed pathogenic factor-disease matrix is used as the loss function L of the encoding and decoding network F2C Net₂；

By a loss function L₁And a loss function L₂Sum L¹As a function of total loss to complement the disease agent-disease matrix;

when the pathogenic factor-physical matrix is completed, the pathogenic factor-physical matrix is completed by adopting the encoding and decoding network F2C Net and the encoding and decoding network D2C Net, for the encoding and decoding network F2C Net, the pathogenic factor-disease matrix is used as an input variable, the self-encoder is adopted to encode the pathogenic factor-disease matrix to generate a reconstructed pathogenic factor-physical examination matrix, the self-decoder is adopted to decode the reconstructed pathogenic factor-physical examination matrix to generate a reconstructed pathogenic factor-disease matrix, taking the square deviation sum loss function of the pathogenic factor-disease matrix and the reconstructed pathogenic factor-disease matrix, and the sum of the dispersion square sum loss functions of the pathogenic factor-physical matrix and the reconstructed pathogenic factor-physical matrix is used as the loss function L of the encoding and decoding network F2C Net.₃；

For the encoding and decoding network D2C Net, a physical examination-disease matrix is used as an input variable, the physical examination-disease matrix is encoded by using an auto-encoder to generate a reconstructed disease-causing-agent matrix, the reconstructed disease-causing-agent matrix is decoded by using an auto-decoder to generate a reconstructed physical examination-disease matrix, and the sum of the dispersion square sum loss functions of the physical examination-disease matrix and the reconstructed physical examination-disease matrix and the dispersion square sum loss function of the disease-causing-agent matrix and the reconstructed disease-causing-agent matrix is used as a loss function L of the encoding and decoding network D2C Net₄；

By a loss function L₃And a loss function L₄Sum L²As a function of the total loss to complement the virulence factor-physical examination matrix;

(5) inputting the physical examination-disease matrix to be completed into a coding and decoding network D2F Net and a coding and decoding network D2C Net, and calculating a completed physical examination-disease matrix by utilizing the completed pathogenic factor-disease matrix, pathogenic factor-physical examination matrix and the coding and decoding network F2C Net;

when the physical examination-disease matrix is complemented, the physical examination-disease matrix is complemented by adopting the coding and decoding network D2C Net and the coding and decoding network D2F Net, for the coding and decoding network D2C Net, the pathogenic factor-physical examination matrix is used as an input variable, the pathogenic factor-physical examination matrix is coded by adopting an auto-coder to generate a reconstructed physical examination-disease matrix, and the reconstructed physical examination-disease matrix is acquiredDecoding the reconstructed physical examination-disease matrix by using a self-decoder to generate a reconstructed pathogenic factor-physical examination matrix, and taking the sum of the squared deviations and the loss functions of the pathogenic factor-physical examination matrix and the reconstructed pathogenic factor-physical examination matrix and the sum of the squared deviations and the loss functions of the physical examination-disease matrix and the reconstructed physical examination-disease matrix as the loss function L of the coding and decoding network D2C Net₅；

For the encoding and decoding network D2F Net, the pathogenic factor-disease matrix is used as an input variable, the pathogenic factor-disease matrix is encoded by using an auto-encoder to generate a reconstructed physical examination-disease matrix, the reconstructed physical examination-disease matrix is decoded by using an auto-decoder to generate a reconstructed pathogenic factor-disease matrix, and the sum of the square deviation sum loss functions of the pathogenic factor-disease matrix and the reconstructed pathogenic factor-disease matrix and the sum of the square deviation sum loss functions of the physical examination-disease matrix and the reconstructed physical examination-disease matrix is used as the loss function L of the encoding and decoding network D2F Net₆；

By a loss function L₅And a loss function L₆Sum L³As a function of the total loss of the completion physical examination-disease matrix.

2. The method for supplementing physical examination data based on side information as set forth in claim 1, wherein in the physical examination-disease matrix, for the physiological characteristics expressed in yin and yang, the detection value corresponding to the physiological characteristic expressed in yang is represented by 1, and the detection value corresponding to the physiological characteristic expressed in yin is represented by 0.

3. A physical examination data completion apparatus based on side information, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that:

the computer memory stores the parameters of the pathogenic factor-disease matrix, the pathogenic factor-physical examination matrix and the codec network D2F Net, the codec network D2C Net and the codec network F2C Net complemented by the side information-based physical examination data complementing method of claim 1 or 2;

the computer processor, when executing the computer program, performs the steps of:

and receiving the input physical examination-disease matrix to be complemented, calculating the physical examination-disease matrix by utilizing the complemented pathogenic factor-disease matrix, the complemented pathogenic factor-physical examination matrix, the coding and decoding network D2F Net, the coding and decoding network D2C Net and the coding and decoding network F2C Net, and outputting the complemented physical examination-disease matrix.

4. Use of the side-information based physical examination data completion apparatus of claim 3 for obtaining disease results, wherein the disease results are obtained by searching the output completed physical examination-disease matrix.

Images