A method for analyzing the recovery status of critically ill patients based on parallel feature extraction

Abstract

The present invention discloses a method for analyzing the recovery status of critically ill patients based on parallel feature extraction, comprising: collecting the features of patients to construct a data set, taking the features of each patient as a piece of medical data; encoding the medical data of each patient into a feature vector; constructing a parallel feature extraction network, inputting the feature vector of the patient into the parallel feature extraction network, and extracting parallel features; mapping the parallel features to obtain a final prediction result; the convolution kernel of the convolution layer of the convolutional neural network in the parallel feature extraction network is a specific rectangle; the present invention can distinguish and encode according to the characteristics of medical data, so that the patient's data can be quantitatively analyzed, and feature analysis is performed in parallel through the extraction features of multiple models, thereby making full use of the potential of the data, evaluating the recovery status of critically ill patients through artificial intelligence technology, and saving medical resources.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a method for analyzing the recovery status of critically ill patients based on parallel feature extraction. Background Art

[0002] In the rehabilitation process of critically ill patients, accurate assessment of postoperative recovery is the key to clinical decision-making and optimization of treatment plans. However, traditional rehabilitation assessment methods often rely on limited physiological indicators and manual observation, lack a comprehensive analysis of the patient's comprehensive status, and may lead to bias and inaccuracy in the assessment results. This limitation not only increases the workload of doctors, but also may affect the patient's rehabilitation effect. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention provides a method for analyzing the recovery status of critically ill patients based on parallel feature extraction, which aims to solve the problems in the background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for analyzing the recovery status of critically ill patients based on parallel feature extraction, comprising the following steps:

[0005] Step S1: Collect the patient's characteristics to construct a data set, and take each patient's characteristics as a piece of medical data;

[0006] Step S2: Encode the medical data of each patient into a vector form, i.e. Patients The eigenvector of ,in, Indicates the index, ranging from 1 to ,in, represents the number of patients;

[0007] Step S3: Construct a parallel feature extraction network to extract the patient's feature vector Input parallel feature extraction network, extract The parallel feature extraction network is composed of a parallel convolutional neural network, a recurrent neural network and a graph neural network;

[0008] Step S4: Map the parallel features to obtain the final prediction result;

[0009] The specific process of encoding each patient's medical data into a vector form is: The features of are divided into low-dimensional embedding representations of continuous features and low-dimensional embedding representations of discrete features:

[0010] The low-dimensional embedding of discrete features is represented as: No. Discrete features are represented as a one-hot encoded vector , The number of dimensions of is equal to the number of its options; In the example, the elements corresponding to the options are set to 1, and the elements other than the elements corresponding to the options are set to 0. Then, the transformation matrix is ​​used to transform Perform dimension conversion to obtain feature space encoding of discrete features , expressed as:

[0011] ;

[0012] In the formula, and are the weight and bias of the transformation matrix respectively; is the activation function of the rectified linear unit ReLU.

[0013] Furthermore, It is expressed as:

[0014] ;

[0015] In the formula, Represents the function independent variable;

[0016] The low-dimensional embedding of continuous features is represented as: No. The continuous features are represented as one-dimensional vectors , the one-dimensional vector Convert the tensor into a vector of a specific dimension, i.e., the feature space encoding of continuous features , expressed as:

[0017] ;

[0018] In the formula, and They are the weight parameters and bias parameters of tensor transformation respectively;

[0019] Enumeration Discrete features from 1 to of and A continuous feature from 1 to of , all and They are connected in series along the column direction to form discrete feature matrices and continuous feature matrices, namely discrete feature aggregation coding. and continuous feature aggregation encoding , expressed as:

[0020] ;

[0021] ;

[0022] In the formula, Represents discrete feature aggregation encoding; Represents continuous feature aggregation encoding; Represents matrix transpose; Indicates from arrive Discrete features of Indicates from arrive Continuous characteristics of is the number of discrete characteristics of the patient, is the number of continuous characteristics of the patient;

[0023] Combine the discrete feature matrix and the continuous feature matrix into a total matrix, namely, patient The eigenvector of , expressed as:

[0024] ;

[0025] In the formula, is the adjustment parameter, which means and exist The proportion of .

[0026] Furthermore, the convolution kernel of the convolutional layer in the convolutional neural network CNN is The rectangle represents the width of the convolution kernel, is the number of characteristics of the patient.

[0027] Furthermore, convolutional neural network extracts The specific process of parallel features is:

[0028] The dimension is of Input convolutional neural network, output convolution extracted features ; is the number of patient characteristics, express Dimensions;

[0029] Among them, the convolutional layer dimension of the convolutional neural network CNN is of The processing process is:

[0030] ;

[0031] In the formula, is the output of the convolutional layer; and are the weight parameters and bias parameters of the convolution kernel respectively; according to the dimension transformation principle in the convolution layer, calculate The first size and the second size , expressed as:

[0032] ;

[0033] ;

[0034] The dimension is ,Will Input to the fully connected layer of the convolutional neural network to generate convolutional extraction features :

[0035] ;

[0036] In the formula, and They are the weight parameters and bias parameters that the convolutional neural network needs to learn.

[0037] Furthermore, graph neural network extraction The specific process of the parallel feature is: After input into the graph neural network for processing, the graph neural network extracted features are obtained , expressed as:

[0038] ;

[0039] In the formula, and Represents the parameters that the graph neural network needs to learn; Represents the degree matrix learned by the graph neural network; It is the adjacency matrix learned by the graph neural network.

[0040] Furthermore, the recurrent neural network extracts The specific process of parallel features is:

[0041] First, set the Within a time node The feature vector :

[0042] ;

[0043] In the formula, Indicates Data feature encoding of each time node;

[0044] Enumeration from arrive , the data feature encoding collected at all a time nodes of each patient; the correlation calculation between nodes is expressed as:

[0045] ;

[0046] In the formula, Indicates the correlation between the feature codes of the data collected at two time nodes; Indicates Feature coding of data collected at each time node; Indicates Feature coding of data collected at each time node; and Represents the weights learned by the recurrent neural network RNN ​​through back propagation;

[0047] Normalize the feature encoding of the data collected at a time node, and use the softmax operator to calculate the normalized weight value :

[0048] ;

[0049] In the formula, is another one from 1 to In the range and not equal to The index number of Indicates a certain time point;

[0050] Based on weight value , make a function:

[0051] ;

[0052] In the formula, and These are weight parameters of the recurrent neural network RNN; is a dimension of The representative vector of

[0053] Aggregate all , calculate Output after inputting the recurrent neural network, i.e. the recurrent neural network extracts features :

[0054] .

[0055] Furthermore, in step S4, the parallel features are mapped to obtain the final prediction result. The specific process is as follows:

[0056] Will , and Converted to the same dimension, expressed as:

[0057] ;

[0058] In the formula, and They are the weight parameters and bias parameters that need to be learned by the parallel feature extraction network respectively; Represents the same dimension after conversion ; Represents the same dimension after conversion ; Represents the same dimension after conversion ;

[0059] Will , and Merge by accumulating each element to get the aggregate feature , expressed as:

[0060] ;

[0061] In the formula, , and They are , and Hyperparameters of weights;

[0062] By adding a linear layer, Mapping the final result is expressed as:

[0063] ;

[0064] In the formula, A score to predict the patient's recovery; is the Sigmoid activation function; Represents a linear layer.

[0065] Furthermore, the feature vector of the patient is extracted in the parallel feature extraction network Before, the parallel feature extraction network is trained using the training set, which is obtained by dividing the data set. The mean square error is used in the parallel feature extraction network training. Perform parameter tuning, expressed as:

[0066] ;

[0067] In the formula, A score indicating actual patient recovery; Represents the number of medical data in the training set.

[0068] Compared with the existing technology, the present invention has the following beneficial effects: the present invention can distinguish and encode according to the characteristics of medical data, so that the patient's data can be quantitatively analyzed, and feature analysis can be performed in parallel by extracting features through multiple models, making full use of the potential of the data, and evaluating the recovery of critically ill patients through artificial intelligence technology, saving medical resources; by using a rectangular convolution kernel, the convolution operation can better capture features in a specific direction, and the aspect ratio of the medical data matrix is ​​relatively large, and the use of a rectangular convolution kernel can also better adapt to the geometric characteristics of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the steps of the present invention.

[0070] Figure 2 Score plots for different models and different model combinations.

[0071] Figure 3 Score graphs for models using rectangular convolution kernels and square convolution kernels. DETAILED DESCRIPTION

[0072] like Figure 1 As shown, the present invention provides a technical solution: a method for analyzing the recovery status of critically ill patients based on parallel feature extraction, comprising the following steps:

[0073] Step S1: Collect the patient's characteristics to construct a data set, and take each patient's characteristics as a piece of medical data;

[0074] Step S2: Encode the medical data of each patient into a vector form, i.e. Patients The eigenvector of ,in, Indicates the index, ranging from 1 to ,in, Indicates the number of patients.

[0075] Specifically, the multi-dimensional patient features are converted into low-dimensional embedding representations; since the parallel feature extraction networks in the subsequent steps use convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph neural networks (GCNs), it is required to input the patient feature vectors of these networks have the same dimensions so that the patient's features can be represented as a matrix in the feature space.

[0076] The patient The features are divided into two types of low-dimensional embedding representations: continuous features and discrete features:

[0077] Discrete features are those whose content is one of multiple options. For example, the attribute "gender" of a person has two possible options: male and female. Therefore, the content of this attribute is one of the two options, and this attribute is considered a discrete feature. For all discrete features, they usually have different numbers of dimensions. Embedding is a tensor transformation operation that converts different dimensions into Convert to a vector with a specific dimension;

[0078] The low-dimensional embedding of discrete features is represented as: No. Discrete features are represented as a one-hot encoded vector , The number of dimensions of is equal to the number of possible options of itself; In the example, the element corresponding to the option is set to 1, and the elements other than the element corresponding to the option are set to 0. Then, Apply a transformation matrix to achieve Dimension transformation to obtain the feature space encoding of discrete features , which can be expressed as:

[0079] (1);

[0080] In the formula, and are the weight and bias of the transformation matrix respectively; is the activation function of the rectified linear unit (ReLU).

[0081] It can be expressed as:

[0082] (2);

[0083] In the formula, Represents a function argument.

[0084] is a vector of intrinsic dimension. In this embodiment, The number of dimensions is 512; enumeration Discrete features from 1 to of , Represents the total number of discrete features, each are of the same dimension, The series connection can construct a discrete feature matrix, that is, discrete feature aggregation coding .

[0085] Continuous features, continuous features mean that their values ​​are within a numerical range, rather than one of several specific options; the number of values ​​that a continuous feature can take is infinite; for example, a person's attribute "height" has infinite possible values, so it is a continuous feature; for all continuous features, they cannot be directly represented as vectors like discrete features because they only have a single value; therefore, the embedding operation of continuous features is different from that of discrete features.

[0086] The low-dimensional embedding of continuous features is represented as: No. A continuous feature is represented as a one-dimensional vector , The format is a vector with only one element; in order to The dimension of the vectorized representation of matches the discrete features, which requires some basic tensor conversion operations; therefore, the embedding operation of continuous features is basically to convert a one-dimensional vector into a vector of a specific dimension, that is, the feature space encoding of the continuous feature , which can be expressed as:

[0087] (3);

[0088] In the formula, and They are the weight parameters and bias parameters of tensor transformation respectively.

[0089] After a specific dimension transformation, we get are vectors of the same dimension; enumerate the A continuous feature from 1 to of , Represents the total number of continuous features, all Have the same dimensions, the number of dimensions is the same as the discrete features, all The concatenation can construct a continuous feature matrix, that is, continuous feature aggregation coding .

[0090] Enumeration Discrete features from 1 to of and A continuous feature from 1 to of , all and They are connected in series along the column direction to form discrete feature matrices and continuous feature matrices, namely discrete feature aggregation coding. and continuous feature aggregation encoding , which can be expressed as:

[0091] (4);

[0092] (5);

[0093] In the formula, Represents discrete feature aggregation encoding; represents continuous feature aggregation encoding; Represents matrix transpose; Indicates from arrive Discrete features of Indicates from arrive Continuous characteristics of is the number of discrete features of the patient, is the number of continuous features of the patient.

[0094] Combine the discrete feature matrix and the continuous feature matrix into a total matrix, namely, patient The eigenvector of , the matrix contains patients All the features of are expressed as:

[0095] (6);

[0096] In the formula, is the adjustment parameter, which means and exist The proportion of .

[0097] Step S3: Construct a parallel feature extraction network to extract the patient's feature vector Input parallel feature extraction network, extract The parallel feature extraction network is composed of a parallel convolutional neural network (CNN), a recurrent neural network (RNN) and a graph neural network (GCN).

[0098] Among them, the convolutional neural network (CNN) consists of convolutional layers, pooling layers, and fully connected layers. Considering that the features in medical data are not complex, only convolutional layers and fully connected layers can be used for processing; the dimension is of Input convolutional neural network (CNN), output convolution extracted features ; is the number of patient characteristics, express Dimension.

[0099] Among them, the convolution kernel of the convolution layer in the convolutional neural network (CNN) is different from the general square convolution kernel. The convolution kernel of the convolution layer is a rectangle; the convolution kernel of the convolution layer in the convolutional neural network (CNN) is designed to be a size of The rectangle Represents the width of the convolution kernel.

[0100] The patient's feature vector The operation after entering the convolutional layer can be expressed as:

[0101] (7);

[0102] In the formula, is the output of the convolutional layer; and are the weight parameters and bias parameters of the convolution kernel respectively; according to the dimension transformation principle in the convolution layer, The first dimension exists and the second size , calculated as follows:

[0103] (8);

[0104] (9);

[0105] The dimension is , and then Input to the fully connected layer to generate a deeper representation, that is, convolution to extract features :

[0106] (10);

[0107] In the formula, and They are the weight parameters and bias parameters that need to be learned by the convolutional neural network (CNN).

[0108] Among them, the graph neural network (GCN) is a neural network model characterized by spectral domain graph convolution operations. Similar to communication signal processing, it performs complex convolution operations in the spectral domain because the convolution operation in the time domain can be converted into a multiplication operation in the spectral domain through Fourier transform; the conversion from time domain to spectral domain is achieved through Fourier transform.

[0109] Will and The operators representing Fourier transform and inverse Fourier transform are expressed as:

[0110] (11);

[0111] In the formula, represents the feature vector; represents the transpose of the eigenvector; Represents the feature matrix after mapping of various physical indicators of the patient.

[0112] Will Input to the graph convolution kernel middle:

[0113] (12);

[0114] In the formula, and Represent graph convolution and matrix dot product respectively.

[0115] Rewrite equation (12) as:

[0116] (13);

[0117] In the formula, is the degree matrix, which represents the degree of each vertex; is the adjacency matrix; represents the adjacency matrix weight, Represents the weight of adjusting the convolution operation; let , Represents the weight parameters that the graph neural network (GCN) needs to learn, and rewrites equation (13) as:

[0118] (14);

[0119] In the formula, represents the degree matrix learned by the graph neural network (GCN), It is the adjacency matrix learned by the graph neural network (GCN).

[0120] Will After input into the graph neural network (GCN) for processing, the graph neural network extracts features , which can be expressed as:

[0121] (15);

[0122] In the formula, Express Perform linear transformation operations; Represents the convolution operation after the graph Perform further linear transformation operations.

[0123] Among them, a recurrent neural network (RNN) is a neural network model implemented on time series data; specifically, the loop operation in the recurrent neural network (RNN) is regarded as using its memory units to capture dependencies in the time series; the memory units in the recurrent neural network (RNN) are similar to the memory in the human brain, representing the ability to remember and process input data at different times.

[0124] set up , which contains Within a time node The data collected at different times are regarded as nodes in the time series data, and the patient's recovery status is inferred by analyzing the relationship between the data time series; first, It is expressed as follows:

[0125] (16);

[0126] In the formula, Indicates Data feature encoding of each time node.

[0127] Enumeration from arrive , the data feature encoding collected at all a time nodes of each patient; the correlation calculation between nodes is expressed as:

[0128] (17);

[0129] In the formula, Indicates the correlation between the feature codes of the data collected at two time nodes; Indicates Feature coding of data collected at each time node; Indicates Feature coding of data collected at each time node; and Represents the weights learned by the recurrent neural network (RNN) through back propagation, which is used to adjust Importance in correlation measures.

[0130] Normalize the feature encoding of the data collected at a time node, which is related to all other features; in order to make the sum of the weight values ​​of all information equal to 1, the normalized weight value is calculated using the softmax operator :

[0131] (18);

[0132] In the formula, is another one from 1 to In the range and not equal to The index number of Indicates a time point.

[0133] Based on weight value , formulate a learnable function to analyze the patient's physical recovery in adjacent time, which can be expressed as:

[0134] (19);

[0135] In the formula, and These are weight parameters of the recurrent neural network (RNN); is a dimension of The representative vector of .

[0136] Aggregate all , calculate Output after inputting the recurrent neural network (RNN), i.e., the recurrent neural network extracts features :

[0137] (20).

[0138] Step S4: Parallel features ( , , ) to map and obtain the final prediction result.

[0139] Specifically, , and The dimensions of ,Will , and The integration is performed using the same set of transformation parameters; first, , and Converted to the same dimension, expressed as:

[0140] (twenty one);

[0141] In the formula, and They are the weight parameters and bias parameters that need to be learned by the parallel feature extraction network respectively; Represents the same dimension after conversion ; Represents the same dimension after conversion ; Represents the same dimension after conversion .

[0142] Will , and Merge by accumulating each element to get the aggregate feature , expressed as:

[0143] (twenty two);

[0144] In the formula, , and They are , and Hyperparameters of weights.

[0145] By adding a linear layer, Mapping the final result is expressed as:

[0146] (twenty three);

[0147] In the formula, A score to predict the patient's recovery; is the Sigmoid activation function, which is used to Restricted to the range (0,1); Represents a linear layer.

[0148] Among them, the feature vector of the patient is extracted in the parallel feature extraction network Before, the parallel feature extraction network is trained using the training set, which is obtained by dividing the data set. The mean square error is used in the parallel feature extraction network training. Perform parameter tuning, expressed as:

[0149] (twenty four);

[0150] In the formula, A score indicating actual patient recovery; Represents the number of medical data in the training set.

[0151] like Figure 2 As shown in the figure, the convolutional neural network (CNN), recurrent neural network (RNN) and graph neural network (GCN) in the parallel feature extraction network are scored individually and in different combinations. The higher the score, the more accurate the model judgment. Figure 2 In the above, CNN represents convolutional neural network (CNN), GCN represents graph neural network (GCN), and RNN represents recurrent neural network (RNN). Figure 2 It can be seen that when convolutional neural network (CNN), recurrent neural network (RNN) and graph neural network (GCN) are combined together, the score is higher than other individual models or combined models.

[0152] like Figure 3 As shown, by comparing the convolutional neural network (CNN) using rectangular convolution kernels with the convolutional neural network (CNN) using normal square convolution kernels, it can be seen that the convolutional neural network (CNN) using rectangular convolution kernels has a higher score than the convolutional neural network (CNN) using normal square convolution kernels.

[0153] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing the recovery status of critically ill patients based on parallel feature extraction, characterized in that: The steps include: Step S1: Collect the patient's characteristics to construct a data set, and take each patient's characteristics as a piece of medical data; Step S2: Encode the medical data of each patient into a vector form, i.e. Patients The eigenvector of ,in, Indicates the index, ranging from 1 to ,in, represents the number of patients; Step S3: Construct a parallel feature extraction network to extract the patient's feature vector Input parallel feature extraction network, extract The parallel feature extraction network is composed of a parallel convolutional neural network, a recurrent neural network and a graph neural network; Step S4: Map the parallel features to obtain the final prediction result; The specific process of encoding each patient's medical data into a vector form is: The features of are divided into low-dimensional embedding representations of continuous features and low-dimensional embedding representations of discrete features: The low-dimensional embedding of discrete features is represented as: No. The discrete features are represented as a one-hot encoded vector , The number of dimensions of is equal to the number of its options; In the example, the elements corresponding to the options are set to 1, and the elements other than the elements corresponding to the options are set to 0. Then, the transformation matrix is ​​used to transform Perform dimension conversion to obtain feature space encoding of discrete features , expressed as: ； In the formula, and are the weight and bias of the transformation matrix respectively; is the activation function of the corrected linear unit ReLU; It is expressed as: ； In the formula, Represents the function independent variable; The low-dimensional embedding of continuous features is represented as: No. Continuous features are represented as one-dimensional vectors , the one-dimensional vector Convert the tensor into a vector of a specific dimension, i.e., the feature space encoding of continuous features , expressed as: ； In the formula, and They are the weight parameters and bias parameters of tensor transformation respectively; Enumeration Discrete features from 1 to of and A continuous feature from 1 to of , all and They are connected in series along the column direction to form discrete feature matrices and continuous feature matrices, namely discrete feature aggregation coding. and continuous feature aggregation encoding , expressed as: ； ； In the formula, Represents discrete feature aggregation encoding; represents continuous feature aggregation encoding; Represents matrix transpose; Indicates from arrive Discrete features of Indicates from arrive Continuous characteristics of is the number of discrete features of the patient, is the number of continuous characteristics of the patient; Combine the discrete feature matrix and the continuous feature matrix into a total matrix, namely, patient The eigenvector of , expressed as: ； In the formula, is the adjustment parameter, which means and exist The proportion of The convolution kernel of the convolutional layer in the convolutional neural network CNN is The rectangle represents the width of the convolution kernel, is the number of characteristics of the patient; Convolutional Neural Network Extraction The specific process of parallel features is: The dimension is of Input convolutional neural network, output convolution extracted features ; is the number of patient characteristics, express Dimensions; Among them, the convolutional layer dimension of the convolutional neural network CNN is of The processing process is: ； In the formula, is the output of the convolutional layer; and are the weight parameters and bias parameters of the convolution kernel respectively; according to the dimension transformation principle in the convolution layer, calculate The first size and the second size , expressed as: ； ； The dimension is ,Will Input to the fully connected layer of the convolutional neural network to generate convolutional extraction features : ； In the formula, and They are the weight parameters and bias parameters that the convolutional neural network needs to learn respectively; Graph Neural Network Extraction The specific process of the parallel feature is: After input into the graph neural network for processing, the graph neural network extracted features are obtained , expressed as: ； In the formula, and Represents the parameters that the graph neural network needs to learn; Represents the degree matrix learned by the graph neural network; is the adjacency matrix learned by the graph neural network; Recurrent Neural Network Extraction The specific process of parallel features is: First, set the Within a time node The feature vector : ； In the formula, Indicates Data feature encoding of each time node; Enumeration from arrive , the data feature encoding collected at all a time nodes of each patient; the correlation calculation between nodes is expressed as: ； In the formula, Indicates the correlation between the feature codes of the data collected at two time nodes; Indicates Feature coding of data collected at each time node; Indicates Feature coding of data collected at each time node; and Represents the weights learned by the recurrent neural network RNN ​​through back propagation; Normalize the feature encoding of the data collected at a time node, and use the softmax operator to calculate the normalized weight value : ； In the formula, is another one from 1 to In the range and not equal to The index number of Indicates a certain time point; Based on weight value , make a function: ； In the formula, and These are weight parameters of the recurrent neural network RNN; is a dimension of The representative vector of Aggregate all , calculate Output after inputting the recurrent neural network, i.e. the recurrent neural network extracts features : ； In step S4, the parallel features are mapped to obtain the final prediction result. The specific process is as follows: Will , and Converted to the same dimension, expressed as: ； In the formula, and They are the weight parameters and bias parameters that need to be learned by the parallel feature extraction network respectively; Represents the same dimension after conversion ; Represents the same dimension after conversion ; Represents the same dimension after conversion ; Will , and Merge by accumulating each element to get the aggregate feature , expressed as: ； In the formula, , and They are , and Hyperparameters of weights; By adding a linear layer, Mapping the final result is expressed as: ； In the formula, A score to predict the patient's recovery; is the Sigmoid activation function; Represents a linear layer.

2. According to claim 1, a method for analyzing the recovery status of critically ill patients based on parallel feature extraction is characterized in that: Extract patient feature vectors in parallel feature extraction network Before, the parallel feature extraction network is trained using the training set, which is obtained by dividing the data set. The mean square error is used in the parallel feature extraction network training. Perform parameter tuning, expressed as: ； In the formula, A score indicating actual patient recovery; Represents the number of medical data in the training set.

Images