Data-driven method for mining risk factors of pulmonary infection after kidney transplantation

Abstract

The application discloses a kind of based on data-driven renal transplantation postoperative pulmonary infection risk factor mining method, first to clinical data set is reduced dimension processing and the comprehensive contribution degree of case pair data set is calculated, then the correlation between candidate infection factor and infection occurrence is calculated, and then the candidate infection factor sorted from high to low according to correlation is searched multiple by neural network prediction model, finally according to the highest prediction accuracy determine the best risk factor set related to pulmonary infection.The method of the application intelligently screens important factors causing pulmonary infection of postoperative patient after renal transplantation from clinical case data in a data-driven manner, selects important factors for the best infection prediction scheme, makes the screening of the best risk factor more accurate and robust, and can be applied to the screening of pulmonary infection factors of postoperative patients after renal transplantation, auxiliary decision-making of postoperative diagnosis and treatment scheme, etc., and can also be expanded to the screening of the best risk factor of other diseases and auxiliary decision-making of clinical diagnosis and treatment.

Description

Technical Field

[0001] This invention belongs to the field of infection treatment technology after kidney transplantation, and specifically relates to a data-driven method for mining risk factors of lung infection after kidney transplantation. Background Technology

[0002] Bacterial infections after kidney transplantation often prolong hospital stays and decrease creatinine clearance, increasing the risk of repeat surgery. Postoperative infection is clinically considered the leading cause of death within one year after kidney transplantation, with pneumonia and severe pneumonia being among the most common fatal infectious diseases following transplantation. Due to weakened immunity after surgery, Pneumocystis carinii easily infects kidney transplant recipients, leading to severe pneumonia, transplant failure, and even death. Domestic studies have also reported high mortality rates from postoperative infections after kidney transplantation, and the drug resistance of the infecting pathogens can be very serious, making clinical treatment difficult. According to publicly available reports, approximately 70% of deaths after kidney transplantation are due to pulmonary infections, highlighting the crucial importance of early prevention and control of postoperative pulmonary infections in reducing the survival rate of transplant recipients.

[0003] Postoperative pulmonary infection is caused by both environmental factors and the use of immunosuppressants and the recipient's overall health. Kidney transplant recipients must use high-dose immunosuppressants long-term to prevent rejection. Clinically used immunosuppressants (such as tacrolimus) have different mechanisms of action, and differences in recipient characteristics (such as sex, age, weight, and genetic polymorphisms, including CYP3A4, CYP3A5, and ABCB1) can lead to varying treatment outcomes and diverse adverse reactions. While immunosuppressants significantly improve surgical success rates and transplant survival, pulmonary infection is a significant factor affecting postoperative survival. Therefore, identifying risk factors for pulmonary infection and recognizing infection signals early are crucial for the prevention and treatment of pulmonary infection and for improving kidney transplant recipient survival.

[0004] The medical community both domestically and internationally has conducted tireless research on predicting the risk of pulmonary infection after kidney transplantation, achieving numerous clinical application results. Current research mainly focuses on the impact of gene polymorphisms such as CYP3A4, CYP3A5, ABCB1, and C3435T in transplant recipients on the blood concentration of immunosuppressants and pulmonary infection, as well as the relationship between complex physical conditions and infection, such as the age of the kidney transplant recipient, the type of kidney received, whether there were other diseases before surgery, and whether long-term dialysis treatment had been received.

[0005] In researching risk models for lung infections after kidney transplantation, classic statistical analysis methods are primarily employed, such as Cox regression analysis and logistic regression models, to establish stratified two-dimensional risk models for predicting rejection and infection in the first year after kidney transplantation. Professor Jordan CL et al. established a postoperative opportunistic infection risk model incorporating six risk factors, including age, EBV serological testing, and CMV levels in both donor and recipient, using binary logistic regression analysis. Manuel Moreno Gonzales et al., building upon the Birmingham risk model scoring and multiple linear regression modeling, improved upon it using the Banff scoring rule, proposing a new Birmingham-Mayo model.

[0006] For screening risk factors for lung infections, in addition to classic statistical methods, another approach is machine learning algorithms based on data mining. Representative methods in this category include the Random Survival Forest (RSF) proposed by Newaz and Pomsuwan, the decision tree method used by Feng-Jen Yang, Rochmawati, and Liu Pei, and the Primary Component Analysis (PCA) method used by Rehman et al. Summary of the Invention

[0007] To address the aforementioned technical challenges, this invention proposes a data-driven method for identifying risk factors for post-kidney transplant pulmonary infection, based on classical principal component analysis, grey relational analysis, and fully connected neural networks. Specifically, it involves how to clean clinically collected data, how to use principal component analysis to reduce the dimensionality of clinical data and calculate the overall contribution of case data, how to use grey relational analysis to calculate the correlation between all possible candidate factors and pulmonary infection status, and how to use a fully connected neural network infection predictor to perform a global multi-factor search to select the optimal risk factor group.

[0008] The technical solution adopted in this invention is: a data-driven method for mining risk factors for pulmonary infection after kidney transplantation, the specific steps of which are as follows:

[0009] S1. Quantify the clinical dataset using principal component analysis;

[0010] S2. Perform dimensionality reduction processing on the quantized data;

[0011] S3. Based on the dimensionality reduction results obtained in step S2, calculate the overall contribution of each case sample to the entire dataset.

[0012] S4. Based on the comprehensive contribution of cases obtained in step S3, calculate the correlation between each candidate risk factor and the occurrence of lung infection, and sort them in descending order.

[0013] S5. Design a lung infection predictor based on a fully connected neural network model;

[0014] S6. Using the predictor designed in step S5, perform incremental multiple search prediction on the candidate risk factors sorted in step S4, and determine the optimal risk factor set based on the highest prediction accuracy.

[0015] Furthermore, in step S1, the entire case dataset is quantified, specifically as follows:

[0016] Assuming a given clinical dataset There are M cases, each containing N possible candidate factors for lung infection, and , Represents the real number field. Original dataset As shown in equation (1), it is quantized into , This indicates that the original dataset matrix has been quantized to obtain the data matrix.

[0017] (1)

[0018] in, and Representing data matrices respectively And the r-th row vector in S, and and These represent the maximum and minimum values ​​of the risk factors in the r-th row vector of the original matrix S, respectively. Representing vectors The mean of all elements in the set, r={0,1, …, N-1}.

[0019] Furthermore, step S2 specifically includes the following:

[0020] S21. Calculate the eigenvalues ​​and eigenmatrix;

[0021] For the quantized row vector Perform zero-mean processing, i.e. ,in, This represents the r-th zero-mean vector obtained. Represents row vectors The mean of all elements in the matrix. Let X be the zero-mean matrix obtained, and calculate the covariance matrix of X. , where T represents the transpose of the matrix.

[0022] Calculate the covariance matrix N eigenvalues and the corresponding feature vectors The constraints are: Where I represents an N-order identity matrix, i = {0, 1, 2, ..., N-1}. (The remaining text appears to be a series of seemingly unrelated characters and symbols, possibly due to OCR errors. A more accurate translation would require the original, coherent text.) According to the corresponding feature value Arrange the vectors in descending order to form an eigenvector sorting matrix P.

[0023] S22. Calculate the cumulative contribution rate;

[0024] First, calculate the contribution rate of each principal component. The calculation rule is as shown in equation (2):

[0025] (2)

[0026] Recalculate the matrix Cumulative contribution rate of n principal components A cumulative contribution rate threshold is preset. ,when The optimal number of principal components n is obtained at this time.

[0027] S23. Calculate the dimensionality reduction result of matrix X;

[0028] Select the first n eigenvectors from matrix P , , … Form a new matrix Based on this, the dimensionality reduction matrix of X is calculated. ,in, .

[0029] Furthermore, step S3 specifically includes the following:

[0030] As shown in equation (3), the overall contribution score of each case data in the overall sample is calculated in the form of matrix operations, forming a column vector F:

[0031] (3)

[0032] Where T denotes the transpose of the matrix. This indicates the contribution rate directly from n principal components. An n-dimensional column vector consisting of (i = 0, 1, 2, ..., n-1).

[0033] Furthermore, step S4 specifically includes the following:

[0034] S41. Parameter initialization;

[0035] Set the column vector F of case contributions calculated in step S3 as the quantization matrix of the clinical dataset. A reference vector for the analysis of all candidate risk factors.

[0036] S42. Calculate the difference matrix D;

[0037] According to the rules Calculate the quantization matrix The difference between the contribution column vector F and the matrix D is calculated and its absolute value is processed. Then, the row vector formed by the maximum values ​​of each column in the difference matrix D is calculated. The row vector formed by the minimum value The min(.) and max(.) functions respectively find the maximum and minimum values ​​of each column of the matrix and return a row vector composed of the results of each column.

[0038] S43. Calculate the grey relational coefficient matrix G;

[0039] Based on the difference matrix D, the grey relational coefficient matrix composed of each candidate risk factor is calculated, as shown in equation (4):

[0040] (4)

[0041] in, This represents the element in the correlation coefficient matrix G with row and column coordinates (i, j). It is a resolution factor that is set, and its value range is limited to 1. .

[0042] S44. Calculate the association between risk factors and lung infection;

[0043] The association between each candidate risk factor in the case dataset and the occurrence of pulmonary infection is calculated according to equation (5):

[0044] (5)

[0045] in, Representing a data matrix The association score between the j-th candidate risk factor and lung infection, j={0,1,…,N-1}, where M still represents the total number of patient cases.

[0046] Furthermore, in step S5, a lung infection predictor is designed based on a neural network, as detailed below:

[0047] This predictor consists of a three-layer fully connected neuron structure: input layer neurons, hidden layer neurons, and output layer neurons.

[0048] Set the parameters of the predictor, where m0 represents the number of neurons in the input layer, and m1 and m2 represent the number of neurons in the hidden layer and the output layer, respectively. and These represent the weight matrices formed by the connections between neurons in the input layer and the hidden layer, and between neurons in the hidden layer and the output layer, respectively.

[0049] In addition, f 1 and b 1 Let f represent the activation function and bias vector of the hidden layer neurons, respectively. 2 and b 2 Then, these represent the activation function and bias vector of the output layer, respectively. x and a represent the feature vector input to the prediction network and the network's prediction result, respectively. and When the prediction network outputs a=0, it indicates that the patient will not develop a lung infection after a kidney transplant, while when the output a=1, it indicates that the patient will develop a lung infection.

[0050] Based on the fully connected pattern of neurons in each layer, the weight matrix and Each condition is satisfied and The output of the prediction network is then calculated using equation (6):

[0051] (6)

[0052] in, f represents the multiplication of a matrix and a vector. 1 For the sigmoid(...) function, f 2 For the Linear(...) function.

[0053] Furthermore, step S6 specifically includes the following:

[0054] S61. Design an incremental search prediction scheme and execute the incremental search;

[0055] First, define the number of risk factors L to be searched and the total number of optimal risk factors L*.

[0056] Using the first L quantified risk factors as input to the predictor, and setting the initial search value of L to L=L0, then initializing the parameters m0~m2 and W of the neural network predictor. 1 W 2 and b 1 and b 2 and put the dataset Divide into training subsets proportionally. (percentage of η) and test subset (The percentage is 1-η).

[0057] The incremental iterative search prediction process is as follows:

[0058] Step (1): Use The Adam optimizer was used to train the lung infection predictor FCNN. After the FCNN training process was completed, it was then used... Test FCNN and record the current prediction accuracy as Acc (L=L0);

[0059] Step (2): Incrementally update L=L+1. When L≤N-1, repeat step (1) and record the infection prediction accuracy Acc(L).

[0060] After the incremental cyclic search prediction process is completed, the optimal total number of lung infection risk factors L is obtained according to equation (7). * At this point, the prediction accuracy Acc(L) is at its maximum value.

[0061] (7)

[0062] Here, arg max(...) is a function that finds the maximum value of a set and returns the parameter k corresponding to the maximum Acc value, k={L0, L0+1,..., N-1}.

[0063] S62. Select the corresponding optimal set of risk factors;

[0064] Determine the total number of optimal risk primes L * Then, based on the sorting by relevance in step S44 in descending order. Filter out the highest L * Identify several risk factors to obtain the optimal risk factor set RF = {rf0, rf1, ..., rf...} L*-1},in, Representing a data matrix Zhongyu One-to-one corresponding risk factors Representing a data matrix The correlation score between the i-th candidate risk factor and lung infection.

[0065] The beneficial effects of this invention are as follows: The method of this invention first performs dimensionality reduction on the clinical dataset and calculates the comprehensive contribution of each case to the dataset. Then, it calculates the correlation between candidate infectious factors and the occurrence of infection. Next, it uses a neural network prediction model to perform multiple searches on the candidate infectious factors sorted from high to low correlation. Finally, it determines the optimal set of risk factors associated with pulmonary infection based on the highest prediction accuracy. This invention intelligently screens important factors causing pulmonary infection in post-kidney transplant patients from clinical case data in a data-driven manner, selecting the optimal number of important factors for the best infection prediction plan. This makes the selection of optimal risk factors more accurate and robust. It can be applied to the screening of pulmonary infection factors in post-kidney transplant patients, assisting in postoperative treatment decision-making, and can also be extended to the screening of optimal risk factors for other diseases and assisting in clinical diagnosis and treatment decision-making. Attached Figure Description

[0066] Figure 1 This is a flowchart of a data-driven method for mining risk factors for lung infection after kidney transplantation according to the present invention.

[0067] Figure 2 This is a sorting chart of the infection correlation calculation results in an embodiment of the present invention.

[0068] Figure 3 This is a structural diagram of the neural network predictor in an embodiment of the present invention.

[0069] Figure 4 The diagram below shows the detailed incremental search prediction technology solution in this embodiment of the invention.

[0070] Figure 5 This is a graph showing the incremental search results in an embodiment of the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0072] In this embodiment, a dataset of 663 clinical follow-up cases of post-kidney transplant pulmonary infection collected and organized by the Department of Urology, West China Hospital of Sichuan University (named SCU-WCH-PI2022) is used as an example. The SCU-WCH-PI2022 dataset contains 663 cases, involving 52 possible clinical candidate risk indicators for post-operative pulmonary infection, such as "donor age," "donor-recipient relationship," and "gender." A small number of follow-up cases in the original data lacked risk indicator factors; after data cleaning and restoration, 608 valid data points were obtained.

[0073] like Figure 1 The flowchart of a data-driven method for identifying risk factors for pulmonary infection after kidney transplantation, as shown in the figure, includes the following specific steps:

[0074] S1. Quantify the clinical dataset using principal component analysis;

[0075] S2. Perform dimensionality reduction processing on the quantized data;

[0076] S3. Based on the dimensionality reduction results obtained in step S2, calculate the overall contribution of each case sample to the entire dataset.

[0077] S4. Based on the comprehensive contribution of cases obtained in step S3, calculate the correlation between each candidate risk factor and the occurrence of lung infection, and sort them in descending order.

[0078] S5. Design a lung infection predictor based on a fully connected neural network model;

[0079] S6. Using the predictor designed in step S5, perform incremental multiple search prediction on the candidate risk factors sorted in step S4, and determine the optimal risk factor set based on the highest prediction accuracy.

[0080] In this embodiment, step S1 involves quantizing the entire case dataset, as follows:

[0081] The valid clinical dataset SCU-WCH-PI2022 (hereinafter referred to as S) contains 608 cases, each case containing 52 possible candidate factors for pulmonary infection, and , Represents the real number field. The original dataset S is quantized as shown in equation (1). , This indicates that the original dataset matrix has been quantized to obtain the data matrix.

[0082] (1)

[0083] in, and Representing data matrices respectively And the r-th row vector in S, and and These represent the maximum and minimum values ​​of the risk factors in the r-th row vector of the original matrix S, respectively. Representing vectors The mean of all elements in the dataset, r={0,1, …, 51}.

[0084] In this embodiment, step S2 specifically includes the following:

[0085] S21. Calculate the eigenvalues ​​and eigenmatrix;

[0086] For the quantized row vector Perform zero-mean processing, i.e. ,in, This represents the r-th zero-mean vector obtained. Represents row vectors The mean of all elements in the matrix. Let X be the zero-mean matrix obtained, and calculate the covariance matrix of X. , where T represents the transpose of the matrix.

[0087] Calculate the covariance matrix 52 eigenvalues and the corresponding feature vectors The constraints are: Where I represents an N-order identity matrix, i = {0, 1, 2, ..., 51}. (The last part, "52", appears to be a typo and can be left as is.) According to the corresponding feature value Arrange the vectors in descending order to form an eigenvector sorting matrix P.

[0088] S22. Calculate the cumulative contribution rate;

[0089] First, calculate the contribution rate of each principal component. The calculation rule is as shown in equation (2):

[0090] (2)

[0091] Recalculate the matrix Cumulative contribution rate of n principal components A pre-set cumulative contribution rate threshold. =0.80, when The optimal number of principal components is n=5.

[0092] S23. Calculate the dimensionality reduction result of matrix X;

[0093] Select the first 5 eigenvectors from matrix P , … Form a new matrix Based on this, the dimensionality reduction matrix of X is calculated. ,in, .

[0094] In this embodiment, step S3 specifically includes the following:

[0095] As shown in equation (3), the overall contribution score of each case data in the overall sample is calculated in the form of matrix operations, forming a column vector F:

[0096] (3)

[0097] Where T denotes the transpose of the matrix. This indicates the contribution rate directly from n principal components. A 5-dimensional column vector consisting of (i = 0, 1, 2, ..., 4).

[0098] In this embodiment, step S4 specifically includes the following:

[0099] S41. Parameter initialization;

[0100] Set the column vector F of case contributions calculated in step S3 as the quantization matrix of the clinical dataset. A reference vector for the analysis of all candidate risk factors.

[0101] S42. Calculate the difference matrix D;

[0102] According to the rules Calculate the quantization matrix The difference between the contribution column vector F and the matrix D is calculated and its absolute value is processed. Then, the row vector formed by the maximum values ​​of each column in the difference matrix D is calculated. The row vector formed by the minimum value The min(.) and max(.) functions respectively find the maximum and minimum values ​​of each column of the matrix and return a row vector composed of the results of each column.

[0103] S43. Calculate the grey relational coefficient matrix G;

[0104] Based on the difference matrix D, the grey relational coefficient matrix composed of each candidate risk factor is calculated, as shown in equation (4):

[0105] (4)

[0106] in, This represents the element with row and column coordinates (i, j) in the correlation coefficient matrix G, representing the resolution coefficient. Set it to 0.5.

[0107] S44. Calculate the association between risk factors and lung infection;

[0108] The association between each candidate risk factor in the case dataset and the occurrence of pulmonary infection is calculated according to equation (5):

[0109] (5)

[0110] in, Representing a data matrix The association score between the j-th candidate risk factor and lung infection, j={0,1,…,51}, M=608 represents the total number of cases in S.

[0111] As shown in equation (5), the calculated risk factors are sorted in descending order of their correlation scores with the occurrence of lung infection, and the results are as follows: Figure 2 As shown, these data can accurately measure the differences in the correlation between lung infection and these risk factors in kidney transplant patients. These correlation data are also an important basis for further screening out the optimal high-risk factor group for lung infection.

[0112] In this embodiment, step S5 involves designing a lung infection predictor based on a neural network, as detailed below:

[0113] The lung infection risk predictor FCNN designed in this embodiment is as follows: Figure 3 As shown.

[0114] This predictor consists of a three-layer fully connected neuron structure: input layer neurons, hidden layer neurons, and output layer neurons.

[0115] Set the parameters of the predictor, where m0 represents the number of neurons in the input layer, and m1=20 and m2=2 represent the number of neurons in the hidden layer and the output layer, respectively. and These represent the weight matrices formed by the connections between neurons in the input layer and the hidden layer, and between neurons in the hidden layer and the output layer, respectively.

[0116] In addition, f 1 and b 1 Let f represent the activation function and bias vector of the hidden layer neurons, respectively. 2 and b 2 Then, these represent the activation function and bias vector of the output layer, respectively. x and a represent the feature vector input to the prediction network and the network's prediction result, respectively. and When the prediction network outputs a=0, it indicates that the patient will not develop a lung infection after a kidney transplant, while when the output a=1, it indicates that the patient will develop a lung infection.

[0117] Based on the fully connected pattern of neurons in each layer, the weight matrix and Each condition is satisfied and The output of the prediction network is then calculated using equation (6):

[0118] (6)

[0119] in, f represents the multiplication of a matrix and a vector. 1 For the sigmoid(...) function, f 2 For the Linear(...) function.

[0120] Furthermore, step S6 specifically includes the following:

[0121] S61. Design an incremental search prediction scheme and execute the incremental search;

[0122] The designed incremental multi-search technology solution is as follows: Figure 4 As shown, we first define the number of risk factors L to be searched and the total number of optimal risk factors L*.

[0123] The first L quantified risk factors are used as input to the predictor. The initial search value of L is set to L=L0=6. Then, the parameters m0~m2 of the neural network predictor are initialized. , and b 1 and b 2 and put the dataset Divide into training subsets proportionally. (η=60%) and test subset (1-η=40%).

[0124] The incremental iterative search prediction process is as follows:

[0125] Step (1): Use The FCNN lung infection predictor was trained using the Adam optimizer, with the following hyperparameters set: batch size = 16, epochs = 500, and learning rate lr = 0.001. After the FCNN training process was completed, the algorithm was then... Test FCNN and record the current prediction accuracy as Acc(6);

[0126] Step (2): Incrementally update L=L+1. When L≤N-1, repeat step (1) and record the current lung infection prediction accuracy Acc(L).

[0127] After the incremental cyclic search prediction process is completed, the optimal total number of lung infection risk factors L is obtained according to equation (7). * =18, corresponding to a prediction accuracy of Acc(18) of 96.2%, as shown below. Figure 5 As shown.

[0128] (7)

[0129] Here, arg max(...) is a function that finds the maximum value of the set and returns the parameter k corresponding to the maximum Acc value, k={L0, L0+1,..., 52}.

[0130] S62. Select the corresponding optimal set of risk factors;

[0131] Determine the total number of optimal risk primes L * Then, based on the descending order of relevance in step S44... Filter out the highest L * =18 risk factors, thus obtaining the optimal risk factor set RF={rf0, rf1,…, rf 17},in, Representing a data matrix Zhongyu The corresponding risk factors.

[0132] To verify the effectiveness, reliability, and advancement of the data-driven method for mining risk factors for post-kidney transplant pulmonary infection proposed in this invention, detailed experiments were conducted on the SCU-WCH-PI2022 dataset. The results of risk factor mining and the accuracy of predicting patient pulmonary infection based on the 18 selected high-risk factors were tested on the test set. The results were also compared with some techniques published in recent years on the SCU-WCH-PI2022 dataset.

[0133] The comparison results of SCU-WCH-PI2022 show that the present invention has better recall, accuracy and prediction precision, as shown in Table 1:

[0134] Table 1

[0135]

[0136] As shown in Table 1, on the SCU-WCH-PI2022 clinical follow-up dataset provided by the Department of Urology, West China Hospital of Sichuan University, the recall, accuracy, and prediction precision obtained by this invention are 93.1%, 96.2%, and 99.5%, respectively, which are 1.5%, 1.2%, and 2.7% higher than the performance metrics achieved by the "LogisticRegression+AdaBoost" method. All methods in the table were tested on the same dataset; therefore, the comparison results of these performance metrics also demonstrate the significant performance advantage of this invention.

[0137] Although illustrative specific embodiments of the present invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the present invention is not limited to the scope of the specific embodiments. Various modifications will be apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and established by the appended claims, and all inventions utilizing the inventive concept are protected.

Claims

1. A data-driven method for mining risk factors of pulmonary infection after kidney transplantation, comprising the following specific steps: S1. Quantifying the clinical data set using principal component analysis; S2. Dimensionality reduction processing of the quantized data; S3. Calculating the comprehensive contribution of each case sample to the entire data set based on the data dimensionality reduction result obtained in step S2; S4. Calculating the correlation degree of each candidate risk factor with respect to the occurrence of pulmonary infection based on the case comprehensive contribution obtained in step S3, and sorting in descending order; S5. Designing a pulmonary infection predictor based on a fully connected neural network model; S6. Using the predictor designed in step S5 to perform incremental multiple search prediction on the candidate risk factors sorted in step S4, and determining the best risk factor set according to the highest prediction accuracy; In step S1, the entire case data set is quantized, specifically as follows: Assume a given clinical data set There are M cases, each case contains N possible lung infection candidate factors, and , denotes the real number field; original data set Quantized into as shown in equation (1) , denotes the data matrix obtained after the original data set matrix is quantized; (1)； wherein, and denote the rth row vector of the data matrix and S, respectively, while and denote the maximum and minimum values of the risk factors in the rth row vector of the original matrix S, respectively, denotes the mean of all elements in the vector r = {0, 1,..., N - 1}. In step S2, the following steps are performed: S21. Calculating eigenvalues and eigenvectors; quantized row vector performing zero mean processing, i.e. wherein denotes the resulting r-th zero mean vector, denotes the mean of all elements in the row vector denotes the resulting zero mean matrix X, the covariance matrix of X is calculated T denotes the transpose of a matrix; Computing a covariance matrix N eigenvalues and corresponding eigenvectors , with the constraint that: where I represents an N order identity matrix, i = {0, 1, 2, …, N-1}; and N are arranged in descending order according to the corresponding eigenvalues to form an eigenvector ordering matrix P; S22. Calculating the cumulative contribution rate; First, the contribution rate of each pivot is calculated The calculation rule is shown in equation (2): (2)； recalculating matrix cumulative contribution rate of n pivots ; preset a cumulative contribution rate threshold , when optimal pivot number n is obtained S23. Calculating the dimensionality reduction result of matrix X; selecting the first n eigenvectors in the matrix P , , , to form a new matrix , from which the reduced dimensionality result matrix X is calculated wherein ; In step S3, the following steps are performed: As shown in equation (3), the comprehensive contribution score of each case data to the entire sample is calculated in the form of matrix operation, forming a column vector F: (3)； where T denotes the transpose of a matrix, denotes a n-dimensional column vector consisting of the direct contribution from n pivots i = 0, 1, 2,..., n - 1; In step S4, the following steps are performed: S41. Initialization of parameters; The case contribution degree column vector F calculated from step S3 is set as a clinical data set quantification matrix a reference vector for analyzing all candidate risk factors in the matrix S42. Calculating the difference matrix D; According to the rules Computing the quantization matrix The difference between the contribution degree column vector F and the absolute value processing, and then calculate the difference matrix D in each column data maximum value of the row vector And the minimum value of the row vector , min(.) and max(.) respectively represent the maximum and minimum value of each column of the matrix, return a row vector composed of the calculation results of each column; S43. Calculating the grey correlation coefficient matrix G; Based on the difference matrix D, the grey correlation coefficient matrix composed of each candidate risk factor is calculated, as shown in equation (4): (4)； wherein denotes the element of the incidence matrix G with row and column coordinates (i, j), is a set resolution coefficient, whose value range is limited to ; S44. Calculating the correlation degree of risk factors with pulmonary infection; According to equation (5), the correlation degree between each candidate risk factor in the case data set and the occurrence of pulmonary infection is calculated: (5)； wherein, representing a data matrix the association score of the jth candidate risk factor with lung infection, j = {0, 1, …, N-1}, M still represents the total number of cases of the patient. 2.The data-driven method for mining risk factors of pulmonary infection after kidney transplantation according to claim 1, characterized in that, In step S5, the pulmonary infection predictor is designed based on the neural network, specifically as follows: This predictor includes a three-layer neuron fully connected structure: input layer neurons, hidden layer neurons, and output layer neurons; Parameters of the predictor are set, where m0represents the number of input layer neurons, m1and m2represent the number of neurons of the hidden layer and the output layer, respectively; and W1and W2represent weight matrices of the connection constitution of the input layer→hidden layer and the hidden layer→output layer neurons, respectively; Moreover, f 1 and b 1 represent the activation function and bias vector of the hidden layer neurons, respectively, f 2 and b 2 represent the activation function and bias vector of the output layer, respectively; x and a represent the feature vector input to the prediction network and the prediction result of the network, respectively, and a = 0 indicates that the patient will not develop a pulmonary infection after the kidney transplant surgery, and a = 1 indicates that the patient will develop a pulmonary infection. According to the full connection mode of neurons in each layer, the weight matrix and satisfy the conditions and respectively; the output of the prediction network is calculated by equation (6): (6)； wherein, denotes matrix multiplication with a vector, f 1 is a sigmoid(…) function, f 2 is a Linear(…) function. 3.The data-driven method for mining risk factors of pulmonary infection after kidney transplantation according to claim 2, characterized in that, In step S6, the following steps are performed: S61. Designing an incremental search prediction scheme and performing incremental search; First, define the number of searched risk factors L and the total number of best risk factors L*; The first L quantized risk factors are taken as the predictor input, the search initial value of L is set as L=L0, and the parameters m0-m2 and W of the neural network predictor are initialized 1 , W 2 , and b 1 , and b 2 , and the is divided into a training subset with a proportion of η and a test subset with a proportion of 1-η ; The incremental cyclic search prediction process is as follows: Step (1): use and train the lung infection predictor FCNN using Adam optimizer; after the FCNN training process is completed, use test the FCNN and record the current prediction accuracy as Acc(L=L0); Step (2): Incrementally updating L = L + 1, and when L ≤ N-1, cyclically executing step (1) to record the infection prediction accuracy Acc(L); After the incremental cyclic search prediction process is completed, the optimal total number of lung infection risk factors L is obtained according to formula (7) * At this time, the prediction accuracy Acc(L) is the maximum value; (7)； Where arg max(...) is a function that finds the maximum value of the set and returns the parameter k corresponding to the maximum Acc value, k = {L0, L0+1,..., N-1}; S62. Screening the corresponding best risk factor set; determining the optimal number of risk factors L * After that, based on the ranking in step S44 in descending order of correlation degree Screening the highest L * Risk factors, so as to obtain the optimal risk factor set RF={rf0, rf1,…, rf L*-1}, Indicates the risk factors in the data matrix Corresponding to , , Indicates the correlation degree score of the i-th candidate risk factor in the data matrix With pulmonary infection.

Images