A combined prediction method for the risk of pulmonary infection after kidney transplantation

Abstract

The present invention discloses a combined prediction method for the risk of pulmonary infection after kidney transplantation, comprising the following steps: S1, quantifying a clinical data set; S2, designing a feature extraction network for preoperative static data and extracting features of the preoperative static data; S3, designing a feature extraction network for postoperative dynamic data sequences and extracting features of the postoperative dynamic data sequences; S4, fusing the features obtained in step S3 and the features obtained in step S4; S5, performing binary classification calculation on the fused features obtained in step S5 using a linear classifier, and inversely obtaining an intelligent prediction result of whether there is an infection from the classification result. The present invention can be applied to predicting in advance the occurrence of the risk of pulmonary infection at each time node after kidney transplantation for patients, and providing clinical auxiliary decision-making support for the diagnosis and treatment plan after kidney transplantation.

Description

Technical Field

[0001] The present invention belongs to the field of infection risk prediction after kidney transplantation, and relates to a prediction method for pulmonary infection after kidney transplantation based on data driving. Background Art

[0002] As the most effective means of treating end-stage renal disease at present, kidney transplantation is widely used to treat end-stage renal failure patients, helping transplant recipients to extend their lives and improve their quality of life. However, at present, some postoperative complications, such as rejection reaction and postoperative infection, still cannot be completely avoided clinically.

[0003] Bacterial infection after kidney transplantation often prolongs the hospital stay of transplant recipients and decreases the creatinine clearance rate, thus increasing the risk of reoperation for kidney transplantation patients. Postoperative infection is clinically considered as the first cause of death within one year after kidney transplantation, and pneumonia and severe pneumonia are also one of the most common fatal infectious diseases after kidney transplantation. Due to the low immunity after surgery, Pneumocystis carinii is extremely likely to infect kidney transplant recipients, leading to severe pneumonia, which makes the transplantation of transplant recipients fail or even die. Some domestic scholars have also reported that the fatality rate of infection after kidney transplantation is relatively high, and the drug resistance of the pathogenic bacteria causing the infection is also very serious, making clinical treatment difficult. According to public reports, among the death cases after kidney transplantation, transplant recipients who died of pulmonary infection account for about 70%. It can be seen that the early prevention and control of postoperative pulmonary infection is crucial for reducing the postoperative survival rate of transplant recipients clinically.

[0004] On the one hand, the cause of postoperative pulmonary infection is the environment, and on the other hand, it depends on the use of postoperative immunosuppressants and the body state of transplant recipients. Kidney transplant recipients must use a large dose of immunosuppressants for a long time after surgery to prevent the occurrence of rejection reaction. The mechanisms of action of clinically used immunosuppressants (such as tacrolimus, etc.) are different. There are differences in the body of transplant recipients, such as gender, age, weight, and gene polymorphisms (such as CYP3A4, CYP3A5, ABCB1), etc., which may produce different treatment effects and cause diverse adverse reactions. Although the use of immunosuppressants can significantly improve the success rate of surgery and the survival rate of transplanted kidneys, the resulting pulmonary infection is an important factor affecting the postoperative survival rate. The factors leading to pulmonary infection after kidney transplantation are very complex. If the pulmonary infection after kidney transplantation can be predicted scientifically and effectively, clinicians can effectively formulate personalized surgical plans for patients, which is of great significance for the prevention and treatment of pulmonary infection clinically and improving the survival rate of kidney transplant recipients.

[0005] When studying the prediction methods for pulmonary infection after kidney transplantation, most existing methods first use statistical methods to analyze the main factors of pulmonary infection, and then use these factors to predict whether there is an infection. The main difference between existing methods lies in mining the risk factors of pulmonary infection. Different statistical analysis methods are used to mine the risk factors of pulmonary infection, and then these high-risk factors are used for prediction through a classifier (such as logistic regression). Hu et al. also used logistic regression analysis to analyze the risks and prognostic factors of severe pulmonary infection after kidney transplantation. It adopted univariate analysis and multiple stepwise logistic regression analysis to find risk factors. The semi-parametric COX regression model (i.e., proportional hazards regression), logistic regression, and grey relational analysis (GRA) are also models widely used in mining the high-risk factors of pulmonary infection. These current classical models mainly rely on statistical analysis for prediction, lack the ability to learn data characteristics, and cannot effectively synergistically fuse pre-operative data and post-operative data, which limits the performance improvement of the prediction system and results in inaccurate prediction results. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a combined prediction method for the risk of pulmonary infection after kidney transplantation, which can be applied to predict in advance the occurrence of the risk of pulmonary infection at each time node after surgery for kidney transplantation patients, and provide clinical auxiliary decision-making support for the diagnosis and treatment plan after kidney transplantation.

[0007] The purpose of the present invention is achieved through the following technical solutions: A combined prediction method for the risk of pulmonary infection after kidney transplantation, comprising the following steps:

[0008] S1. Quantify the clinical data set;

[0009] S2. Design a feature extraction network for pre-operative static data and extract the features of pre-operative static data;

[0010] S3. Design a feature extraction network for post-operative dynamic data sequence and extract the features of post-operative dynamic data sequence;

[0011] S4. Fuse the features obtained in step S3 and the features obtained in step S4;

[0012] S5. Use a linear classifier to perform binary classification calculation on the fused features obtained in step S5, and inversely obtain an intelligent prediction result of whether there is an infection from the classification result.

[0013] The specific implementation method of step S1 is as follows: Assume that the given clinical data set S has M cases, each case contains N possible candidate factors for pulmonary infection, and represents the real number field; Quantify the original data set S into S as shown in formula (1) norm ∈[-1, 1]N×M :

[0014]

[0015] Among them, S norm represents the data matrix obtained after quantization of the original data set matrix, and S r respectively represent the r-th row vectors in the data matrices S norm and S, while and respectively represent the maximum and minimum values of the risk factors in the r-th row vector of the original matrix S, represents the mean value of all elements in the vector S r , r = {0, 1,..., N - 1};

[0016] The data of each sample in the original data set is divided into two parts: preoperative static data and postoperative dynamic data sequences, and these two types of data are respectively labeled as X static and X dynamic .

[0017] The specific implementation method of the step S2 is as follows: The overall structure of the preoperative static data feature extraction network is a multi-layer perceptron network; this feature extraction network includes a three-layer neuron fully connected structure: an input layer neuron, a hidden layer neuron, and an output layer neuron; among them, m0 represents the number of input layer neurons, and m1 and m2 respectively represent the number of neurons in the hidden layer and the output layer; W 1 and W 2 respectively represent the weight matrices formed by the connections of neurons from the input layer → hidden layer → output layer; f 1 and b 1 respectively represent the activation function and bias vector of the hidden layer neurons, and b 2 respectively represent the bias vectors of the output layer; F static represents the static features obtained by the feature extraction network. According to the fully connected mode of neurons in each layer, the weight matrices W 1 and W 2 respectively satisfy the conditions and The output of the preoperative static data feature extraction network is calculated by Equation (2):

[0018] F static = W 2 * f 1 (W 1 * x + b 1 ) + b 2 (2)

[0019] Among them, x represents the preoperative static data after quantization, and the symbol "*" represents the multiplication of a matrix and a vector, f1 (…) is a standard ReLu() activation function.

[0020] The specific implementation method of step S3 is as follows: The overall structural model of the feature extraction network for the postoperative dynamic data sequence is a gated recurrent neural network, and this feature extraction network is a standard single-layer unidirectional GRU; the specific structure includes an update gate and a reset gate, and both the update gate and the reset gate contain weight matrices W z and W r and bias b z and b r ; the input postoperative dynamic data X dynamic is a data sequence composed of inspection data at n time nodes, denoted as X dynamic ={X1,..., X n}; the data at each time node is a vector;

[0021] Specifically, the dynamic feature extraction network calculates the final dynamic features according to the following formula:

[0022] z t =σ(W z *[X t , h t-1 ) + b z (3)

[0023] r t =σ(W r *[X t , h t-1 ) + b r (4)

[0024]

[0025]

[0026] F dynamic =h last (7)

[0027] In formulas (3) to (6), t = {1,..., n}, and the operation symbol * represents the multiplication of a matrix and a vector, × represents multiplication, σ(...) represents the adoption of a sigmoid(...) activation function, z t is the output feature of the update gate, r t is the output feature of the reset gate, is the candidate hidden feature, h t is the hidden feature finally output by the GRU. When t = 1, h t-1 =h0 is directly initialized by the system;

[0028] The calculation process of dynamic feature extraction is iteratively performed until the input data sequence X dynamic The data of its last time node is calculated through formulas (3) to (6) to obtain the output hidden feature h n ; h last represents the hidden feature of the last output, that is, h last = h n , and this output is used as the dynamic feature extracted by the GRU network from the postoperative data sequence.

[0029] The specific implementation method of the step S4 is as follows: the features obtained in step S3 and the feature vector obtained in step S4 are further fused by residual connection, and the specific fusion calculation rules are as follows:

[0030] F sum = F static + F dynamic (8).

[0031] The specific implementation method of the step S5 is as follows: the pulmonary infection predictor is a typical linear classifier, including a fully connected structure and an activation function; the output of the prediction network is calculated by formula (9):

[0032] a = f(W * F sum )(9)

[0033] The symbol "*" represents the multiplication of a matrix and a vector, f is a sigmoid(…) activation function, W is the weight of the fully connected structure, and F sum is the fusion feature obtained in step S4; where a represents the prediction result of the prediction network, and a ∈ {0, 1}; when the prediction network outputs a = 0, it means that it is predicted that the patient has no pulmonary infection after kidney transplantation surgery, and when the output a = 1, it means that the predicted risk of pulmonary infection in the patient is high risk.

[0034] The beneficial effects of the present invention are: the present invention intelligently learns and predicts the risk of pulmonary infection in patients after kidney transplantation from clinical case data in a data-driven manner, and jointly applies the static examination data of the patient before surgery and the examination data of each time node after surgery, which can be applied to predict in advance the risk of pulmonary infection in patients after kidney transplantation at each time node after surgery, provide clinical auxiliary decision-making support for the diagnosis and treatment plan after kidney transplantation, and can also be used as a technical reference for the risk prediction methods of other diseases. Brief Description of the Drawings

[0035] Figure 1 is a flowchart of the joint prediction method for the risk of pulmonary infection after kidney transplantation of the present invention;

[0036] Figure 2It is the overall technical structure diagram of the joint prediction method in this embodiment;

[0037] Figure 3 It is the feature extraction network structure of the preoperative static examination data in this embodiment;

[0038] Figure 4 It is the internal structure diagram of the gated recurrent neural network node in this embodiment;

[0039] Figure 5 It is the neural network structure diagram of the pulmonary infection risk predictor in this embodiment. Detailed implementation manners

[0040] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0041] In this embodiment, taking the clinical follow-up case dataset of 656 cases of pulmonary infection after kidney transplantation collected and sorted out by the Urology Department of West China Hospital of Sichuan University (named SCU-WCH-PI2022) as an example, the SCU-WCH-PI2022 dataset contains a total of 663 cases, involving 52 clinical candidate risk index factors for postoperative pulmonary infection such as "donor age", "relationship between donor and recipient", and "gender". There are a small number of missing risk index factors in the follow-up cases in the original data. After data cleaning, restoration and sorting, a total of 373 valid data are obtained, using 7 preoperative risk index factors and 3 postoperative risk index factors. The 5 biochemical examination time nodes after surgery include: 3 days, 7 days, 14 days, 21 days, and 30 days.

[0042] As Figure 1 shown, a joint prediction method for the risk of pulmonary infection after kidney transplantation according to the present invention includes the following steps:

[0043] S1. Quantify the clinical dataset; the specific implementation method is as follows: The effective clinical dataset SCU-WCH-PI2022 (hereinafter referred to as S) has 656 cases, and each case contains 52 possible pulmonary infection candidate factors. There are a small number of missing risk index factors in the follow-up cases in the original data. After data cleaning, restoration and sorting, a total of 373 valid data are obtained for network prediction. Therefore represents the real number field. These 373 data use 7 preoperative risk index factors and 3 postoperative risk index factors.

[0044] Quantify the original dataset S into S as shown in formula (1) norm ∈[-1, 1] 52×373 :

[0045]

[0046] Among them, S norm represents the data matrix obtained after quantization of the original data set matrix, and S r respectively represent the r-th row vectors in the data matrices S norm and S, while and respectively represent the maximum and minimum values of the risk factors in the r-th row vector of the original matrix S, represents the mean value of all elements in the vector S r , r = {0, 1, …, N - 1};

[0047] The data of each sample in the original data set is divided into two parts: preoperative static data and postoperative dynamic data sequences. These two types of data are respectively labeled as X static ∈[-1, 1] 7×373 and X dynamic ∈[-1, 1] 3×5×373 . X dynamic is a data sequence composed of 3 risk index factors at 5 time nodes after surgery.

[0048] The overall structure of the joint prediction system designed in this embodiment is as Figure 2 shown. It mainly includes a static feature extraction part, a dynamic feature extraction part, a static and dynamic feature fusion part, and a pulmonary infection predictor part.

[0049] S2. Design a feature extraction network for preoperative static data to extract the features of preoperative static data; the specific implementation method is as follows: The overall structure of the preoperative static data feature extraction network is a multi-layer perceptron network, as Figure 3 shown; this feature extraction network includes a three-layer neuron fully connected structure: an input layer neuron, a hidden layer neuron, and an output layer neuron; among them, m0 = 7 represents the number of input layer neurons, m1 = 10 and m2 = 1 respectively represent the number of neurons in the hidden layer and the output layer; W 1 and W 2 respectively represent the weight matrices formed by the connections of the input layer → hidden layer → output layer neurons; f 1 and b 1 respectively represent the activation function and the bias vector of the hidden layer neurons, and b 2 respectively represent the bias vector of the output layer; F static represents the static features obtained by the feature extraction network. According to the fully connected mode of each layer of neurons, the weight matrices W 1 and W 2 respectively satisfy the conditions and The output of the preoperative static data feature extraction network is calculated by Equation (2):

[0050] F static = W 2 * f 1 (W 1 * x + b 1 ) + b 2 (2)

[0051] where x represents the pre - operative static data after quantization, the symbol "*" represents the multiplication of a matrix and a vector, and f 1 (...) is a standard ReLu() activation function.

[0052] S3. Design a feature extraction network for the post - operative dynamic data sequence to extract the features of the post - operative dynamic data sequence. The specific implementation method is as follows: The overall structural model of the feature extraction network for the post - operative dynamic data sequence is a gated recurrent neural network (i.e., GRU). This feature extraction network is a standard single - layer unidirectional GRU, as Figure 4 shown. The number of neurons in the hidden layer is 10 and the number of layers is 1. The specific structure includes an update gate and a reset gate. The update gate and the reset gate each contain weight matrices W z and W r as well as biases b z and b r . The input post - operative dynamic data X dynamic contains sequence data composed of 5 time nodes, that is, the dynamic data sequence X dynamic = {X1, X2,..., X5}; the data of each time node is a 3 - dimensional vector.

[0053] Specifically, the dynamic feature extraction network calculates the final dynamic features according to the following formula:[[]]

[0054] z t = σ(W z * [X t , h t-1 + b z ) (3)[[]]

[0055] r t = σ(W r * [X t , h t-1 + b r ) (4)[[]]

[0056]

[0057]

[0058] F dynamic = h last (7)[[]]

[0059] In Formulas (3) to (6), t = {1, 2,..., 5}, and the operation symbol * represents the multiplication of a matrix and a vector, × represents multiplication, σ(...) represents the use of a sigmoid(...) activation function, and z t is the output feature of the update gate, and r t is the output feature of the reset gate, is the candidate hidden feature, and h t is the finally output hidden feature of the GRU. When t = 1, h t-1 = h0 is directly initialized by the system;

[0060] The calculation process of dynamic feature extraction is iteratively performed until the input data sequence X dynamic The data at its last time node (i.e., when t = 5) is calculated through Formulas (3) to (6) to obtain the output hidden feature h5; h last represents the finally output hidden feature, that is, h last = h5, and this output is used as the dynamic feature extracted by the GRU network from the postoperative data sequence.

[0061] S4. Fuse the features obtained in step S3 and the features obtained in step S4; the specific implementation method is as follows: Use residual connection to further fuse the features obtained in step S3 and the feature vectors obtained in step S4. The specific fusion calculation rules are as follows:

[0062] F sum = F static + F dynamic (8)

[0063] where the operation symbol "+" represents vector addition.

[0064] S5. Use a linear classifier to perform binary classification calculation on the fused features obtained in step S5, and inversely derive the intelligent prediction result of whether there is an infection from the classification result; the specific implementation method is as follows: Use a pulmonary infection predictor as a typical linear classifier, including a fully connected structure (i.e., FCN) and an activation function, and its structure is as Figure 5 shown; the output of the prediction network is calculated by Formula (9):

[0065] a = f(W * F sum ) (9)

[0066] The symbol "*" represents the multiplication of a matrix and a vector, f is a sigmoid(...) activation function, W is the weight of the fully connected structure, and F sumIt is the fused feature obtained in step S4; where a represents the prediction result of the prediction network, and a ∈ {0, 1}; when the prediction network outputs a = 0, it indicates that there is no pulmonary infection after the patient's kidney transplantation surgery, and when the output is a = 1, it indicates that the risk of pulmonary infection in the patient is high risk.

[0067] To verify the technical effectiveness, reliability, and advancement of the present invention, detailed experiments were conducted on the dataset SCU-WCH-PI2022 in this embodiment. The prediction results of pulmonary infection after kidney transplantation were tested on the test set and compared with some publicly available technical methods in recent years on the dataset SCU-WCH-PI2022. The comparison results on SCU-WCH-PI2022 show that the present invention has advantages in recall rate, accuracy, and prediction accuracy, as shown in Table 1:

[0068] Table 1

[0069]

[0070] As shown in Table 1, on the clinical follow-up dataset SCU-WCH-PI2022 provided by the Department of Urology, West China Hospital, Sichuan University, the recall rate, accuracy, and prediction accuracy obtained by the present invention are 95.7%, 99.1%, and 99.9% respectively, which are 4.5%, 4.1%, and 2.1% higher than the performance indicators achieved by the "Logistic Regression + AdaBoost" method. All the technical methods compared in Table 1 were tested on the same dataset, so the comparison results of the performance indicators indicate that the present invention has obvious technical performance advantages.

[0071] Those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present invention and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A combined prediction method for the risk of pulmonary infection after kidney transplantation, characterized in that, Including the following steps: S1. Quantify the clinical dataset. The specific implementation method is as follows: Assume that the given clinical dataset S has M cases, and each case contains N possible candidate factors for lung infection, and represents the real number field; Quantify the original dataset S into S as shown in Equation (1) norm ∈[-1,1] N×M : Among them, S norm represents the data matrix obtained after quantization of the original data set matrix, and S r respectively represent the r-th row vectors in the data matrices S norm and S, and and respectively represent the maximum and minimum values of the risk factors in the r-th row vector of the original matrix S, represents the mean value of all elements in the vector S r , where r = {0, 1, …, N - 1}; The data of each sample in the original dataset is divided into two parts: preoperative static data and postoperative dynamic data sequences. These two types of data are respectively labeled as X static and X dynamic ; S2. Design a feature extraction network for preoperative static data to extract the features of preoperative static data. The specific implementation method is as follows: The overall structure of the preoperative static data feature extraction network is a multi-layer perceptron network. This feature extraction network includes a three-layer fully connected structure of neurons: input layer neurons, hidden layer neurons, and output layer neurons. Among them, m0 represents the number of input layer neurons, and m1 and m2 represent the number of neurons in the hidden layer and output layer respectively; W 1 and W 2 respectively represent the weight matrices formed by the connections of neurons from the input layer → hidden layer → output layer; f 1 and b 1 respectively represent the activation function and bias vector of the hidden layer neurons, and b 2 respectively represent the bias vectors of the output layer; F static represents the static features obtained by the feature extraction network. According to the fully connected mode of neurons in each layer, the weight matrices W 1 and W 2 respectively satisfy the conditions and The output of the preoperative static data feature extraction network is calculated by Equation (2): F static = W 2* f 1 (W 1* x + b 1 ) + b 2 (2) Among them, x represents the pre-operative static data after quantization, the symbol "*" represents the multiplication of a matrix and a vector, and f 1 (…) is a standard ReLu() activation function; S3. Design a feature extraction network for the postoperative dynamic data sequence to extract the features of the postoperative dynamic data sequence; S4. Fuse the features obtained in step S3 and the features obtained in step S4; S5. Use a linear classifier to perform binary classification calculation on the fused features obtained in step S5, and inversely obtain the intelligent prediction result of whether there is an infection from the classification result.

2. The combined prediction method for the risk of pulmonary infection after kidney transplantation according to claim 1, wherein The specific implementation method of the step S3 is as follows: The overall structural model of the feature extraction network for the postoperative dynamic data sequence is a gated recurrent neural network, and this feature extraction network is a standard single-layer unidirectional GRU; the specific structure includes an update gate and a reset gate, and both the update gate and the reset gate contain weight matrices W z and W r as well as biases b z and b r ; the input postoperative dynamic data X dynamic is a data sequence composed of inspection data at n time nodes, denoted as X dynamic ={X1,…,X n}; the data at each time node is a vector; Specifically, the dynamic feature extraction network calculates the final dynamic features according to the following formula: z t = σ(W z * [X t , h t-1 + b z ) (3) r t = σ(W r * [X t , h t-1 + b r ) (4) F dynamic = h last (7) In formulas (3) to (6), t = {1, …, n}, and the operation symbol * represents matrix-vector multiplication, × represents multiplication, σ(…) represents the use of a sigmoid(…) activation function, and z t is the output feature of the update gate, r t is the output feature of the reset gate, is the candidate hidden feature, h t is the finally output hidden feature of the GRU. When t = 1, h t-1 = h0 is directly initialized by the system; The calculation process of dynamic feature extraction is iteratively performed until the input data sequence X dynamic The data of its last time node is calculated through formulas (3) to (6) to obtain the output hidden feature h n ; h last represents the hidden feature of the last output, that is, h last = h n , and this output is used as the dynamic feature extracted by the GRU network from the postoperative data sequence.

3. The combined prediction method for the risk of pulmonary infection after kidney transplantation according to claim 1, wherein, The specific implementation method of step S4 is as follows: Use residual connection to further fuse the features obtained in step S3 and the feature vectors obtained in step S4, and the specific fusion calculation rules are as follows: F sum = F static + F dynamic (8).

4. The combined prediction method for the risk of pulmonary infection after kidney transplantation according to claim 1, wherein The specific implementation method of step S5 is as follows: Use a pulmonary infection predictor as a typical linear classifier, including a fully connected structure and an activation function; the output of the prediction network is calculated by formula (9): a = f(W * F sum ) (9) The symbol "*" represents the multiplication of a matrix and a vector, f is a sigmoid(…) activation function, W is the weight of the fully connected structure, and F sum is the fused feature obtained in step S4; where a represents the prediction result of the prediction network, and a ∈ {0, 1}; when the prediction network outputs a = 0, it means that no pulmonary infection occurs after the patient's kidney transplantation surgery, and when the output a = 1, it means that the risk of pulmonary infection in the patient is high risk.

Images